5 #include "lib/output_buffer.h"
7 struct group_run_stats {
8 uint64_t max_run[DDIR_RWDIR_CNT], min_run[DDIR_RWDIR_CNT];
9 uint64_t max_bw[DDIR_RWDIR_CNT], min_bw[DDIR_RWDIR_CNT];
10 uint64_t iobytes[DDIR_RWDIR_CNT];
11 uint64_t agg[DDIR_RWDIR_CNT];
15 uint32_t unified_rw_rep;
16 } __attribute__((packed));
19 * How many depth levels to log
21 #define FIO_IO_U_MAP_NR 7
22 #define FIO_IO_U_LAT_N_NR 10
23 #define FIO_IO_U_LAT_U_NR 10
24 #define FIO_IO_U_LAT_M_NR 12
27 * Aggregate clat samples to report percentile(s) of them.
31 * FIO_IO_U_PLAT_BITS determines the maximum statistical error on the
32 * value of resulting percentiles. The error will be approximately
33 * 1/2^(FIO_IO_U_PLAT_BITS+1) of the value.
35 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the maximum
36 * range being tracked for latency samples. The maximum value tracked
37 * accurately will be 2^(GROUP_NR + PLAT_BITS -1) microseconds.
39 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the memory
40 * requirement of storing those aggregate counts. The memory used will
41 * be (FIO_IO_U_PLAT_GROUP_NR * 2^FIO_IO_U_PLAT_BITS) * sizeof(int)
44 * FIO_IO_U_PLAT_NR is the total number of buckets.
48 * Suppose the clat varies from 0 to 999 (usec), the straightforward
49 * method is to keep an array of (999 + 1) buckets, in which a counter
50 * keeps the count of samples which fall in the bucket, e.g.,
51 * {[0],[1],...,[999]}. However this consumes a huge amount of space,
52 * and can be avoided if an approximation is acceptable.
54 * One such method is to let the range of the bucket to be greater
55 * than one. This method has low accuracy when the value is small. For
56 * example, let the buckets be {[0,99],[100,199],...,[900,999]}, and
57 * the represented value of each bucket be the mean of the range. Then
58 * a value 0 has an round-off error of 49.5. To improve on this, we
59 * use buckets with non-uniform ranges, while bounding the error of
60 * each bucket within a ratio of the sample value. A simple example
61 * would be when error_bound = 0.005, buckets are {
62 * {[0],[1],...,[99]}, {[100,101],[102,103],...,[198,199]},..,
63 * {[900,909],[910,919]...} }. The total range is partitioned into
64 * groups with different ranges, then buckets with uniform ranges. An
65 * upper bound of the error is (range_of_bucket/2)/value_of_bucket
67 * For better efficiency, we implement this using base two. We group
68 * samples by their Most Significant Bit (MSB), extract the next M bit
69 * of them as an index within the group, and discard the rest of the
72 * E.g., assume a sample 'x' whose MSB is bit n (starting from bit 0),
73 * and use M bit for indexing
75 * | n | M bits | bit (n-M-1) ... bit 0 |
77 * Because x is at least 2^n, and bit 0 to bit (n-M-1) is at most
78 * (2^(n-M) - 1), discarding bit 0 to (n-M-1) makes the round-off
82 * e <= --------- <= ------- = ---
85 * Furthermore, we use "mean" of the range to represent the bucket,
86 * the error e can be lowered by half to 1 / 2^(M+1). By using M bits
87 * as the index, each group must contains 2^M buckets.
89 * E.g. Let M (FIO_IO_U_PLAT_BITS) be 6
90 * Error bound is 1/2^(6+1) = 0.0078125 (< 1%)
92 * Group MSB #discarded range of #buckets
94 * ----------------------------------------------------------------
100 * ... ... ... [...,...] ...
101 * 18 23 17 [8838608,+inf]** 64
103 * * Special cases: when n < (M-1) or when n == (M-1), in both cases,
104 * the value cannot be rounded off. Use all bits of the sample as
107 * ** If a sample's MSB is greater than 23, it will be counted as 23.
110 #define FIO_IO_U_PLAT_BITS 6
111 #define FIO_IO_U_PLAT_VAL (1 << FIO_IO_U_PLAT_BITS)
112 #define FIO_IO_U_PLAT_GROUP_NR 29
113 #define FIO_IO_U_PLAT_NR (FIO_IO_U_PLAT_GROUP_NR * FIO_IO_U_PLAT_VAL)
114 #define FIO_IO_U_LIST_MAX_LEN 20 /* The size of the default and user-specified
115 list of percentiles */
118 * Trim cycle count measurements
120 #define MAX_NR_BLOCK_INFOS 8192
121 #define BLOCK_INFO_STATE_SHIFT 29
122 #define BLOCK_INFO_TRIMS(block_info) \
123 ((block_info) & ((1 << BLOCK_INFO_STATE_SHIFT) - 1))
124 #define BLOCK_INFO_STATE(block_info) \
125 ((block_info) >> BLOCK_INFO_STATE_SHIFT)
126 #define BLOCK_INFO(state, trim_cycles) \
127 ((trim_cycles) | ((unsigned int) (state) << BLOCK_INFO_STATE_SHIFT))
128 #define BLOCK_INFO_SET_STATE(block_info, state) \
129 BLOCK_INFO(state, BLOCK_INFO_TRIMS(block_info))
130 enum block_info_state {
134 BLOCK_STATE_TRIM_FAILURE,
135 BLOCK_STATE_WRITE_FAILURE,
139 #define MAX_PATTERN_SIZE 512
140 #define FIO_JOBNAME_SIZE 128
141 #define FIO_JOBDESC_SIZE 256
142 #define FIO_VERROR_SIZE 128
145 char name[FIO_JOBNAME_SIZE];
146 char verror[FIO_VERROR_SIZE];
148 uint32_t thread_number;
151 char description[FIO_JOBDESC_SIZE];
153 uint32_t unified_rw_rep;
156 * bandwidth and latency stats
158 struct io_stat clat_stat[DDIR_RWDIR_CNT]; /* completion latency */
159 struct io_stat slat_stat[DDIR_RWDIR_CNT]; /* submission latency */
160 struct io_stat lat_stat[DDIR_RWDIR_CNT]; /* total latency */
161 struct io_stat bw_stat[DDIR_RWDIR_CNT]; /* bandwidth stats */
162 struct io_stat iops_stat[DDIR_RWDIR_CNT]; /* IOPS stats */
165 * fio system usage accounting
173 * IO depth and latency stats
175 uint64_t clat_percentiles;
176 uint64_t percentile_precision;
177 fio_fp64_t percentile_list[FIO_IO_U_LIST_MAX_LEN];
179 uint32_t io_u_map[FIO_IO_U_MAP_NR];
180 uint32_t io_u_submit[FIO_IO_U_MAP_NR];
181 uint32_t io_u_complete[FIO_IO_U_MAP_NR];
182 uint32_t io_u_lat_n[FIO_IO_U_LAT_N_NR];
183 uint32_t io_u_lat_u[FIO_IO_U_LAT_U_NR];
184 uint32_t io_u_lat_m[FIO_IO_U_LAT_M_NR];
185 uint32_t io_u_plat[DDIR_RWDIR_CNT][FIO_IO_U_PLAT_NR];
188 uint64_t total_io_u[DDIR_RWDIR_CNT];
189 uint64_t short_io_u[DDIR_RWDIR_CNT];
190 uint64_t drop_io_u[DDIR_RWDIR_CNT];
191 uint64_t total_submit;
192 uint64_t total_complete;
194 uint64_t io_bytes[DDIR_RWDIR_CNT];
195 uint64_t runtime[DDIR_RWDIR_CNT];
196 uint64_t total_run_time;
199 * IO Error related stats
202 uint16_t continue_on_error;
205 uint32_t first_error;
206 uint64_t total_err_count;
208 uint64_t nr_block_infos;
209 uint32_t block_infos[MAX_NR_BLOCK_INFOS];
214 uint32_t latency_depth;
216 uint64_t latency_target;
217 fio_fp64_t latency_percentile;
218 uint64_t latency_window;
226 fio_fp64_t ss_deviation;
227 fio_fp64_t ss_criterion;
230 uint64_t *ss_iops_data;
235 uint64_t *ss_bw_data;
238 } __attribute__((packed));
245 uint32_t nr_setting_up;
247 uint64_t m_rate[DDIR_RWDIR_CNT], t_rate[DDIR_RWDIR_CNT];
248 uint64_t rate[DDIR_RWDIR_CNT];
249 uint32_t m_iops[DDIR_RWDIR_CNT], t_iops[DDIR_RWDIR_CNT];
250 uint32_t iops[DDIR_RWDIR_CNT];
251 uint64_t elapsed_sec;
259 * Network 'copy' of run_str[]
263 } __attribute__((packed));
265 struct io_u_plat_entry {
266 struct flist_head list;
267 unsigned int io_u_plat[FIO_IO_U_PLAT_NR];
270 extern struct fio_mutex *stat_mutex;
272 extern struct jobs_eta *get_jobs_eta(bool force, size_t *size);
274 extern void stat_init(void);
275 extern void stat_exit(void);
277 extern struct json_object * show_thread_status(struct thread_stat *ts, struct group_run_stats *rs, struct flist_head *, struct buf_output *);
278 extern void show_group_stats(struct group_run_stats *rs, struct buf_output *);
279 extern bool calc_thread_status(struct jobs_eta *je, int force);
280 extern void display_thread_status(struct jobs_eta *je);
281 extern void show_run_stats(void);
282 extern void __show_run_stats(void);
283 extern void __show_running_run_stats(void);
284 extern void show_running_run_stats(void);
285 extern void check_for_running_stats(void);
286 extern void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, bool first);
287 extern void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src);
288 extern void init_thread_stat(struct thread_stat *ts);
289 extern void init_group_run_stat(struct group_run_stats *gs);
290 extern void eta_to_str(char *str, unsigned long eta_sec);
291 extern bool calc_lat(struct io_stat *is, unsigned long long *min, unsigned long long *max, double *mean, double *dev);
292 extern unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr, fio_fp64_t *plist, unsigned long long **output, unsigned long long *maxv, unsigned long long *minv);
293 extern void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat);
294 extern void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat);
295 extern void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat);
296 extern void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist);
297 extern void reset_io_stats(struct thread_data *);
298 extern void update_rusage_stat(struct thread_data *);
299 extern void clear_rusage_stat(struct thread_data *);
301 extern void add_lat_sample(struct thread_data *, enum fio_ddir, unsigned long long,
302 unsigned int, uint64_t);
303 extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long long,
304 unsigned int, uint64_t);
305 extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long,
306 unsigned int, uint64_t);
307 extern void add_agg_sample(union io_sample_data, enum fio_ddir, unsigned int);
308 extern void add_iops_sample(struct thread_data *, struct io_u *,
310 extern void add_bw_sample(struct thread_data *, struct io_u *,
311 unsigned int, unsigned long long);
312 extern int calc_log_samples(void);
314 extern struct io_log *agg_io_log[DDIR_RWDIR_CNT];
315 extern int write_bw_log;
317 static inline bool nsec_to_usec(unsigned long long *min,
318 unsigned long long *max, double *mean,
321 if (*min > 2000 && *max > 99999 && *dev > 1000.0) {
332 static inline bool nsec_to_msec(unsigned long long *min,
333 unsigned long long *max, double *mean,
336 if (*min > 2000000 && *max > 99999999ULL && *dev > 1000000.0) {
348 * Worst level condensing would be 1:5, so allow enough room for that
350 #define __THREAD_RUNSTR_SZ(nr) ((nr) * 5)
351 #define THREAD_RUNSTR_SZ __THREAD_RUNSTR_SZ(thread_number)
353 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u);