1 #ifndef FIO_STAT_H
2 #define FIO_STAT_H
3
4 #include "iolog.h"
5 #include "lib/output_buffer.h"
6 #include "diskutil.h"
7 #include "json.h"
8
9 struct group_run_stats {
10 uint64_t max_run[DDIR_RWDIR_CNT], min_run[DDIR_RWDIR_CNT];
11 uint64_t max_bw[DDIR_RWDIR_CNT], min_bw[DDIR_RWDIR_CNT];
12 uint64_t iobytes[DDIR_RWDIR_CNT];
13 uint64_t agg[DDIR_RWDIR_CNT];
14 uint32_t kb_base;
15 uint32_t unit_base;
16 uint32_t sig_figs;
17 uint32_t groupid;
18 uint32_t unified_rw_rep;
19 } __attribute__((packed));
20
21 /*
22 * How many depth levels to log
23 */
24 #define FIO_IO_U_MAP_NR 7
25 #define FIO_IO_U_LAT_N_NR 10
26 #define FIO_IO_U_LAT_U_NR 10
27 #define FIO_IO_U_LAT_M_NR 12
28
29 /*
30 * Constants for clat percentiles
31 */
32 #define FIO_IO_U_PLAT_BITS 6
33 #define FIO_IO_U_PLAT_VAL (1 << FIO_IO_U_PLAT_BITS)
34 #define FIO_IO_U_PLAT_GROUP_NR 29
35 #define FIO_IO_U_PLAT_NR (FIO_IO_U_PLAT_GROUP_NR * FIO_IO_U_PLAT_VAL)
36 #define FIO_IO_U_LIST_MAX_LEN 20 /* The size of the default and user-specified
37 list of percentiles */
38
39 /*
40 * Aggregate latency samples for reporting percentile(s).
41 *
42 * EXECUTIVE SUMMARY
43 *
44 * FIO_IO_U_PLAT_BITS determines the maximum statistical error on the
45 * value of resulting percentiles. The error will be approximately
46 * 1/2^(FIO_IO_U_PLAT_BITS+1) of the value.
47 *
48 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the maximum
49 * range being tracked for latency samples. The maximum value tracked
50 * accurately will be 2^(GROUP_NR + PLAT_BITS - 1) nanoseconds.
51 *
52 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the memory
53 * requirement of storing those aggregate counts. The memory used will
54 * be (FIO_IO_U_PLAT_GROUP_NR * 2^FIO_IO_U_PLAT_BITS) * sizeof(int)
55 * bytes.
56 *
57 * FIO_IO_U_PLAT_NR is the total number of buckets.
58 *
59 * DETAILS
60 *
61 * Suppose the lat varies from 0 to 999 (usec), the straightforward
62 * method is to keep an array of (999 + 1) buckets, in which a counter
63 * keeps the count of samples which fall in the bucket, e.g.,
64 * {[0],[1],...,[999]}. However this consumes a huge amount of space,
65 * and can be avoided if an approximation is acceptable.
66 *
67 * One such method is to let the range of the bucket to be greater
68 * than one. This method has low accuracy when the value is small. For
69 * example, let the buckets be {[0,99],[100,199],...,[900,999]}, and
70 * the represented value of each bucket be the mean of the range. Then
71 * a value 0 has an round-off error of 49.5. To improve on this, we
72 * use buckets with non-uniform ranges, while bounding the error of
73 * each bucket within a ratio of the sample value. A simple example
74 * would be when error_bound = 0.005, buckets are {
75 * {[0],[1],...,[99]}, {[100,101],[102,103],...,[198,199]},..,
76 * {[900,909],[910,919]...} }. The total range is partitioned into
77 * groups with different ranges, then buckets with uniform ranges. An
78 * upper bound of the error is (range_of_bucket/2)/value_of_bucket
79 *
80 * For better efficiency, we implement this using base two. We group
81 * samples by their Most Significant Bit (MSB), extract the next M bit
82 * of them as an index within the group, and discard the rest of the
83 * bits.
84 *
85 * E.g., assume a sample 'x' whose MSB is bit n (starting from bit 0),
86 * and use M bit for indexing
87 *
88 * | n | M bits | bit (n-M-1) ... bit 0 |
89 *
90 * Because x is at least 2^n, and bit 0 to bit (n-M-1) is at most
91 * (2^(n-M) - 1), discarding bit 0 to (n-M-1) makes the round-off
92 * error
93 *
94 * 2^(n-M)-1 2^(n-M) 1
95 * e <= --------- <= ------- = ---
96 * 2^n 2^n 2^M
97 *
98 * Furthermore, we use "mean" of the range to represent the bucket,
99 * the error e can be lowered by half to 1 / 2^(M+1). By using M bits
100 * as the index, each group must contains 2^M buckets.
101 *
102 * E.g. Let M (FIO_IO_U_PLAT_BITS) be 6
103 * Error bound is 1/2^(6+1) = 0.0078125 (< 1%)
104 *
105 * Group MSB #discarded range of #buckets
106 * error_bits value
107 * ----------------------------------------------------------------
108 * 0* 0~5 0 [0,63] 64
109 * 1* 6 0 [64,127] 64
110 * 2 7 1 [128,255] 64
111 * 3 8 2 [256,511] 64
112 * 4 9 3 [512,1023] 64
113 * ... ... ... [...,...] ...
114 * 28 33 27 [8589934592,+inf]** 64
115 *
116 * * Special cases: when n < (M-1) or when n == (M-1), in both cases,
117 * the value cannot be rounded off. Use all bits of the sample as
118 * index.
119 *
120 * ** If a sample's MSB is greater than 33, it will be counted as 33.
121 */
122
123 /*
124 * Trim cycle count measurements
125 */
126 #define MAX_NR_BLOCK_INFOS 8192
127 #define BLOCK_INFO_STATE_SHIFT 29
128 #define BLOCK_INFO_TRIMS(block_info) \
129 ((block_info) & ((1 << BLOCK_INFO_STATE_SHIFT) - 1))
130 #define BLOCK_INFO_STATE(block_info) \
131 ((block_info) >> BLOCK_INFO_STATE_SHIFT)
132 #define BLOCK_INFO(state, trim_cycles) \
133 ((trim_cycles) | ((unsigned int) (state) << BLOCK_INFO_STATE_SHIFT))
134 #define BLOCK_INFO_SET_STATE(block_info, state) \
135 BLOCK_INFO(state, BLOCK_INFO_TRIMS(block_info))
136 enum block_info_state {
137 BLOCK_STATE_UNINIT,
138 BLOCK_STATE_TRIMMED,
139 BLOCK_STATE_WRITTEN,
140 BLOCK_STATE_TRIM_FAILURE,
141 BLOCK_STATE_WRITE_FAILURE,
142 BLOCK_STATE_COUNT,
143 };
144
145 #define MAX_PATTERN_SIZE 512
146 #define FIO_JOBNAME_SIZE 128
147 #define FIO_JOBDESC_SIZE 256
148 #define FIO_VERROR_SIZE 128
149 #define UNIFIED_SPLIT 0
150 #define UNIFIED_MIXED 1
151 #define UNIFIED_BOTH 2
152
153 enum fio_lat {
154 FIO_SLAT = 0,
155 FIO_CLAT,
156 FIO_LAT,
157
158 FIO_LAT_CNT = 3,
159 };
160
161 struct thread_stat {
162 char name[FIO_JOBNAME_SIZE];
163 char verror[FIO_VERROR_SIZE];
164 uint32_t error;
165 uint32_t thread_number;
166 uint32_t groupid;
167 uint32_t pid;
168 char description[FIO_JOBDESC_SIZE];
169 uint32_t members;
170 uint32_t unified_rw_rep;
171
172 /*
173 * bandwidth and latency stats
174 */
175 struct io_stat sync_stat __attribute__((aligned(8)));/* fsync etc stats */
176 struct io_stat clat_stat[DDIR_RWDIR_CNT]; /* completion latency */
177 struct io_stat slat_stat[DDIR_RWDIR_CNT]; /* submission latency */
178 struct io_stat lat_stat[DDIR_RWDIR_CNT]; /* total latency */
179 struct io_stat bw_stat[DDIR_RWDIR_CNT]; /* bandwidth stats */
180 struct io_stat iops_stat[DDIR_RWDIR_CNT]; /* IOPS stats */
181
182 /*
183 * fio system usage accounting
184 */
185 uint64_t usr_time;
186 uint64_t sys_time;
187 uint64_t ctx;
188 uint64_t minf, majf;
189
190 /*
191 * IO depth and latency stats
192 */
193 uint32_t clat_percentiles;
194 uint32_t lat_percentiles;
195 uint32_t slat_percentiles;
196 uint32_t pad;
197 uint64_t percentile_precision;
198 fio_fp64_t percentile_list[FIO_IO_U_LIST_MAX_LEN];
199
200 uint64_t io_u_map[FIO_IO_U_MAP_NR];
201 uint64_t io_u_submit[FIO_IO_U_MAP_NR];
202 uint64_t io_u_complete[FIO_IO_U_MAP_NR];
203 uint64_t io_u_lat_n[FIO_IO_U_LAT_N_NR];
204 uint64_t io_u_lat_u[FIO_IO_U_LAT_U_NR];
205 uint64_t io_u_lat_m[FIO_IO_U_LAT_M_NR];
206 uint64_t io_u_plat[FIO_LAT_CNT][DDIR_RWDIR_CNT][FIO_IO_U_PLAT_NR];
207 uint64_t io_u_sync_plat[FIO_IO_U_PLAT_NR];
208
209 uint64_t total_io_u[DDIR_RWDIR_SYNC_CNT];
210 uint64_t short_io_u[DDIR_RWDIR_CNT];
211 uint64_t drop_io_u[DDIR_RWDIR_CNT];
212 uint64_t total_submit;
213 uint64_t total_complete;
214
215 uint64_t io_bytes[DDIR_RWDIR_CNT];
216 uint64_t runtime[DDIR_RWDIR_CNT];
217 uint64_t total_run_time;
218
219 /*
220 * IO Error related stats
221 */
222 union {
223 uint16_t continue_on_error;
224 uint32_t pad2;
225 };
226 uint32_t first_error;
227 uint64_t total_err_count;
228
229 /* ZBD stats */
230 uint64_t nr_zone_resets;
231
232 uint64_t nr_block_infos;
233 uint32_t block_infos[MAX_NR_BLOCK_INFOS];
234
235 uint32_t kb_base;
236 uint32_t unit_base;
237
238 uint32_t latency_depth;
239 uint32_t pad3;
240 uint64_t latency_target;
241 fio_fp64_t latency_percentile;
242 uint64_t latency_window;
243
244 uint32_t sig_figs;
245
246 uint64_t ss_dur;
247 uint32_t ss_state;
248 uint32_t ss_head;
249
250 fio_fp64_t ss_limit;
251 fio_fp64_t ss_slope;
252 fio_fp64_t ss_deviation;
253 fio_fp64_t ss_criterion;
254
255 uint64_t io_u_plat_high_prio[DDIR_RWDIR_CNT][FIO_IO_U_PLAT_NR] __attribute__((aligned(8)));;
256 uint64_t io_u_plat_low_prio[DDIR_RWDIR_CNT][FIO_IO_U_PLAT_NR];
257 struct io_stat clat_high_prio_stat[DDIR_RWDIR_CNT] __attribute__((aligned(8)));
258 struct io_stat clat_low_prio_stat[DDIR_RWDIR_CNT];
259
260 union {
261 uint64_t *ss_iops_data;
262 uint64_t pad4;
263 };
264
265 union {
266 uint64_t *ss_bw_data;
267 uint64_t pad5;
268 };
269
270 uint64_t cachehit;
271 uint64_t cachemiss;
272 } __attribute__((packed));
273
274 #define JOBS_ETA { \
275 uint32_t nr_running; \
276 uint32_t nr_ramp; \
277 \
278 uint32_t nr_pending; \
279 uint32_t nr_setting_up; \
280 \
281 uint64_t m_rate[DDIR_RWDIR_CNT]; \
282 uint64_t t_rate[DDIR_RWDIR_CNT]; \
283 uint64_t rate[DDIR_RWDIR_CNT]; \
284 uint32_t m_iops[DDIR_RWDIR_CNT]; \
285 uint32_t t_iops[DDIR_RWDIR_CNT]; \
286 uint32_t iops[DDIR_RWDIR_CNT]; \
287 uint32_t pad; \
288 uint64_t elapsed_sec; \
289 uint64_t eta_sec; \
290 uint32_t is_pow2; \
291 uint32_t unit_base; \
292 \
293 uint32_t sig_figs; \
294 \
295 uint32_t files_open; \
296 \
297 /* \
298 * Network 'copy' of run_str[] \
299 */ \
300 uint32_t nr_threads; \
301 uint32_t pad2; \
302 uint8_t run_str[]; \
303 }
304
305 struct jobs_eta JOBS_ETA;
306 struct jobs_eta_packed JOBS_ETA __attribute__((packed));
307
308 struct io_u_plat_entry {
309 struct flist_head list;
310 uint64_t io_u_plat[FIO_IO_U_PLAT_NR];
311 };
312
313 extern struct fio_sem *stat_sem;
314
315 extern struct jobs_eta *get_jobs_eta(bool force, size_t *size);
316
317 extern void stat_init(void);
318 extern void stat_exit(void);
319
320 extern struct json_object * show_thread_status(struct thread_stat *ts, struct group_run_stats *rs, struct flist_head *, struct buf_output *);
321 extern void show_group_stats(struct group_run_stats *rs, struct buf_output *);
322 extern bool calc_thread_status(struct jobs_eta *je, int force);
323 extern void display_thread_status(struct jobs_eta *je);
324 extern void __show_run_stats(void);
325 extern int __show_running_run_stats(void);
326 extern void show_running_run_stats(void);
327 extern void check_for_running_stats(void);
328 extern void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, bool first);
329 extern void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src);
330 extern void init_thread_stat_min_vals(struct thread_stat *ts);
331 extern void init_thread_stat(struct thread_stat *ts);
332 extern void init_group_run_stat(struct group_run_stats *gs);
333 extern void eta_to_str(char *str, unsigned long eta_sec);
334 extern bool calc_lat(struct io_stat *is, unsigned long long *min, unsigned long long *max, double *mean, double *dev);
335 extern unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr, fio_fp64_t *plist, unsigned long long **output, unsigned long long *maxv, unsigned long long *minv);
336 extern void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat);
337 extern void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat);
338 extern void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat);
339 extern void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist);
340 extern void reset_io_stats(struct thread_data *);
341 extern void update_rusage_stat(struct thread_data *);
342 extern void clear_rusage_stat(struct thread_data *);
343
344 extern void add_lat_sample(struct thread_data *, enum fio_ddir, unsigned long long,
345 unsigned long long, uint64_t, unsigned int, bool);
346 extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long long,
347 unsigned long long, uint64_t, unsigned int, bool);
348 extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long long,
349 unsigned long long, uint64_t, unsigned int);
350 extern void add_agg_sample(union io_sample_data, enum fio_ddir, unsigned long long);
351 extern void add_iops_sample(struct thread_data *, struct io_u *,
352 unsigned int);
353 extern void add_bw_sample(struct thread_data *, struct io_u *,
354 unsigned int, unsigned long long);
355 extern void add_sync_clat_sample(struct thread_stat *ts,
356 unsigned long long nsec);
357 extern int calc_log_samples(void);
358
359 extern void print_disk_util(struct disk_util_stat *, struct disk_util_agg *, int terse, struct buf_output *);
360 extern void json_array_add_disk_util(struct disk_util_stat *dus,
361 struct disk_util_agg *agg, struct json_array *parent);
362
363 extern struct io_log *agg_io_log[DDIR_RWDIR_CNT];
364 extern bool write_bw_log;
365
nsec_to_usec(unsigned long long * min,unsigned long long * max,double * mean,double * dev)366 static inline bool nsec_to_usec(unsigned long long *min,
367 unsigned long long *max, double *mean,
368 double *dev)
369 {
370 if (*min > 2000 && *max > 99999 && *dev > 1000.0) {
371 *min /= 1000;
372 *max /= 1000;
373 *mean /= 1000.0;
374 *dev /= 1000.0;
375 return true;
376 }
377
378 return false;
379 }
380
nsec_to_msec(unsigned long long * min,unsigned long long * max,double * mean,double * dev)381 static inline bool nsec_to_msec(unsigned long long *min,
382 unsigned long long *max, double *mean,
383 double *dev)
384 {
385 if (*min > 2000000 && *max > 99999999ULL && *dev > 1000000.0) {
386 *min /= 1000000;
387 *max /= 1000000;
388 *mean /= 1000000.0;
389 *dev /= 1000000.0;
390 return true;
391 }
392
393 return false;
394 }
395
396 /*
397 * Worst level condensing would be 1:5, so allow enough room for that
398 */
399 #define __THREAD_RUNSTR_SZ(nr) ((nr) * 5)
400 #define THREAD_RUNSTR_SZ __THREAD_RUNSTR_SZ(thread_number)
401
402 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u);
403
404 #endif
405