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};
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#define MAX_PATTERN_SIZE	512
116#define FIO_JOBNAME_SIZE	128
117#define FIO_JOBDESC_SIZE	256
118#define FIO_VERROR_SIZE		128
119
120struct thread_stat {
121	char name[FIO_JOBNAME_SIZE];
122	char verror[FIO_VERROR_SIZE];
123	uint32_t error;
124	uint32_t thread_number;
125	uint32_t groupid;
126	uint32_t pid;
127	char description[FIO_JOBDESC_SIZE];
128	uint32_t members;
129	uint32_t unified_rw_rep;
130
131	/*
132	 * bandwidth and latency stats
133	 */
134	struct io_stat clat_stat[DDIR_RWDIR_CNT]; /* completion latency */
135	struct io_stat slat_stat[DDIR_RWDIR_CNT]; /* submission latency */
136	struct io_stat lat_stat[DDIR_RWDIR_CNT]; /* total latency */
137	struct io_stat bw_stat[DDIR_RWDIR_CNT]; /* bandwidth stats */
138	struct io_stat iops_stat[DDIR_RWDIR_CNT]; /* IOPS stats */
139
140	/*
141	 * fio system usage accounting
142	 */
143	uint64_t usr_time;
144	uint64_t sys_time;
145	uint64_t ctx;
146	uint64_t minf, majf;
147
148	/*
149	 * IO depth and latency stats
150	 */
151	uint64_t clat_percentiles;
152	uint64_t percentile_precision;
153	fio_fp64_t percentile_list[FIO_IO_U_LIST_MAX_LEN];
154
155	uint32_t io_u_map[FIO_IO_U_MAP_NR];
156	uint32_t io_u_submit[FIO_IO_U_MAP_NR];
157	uint32_t io_u_complete[FIO_IO_U_MAP_NR];
158	uint32_t io_u_lat_u[FIO_IO_U_LAT_U_NR];
159	uint32_t io_u_lat_m[FIO_IO_U_LAT_M_NR];
160	uint32_t io_u_plat[DDIR_RWDIR_CNT][FIO_IO_U_PLAT_NR];
161	uint64_t total_io_u[3];
162	uint64_t short_io_u[3];
163	uint64_t total_submit;
164	uint64_t total_complete;
165
166	uint64_t io_bytes[DDIR_RWDIR_CNT];
167	uint64_t runtime[DDIR_RWDIR_CNT];
168	uint64_t total_run_time;
169
170	/*
171	 * IO Error related stats
172	 */
173	uint16_t continue_on_error;
174	uint64_t total_err_count;
175	uint32_t first_error;
176
177	uint32_t kb_base;
178	uint32_t unit_base;
179
180	uint32_t latency_depth;
181	uint64_t latency_target;
182	fio_fp64_t latency_percentile;
183	uint64_t latency_window;
184};
185
186struct jobs_eta {
187	uint32_t nr_running;
188	uint32_t nr_ramp;
189	uint32_t nr_pending;
190	uint32_t nr_setting_up;
191	uint32_t files_open;
192	uint32_t m_rate[DDIR_RWDIR_CNT], t_rate[DDIR_RWDIR_CNT];
193	uint32_t m_iops[DDIR_RWDIR_CNT], t_iops[DDIR_RWDIR_CNT];
194	uint32_t rate[DDIR_RWDIR_CNT];
195	uint32_t iops[DDIR_RWDIR_CNT];
196	uint64_t elapsed_sec;
197	uint64_t eta_sec;
198	uint32_t is_pow2;
199	uint32_t unit_base;
200
201	/*
202	 * Network 'copy' of run_str[]
203	 */
204	uint32_t nr_threads;
205	uint8_t run_str[];
206};
207
208extern void stat_init(void);
209extern void stat_exit(void);
210
211extern struct json_object * show_thread_status(struct thread_stat *ts, struct group_run_stats *rs);
212extern void show_group_stats(struct group_run_stats *rs);
213extern int calc_thread_status(struct jobs_eta *je, int force);
214extern void display_thread_status(struct jobs_eta *je);
215extern void show_run_stats(void);
216extern void show_running_run_stats(void);
217extern void check_for_running_stats(void);
218extern void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr);
219extern void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src);
220extern void init_thread_stat(struct thread_stat *ts);
221extern void init_group_run_stat(struct group_run_stats *gs);
222extern void eta_to_str(char *str, unsigned long eta_sec);
223extern int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max, double *mean, double *dev);
224extern 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);
225extern void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat);
226extern void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat);
227extern void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist);
228extern void reset_io_stats(struct thread_data *);
229
230static inline int usec_to_msec(unsigned long *min, unsigned long *max,
231			       double *mean, double *dev)
232{
233	if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) {
234		*min /= 1000;
235		*max /= 1000;
236		*mean /= 1000.0;
237		*dev /= 1000.0;
238		return 0;
239	}
240
241	return 1;
242}
243
244#endif
245