8 #include <sys/resource.h>
19 #include "compiler/compiler.h"
23 #include "arch/arch.h"
36 #include "lib/getopt.h"
43 #ifdef FIO_HAVE_SOLARISAIO
44 #include <sys/asynch.h>
47 struct group_run_stats {
48 unsigned long long max_run[2], min_run[2];
49 unsigned long long max_bw[2], min_bw[2];
50 unsigned long long io_kb[2];
51 unsigned long long agg[2];
56 * What type of allocation to use for io buffers
59 MEM_MALLOC = 0, /* ordinary malloc */
60 MEM_SHM, /* use shared memory segments */
61 MEM_SHMHUGE, /* use shared memory segments with huge pages */
62 MEM_MMAP, /* use anonynomous mmap */
63 MEM_MMAPHUGE, /* memory mapped huge file */
67 * offset generator types
75 * How many depth levels to log
77 #define FIO_IO_U_MAP_NR 7
78 #define FIO_IO_U_LAT_U_NR 10
79 #define FIO_IO_U_LAT_M_NR 12
82 * Aggregate clat samples to report percentile(s) of them.
86 * FIO_IO_U_PLAT_BITS determines the maximum statistical error on the
87 * value of resulting percentiles. The error will be approximately
88 * 1/2^(FIO_IO_U_PLAT_BITS+1) of the value.
90 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the maximum
91 * range being tracked for latency samples. The maximum value tracked
92 * accurately will be 2^(GROUP_NR + PLAT_BITS -1) microseconds.
94 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the memory
95 * requirement of storing those aggregate counts. The memory used will
96 * be (FIO_IO_U_PLAT_GROUP_NR * 2^FIO_IO_U_PLAT_BITS) * sizeof(int)
99 * FIO_IO_U_PLAT_NR is the total number of buckets.
103 * Suppose the clat varies from 0 to 999 (usec), the straightforward
104 * method is to keep an array of (999 + 1) buckets, in which a counter
105 * keeps the count of samples which fall in the bucket, e.g.,
106 * {[0],[1],...,[999]}. However this consumes a huge amount of space,
107 * and can be avoided if an approximation is acceptable.
109 * One such method is to let the range of the bucket to be greater
110 * than one. This method has low accuracy when the value is small. For
111 * example, let the buckets be {[0,99],[100,199],...,[900,999]}, and
112 * the represented value of each bucket be the mean of the range. Then
113 * a value 0 has an round-off error of 49.5. To improve on this, we
114 * use buckets with non-uniform ranges, while bounding the error of
115 * each bucket within a ratio of the sample value. A simple example
116 * would be when error_bound = 0.005, buckets are {
117 * {[0],[1],...,[99]}, {[100,101],[102,103],...,[198,199]},..,
118 * {[900,909],[910,919]...} }. The total range is partitioned into
119 * groups with different ranges, then buckets with uniform ranges. An
120 * upper bound of the error is (range_of_bucket/2)/value_of_bucket
122 * For better efficiency, we implement this using base two. We group
123 * samples by their Most Significant Bit (MSB), extract the next M bit
124 * of them as an index within the group, and discard the rest of the
127 * E.g., assume a sample 'x' whose MSB is bit n (starting from bit 0),
128 * and use M bit for indexing
130 * | n | M bits | bit (n-M-1) ... bit 0 |
132 * Because x is at least 2^n, and bit 0 to bit (n-M-1) is at most
133 * (2^(n-M) - 1), discarding bit 0 to (n-M-1) makes the round-off
136 * 2^(n-M)-1 2^(n-M) 1
137 * e <= --------- <= ------- = ---
140 * Furthermore, we use "mean" of the range to represent the bucket,
141 * the error e can be lowered by half to 1 / 2^(M+1). By using M bits
142 * as the index, each group must contains 2^M buckets.
144 * E.g. Let M (FIO_IO_U_PLAT_BITS) be 6
145 * Error bound is 1/2^(6+1) = 0.0078125 (< 1%)
147 * Group MSB #discarded range of #buckets
149 * ----------------------------------------------------------------
154 * 4 9 3 [512,1023] 64
155 * ... ... ... [...,...] ...
156 * 18 23 17 [8838608,+inf]** 64
158 * * Special cases: when n < (M-1) or when n == (M-1), in both cases,
159 * the value cannot be rounded off. Use all bits of the sample as
162 * ** If a sample's MSB is greater than 23, it will be counted as 23.
165 #define FIO_IO_U_PLAT_BITS 6
166 #define FIO_IO_U_PLAT_VAL (1 << FIO_IO_U_PLAT_BITS)
167 #define FIO_IO_U_PLAT_GROUP_NR 19
168 #define FIO_IO_U_PLAT_NR (FIO_IO_U_PLAT_GROUP_NR * FIO_IO_U_PLAT_VAL)
169 #define FIO_IO_U_LIST_MAX_LEN 20 /* The size of the default and user-specified
170 list of percentiles */
172 #define MAX_PATTERN_SIZE 512
183 struct io_log *slat_log;
184 struct io_log *clat_log;
185 struct io_log *lat_log;
186 struct io_log *bw_log;
189 * bandwidth and latency stats
191 struct io_stat clat_stat[2]; /* completion latency */
192 struct io_stat slat_stat[2]; /* submission latency */
193 struct io_stat lat_stat[2]; /* total latency */
194 struct io_stat bw_stat[2]; /* bandwidth stats */
196 unsigned long long stat_io_bytes[2];
197 struct timeval stat_sample_time[2];
200 * fio system usage accounting
202 struct rusage ru_start;
203 struct rusage ru_end;
204 unsigned long usr_time;
205 unsigned long sys_time;
207 unsigned long minf, majf;
210 * IO depth and latency stats
212 unsigned int clat_percentiles;
213 double* percentile_list;
215 unsigned int io_u_map[FIO_IO_U_MAP_NR];
216 unsigned int io_u_submit[FIO_IO_U_MAP_NR];
217 unsigned int io_u_complete[FIO_IO_U_MAP_NR];
218 unsigned int io_u_lat_u[FIO_IO_U_LAT_U_NR];
219 unsigned int io_u_lat_m[FIO_IO_U_LAT_M_NR];
220 unsigned int io_u_plat[3][FIO_IO_U_PLAT_NR];
221 unsigned long total_io_u[3];
222 unsigned long short_io_u[3];
223 unsigned long total_submit;
224 unsigned long total_complete;
226 unsigned long long io_bytes[2];
227 unsigned long long runtime[2];
228 unsigned long total_run_time;
231 * IO Error related stats
233 unsigned continue_on_error;
234 unsigned long total_err_count;
237 unsigned int kb_base;
245 struct thread_options {
253 enum td_ddir td_ddir;
255 unsigned int kb_base;
256 unsigned int ddir_seq_nr;
257 unsigned int iodepth;
258 unsigned int iodepth_low;
259 unsigned int iodepth_batch;
260 unsigned int iodepth_batch_complete;
262 unsigned long long size;
263 unsigned int size_percent;
264 unsigned int fill_device;
265 unsigned long long file_size_low;
266 unsigned long long file_size_high;
267 unsigned long long start_offset;
271 unsigned int min_bs[2];
272 unsigned int max_bs[2];
273 struct bssplit *bssplit[2];
274 unsigned int bssplit_nr[2];
276 unsigned int nr_files;
277 unsigned int open_files;
278 enum file_lock_mode file_lock_mode;
279 unsigned int lockfile_batch;
281 unsigned int odirect;
282 unsigned int invalidate_cache;
283 unsigned int create_serialize;
284 unsigned int create_fsync;
285 unsigned int create_on_open;
286 unsigned int end_fsync;
287 unsigned int pre_read;
288 unsigned int sync_io;
290 unsigned int do_verify;
291 unsigned int verifysort;
292 unsigned int verify_interval;
293 unsigned int verify_offset;
294 char verify_pattern[MAX_PATTERN_SIZE];
295 unsigned int verify_pattern_bytes;
296 unsigned int verify_fatal;
297 unsigned int verify_dump;
298 unsigned int verify_async;
299 unsigned long long verify_backlog;
300 unsigned int verify_batch;
301 unsigned int use_thread;
303 unsigned int do_disk_util;
304 unsigned int override_sync;
305 unsigned int rand_repeatable;
306 unsigned int use_os_rand;
307 unsigned int write_lat_log;
308 unsigned int write_bw_log;
309 unsigned int norandommap;
310 unsigned int softrandommap;
311 unsigned int bs_unaligned;
312 unsigned int fsync_on_close;
314 unsigned int hugepage_size;
315 unsigned int rw_min_bs;
316 unsigned int thinktime;
317 unsigned int thinktime_spin;
318 unsigned int thinktime_blocks;
319 unsigned int fsync_blocks;
320 unsigned int fdatasync_blocks;
321 unsigned int barrier_blocks;
322 unsigned long long start_delay;
323 unsigned long long timeout;
324 unsigned long long ramp_time;
325 unsigned int overwrite;
326 unsigned int bw_avg_time;
328 unsigned long long zone_size;
329 unsigned long long zone_skip;
330 enum fio_memtype mem_type;
331 unsigned int mem_align;
333 unsigned int stonewall;
334 unsigned int new_group;
335 unsigned int numjobs;
336 os_cpu_mask_t cpumask;
337 unsigned int cpumask_set;
338 os_cpu_mask_t verify_cpumask;
339 unsigned int verify_cpumask_set;
341 unsigned int rwmixcycle;
342 unsigned int rwmix[2];
344 unsigned int file_service_type;
345 unsigned int group_reporting;
346 unsigned int fadvise_hint;
347 enum fio_fallocate_mode fallocate_mode;
348 unsigned int zero_buffers;
349 unsigned int refill_buffers;
350 unsigned int time_based;
351 unsigned int disable_lat;
352 unsigned int disable_clat;
353 unsigned int disable_slat;
354 unsigned int disable_bw;
355 unsigned int gtod_reduce;
356 unsigned int gtod_cpu;
357 unsigned int gtod_offload;
358 enum fio_cs clocksource;
359 unsigned int no_stall;
360 unsigned int trim_percentage;
361 unsigned int trim_batch;
362 unsigned int trim_zero;
363 unsigned long long trim_backlog;
364 unsigned int clat_percentiles;
365 unsigned int overwrite_plist;
366 double percentile_list[FIO_IO_U_LIST_MAX_LEN];
368 char *read_iolog_file;
369 char *write_iolog_file;
372 char *replay_redirect;
375 * Pre-run and post-run shell
380 unsigned int rate[2];
381 unsigned int ratemin[2];
382 unsigned int ratecycle;
383 unsigned int rate_iops[2];
384 unsigned int rate_iops_min[2];
389 * CPU "io" cycle burner
391 unsigned int cpuload;
392 unsigned int cpucycle;
397 unsigned int continue_on_error;
400 * Benchmark profile type
405 * blkio cgroup support
408 unsigned int cgroup_weight;
409 unsigned int cgroup_nodelete;
414 unsigned int sync_file_range;
417 #define FIO_VERROR_SIZE 128
420 * This describes a single thread/process executing a fio job.
423 struct thread_options o;
424 char verror[FIO_VERROR_SIZE];
428 struct thread_stat ts;
429 struct fio_file **files;
430 unsigned int files_size;
431 unsigned int files_index;
432 unsigned int nr_open_files;
433 unsigned int nr_done_files;
434 unsigned int nr_normal_files;
436 unsigned int next_file;
437 os_random_state_t next_file_state;
438 struct frand_state __next_file_state;
444 size_t orig_buffer_size;
445 volatile int terminate;
446 volatile int runstate;
448 unsigned int ioprio_set;
449 unsigned int last_was_sync;
450 enum fio_ddir last_ddir;
460 unsigned long rand_seeds[7];
463 os_random_state_t bsrange_state;
464 struct frand_state __bsrange_state;
467 os_random_state_t verify_state;
468 struct frand_state __verify_state;
471 os_random_state_t trim_state;
472 struct frand_state __trim_state;
475 unsigned int verify_batch;
476 unsigned int trim_batch;
481 * IO engine hooks, contains everything needed to submit an io_u
482 * to any of the available IO engines.
484 struct ioengine_ops *io_ops;
487 * Current IO depth and list of free and busy io_u's.
489 unsigned int cur_depth;
490 unsigned int io_u_queued;
491 struct flist_head io_u_freelist;
492 struct flist_head io_u_busylist;
493 struct flist_head io_u_requeues;
494 pthread_mutex_t io_u_lock;
495 pthread_cond_t free_cond;
498 * async verify offload
500 struct flist_head verify_list;
501 pthread_t *verify_threads;
502 unsigned int nr_verify_threads;
503 pthread_cond_t verify_cond;
504 int verify_thread_exit;
509 unsigned long rate_nsec_cycle[2];
510 long rate_pending_usleep[2];
511 unsigned long rate_bytes[2];
512 unsigned long rate_blocks[2];
513 struct timeval lastrate[2];
515 unsigned long long total_io_size;
516 unsigned long long fill_device_size;
518 unsigned long io_issues[2];
519 unsigned long long io_blocks[2];
520 unsigned long long io_bytes[2];
521 unsigned long long io_skip_bytes;
522 unsigned long long this_io_bytes[2];
523 unsigned long long zone_bytes;
524 struct fio_mutex *mutex;
527 * State for random io, a bitmap of blocks done vs not done
530 os_random_state_t random_state;
531 struct frand_state __random_state;
534 struct timeval start; /* start of this loop */
535 struct timeval epoch; /* time job was started */
536 struct timeval last_issue;
537 struct timeval tv_cache;
538 unsigned int tv_cache_nr;
539 unsigned int tv_cache_mask;
540 unsigned int ramp_time_over;
543 * read/write mixed workload state
546 os_random_state_t rwmix_state;
547 struct frand_state __rwmix_state;
549 unsigned long rwmix_issues;
550 enum fio_ddir rwmix_ddir;
551 unsigned int ddir_seq_nr;
554 * IO history logs for verification. We use a tree for sorting,
555 * if we are overwriting. Otherwise just use a fifo.
557 struct rb_root io_hist_tree;
558 struct flist_head io_hist_list;
559 unsigned long io_hist_len;
564 struct flist_head io_log_list;
567 * For tracking/handling discards
569 struct flist_head trim_list;
570 unsigned long trim_entries;
573 * for fileservice, how often to switch to a new file
575 unsigned int file_service_nr;
576 unsigned int file_service_left;
577 struct fio_file *file_service_file;
579 unsigned int sync_file_range_nr;
582 * For generating file sizes
585 os_random_state_t file_size_state;
586 struct frand_state __file_size_state;
592 unsigned int total_err_count;
596 * Can be overloaded by profiles
598 struct prof_io_ops prof_io_ops;
603 * when should interactive ETA output be generated
611 #define __td_verror(td, err, msg, func) \
617 if (!(td)->first_error) \
618 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
622 #define td_clear_error(td) \
624 #define td_verror(td, err, func) \
625 __td_verror((td), (err), strerror((err)), (func))
626 #define td_vmsg(td, err, msg, func) \
627 __td_verror((td), (err), (msg), (func))
629 extern int exitall_on_terminate;
630 extern int thread_number;
631 extern int nr_process, nr_thread;
634 extern int terse_output;
635 extern int temp_stall_ts;
636 extern unsigned long long mlock_size;
637 extern unsigned long page_mask, page_size;
638 extern int read_only;
639 extern int eta_print;
640 extern unsigned long done_secs;
641 extern char *job_section;
642 extern int fio_gtod_offload;
643 extern int fio_gtod_cpu;
644 extern enum fio_cs fio_clock_source;
645 extern int warnings_fatal;
647 extern struct thread_data *threads;
649 static inline void fio_ro_check(struct thread_data *td, struct io_u *io_u)
651 assert(!(io_u->ddir == DDIR_WRITE && !td_write(td)));
654 #define BLOCKS_PER_MAP (8 * sizeof(unsigned long))
655 #define TO_MAP_BLOCK(f, b) (b)
656 #define RAND_MAP_IDX(f, b) (TO_MAP_BLOCK(f, b) / BLOCKS_PER_MAP)
657 #define RAND_MAP_BIT(f, b) (TO_MAP_BLOCK(f, b) & (BLOCKS_PER_MAP - 1))
659 #define REAL_MAX_JOBS 2048
661 #define td_non_fatal_error(e) ((e) == EIO || (e) == EILSEQ)
663 static inline void update_error_count(struct thread_data *td, int err)
665 td->total_err_count++;
666 if (td->total_err_count == 1)
667 td->first_error = err;
670 static inline int should_fsync(struct thread_data *td)
672 if (td->last_was_sync)
676 if (td_write(td) || td_rw(td) || td->o.override_sync)
683 * Init/option functions
685 extern int __must_check parse_options(int, char **);
686 extern int fio_options_parse(struct thread_data *, char **, int);
687 extern void fio_keywords_init(void);
688 extern int fio_cmd_option_parse(struct thread_data *, const char *, char *);
689 extern void fio_fill_default_options(struct thread_data *);
690 extern int fio_show_option_help(const char *);
691 extern void fio_options_dup_and_init(struct option *);
692 extern void options_mem_dupe(struct thread_data *);
693 extern void options_mem_free(struct thread_data *);
694 extern void td_fill_rand_seeds(struct thread_data *);
695 extern void add_job_opts(const char **);
696 extern char *num2str(unsigned long, int, int, int);
698 #define FIO_GETOPT_JOB 0x89988998
699 #define FIO_NR_OPTIONS (FIO_MAX_OPTS + 128)
704 extern void print_thread_status(void);
705 extern void print_status_init(int);
708 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
709 * will never back again. It may cycle between running/verififying/fsyncing.
710 * Once the thread reaches TD_EXITED, it is just waiting for the core to
726 extern void td_set_runstate(struct thread_data *, int);
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 *);
737 * Reset stats after ramp time completes
739 extern void reset_all_stats(struct thread_data *);
744 #ifdef FIO_HAVE_BLKTRACE
745 extern int is_blktrace(const char *);
746 extern int load_blktrace(struct thread_data *, const char *);
750 * Mark unused variables passed to ops functions as unused, to silence gcc
752 #define fio_unused __attribute((__unused__))
753 #define fio_init __attribute__((constructor))
754 #define fio_exit __attribute__((destructor))
756 #define for_each_td(td, i) \
757 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
758 #define for_each_file(td, f, i) \
759 if ((td)->files_index) \
760 for ((i) = 0, (f) = (td)->files[0]; \
761 (i) < (td)->o.nr_files && ((f) = (td)->files[i]) != NULL; \
764 #define fio_assert(td, cond) do { \
767 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
768 td_set_runstate((td), TD_EXITED); \
769 (td)->error = EFAULT; \
774 static inline int fio_fill_issue_time(struct thread_data *td)
776 if (td->o.read_iolog_file ||
777 !td->o.disable_clat || !td->o.disable_slat || !td->o.disable_bw)
783 static inline int __should_check_rate(struct thread_data *td,
786 struct thread_options *o = &td->o;
789 * If some rate setting was given, we need to check it
791 if (o->rate[ddir] || o->ratemin[ddir] || o->rate_iops[ddir] ||
792 o->rate_iops_min[ddir])
798 static inline int should_check_rate(struct thread_data *td,
799 unsigned long *bytes_done)
804 ret |= __should_check_rate(td, 0);
806 ret |= __should_check_rate(td, 1);
811 static inline int is_power_of_2(unsigned int val)
813 return (val != 0 && ((val & (val - 1)) == 0));
817 * We currently only need to do locking if we have verifier threads
818 * accessing our internal structures too
820 static inline void td_io_u_lock(struct thread_data *td)
822 if (td->o.verify_async)
823 pthread_mutex_lock(&td->io_u_lock);
826 static inline void td_io_u_unlock(struct thread_data *td)
828 if (td->o.verify_async)
829 pthread_mutex_unlock(&td->io_u_lock);
832 static inline void td_io_u_free_notify(struct thread_data *td)
834 if (td->o.verify_async)
835 pthread_cond_signal(&td->free_cond);