Use PATH_MAX instead of _POSIX_HOST_NAME_MAX
[fio.git] / stat.h
... / ...
CommitLineData
1#ifndef FIO_STAT_H
2#define FIO_STAT_H
3
4#include "iolog.h"
5
6struct group_run_stats {
7 uint64_t max_run[DDIR_RWDIR_CNT], min_run[DDIR_RWDIR_CNT];
8 uint64_t max_bw[DDIR_RWDIR_CNT], min_bw[DDIR_RWDIR_CNT];
9 uint64_t io_kb[DDIR_RWDIR_CNT];
10 uint64_t agg[DDIR_RWDIR_CNT];
11 uint32_t kb_base;
12 uint32_t unit_base;
13 uint32_t groupid;
14 uint32_t unified_rw_rep;
15} __attribute__((packed));
16
17/*
18 * How many depth levels to log
19 */
20#define FIO_IO_U_MAP_NR 7
21#define FIO_IO_U_LAT_U_NR 10
22#define FIO_IO_U_LAT_M_NR 12
23
24/*
25 * Aggregate clat samples to report percentile(s) of them.
26 *
27 * EXECUTIVE SUMMARY
28 *
29 * FIO_IO_U_PLAT_BITS determines the maximum statistical error on the
30 * value of resulting percentiles. The error will be approximately
31 * 1/2^(FIO_IO_U_PLAT_BITS+1) of the value.
32 *
33 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the maximum
34 * range being tracked for latency samples. The maximum value tracked
35 * accurately will be 2^(GROUP_NR + PLAT_BITS -1) microseconds.
36 *
37 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the memory
38 * requirement of storing those aggregate counts. The memory used will
39 * be (FIO_IO_U_PLAT_GROUP_NR * 2^FIO_IO_U_PLAT_BITS) * sizeof(int)
40 * bytes.
41 *
42 * FIO_IO_U_PLAT_NR is the total number of buckets.
43 *
44 * DETAILS
45 *
46 * Suppose the clat varies from 0 to 999 (usec), the straightforward
47 * method is to keep an array of (999 + 1) buckets, in which a counter
48 * keeps the count of samples which fall in the bucket, e.g.,
49 * {[0],[1],...,[999]}. However this consumes a huge amount of space,
50 * and can be avoided if an approximation is acceptable.
51 *
52 * One such method is to let the range of the bucket to be greater
53 * than one. This method has low accuracy when the value is small. For
54 * example, let the buckets be {[0,99],[100,199],...,[900,999]}, and
55 * the represented value of each bucket be the mean of the range. Then
56 * a value 0 has an round-off error of 49.5. To improve on this, we
57 * use buckets with non-uniform ranges, while bounding the error of
58 * each bucket within a ratio of the sample value. A simple example
59 * would be when error_bound = 0.005, buckets are {
60 * {[0],[1],...,[99]}, {[100,101],[102,103],...,[198,199]},..,
61 * {[900,909],[910,919]...} }. The total range is partitioned into
62 * groups with different ranges, then buckets with uniform ranges. An
63 * upper bound of the error is (range_of_bucket/2)/value_of_bucket
64 *
65 * For better efficiency, we implement this using base two. We group
66 * samples by their Most Significant Bit (MSB), extract the next M bit
67 * of them as an index within the group, and discard the rest of the
68 * bits.
69 *
70 * E.g., assume a sample 'x' whose MSB is bit n (starting from bit 0),
71 * and use M bit for indexing
72 *
73 * | n | M bits | bit (n-M-1) ... bit 0 |
74 *
75 * Because x is at least 2^n, and bit 0 to bit (n-M-1) is at most
76 * (2^(n-M) - 1), discarding bit 0 to (n-M-1) makes the round-off
77 * error
78 *
79 * 2^(n-M)-1 2^(n-M) 1
80 * e <= --------- <= ------- = ---
81 * 2^n 2^n 2^M
82 *
83 * Furthermore, we use "mean" of the range to represent the bucket,
84 * the error e can be lowered by half to 1 / 2^(M+1). By using M bits
85 * as the index, each group must contains 2^M buckets.
86 *
87 * E.g. Let M (FIO_IO_U_PLAT_BITS) be 6
88 * Error bound is 1/2^(6+1) = 0.0078125 (< 1%)
89 *
90 * Group MSB #discarded range of #buckets
91 * error_bits value
92 * ----------------------------------------------------------------
93 * 0* 0~5 0 [0,63] 64
94 * 1* 6 0 [64,127] 64
95 * 2 7 1 [128,255] 64
96 * 3 8 2 [256,511] 64
97 * 4 9 3 [512,1023] 64
98 * ... ... ... [...,...] ...
99 * 18 23 17 [8838608,+inf]** 64
100 *
101 * * Special cases: when n < (M-1) or when n == (M-1), in both cases,
102 * the value cannot be rounded off. Use all bits of the sample as
103 * index.
104 *
105 * ** If a sample's MSB is greater than 23, it will be counted as 23.
106 */
107
108#define FIO_IO_U_PLAT_BITS 6
109#define FIO_IO_U_PLAT_VAL (1 << FIO_IO_U_PLAT_BITS)
110#define FIO_IO_U_PLAT_GROUP_NR 19
111#define FIO_IO_U_PLAT_NR (FIO_IO_U_PLAT_GROUP_NR * FIO_IO_U_PLAT_VAL)
112#define FIO_IO_U_LIST_MAX_LEN 20 /* The size of the default and user-specified
113 list of percentiles */
114
115/*
116 * Trim cycle count measurements
117 */
118#define MAX_NR_BLOCK_INFOS 8192
119#define BLOCK_INFO_STATE_SHIFT 29
120#define BLOCK_INFO_TRIMS(block_info) \
121 ((block_info) & ((1 << BLOCK_INFO_STATE_SHIFT) - 1))
122#define BLOCK_INFO_STATE(block_info) \
123 ((block_info) >> BLOCK_INFO_STATE_SHIFT)
124#define BLOCK_INFO(state, trim_cycles) \
125 ((trim_cycles) | ((state) << BLOCK_INFO_STATE_SHIFT))
126#define BLOCK_INFO_SET_STATE(block_info, state) \
127 BLOCK_INFO(state, BLOCK_INFO_TRIMS(block_info))
128enum block_info_state {
129 BLOCK_STATE_UNINIT,
130 BLOCK_STATE_TRIMMED,
131 BLOCK_STATE_WRITTEN,
132 BLOCK_STATE_TRIM_FAILURE,
133 BLOCK_STATE_WRITE_FAILURE,
134 BLOCK_STATE_COUNT,
135};
136
137#define MAX_PATTERN_SIZE 512
138#define FIO_JOBNAME_SIZE 128
139#define FIO_JOBDESC_SIZE 256
140#define FIO_VERROR_SIZE 128
141
142struct thread_stat {
143 char name[FIO_JOBNAME_SIZE];
144 char verror[FIO_VERROR_SIZE];
145 uint32_t error;
146 uint32_t thread_number;
147 uint32_t groupid;
148 uint32_t pid;
149 char description[FIO_JOBDESC_SIZE];
150 uint32_t members;
151 uint32_t unified_rw_rep;
152
153 /*
154 * bandwidth and latency stats
155 */
156 struct io_stat clat_stat[DDIR_RWDIR_CNT]; /* completion latency */
157 struct io_stat slat_stat[DDIR_RWDIR_CNT]; /* submission latency */
158 struct io_stat lat_stat[DDIR_RWDIR_CNT]; /* total latency */
159 struct io_stat bw_stat[DDIR_RWDIR_CNT]; /* bandwidth stats */
160 struct io_stat iops_stat[DDIR_RWDIR_CNT]; /* IOPS stats */
161
162 /*
163 * fio system usage accounting
164 */
165 uint64_t usr_time;
166 uint64_t sys_time;
167 uint64_t ctx;
168 uint64_t minf, majf;
169
170 /*
171 * IO depth and latency stats
172 */
173 uint64_t clat_percentiles;
174 uint64_t percentile_precision;
175 fio_fp64_t percentile_list[FIO_IO_U_LIST_MAX_LEN];
176
177 uint32_t io_u_map[FIO_IO_U_MAP_NR];
178 uint32_t io_u_submit[FIO_IO_U_MAP_NR];
179 uint32_t io_u_complete[FIO_IO_U_MAP_NR];
180 uint32_t io_u_lat_u[FIO_IO_U_LAT_U_NR];
181 uint32_t io_u_lat_m[FIO_IO_U_LAT_M_NR];
182 uint32_t io_u_plat[DDIR_RWDIR_CNT][FIO_IO_U_PLAT_NR];
183 uint32_t pad;
184
185 uint64_t total_io_u[3];
186 uint64_t short_io_u[3];
187 uint64_t drop_io_u[3];
188 uint64_t total_submit;
189 uint64_t total_complete;
190
191 uint64_t io_bytes[DDIR_RWDIR_CNT];
192 uint64_t runtime[DDIR_RWDIR_CNT];
193 uint64_t total_run_time;
194
195 /*
196 * IO Error related stats
197 */
198 union {
199 uint16_t continue_on_error;
200 uint64_t pad2;
201 };
202 uint64_t total_err_count;
203 uint32_t first_error;
204
205 uint64_t nr_block_infos;
206 uint32_t block_infos[MAX_NR_BLOCK_INFOS];
207
208 uint32_t kb_base;
209 uint32_t unit_base;
210
211 uint32_t latency_depth;
212 uint64_t latency_target;
213 fio_fp64_t latency_percentile;
214 uint64_t latency_window;
215} __attribute__((packed));
216
217struct jobs_eta {
218 uint32_t nr_running;
219 uint32_t nr_ramp;
220
221 uint32_t nr_pending;
222 uint32_t nr_setting_up;
223
224 uint32_t files_open;
225
226 uint32_t m_rate[DDIR_RWDIR_CNT], t_rate[DDIR_RWDIR_CNT];
227 uint32_t m_iops[DDIR_RWDIR_CNT], t_iops[DDIR_RWDIR_CNT];
228 uint32_t rate[DDIR_RWDIR_CNT];
229 uint32_t iops[DDIR_RWDIR_CNT];
230 uint64_t elapsed_sec;
231 uint64_t eta_sec;
232 uint32_t is_pow2;
233 uint32_t unit_base;
234
235 /*
236 * Network 'copy' of run_str[]
237 */
238 uint32_t nr_threads;
239 uint8_t run_str[];
240} __attribute__((packed));
241
242extern struct fio_mutex *stat_mutex;
243
244extern struct jobs_eta *get_jobs_eta(int force, size_t *size);
245
246extern void stat_init(void);
247extern void stat_exit(void);
248
249extern struct json_object * show_thread_status(struct thread_stat *ts, struct group_run_stats *rs);
250extern void show_group_stats(struct group_run_stats *rs);
251extern int calc_thread_status(struct jobs_eta *je, int force);
252extern void display_thread_status(struct jobs_eta *je);
253extern void show_run_stats(void);
254extern void __show_run_stats(void);
255extern void __show_running_run_stats(void);
256extern void show_running_run_stats(void);
257extern void check_for_running_stats(void);
258extern void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr);
259extern void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src);
260extern void init_thread_stat(struct thread_stat *ts);
261extern void init_group_run_stat(struct group_run_stats *gs);
262extern void eta_to_str(char *str, unsigned long eta_sec);
263extern int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max, double *mean, double *dev);
264extern unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr, fio_fp64_t *plist, unsigned int **output, unsigned int *maxv, unsigned int *minv);
265extern void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat);
266extern void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat);
267extern void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist);
268extern void reset_io_stats(struct thread_data *);
269
270static inline int usec_to_msec(unsigned long *min, unsigned long *max,
271 double *mean, double *dev)
272{
273 if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) {
274 *min /= 1000;
275 *max /= 1000;
276 *mean /= 1000.0;
277 *dev /= 1000.0;
278 return 0;
279 }
280
281 return 1;
282}
283/*
284 * Worst level condensing would be 1:5, so allow enough room for that
285 */
286#define __THREAD_RUNSTR_SZ(nr) ((nr) * 5)
287#define THREAD_RUNSTR_SZ __THREAD_RUNSTR_SZ(thread_number)
288
289uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u);
290
291#endif