projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Merge branch 'master' of https://github.com/celestinechen/fio
[fio.git] / stat.c
... / ...
CommitLineData
1#include <stdio.h>
2#include <string.h>
3#include <stdlib.h>
4#include <sys/time.h>
5#include <sys/stat.h>
6#include <math.h>
7
8#include "fio.h"
9#include "diskutil.h"
10#include "lib/ieee754.h"
11#include "json.h"
12#include "lib/getrusage.h"
13#include "idletime.h"
14#include "lib/pow2.h"
15#include "lib/output_buffer.h"
16#include "helper_thread.h"
17#include "smalloc.h"
18#include "zbd.h"
19#include "oslib/asprintf.h"
20
21#ifdef WIN32
22#define LOG_MSEC_SLACK 2
23#else
24#define LOG_MSEC_SLACK 1
25#endif
26
27struct fio_sem *stat_sem;
28
29void clear_rusage_stat(struct thread_data *td)
30{
31 struct thread_stat *ts = &td->ts;
32
33 fio_getrusage(&td->ru_start);
34 ts->usr_time = ts->sys_time = 0;
35 ts->ctx = 0;
36 ts->minf = ts->majf = 0;
37}
38
39void update_rusage_stat(struct thread_data *td)
40{
41 struct thread_stat *ts = &td->ts;
42
43 fio_getrusage(&td->ru_end);
44 ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
45 &td->ru_end.ru_utime);
46 ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
47 &td->ru_end.ru_stime);
48 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
49 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
50 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
51 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
52
53 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
54}
55
56/*
57 * Given a latency, return the index of the corresponding bucket in
58 * the structure tracking percentiles.
59 *
60 * (1) find the group (and error bits) that the value (latency)
61 * belongs to by looking at its MSB. (2) find the bucket number in the
62 * group by looking at the index bits.
63 *
64 */
65static unsigned int plat_val_to_idx(unsigned long long val)
66{
67 unsigned int msb, error_bits, base, offset, idx;
68
69 /* Find MSB starting from bit 0 */
70 if (val == 0)
71 msb = 0;
72 else
73 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
74
75 /*
76 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
77 * all bits of the sample as index
78 */
79 if (msb <= FIO_IO_U_PLAT_BITS)
80 return val;
81
82 /* Compute the number of error bits to discard*/
83 error_bits = msb - FIO_IO_U_PLAT_BITS;
84
85 /* Compute the number of buckets before the group */
86 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
87
88 /*
89 * Discard the error bits and apply the mask to find the
90 * index for the buckets in the group
91 */
92 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
93
94 /* Make sure the index does not exceed (array size - 1) */
95 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
96 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
97
98 return idx;
99}
100
101/*
102 * Convert the given index of the bucket array to the value
103 * represented by the bucket
104 */
105static unsigned long long plat_idx_to_val(unsigned int idx)
106{
107 unsigned int error_bits;
108 unsigned long long k, base;
109
110 assert(idx < FIO_IO_U_PLAT_NR);
111
112 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
113 * all bits of the sample as index */
114 if (idx < (FIO_IO_U_PLAT_VAL << 1))
115 return idx;
116
117 /* Find the group and compute the minimum value of that group */
118 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
119 base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
120
121 /* Find its bucket number of the group */
122 k = idx % FIO_IO_U_PLAT_VAL;
123
124 /* Return the mean of the range of the bucket */
125 return base + ((k + 0.5) * (1 << error_bits));
126}
127
128static int double_cmp(const void *a, const void *b)
129{
130 const fio_fp64_t fa = *(const fio_fp64_t *) a;
131 const fio_fp64_t fb = *(const fio_fp64_t *) b;
132 int cmp = 0;
133
134 if (fa.u.f > fb.u.f)
135 cmp = 1;
136 else if (fa.u.f < fb.u.f)
137 cmp = -1;
138
139 return cmp;
140}
141
142unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
143 fio_fp64_t *plist, unsigned long long **output,
144 unsigned long long *maxv, unsigned long long *minv)
145{
146 unsigned long long sum = 0;
147 unsigned int len, i, j = 0;
148 unsigned long long *ovals = NULL;
149 bool is_last;
150
151 *minv = -1ULL;
152 *maxv = 0;
153
154 len = 0;
155 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
156 len++;
157
158 if (!len)
159 return 0;
160
161 /*
162 * Sort the percentile list. Note that it may already be sorted if
163 * we are using the default values, but since it's a short list this
164 * isn't a worry. Also note that this does not work for NaN values.
165 */
166 if (len > 1)
167 qsort(plist, len, sizeof(plist[0]), double_cmp);
168
169 ovals = malloc(len * sizeof(*ovals));
170 if (!ovals)
171 return 0;
172
173 /*
174 * Calculate bucket values, note down max and min values
175 */
176 is_last = false;
177 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
178 sum += io_u_plat[i];
179 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
180 assert(plist[j].u.f <= 100.0);
181
182 ovals[j] = plat_idx_to_val(i);
183 if (ovals[j] < *minv)
184 *minv = ovals[j];
185 if (ovals[j] > *maxv)
186 *maxv = ovals[j];
187
188 is_last = (j == len - 1) != 0;
189 if (is_last)
190 break;
191
192 j++;
193 }
194 }
195
196 if (!is_last)
197 log_err("fio: error calculating latency percentiles\n");
198
199 *output = ovals;
200 return len;
201}
202
203/*
204 * Find and display the p-th percentile of clat
205 */
206static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
207 fio_fp64_t *plist, unsigned int precision,
208 const char *pre, struct buf_output *out)
209{
210 unsigned int divisor, len, i, j = 0;
211 unsigned long long minv, maxv;
212 unsigned long long *ovals;
213 int per_line, scale_down, time_width;
214 bool is_last;
215 char fmt[32];
216
217 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
218 if (!len || !ovals)
219 return;
220
221 /*
222 * We default to nsecs, but if the value range is such that we
223 * should scale down to usecs or msecs, do that.
224 */
225 if (minv > 2000000 && maxv > 99999999ULL) {
226 scale_down = 2;
227 divisor = 1000000;
228 log_buf(out, " %s percentiles (msec):\n |", pre);
229 } else if (minv > 2000 && maxv > 99999) {
230 scale_down = 1;
231 divisor = 1000;
232 log_buf(out, " %s percentiles (usec):\n |", pre);
233 } else {
234 scale_down = 0;
235 divisor = 1;
236 log_buf(out, " %s percentiles (nsec):\n |", pre);
237 }
238
239
240 time_width = max(5, (int) (log10(maxv / divisor) + 1));
241 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
242 precision, time_width);
243 /* fmt will be something like " %5.2fth=[%4llu]%c" */
244 per_line = (80 - 7) / (precision + 10 + time_width);
245
246 for (j = 0; j < len; j++) {
247 /* for formatting */
248 if (j != 0 && (j % per_line) == 0)
249 log_buf(out, " |");
250
251 /* end of the list */
252 is_last = (j == len - 1) != 0;
253
254 for (i = 0; i < scale_down; i++)
255 ovals[j] = (ovals[j] + 999) / 1000;
256
257 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
258
259 if (is_last)
260 break;
261
262 if ((j % per_line) == per_line - 1) /* for formatting */
263 log_buf(out, "\n");
264 }
265
266 free(ovals);
267}
268
269static int get_nr_prios_with_samples(struct thread_stat *ts, enum fio_ddir ddir)
270{
271 int i, nr_prios_with_samples = 0;
272
273 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
274 if (ts->clat_prio[ddir][i].clat_stat.samples)
275 nr_prios_with_samples++;
276 }
277
278 return nr_prios_with_samples;
279}
280
281bool calc_lat(struct io_stat *is, unsigned long long *min,
282 unsigned long long *max, double *mean, double *dev)
283{
284 double n = (double) is->samples;
285
286 if (n == 0)
287 return false;
288
289 *min = is->min_val;
290 *max = is->max_val;
291 *mean = is->mean.u.f;
292
293 if (n > 1.0)
294 *dev = sqrt(is->S.u.f / (n - 1.0));
295 else
296 *dev = 0;
297
298 return true;
299}
300
301void show_mixed_group_stats(struct group_run_stats *rs, struct buf_output *out)
302{
303 char *io, *agg, *min, *max;
304 char *ioalt, *aggalt, *minalt, *maxalt;
305 uint64_t io_mix = 0, agg_mix = 0, min_mix = -1, max_mix = 0;
306 uint64_t min_run = -1, max_run = 0;
307 const int i2p = is_power_of_2(rs->kb_base);
308 int i;
309
310 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
311 if (!rs->max_run[i])
312 continue;
313 io_mix += rs->iobytes[i];
314 agg_mix += rs->agg[i];
315 min_mix = min_mix < rs->min_bw[i] ? min_mix : rs->min_bw[i];
316 max_mix = max_mix > rs->max_bw[i] ? max_mix : rs->max_bw[i];
317 min_run = min_run < rs->min_run[i] ? min_run : rs->min_run[i];
318 max_run = max_run > rs->max_run[i] ? max_run : rs->max_run[i];
319 }
320 io = num2str(io_mix, rs->sig_figs, 1, i2p, N2S_BYTE);
321 ioalt = num2str(io_mix, rs->sig_figs, 1, !i2p, N2S_BYTE);
322 agg = num2str(agg_mix, rs->sig_figs, 1, i2p, rs->unit_base);
323 aggalt = num2str(agg_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
324 min = num2str(min_mix, rs->sig_figs, 1, i2p, rs->unit_base);
325 minalt = num2str(min_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
326 max = num2str(max_mix, rs->sig_figs, 1, i2p, rs->unit_base);
327 maxalt = num2str(max_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
328 log_buf(out, " MIXED: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
329 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
330 (unsigned long long) min_run,
331 (unsigned long long) max_run);
332 free(io);
333 free(agg);
334 free(min);
335 free(max);
336 free(ioalt);
337 free(aggalt);
338 free(minalt);
339 free(maxalt);
340}
341
342void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
343{
344 char *io, *agg, *min, *max;
345 char *ioalt, *aggalt, *minalt, *maxalt;
346 const char *str[] = { " READ", " WRITE" , " TRIM"};
347 int i;
348
349 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
350
351 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
352 const int i2p = is_power_of_2(rs->kb_base);
353
354 if (!rs->max_run[i])
355 continue;
356
357 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
358 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
359 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
360 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
361 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
362 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
363 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
364 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
365 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
366 (rs->unified_rw_rep == UNIFIED_MIXED) ? " MIXED" : str[i],
367 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
368 (unsigned long long) rs->min_run[i],
369 (unsigned long long) rs->max_run[i]);
370
371 free(io);
372 free(agg);
373 free(min);
374 free(max);
375 free(ioalt);
376 free(aggalt);
377 free(minalt);
378 free(maxalt);
379 }
380
381 /* Need to aggregate statistics to show mixed values */
382 if (rs->unified_rw_rep == UNIFIED_BOTH)
383 show_mixed_group_stats(rs, out);
384}
385
386void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
387{
388 int i;
389
390 /*
391 * Do depth distribution calculations
392 */
393 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
394 if (total) {
395 io_u_dist[i] = (double) map[i] / (double) total;
396 io_u_dist[i] *= 100.0;
397 if (io_u_dist[i] < 0.1 && map[i])
398 io_u_dist[i] = 0.1;
399 } else
400 io_u_dist[i] = 0.0;
401 }
402}
403
404static void stat_calc_lat(struct thread_stat *ts, double *dst,
405 uint64_t *src, int nr)
406{
407 unsigned long total = ddir_rw_sum(ts->total_io_u);
408 int i;
409
410 /*
411 * Do latency distribution calculations
412 */
413 for (i = 0; i < nr; i++) {
414 if (total) {
415 dst[i] = (double) src[i] / (double) total;
416 dst[i] *= 100.0;
417 if (dst[i] < 0.01 && src[i])
418 dst[i] = 0.01;
419 } else
420 dst[i] = 0.0;
421 }
422}
423
424/*
425 * To keep the terse format unaltered, add all of the ns latency
426 * buckets to the first us latency bucket
427 */
428static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
429{
430 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
431 int i;
432
433 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
434
435 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
436 ntotal += ts->io_u_lat_n[i];
437
438 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
439}
440
441void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
442{
443 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
444}
445
446void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
447{
448 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
449}
450
451void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
452{
453 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
454}
455
456static void display_lat(const char *name, unsigned long long min,
457 unsigned long long max, double mean, double dev,
458 struct buf_output *out)
459{
460 const char *base = "(nsec)";
461 char *minp, *maxp;
462
463 if (nsec_to_msec(&min, &max, &mean, &dev))
464 base = "(msec)";
465 else if (nsec_to_usec(&min, &max, &mean, &dev))
466 base = "(usec)";
467
468 minp = num2str(min, 6, 1, 0, N2S_NONE);
469 maxp = num2str(max, 6, 1, 0, N2S_NONE);
470
471 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
472 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
473
474 free(minp);
475 free(maxp);
476}
477
478static struct thread_stat *gen_mixed_ddir_stats_from_ts(struct thread_stat *ts)
479{
480 struct thread_stat *ts_lcl;
481
482 /*
483 * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
484 * Trims (ddir = 2)
485 */
486 ts_lcl = malloc(sizeof(struct thread_stat));
487 if (!ts_lcl) {
488 log_err("fio: failed to allocate local thread stat\n");
489 return NULL;
490 }
491
492 init_thread_stat(ts_lcl);
493
494 /* calculate mixed stats */
495 ts_lcl->unified_rw_rep = UNIFIED_MIXED;
496 ts_lcl->lat_percentiles = ts->lat_percentiles;
497 ts_lcl->clat_percentiles = ts->clat_percentiles;
498 ts_lcl->slat_percentiles = ts->slat_percentiles;
499 ts_lcl->percentile_precision = ts->percentile_precision;
500 memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
501
502 sum_thread_stats(ts_lcl, ts);
503
504 return ts_lcl;
505}
506
507static double convert_agg_kbytes_percent(struct group_run_stats *rs,
508 enum fio_ddir ddir, int mean)
509{
510 double p_of_agg = 100.0;
511 if (rs && rs->agg[ddir] > 1024) {
512 p_of_agg = mean * 100.0 / (double) (rs->agg[ddir] / 1024.0);
513
514 if (p_of_agg > 100.0)
515 p_of_agg = 100.0;
516 }
517 return p_of_agg;
518}
519
520static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
521 enum fio_ddir ddir, struct buf_output *out)
522{
523 unsigned long runt;
524 unsigned long long min, max, bw, iops;
525 double mean, dev;
526 char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
527 int i2p, i;
528 const char *clat_type = ts->lat_percentiles ? "lat" : "clat";
529
530 if (ddir_sync(ddir)) {
531 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
532 log_buf(out, " %s:\n", "fsync/fdatasync/sync_file_range");
533 display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
534 show_clat_percentiles(ts->io_u_sync_plat,
535 ts->sync_stat.samples,
536 ts->percentile_list,
537 ts->percentile_precision,
538 io_ddir_name(ddir), out);
539 }
540 return;
541 }
542
543 assert(ddir_rw(ddir));
544
545 if (!ts->runtime[ddir])
546 return;
547
548 i2p = is_power_of_2(rs->kb_base);
549 runt = ts->runtime[ddir];
550
551 bw = (1000 * ts->io_bytes[ddir]) / runt;
552 io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
553 bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
554 bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
555
556 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
557 iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
558 if (ddir == DDIR_WRITE || ddir == DDIR_TRIM)
559 post_st = zbd_write_status(ts);
560 else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
561 uint64_t total;
562 double hit;
563
564 total = ts->cachehit + ts->cachemiss;
565 hit = (double) ts->cachehit / (double) total;
566 hit *= 100.0;
567 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
568 post_st = NULL;
569 }
570
571 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
572 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir),
573 iops_p, bw_p, bw_p_alt, io_p,
574 (unsigned long long) ts->runtime[ddir],
575 post_st ? : "");
576
577 free(post_st);
578 free(io_p);
579 free(bw_p);
580 free(bw_p_alt);
581 free(iops_p);
582
583 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
584 display_lat("slat", min, max, mean, dev, out);
585 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
586 display_lat("clat", min, max, mean, dev, out);
587 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
588 display_lat(" lat", min, max, mean, dev, out);
589
590 /* Only print per prio stats if there are >= 2 prios with samples */
591 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
592 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
593 char buf[64];
594
595 if (!calc_lat(&ts->clat_prio[ddir][i].clat_stat, &min,
596 &max, &mean, &dev))
597 continue;
598
599 snprintf(buf, sizeof(buf),
600 "%s prio %u/%u/%u",
601 clat_type,
602 ioprio_class(ts->clat_prio[ddir][i].ioprio),
603 ioprio(ts->clat_prio[ddir][i].ioprio),
604 ioprio_hint(ts->clat_prio[ddir][i].ioprio));
605 display_lat(buf, min, max, mean, dev, out);
606 }
607 }
608
609 if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
610 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
611 ts->slat_stat[ddir].samples,
612 ts->percentile_list,
613 ts->percentile_precision, "slat", out);
614 if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
615 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
616 ts->clat_stat[ddir].samples,
617 ts->percentile_list,
618 ts->percentile_precision, "clat", out);
619 if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
620 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
621 ts->lat_stat[ddir].samples,
622 ts->percentile_list,
623 ts->percentile_precision, "lat", out);
624
625 if (ts->clat_percentiles || ts->lat_percentiles) {
626 char prio_name[64];
627 uint64_t samples;
628
629 if (ts->lat_percentiles)
630 samples = ts->lat_stat[ddir].samples;
631 else
632 samples = ts->clat_stat[ddir].samples;
633
634 /* Only print per prio stats if there are >= 2 prios with samples */
635 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
636 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
637 uint64_t prio_samples =
638 ts->clat_prio[ddir][i].clat_stat.samples;
639
640 if (!prio_samples)
641 continue;
642
643 snprintf(prio_name, sizeof(prio_name),
644 "%s prio %u/%u/%u (%.2f%% of IOs)",
645 clat_type,
646 ioprio_class(ts->clat_prio[ddir][i].ioprio),
647 ioprio(ts->clat_prio[ddir][i].ioprio),
648 ioprio_hint(ts->clat_prio[ddir][i].ioprio),
649 100. * (double) prio_samples / (double) samples);
650 show_clat_percentiles(ts->clat_prio[ddir][i].io_u_plat,
651 prio_samples, ts->percentile_list,
652 ts->percentile_precision,
653 prio_name, out);
654 }
655 }
656 }
657
658 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
659 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
660 const char *bw_str;
661
662 if ((rs->unit_base == 1) && i2p)
663 bw_str = "Kibit";
664 else if (rs->unit_base == 1)
665 bw_str = "kbit";
666 else if (i2p)
667 bw_str = "KiB";
668 else
669 bw_str = "kB";
670
671 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
672
673 if (rs->unit_base == 1) {
674 min *= 8.0;
675 max *= 8.0;
676 mean *= 8.0;
677 dev *= 8.0;
678 }
679
680 if (mean > fkb_base * fkb_base) {
681 min /= fkb_base;
682 max /= fkb_base;
683 mean /= fkb_base;
684 dev /= fkb_base;
685 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
686 }
687
688 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
689 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
690 bw_str, min, max, p_of_agg, mean, dev,
691 (&ts->bw_stat[ddir])->samples);
692 }
693 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
694 log_buf(out, " iops : min=%5llu, max=%5llu, "
695 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
696 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
697 }
698}
699
700static void show_mixed_ddir_status(struct group_run_stats *rs,
701 struct thread_stat *ts,
702 struct buf_output *out)
703{
704 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
705
706 if (ts_lcl)
707 show_ddir_status(rs, ts_lcl, DDIR_READ, out);
708
709 free_clat_prio_stats(ts_lcl);
710 free(ts_lcl);
711}
712
713static bool show_lat(double *io_u_lat, int nr, const char **ranges,
714 const char *msg, struct buf_output *out)
715{
716 bool new_line = true, shown = false;
717 int i, line = 0;
718
719 for (i = 0; i < nr; i++) {
720 if (io_u_lat[i] <= 0.0)
721 continue;
722 shown = true;
723 if (new_line) {
724 if (line)
725 log_buf(out, "\n");
726 log_buf(out, " lat (%s) : ", msg);
727 new_line = false;
728 line = 0;
729 }
730 if (line)
731 log_buf(out, ", ");
732 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
733 line++;
734 if (line == 5)
735 new_line = true;
736 }
737
738 if (shown)
739 log_buf(out, "\n");
740
741 return true;
742}
743
744static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
745{
746 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
747 "250=", "500=", "750=", "1000=", };
748
749 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
750}
751
752static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
753{
754 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
755 "250=", "500=", "750=", "1000=", };
756
757 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
758}
759
760static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
761{
762 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
763 "250=", "500=", "750=", "1000=", "2000=",
764 ">=2000=", };
765
766 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
767}
768
769static void show_latencies(struct thread_stat *ts, struct buf_output *out)
770{
771 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
772 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
773 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
774
775 stat_calc_lat_n(ts, io_u_lat_n);
776 stat_calc_lat_u(ts, io_u_lat_u);
777 stat_calc_lat_m(ts, io_u_lat_m);
778
779 show_lat_n(io_u_lat_n, out);
780 show_lat_u(io_u_lat_u, out);
781 show_lat_m(io_u_lat_m, out);
782}
783
784static int block_state_category(int block_state)
785{
786 switch (block_state) {
787 case BLOCK_STATE_UNINIT:
788 return 0;
789 case BLOCK_STATE_TRIMMED:
790 case BLOCK_STATE_WRITTEN:
791 return 1;
792 case BLOCK_STATE_WRITE_FAILURE:
793 case BLOCK_STATE_TRIM_FAILURE:
794 return 2;
795 default:
796 /* Silence compile warning on some BSDs and have a return */
797 assert(0);
798 return -1;
799 }
800}
801
802static int compare_block_infos(const void *bs1, const void *bs2)
803{
804 uint64_t block1 = *(uint64_t *)bs1;
805 uint64_t block2 = *(uint64_t *)bs2;
806 int state1 = BLOCK_INFO_STATE(block1);
807 int state2 = BLOCK_INFO_STATE(block2);
808 int bscat1 = block_state_category(state1);
809 int bscat2 = block_state_category(state2);
810 int cycles1 = BLOCK_INFO_TRIMS(block1);
811 int cycles2 = BLOCK_INFO_TRIMS(block2);
812
813 if (bscat1 < bscat2)
814 return -1;
815 if (bscat1 > bscat2)
816 return 1;
817
818 if (cycles1 < cycles2)
819 return -1;
820 if (cycles1 > cycles2)
821 return 1;
822
823 if (state1 < state2)
824 return -1;
825 if (state1 > state2)
826 return 1;
827
828 assert(block1 == block2);
829 return 0;
830}
831
832static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
833 fio_fp64_t *plist, unsigned int **percentiles,
834 unsigned int *types)
835{
836 int len = 0;
837 int i, nr_uninit;
838
839 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
840
841 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
842 len++;
843
844 if (!len)
845 return 0;
846
847 /*
848 * Sort the percentile list. Note that it may already be sorted if
849 * we are using the default values, but since it's a short list this
850 * isn't a worry. Also note that this does not work for NaN values.
851 */
852 if (len > 1)
853 qsort(plist, len, sizeof(plist[0]), double_cmp);
854
855 /* Start only after the uninit entries end */
856 for (nr_uninit = 0;
857 nr_uninit < nr_block_infos
858 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
859 nr_uninit ++)
860 ;
861
862 if (nr_uninit == nr_block_infos)
863 return 0;
864
865 *percentiles = calloc(len, sizeof(**percentiles));
866
867 for (i = 0; i < len; i++) {
868 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
869 + nr_uninit;
870 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
871 }
872
873 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
874 for (i = 0; i < nr_block_infos; i++)
875 types[BLOCK_INFO_STATE(block_infos[i])]++;
876
877 return len;
878}
879
880static const char *block_state_names[] = {
881 [BLOCK_STATE_UNINIT] = "unwritten",
882 [BLOCK_STATE_TRIMMED] = "trimmed",
883 [BLOCK_STATE_WRITTEN] = "written",
884 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
885 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
886};
887
888static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
889 fio_fp64_t *plist, struct buf_output *out)
890{
891 int len, pos, i;
892 unsigned int *percentiles = NULL;
893 unsigned int block_state_counts[BLOCK_STATE_COUNT];
894
895 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
896 &percentiles, block_state_counts);
897
898 log_buf(out, " block lifetime percentiles :\n |");
899 pos = 0;
900 for (i = 0; i < len; i++) {
901 uint32_t block_info = percentiles[i];
902#define LINE_LENGTH 75
903 char str[LINE_LENGTH];
904 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
905 plist[i].u.f, block_info,
906 i == len - 1 ? '\n' : ',');
907 assert(strln < LINE_LENGTH);
908 if (pos + strln > LINE_LENGTH) {
909 pos = 0;
910 log_buf(out, "\n |");
911 }
912 log_buf(out, "%s", str);
913 pos += strln;
914#undef LINE_LENGTH
915 }
916 if (percentiles)
917 free(percentiles);
918
919 log_buf(out, " states :");
920 for (i = 0; i < BLOCK_STATE_COUNT; i++)
921 log_buf(out, " %s=%u%c",
922 block_state_names[i], block_state_counts[i],
923 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
924}
925
926static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
927{
928 char *p1, *p1alt, *p2;
929 unsigned long long bw_mean, iops_mean;
930 const int i2p = is_power_of_2(ts->kb_base);
931
932 if (!ts->ss_dur)
933 return;
934
935 bw_mean = steadystate_bw_mean(ts);
936 iops_mean = steadystate_iops_mean(ts);
937
938 p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
939 p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
940 p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
941
942 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
943 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
944 p1, p1alt, p2,
945 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
946 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
947 ts->ss_criterion.u.f,
948 ts->ss_state & FIO_SS_PCT ? "%" : "");
949
950 free(p1);
951 free(p1alt);
952 free(p2);
953}
954
955static void show_agg_stats(struct disk_util_agg *agg, int terse,
956 struct buf_output *out)
957{
958 if (!agg->slavecount)
959 return;
960
961 if (!terse) {
962 log_buf(out, ", aggrios=%llu/%llu, aggsectors=%llu/%llu, "
963 "aggrmerge=%llu/%llu, aggrticks=%llu/%llu, "
964 "aggrin_queue=%llu, aggrutil=%3.2f%%",
965 (unsigned long long) agg->ios[0] / agg->slavecount,
966 (unsigned long long) agg->ios[1] / agg->slavecount,
967 (unsigned long long) agg->sectors[0] / agg->slavecount,
968 (unsigned long long) agg->sectors[1] / agg->slavecount,
969 (unsigned long long) agg->merges[0] / agg->slavecount,
970 (unsigned long long) agg->merges[1] / agg->slavecount,
971 (unsigned long long) agg->ticks[0] / agg->slavecount,
972 (unsigned long long) agg->ticks[1] / agg->slavecount,
973 (unsigned long long) agg->time_in_queue / agg->slavecount,
974 agg->max_util.u.f);
975 } else {
976 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
977 (unsigned long long) agg->ios[0] / agg->slavecount,
978 (unsigned long long) agg->ios[1] / agg->slavecount,
979 (unsigned long long) agg->merges[0] / agg->slavecount,
980 (unsigned long long) agg->merges[1] / agg->slavecount,
981 (unsigned long long) agg->ticks[0] / agg->slavecount,
982 (unsigned long long) agg->ticks[1] / agg->slavecount,
983 (unsigned long long) agg->time_in_queue / agg->slavecount,
984 agg->max_util.u.f);
985 }
986}
987
988static void aggregate_slaves_stats(struct disk_util *masterdu)
989{
990 struct disk_util_agg *agg = &masterdu->agg;
991 struct disk_util_stat *dus;
992 struct flist_head *entry;
993 struct disk_util *slavedu;
994 double util;
995
996 flist_for_each(entry, &masterdu->slaves) {
997 slavedu = flist_entry(entry, struct disk_util, slavelist);
998 dus = &slavedu->dus;
999 agg->ios[0] += dus->s.ios[0];
1000 agg->ios[1] += dus->s.ios[1];
1001 agg->merges[0] += dus->s.merges[0];
1002 agg->merges[1] += dus->s.merges[1];
1003 agg->sectors[0] += dus->s.sectors[0];
1004 agg->sectors[1] += dus->s.sectors[1];
1005 agg->ticks[0] += dus->s.ticks[0];
1006 agg->ticks[1] += dus->s.ticks[1];
1007 agg->time_in_queue += dus->s.time_in_queue;
1008 agg->slavecount++;
1009
1010 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
1011 /* System utilization is the utilization of the
1012 * component with the highest utilization.
1013 */
1014 if (util > agg->max_util.u.f)
1015 agg->max_util.u.f = util;
1016
1017 }
1018
1019 if (agg->max_util.u.f > 100.0)
1020 agg->max_util.u.f = 100.0;
1021}
1022
1023void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
1024 int terse, struct buf_output *out)
1025{
1026 double util = 0;
1027
1028 if (dus->s.msec)
1029 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1030 if (util > 100.0)
1031 util = 100.0;
1032
1033 if (!terse) {
1034 if (agg->slavecount)
1035 log_buf(out, " ");
1036
1037 log_buf(out, " %s: ios=%llu/%llu, sectors=%llu/%llu, "
1038 "merge=%llu/%llu, ticks=%llu/%llu, in_queue=%llu, "
1039 "util=%3.2f%%",
1040 dus->name,
1041 (unsigned long long) dus->s.ios[0],
1042 (unsigned long long) dus->s.ios[1],
1043 (unsigned long long) dus->s.sectors[0],
1044 (unsigned long long) dus->s.sectors[1],
1045 (unsigned long long) dus->s.merges[0],
1046 (unsigned long long) dus->s.merges[1],
1047 (unsigned long long) dus->s.ticks[0],
1048 (unsigned long long) dus->s.ticks[1],
1049 (unsigned long long) dus->s.time_in_queue,
1050 util);
1051 } else {
1052 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1053 dus->name,
1054 (unsigned long long) dus->s.ios[0],
1055 (unsigned long long) dus->s.ios[1],
1056 (unsigned long long) dus->s.merges[0],
1057 (unsigned long long) dus->s.merges[1],
1058 (unsigned long long) dus->s.ticks[0],
1059 (unsigned long long) dus->s.ticks[1],
1060 (unsigned long long) dus->s.time_in_queue,
1061 util);
1062 }
1063
1064 /*
1065 * If the device has slaves, aggregate the stats for
1066 * those slave devices also.
1067 */
1068 show_agg_stats(agg, terse, out);
1069
1070 if (!terse)
1071 log_buf(out, "\n");
1072}
1073
1074void json_array_add_disk_util(struct disk_util_stat *dus,
1075 struct disk_util_agg *agg, struct json_array *array)
1076{
1077 struct json_object *obj;
1078 double util = 0;
1079
1080 if (dus->s.msec)
1081 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1082 if (util > 100.0)
1083 util = 100.0;
1084
1085 obj = json_create_object();
1086 json_array_add_value_object(array, obj);
1087
1088 json_object_add_value_string(obj, "name", (const char *)dus->name);
1089 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1090 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1091 json_object_add_value_int(obj, "read_sectors", dus->s.sectors[0]);
1092 json_object_add_value_int(obj, "write_sectors", dus->s.sectors[1]);
1093 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1094 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1095 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1096 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1097 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1098 json_object_add_value_float(obj, "util", util);
1099
1100 /*
1101 * If the device has slaves, aggregate the stats for
1102 * those slave devices also.
1103 */
1104 if (!agg->slavecount)
1105 return;
1106 json_object_add_value_int(obj, "aggr_read_ios",
1107 agg->ios[0] / agg->slavecount);
1108 json_object_add_value_int(obj, "aggr_write_ios",
1109 agg->ios[1] / agg->slavecount);
1110 json_object_add_value_int(obj, "aggr_read_sectors",
1111 agg->sectors[0] / agg->slavecount);
1112 json_object_add_value_int(obj, "aggr_write_sectors",
1113 agg->sectors[1] / agg->slavecount);
1114 json_object_add_value_int(obj, "aggr_read_merges",
1115 agg->merges[0] / agg->slavecount);
1116 json_object_add_value_int(obj, "aggr_write_merge",
1117 agg->merges[1] / agg->slavecount);
1118 json_object_add_value_int(obj, "aggr_read_ticks",
1119 agg->ticks[0] / agg->slavecount);
1120 json_object_add_value_int(obj, "aggr_write_ticks",
1121 agg->ticks[1] / agg->slavecount);
1122 json_object_add_value_int(obj, "aggr_in_queue",
1123 agg->time_in_queue / agg->slavecount);
1124 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1125}
1126
1127static void json_object_add_disk_utils(struct json_object *obj,
1128 struct flist_head *head)
1129{
1130 struct json_array *array = json_create_array();
1131 struct flist_head *entry;
1132 struct disk_util *du;
1133
1134 json_object_add_value_array(obj, "disk_util", array);
1135
1136 flist_for_each(entry, head) {
1137 du = flist_entry(entry, struct disk_util, list);
1138
1139 aggregate_slaves_stats(du);
1140 json_array_add_disk_util(&du->dus, &du->agg, array);
1141 }
1142}
1143
1144void show_disk_util(int terse, struct json_object *parent,
1145 struct buf_output *out)
1146{
1147 struct flist_head *entry;
1148 struct disk_util *du;
1149 bool do_json;
1150
1151 if (!is_running_backend())
1152 return;
1153
1154 if (flist_empty(&disk_list))
1155 return;
1156
1157 if ((output_format & FIO_OUTPUT_JSON) && parent)
1158 do_json = true;
1159 else
1160 do_json = false;
1161
1162 if (!terse && !do_json)
1163 log_buf(out, "\nDisk stats (read/write):\n");
1164
1165 if (do_json) {
1166 json_object_add_disk_utils(parent, &disk_list);
1167 } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1168 flist_for_each(entry, &disk_list) {
1169 du = flist_entry(entry, struct disk_util, list);
1170
1171 aggregate_slaves_stats(du);
1172 print_disk_util(&du->dus, &du->agg, terse, out);
1173 }
1174 }
1175}
1176
1177static void show_thread_status_normal(struct thread_stat *ts,
1178 struct group_run_stats *rs,
1179 struct buf_output *out)
1180{
1181 double usr_cpu, sys_cpu;
1182 unsigned long runtime;
1183 double io_u_dist[FIO_IO_U_MAP_NR];
1184 time_t time_p;
1185 char time_buf[32];
1186
1187 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1188 return;
1189
1190 memset(time_buf, 0, sizeof(time_buf));
1191
1192 time(&time_p);
1193 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1194
1195 if (!ts->error) {
1196 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1197 ts->name, ts->groupid, ts->members,
1198 ts->error, (int) ts->pid, time_buf);
1199 } else {
1200 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1201 ts->name, ts->groupid, ts->members,
1202 ts->error, ts->verror, (int) ts->pid,
1203 time_buf);
1204 }
1205
1206 if (strlen(ts->description))
1207 log_buf(out, " Description : [%s]\n", ts->description);
1208
1209 for_each_rw_ddir(ddir) {
1210 if (ts->io_bytes[ddir])
1211 show_ddir_status(rs, ts, ddir, out);
1212 }
1213
1214 if (ts->unified_rw_rep == UNIFIED_BOTH)
1215 show_mixed_ddir_status(rs, ts, out);
1216
1217 show_latencies(ts, out);
1218
1219 if (ts->sync_stat.samples)
1220 show_ddir_status(rs, ts, DDIR_SYNC, out);
1221
1222 runtime = ts->total_run_time;
1223 if (runtime) {
1224 double runt = (double) runtime;
1225
1226 usr_cpu = (double) ts->usr_time * 100 / runt;
1227 sys_cpu = (double) ts->sys_time * 100 / runt;
1228 } else {
1229 usr_cpu = 0;
1230 sys_cpu = 0;
1231 }
1232
1233 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1234 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1235 (unsigned long long) ts->ctx,
1236 (unsigned long long) ts->majf,
1237 (unsigned long long) ts->minf);
1238
1239 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1240 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1241 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1242 io_u_dist[1], io_u_dist[2],
1243 io_u_dist[3], io_u_dist[4],
1244 io_u_dist[5], io_u_dist[6]);
1245
1246 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1247 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1248 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1249 io_u_dist[1], io_u_dist[2],
1250 io_u_dist[3], io_u_dist[4],
1251 io_u_dist[5], io_u_dist[6]);
1252 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1253 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1254 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1255 io_u_dist[1], io_u_dist[2],
1256 io_u_dist[3], io_u_dist[4],
1257 io_u_dist[5], io_u_dist[6]);
1258 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
1259 " short=%llu,%llu,%llu,0"
1260 " dropped=%llu,%llu,%llu,0\n",
1261 (unsigned long long) ts->total_io_u[0],
1262 (unsigned long long) ts->total_io_u[1],
1263 (unsigned long long) ts->total_io_u[2],
1264 (unsigned long long) ts->total_io_u[3],
1265 (unsigned long long) ts->short_io_u[0],
1266 (unsigned long long) ts->short_io_u[1],
1267 (unsigned long long) ts->short_io_u[2],
1268 (unsigned long long) ts->drop_io_u[0],
1269 (unsigned long long) ts->drop_io_u[1],
1270 (unsigned long long) ts->drop_io_u[2]);
1271 if (ts->continue_on_error) {
1272 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
1273 (unsigned long long)ts->total_err_count,
1274 ts->first_error,
1275 strerror(ts->first_error));
1276 }
1277 if (ts->latency_depth) {
1278 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1279 (unsigned long long)ts->latency_target,
1280 (unsigned long long)ts->latency_window,
1281 ts->latency_percentile.u.f,
1282 ts->latency_depth);
1283 }
1284
1285 if (ts->nr_block_infos)
1286 show_block_infos(ts->nr_block_infos, ts->block_infos,
1287 ts->percentile_list, out);
1288
1289 if (ts->ss_dur)
1290 show_ss_normal(ts, out);
1291}
1292
1293static void show_ddir_status_terse(struct thread_stat *ts,
1294 struct group_run_stats *rs,
1295 enum fio_ddir ddir, int ver,
1296 struct buf_output *out)
1297{
1298 unsigned long long min, max, minv, maxv, bw, iops;
1299 unsigned long long *ovals = NULL;
1300 double mean, dev;
1301 unsigned int len;
1302 int i, bw_stat;
1303
1304 assert(ddir_rw(ddir));
1305
1306 iops = bw = 0;
1307 if (ts->runtime[ddir]) {
1308 uint64_t runt = ts->runtime[ddir];
1309
1310 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1311 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1312 }
1313
1314 log_buf(out, ";%llu;%llu;%llu;%llu",
1315 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1316 (unsigned long long) ts->runtime[ddir]);
1317
1318 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1319 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1320 else
1321 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1322
1323 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1324 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1325 else
1326 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1327
1328 if (ts->lat_percentiles) {
1329 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1330 ts->lat_stat[ddir].samples,
1331 ts->percentile_list, &ovals, &maxv,
1332 &minv);
1333 } else if (ts->clat_percentiles) {
1334 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1335 ts->clat_stat[ddir].samples,
1336 ts->percentile_list, &ovals, &maxv,
1337 &minv);
1338 } else {
1339 len = 0;
1340 }
1341
1342 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1343 if (i >= len) {
1344 log_buf(out, ";0%%=0");
1345 continue;
1346 }
1347 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1348 }
1349
1350 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1351 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1352 else
1353 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1354
1355 free(ovals);
1356
1357 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1358 if (bw_stat) {
1359 double p_of_agg = 100.0;
1360
1361 if (rs->agg[ddir]) {
1362 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1363 if (p_of_agg > 100.0)
1364 p_of_agg = 100.0;
1365 }
1366
1367 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1368 } else {
1369 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1370 }
1371
1372 if (ver == 5) {
1373 if (bw_stat)
1374 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1375 else
1376 log_buf(out, ";%lu", 0UL);
1377
1378 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1379 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1380 mean, dev, (&ts->iops_stat[ddir])->samples);
1381 else
1382 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1383 }
1384}
1385
1386static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1387 struct group_run_stats *rs,
1388 int ver, struct buf_output *out)
1389{
1390 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1391
1392 if (ts_lcl)
1393 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1394
1395 free_clat_prio_stats(ts_lcl);
1396 free(ts_lcl);
1397}
1398
1399static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1400 uint32_t percentiles,
1401 struct io_stat *lat_stat,
1402 uint64_t *io_u_plat)
1403{
1404 char buf[120];
1405 double mean, dev;
1406 unsigned int i, len;
1407 struct json_object *lat_object, *percentile_object, *clat_bins_object;
1408 unsigned long long min, max, maxv, minv, *ovals = NULL;
1409
1410 if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1411 min = max = 0;
1412 mean = dev = 0.0;
1413 }
1414 lat_object = json_create_object();
1415 json_object_add_value_int(lat_object, "min", min);
1416 json_object_add_value_int(lat_object, "max", max);
1417 json_object_add_value_float(lat_object, "mean", mean);
1418 json_object_add_value_float(lat_object, "stddev", dev);
1419 json_object_add_value_int(lat_object, "N", lat_stat->samples);
1420
1421 if (percentiles && lat_stat->samples) {
1422 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1423 ts->percentile_list, &ovals, &maxv, &minv);
1424
1425 if (len > FIO_IO_U_LIST_MAX_LEN)
1426 len = FIO_IO_U_LIST_MAX_LEN;
1427
1428 percentile_object = json_create_object();
1429 json_object_add_value_object(lat_object, "percentile", percentile_object);
1430 for (i = 0; i < len; i++) {
1431 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1432 json_object_add_value_int(percentile_object, buf, ovals[i]);
1433 }
1434 free(ovals);
1435
1436 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1437 clat_bins_object = json_create_object();
1438 json_object_add_value_object(lat_object, "bins", clat_bins_object);
1439
1440 for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1441 if (io_u_plat[i]) {
1442 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1443 json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1444 }
1445 }
1446 }
1447
1448 return lat_object;
1449}
1450
1451static void add_ddir_status_json(struct thread_stat *ts,
1452 struct group_run_stats *rs, enum fio_ddir ddir,
1453 struct json_object *parent)
1454{
1455 unsigned long long min, max;
1456 unsigned long long bw_bytes, bw;
1457 double mean, dev, iops;
1458 struct json_object *dir_object, *tmp_object;
1459 double p_of_agg = 100.0;
1460
1461 assert(ddir_rw(ddir) || ddir_sync(ddir));
1462
1463 if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1464 return;
1465
1466 dir_object = json_create_object();
1467 json_object_add_value_object(parent,
1468 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1469
1470 if (ddir_rw(ddir)) {
1471 bw_bytes = 0;
1472 bw = 0;
1473 iops = 0.0;
1474 if (ts->runtime[ddir]) {
1475 uint64_t runt = ts->runtime[ddir];
1476
1477 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1478 bw = bw_bytes / 1024; /* KiB/s */
1479 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1480 }
1481
1482 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1483 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1484 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1485 json_object_add_value_int(dir_object, "bw", bw);
1486 json_object_add_value_float(dir_object, "iops", iops);
1487 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1488 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1489 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1490 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1491
1492 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1493 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1494 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1495
1496 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1497 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1498 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1499
1500 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1501 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1502 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1503 } else {
1504 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1505 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1506 &ts->sync_stat, ts->io_u_sync_plat);
1507 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1508 }
1509
1510 if (!ddir_rw(ddir))
1511 return;
1512
1513 /* Only include per prio stats if there are >= 2 prios with samples */
1514 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
1515 struct json_array *array = json_create_array();
1516 const char *obj_name;
1517 int i;
1518
1519 if (ts->lat_percentiles)
1520 obj_name = "lat_ns";
1521 else
1522 obj_name = "clat_ns";
1523
1524 json_object_add_value_array(dir_object, "prios", array);
1525
1526 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
1527 struct json_object *obj;
1528
1529 if (!ts->clat_prio[ddir][i].clat_stat.samples)
1530 continue;
1531
1532 obj = json_create_object();
1533
1534 json_object_add_value_int(obj, "prioclass",
1535 ioprio_class(ts->clat_prio[ddir][i].ioprio));
1536 json_object_add_value_int(obj, "prio",
1537 ioprio(ts->clat_prio[ddir][i].ioprio));
1538 json_object_add_value_int(obj, "priohint",
1539 ioprio_hint(ts->clat_prio[ddir][i].ioprio));
1540
1541 tmp_object = add_ddir_lat_json(ts,
1542 ts->clat_percentiles | ts->lat_percentiles,
1543 &ts->clat_prio[ddir][i].clat_stat,
1544 ts->clat_prio[ddir][i].io_u_plat);
1545 json_object_add_value_object(obj, obj_name, tmp_object);
1546 json_array_add_value_object(array, obj);
1547 }
1548 }
1549
1550 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1551 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1552 } else {
1553 min = max = 0;
1554 p_of_agg = mean = dev = 0.0;
1555 }
1556
1557 json_object_add_value_int(dir_object, "bw_min", min);
1558 json_object_add_value_int(dir_object, "bw_max", max);
1559 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1560 json_object_add_value_float(dir_object, "bw_mean", mean);
1561 json_object_add_value_float(dir_object, "bw_dev", dev);
1562 json_object_add_value_int(dir_object, "bw_samples",
1563 (&ts->bw_stat[ddir])->samples);
1564
1565 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1566 min = max = 0;
1567 mean = dev = 0.0;
1568 }
1569 json_object_add_value_int(dir_object, "iops_min", min);
1570 json_object_add_value_int(dir_object, "iops_max", max);
1571 json_object_add_value_float(dir_object, "iops_mean", mean);
1572 json_object_add_value_float(dir_object, "iops_stddev", dev);
1573 json_object_add_value_int(dir_object, "iops_samples",
1574 (&ts->iops_stat[ddir])->samples);
1575
1576 if (ts->cachehit + ts->cachemiss) {
1577 uint64_t total;
1578 double hit;
1579
1580 total = ts->cachehit + ts->cachemiss;
1581 hit = (double) ts->cachehit / (double) total;
1582 hit *= 100.0;
1583 json_object_add_value_float(dir_object, "cachehit", hit);
1584 }
1585}
1586
1587static void add_mixed_ddir_status_json(struct thread_stat *ts,
1588 struct group_run_stats *rs, struct json_object *parent)
1589{
1590 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1591
1592 /* add the aggregated stats to json parent */
1593 if (ts_lcl)
1594 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1595
1596 free_clat_prio_stats(ts_lcl);
1597 free(ts_lcl);
1598}
1599
1600static void show_thread_status_terse_all(struct thread_stat *ts,
1601 struct group_run_stats *rs, int ver,
1602 struct buf_output *out)
1603{
1604 double io_u_dist[FIO_IO_U_MAP_NR];
1605 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1606 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1607 double usr_cpu, sys_cpu;
1608 int i;
1609
1610 /* General Info */
1611 if (ver == 2)
1612 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1613 else
1614 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1615 ts->name, ts->groupid, ts->error);
1616
1617 /* Log Read Status, or mixed if unified_rw_rep = 1 */
1618 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1619 if (ts->unified_rw_rep != UNIFIED_MIXED) {
1620 /* Log Write Status */
1621 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1622 /* Log Trim Status */
1623 if (ver == 2 || ver == 4 || ver == 5)
1624 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1625 }
1626 if (ts->unified_rw_rep == UNIFIED_BOTH)
1627 show_mixed_ddir_status_terse(ts, rs, ver, out);
1628 /* CPU Usage */
1629 if (ts->total_run_time) {
1630 double runt = (double) ts->total_run_time;
1631
1632 usr_cpu = (double) ts->usr_time * 100 / runt;
1633 sys_cpu = (double) ts->sys_time * 100 / runt;
1634 } else {
1635 usr_cpu = 0;
1636 sys_cpu = 0;
1637 }
1638
1639 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1640 (unsigned long long) ts->ctx,
1641 (unsigned long long) ts->majf,
1642 (unsigned long long) ts->minf);
1643
1644 /* Calc % distribution of IO depths, usecond, msecond latency */
1645 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1646 stat_calc_lat_nu(ts, io_u_lat_u);
1647 stat_calc_lat_m(ts, io_u_lat_m);
1648
1649 /* Only show fixed 7 I/O depth levels*/
1650 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1651 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1652 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1653
1654 /* Microsecond latency */
1655 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1656 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1657 /* Millisecond latency */
1658 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1659 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1660
1661 /* disk util stats, if any */
1662 if (ver >= 3 && is_running_backend())
1663 show_disk_util(1, NULL, out);
1664
1665 /* Additional output if continue_on_error set - default off*/
1666 if (ts->continue_on_error)
1667 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1668
1669 /* Additional output if description is set */
1670 if (strlen(ts->description)) {
1671 if (ver == 2)
1672 log_buf(out, "\n");
1673 log_buf(out, ";%s", ts->description);
1674 }
1675
1676 log_buf(out, "\n");
1677}
1678
1679static void json_add_job_opts(struct json_object *root, const char *name,
1680 struct flist_head *opt_list)
1681{
1682 struct json_object *dir_object;
1683 struct flist_head *entry;
1684 struct print_option *p;
1685
1686 if (flist_empty(opt_list))
1687 return;
1688
1689 dir_object = json_create_object();
1690 json_object_add_value_object(root, name, dir_object);
1691
1692 flist_for_each(entry, opt_list) {
1693 p = flist_entry(entry, struct print_option, list);
1694 json_object_add_value_string(dir_object, p->name, p->value);
1695 }
1696}
1697
1698static struct json_object *show_thread_status_json(struct thread_stat *ts,
1699 struct group_run_stats *rs,
1700 struct flist_head *opt_list)
1701{
1702 struct json_object *root, *tmp;
1703 struct jobs_eta *je;
1704 double io_u_dist[FIO_IO_U_MAP_NR];
1705 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1706 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1707 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1708 double usr_cpu, sys_cpu;
1709 int i;
1710 size_t size;
1711
1712 root = json_create_object();
1713 json_object_add_value_string(root, "jobname", ts->name);
1714 json_object_add_value_int(root, "groupid", ts->groupid);
1715 json_object_add_value_int(root, "job_start", ts->job_start);
1716 json_object_add_value_int(root, "error", ts->error);
1717
1718 /* ETA Info */
1719 je = get_jobs_eta(true, &size);
1720 if (je) {
1721 json_object_add_value_int(root, "eta", je->eta_sec);
1722 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1723 free(je);
1724 }
1725
1726 if (opt_list)
1727 json_add_job_opts(root, "job options", opt_list);
1728
1729 add_ddir_status_json(ts, rs, DDIR_READ, root);
1730 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1731 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1732 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1733
1734 if (ts->unified_rw_rep == UNIFIED_BOTH)
1735 add_mixed_ddir_status_json(ts, rs, root);
1736
1737 /* CPU Usage */
1738 if (ts->total_run_time) {
1739 double runt = (double) ts->total_run_time;
1740
1741 usr_cpu = (double) ts->usr_time * 100 / runt;
1742 sys_cpu = (double) ts->sys_time * 100 / runt;
1743 } else {
1744 usr_cpu = 0;
1745 sys_cpu = 0;
1746 }
1747 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1748 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1749 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1750 json_object_add_value_int(root, "ctx", ts->ctx);
1751 json_object_add_value_int(root, "majf", ts->majf);
1752 json_object_add_value_int(root, "minf", ts->minf);
1753
1754 /* Calc % distribution of IO depths */
1755 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1756 tmp = json_create_object();
1757 json_object_add_value_object(root, "iodepth_level", tmp);
1758 /* Only show fixed 7 I/O depth levels*/
1759 for (i = 0; i < 7; i++) {
1760 char name[20];
1761 if (i < 6)
1762 snprintf(name, 20, "%d", 1 << i);
1763 else
1764 snprintf(name, 20, ">=%d", 1 << i);
1765 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1766 }
1767
1768 /* Calc % distribution of submit IO depths */
1769 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1770 tmp = json_create_object();
1771 json_object_add_value_object(root, "iodepth_submit", tmp);
1772 /* Only show fixed 7 I/O depth levels*/
1773 for (i = 0; i < 7; i++) {
1774 char name[20];
1775 if (i == 0)
1776 snprintf(name, 20, "0");
1777 else if (i < 6)
1778 snprintf(name, 20, "%d", 1 << (i+1));
1779 else
1780 snprintf(name, 20, ">=%d", 1 << i);
1781 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1782 }
1783
1784 /* Calc % distribution of completion IO depths */
1785 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1786 tmp = json_create_object();
1787 json_object_add_value_object(root, "iodepth_complete", tmp);
1788 /* Only show fixed 7 I/O depth levels*/
1789 for (i = 0; i < 7; i++) {
1790 char name[20];
1791 if (i == 0)
1792 snprintf(name, 20, "0");
1793 else if (i < 6)
1794 snprintf(name, 20, "%d", 1 << (i+1));
1795 else
1796 snprintf(name, 20, ">=%d", 1 << i);
1797 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1798 }
1799
1800 /* Calc % distribution of nsecond, usecond, msecond latency */
1801 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1802 stat_calc_lat_n(ts, io_u_lat_n);
1803 stat_calc_lat_u(ts, io_u_lat_u);
1804 stat_calc_lat_m(ts, io_u_lat_m);
1805
1806 /* Nanosecond latency */
1807 tmp = json_create_object();
1808 json_object_add_value_object(root, "latency_ns", tmp);
1809 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1810 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1811 "250", "500", "750", "1000", };
1812 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1813 }
1814 /* Microsecond latency */
1815 tmp = json_create_object();
1816 json_object_add_value_object(root, "latency_us", tmp);
1817 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1818 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1819 "250", "500", "750", "1000", };
1820 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1821 }
1822 /* Millisecond latency */
1823 tmp = json_create_object();
1824 json_object_add_value_object(root, "latency_ms", tmp);
1825 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1826 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1827 "250", "500", "750", "1000", "2000",
1828 ">=2000", };
1829 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1830 }
1831
1832 /* Additional output if continue_on_error set - default off*/
1833 if (ts->continue_on_error) {
1834 json_object_add_value_int(root, "total_err", ts->total_err_count);
1835 json_object_add_value_int(root, "first_error", ts->first_error);
1836 }
1837
1838 if (ts->latency_depth) {
1839 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1840 json_object_add_value_int(root, "latency_target", ts->latency_target);
1841 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1842 json_object_add_value_int(root, "latency_window", ts->latency_window);
1843 }
1844
1845 /* Additional output if description is set */
1846 if (strlen(ts->description))
1847 json_object_add_value_string(root, "desc", ts->description);
1848
1849 if (ts->nr_block_infos) {
1850 /* Block error histogram and types */
1851 int len;
1852 unsigned int *percentiles = NULL;
1853 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1854
1855 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1856 ts->percentile_list,
1857 &percentiles, block_state_counts);
1858
1859 if (len) {
1860 struct json_object *block, *percentile_object, *states;
1861 int state;
1862 block = json_create_object();
1863 json_object_add_value_object(root, "block", block);
1864
1865 percentile_object = json_create_object();
1866 json_object_add_value_object(block, "percentiles",
1867 percentile_object);
1868 for (i = 0; i < len; i++) {
1869 char buf[20];
1870 snprintf(buf, sizeof(buf), "%f",
1871 ts->percentile_list[i].u.f);
1872 json_object_add_value_int(percentile_object,
1873 buf,
1874 percentiles[i]);
1875 }
1876
1877 states = json_create_object();
1878 json_object_add_value_object(block, "states", states);
1879 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1880 json_object_add_value_int(states,
1881 block_state_names[state],
1882 block_state_counts[state]);
1883 }
1884 free(percentiles);
1885 }
1886 }
1887
1888 if (ts->ss_dur) {
1889 struct json_object *data;
1890 struct json_array *iops, *bw;
1891 int j, k, l;
1892 char ss_buf[64];
1893 int intervals = ts->ss_dur / (ss_check_interval / 1000L);
1894
1895 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1896 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1897 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1898 (float) ts->ss_limit.u.f,
1899 ts->ss_state & FIO_SS_PCT ? "%" : "");
1900
1901 tmp = json_create_object();
1902 json_object_add_value_object(root, "steadystate", tmp);
1903 json_object_add_value_string(tmp, "ss", ss_buf);
1904 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1905 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1906
1907 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1908 ts->ss_state & FIO_SS_PCT ? "%" : "");
1909 json_object_add_value_string(tmp, "criterion", ss_buf);
1910 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1911 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1912
1913 data = json_create_object();
1914 json_object_add_value_object(tmp, "data", data);
1915 bw = json_create_array();
1916 iops = json_create_array();
1917
1918 /*
1919 ** if ss was attained or the buffer is not full,
1920 ** ss->head points to the first element in the list.
1921 ** otherwise it actually points to the second element
1922 ** in the list
1923 */
1924 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1925 j = ts->ss_head;
1926 else
1927 j = ts->ss_head == 0 ? intervals - 1 : ts->ss_head - 1;
1928 for (l = 0; l < intervals; l++) {
1929 k = (j + l) % intervals;
1930 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1931 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1932 }
1933 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1934 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1935 json_object_add_value_array(data, "iops", iops);
1936 json_object_add_value_array(data, "bw", bw);
1937 }
1938
1939 return root;
1940}
1941
1942static void show_thread_status_terse(struct thread_stat *ts,
1943 struct group_run_stats *rs,
1944 struct buf_output *out)
1945{
1946 if (terse_version >= 2 && terse_version <= 5)
1947 show_thread_status_terse_all(ts, rs, terse_version, out);
1948 else
1949 log_err("fio: bad terse version!? %d\n", terse_version);
1950}
1951
1952struct json_object *show_thread_status(struct thread_stat *ts,
1953 struct group_run_stats *rs,
1954 struct flist_head *opt_list,
1955 struct buf_output *out)
1956{
1957 struct json_object *ret = NULL;
1958
1959 if (output_format & FIO_OUTPUT_TERSE)
1960 show_thread_status_terse(ts, rs, out);
1961 if (output_format & FIO_OUTPUT_JSON)
1962 ret = show_thread_status_json(ts, rs, opt_list);
1963 if (output_format & FIO_OUTPUT_NORMAL)
1964 show_thread_status_normal(ts, rs, out);
1965
1966 return ret;
1967}
1968
1969static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1970{
1971 double mean, S;
1972
1973 dst->min_val = min(dst->min_val, src->min_val);
1974 dst->max_val = max(dst->max_val, src->max_val);
1975
1976 /*
1977 * Compute new mean and S after the merge
1978 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1979 * #Parallel_algorithm>
1980 */
1981 if (first) {
1982 mean = src->mean.u.f;
1983 S = src->S.u.f;
1984 } else {
1985 double delta = src->mean.u.f - dst->mean.u.f;
1986
1987 mean = ((src->mean.u.f * src->samples) +
1988 (dst->mean.u.f * dst->samples)) /
1989 (dst->samples + src->samples);
1990
1991 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1992 (dst->samples * src->samples) /
1993 (dst->samples + src->samples);
1994 }
1995
1996 dst->samples += src->samples;
1997 dst->mean.u.f = mean;
1998 dst->S.u.f = S;
1999
2000}
2001
2002/*
2003 * We sum two kinds of stats - one that is time based, in which case we
2004 * apply the proper summing technique, and then one that is iops/bw
2005 * numbers. For group_reporting, we should just add those up, not make
2006 * them the mean of everything.
2007 */
2008static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
2009{
2010 bool first = dst->samples == 0;
2011
2012 if (src->samples == 0)
2013 return;
2014
2015 if (!pure_sum) {
2016 __sum_stat(dst, src, first);
2017 return;
2018 }
2019
2020 if (first) {
2021 dst->min_val = src->min_val;
2022 dst->max_val = src->max_val;
2023 dst->samples = src->samples;
2024 dst->mean.u.f = src->mean.u.f;
2025 dst->S.u.f = src->S.u.f;
2026 } else {
2027 dst->min_val += src->min_val;
2028 dst->max_val += src->max_val;
2029 dst->samples += src->samples;
2030 dst->mean.u.f += src->mean.u.f;
2031 dst->S.u.f += src->S.u.f;
2032 }
2033}
2034
2035void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
2036{
2037 int i;
2038
2039 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2040 if (dst->max_run[i] < src->max_run[i])
2041 dst->max_run[i] = src->max_run[i];
2042 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
2043 dst->min_run[i] = src->min_run[i];
2044 if (dst->max_bw[i] < src->max_bw[i])
2045 dst->max_bw[i] = src->max_bw[i];
2046 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
2047 dst->min_bw[i] = src->min_bw[i];
2048
2049 dst->iobytes[i] += src->iobytes[i];
2050 dst->agg[i] += src->agg[i];
2051 }
2052
2053 if (!dst->kb_base)
2054 dst->kb_base = src->kb_base;
2055 if (!dst->unit_base)
2056 dst->unit_base = src->unit_base;
2057 if (!dst->sig_figs)
2058 dst->sig_figs = src->sig_figs;
2059}
2060
2061/*
2062 * Free the clat_prio_stat arrays allocated by alloc_clat_prio_stat_ddir().
2063 */
2064void free_clat_prio_stats(struct thread_stat *ts)
2065{
2066 enum fio_ddir ddir;
2067
2068 if (!ts)
2069 return;
2070
2071 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2072 sfree(ts->clat_prio[ddir]);
2073 ts->clat_prio[ddir] = NULL;
2074 ts->nr_clat_prio[ddir] = 0;
2075 }
2076}
2077
2078/*
2079 * Allocate a clat_prio_stat array. The array has to be allocated/freed using
2080 * smalloc/sfree, so that it is accessible by the process/thread summing the
2081 * thread_stats.
2082 */
2083int alloc_clat_prio_stat_ddir(struct thread_stat *ts, enum fio_ddir ddir,
2084 int nr_prios)
2085{
2086 struct clat_prio_stat *clat_prio;
2087 int i;
2088
2089 clat_prio = scalloc(nr_prios, sizeof(*ts->clat_prio[ddir]));
2090 if (!clat_prio) {
2091 log_err("fio: failed to allocate ts clat data\n");
2092 return 1;
2093 }
2094
2095 for (i = 0; i < nr_prios; i++)
2096 clat_prio[i].clat_stat.min_val = ULONG_MAX;
2097
2098 ts->clat_prio[ddir] = clat_prio;
2099 ts->nr_clat_prio[ddir] = nr_prios;
2100
2101 return 0;
2102}
2103
2104static int grow_clat_prio_stat(struct thread_stat *dst, enum fio_ddir ddir)
2105{
2106 int curr_len = dst->nr_clat_prio[ddir];
2107 void *new_arr;
2108
2109 new_arr = scalloc(curr_len + 1, sizeof(*dst->clat_prio[ddir]));
2110 if (!new_arr) {
2111 log_err("fio: failed to grow clat prio array\n");
2112 return 1;
2113 }
2114
2115 memcpy(new_arr, dst->clat_prio[ddir],
2116 curr_len * sizeof(*dst->clat_prio[ddir]));
2117 sfree(dst->clat_prio[ddir]);
2118
2119 dst->clat_prio[ddir] = new_arr;
2120 dst->clat_prio[ddir][curr_len].clat_stat.min_val = ULONG_MAX;
2121 dst->nr_clat_prio[ddir]++;
2122
2123 return 0;
2124}
2125
2126static int find_clat_prio_index(struct thread_stat *dst, enum fio_ddir ddir,
2127 uint32_t ioprio)
2128{
2129 int i, nr_prios = dst->nr_clat_prio[ddir];
2130
2131 for (i = 0; i < nr_prios; i++) {
2132 if (dst->clat_prio[ddir][i].ioprio == ioprio)
2133 return i;
2134 }
2135
2136 return -1;
2137}
2138
2139static int alloc_or_get_clat_prio_index(struct thread_stat *dst,
2140 enum fio_ddir ddir, uint32_t ioprio,
2141 int *idx)
2142{
2143 int index = find_clat_prio_index(dst, ddir, ioprio);
2144
2145 if (index == -1) {
2146 index = dst->nr_clat_prio[ddir];
2147
2148 if (grow_clat_prio_stat(dst, ddir))
2149 return 1;
2150
2151 dst->clat_prio[ddir][index].ioprio = ioprio;
2152 }
2153
2154 *idx = index;
2155
2156 return 0;
2157}
2158
2159static int clat_prio_stats_copy(struct thread_stat *dst, struct thread_stat *src,
2160 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2161{
2162 size_t sz = sizeof(*src->clat_prio[src_ddir]) *
2163 src->nr_clat_prio[src_ddir];
2164
2165 dst->clat_prio[dst_ddir] = smalloc(sz);
2166 if (!dst->clat_prio[dst_ddir]) {
2167 log_err("fio: failed to alloc clat prio array\n");
2168 return 1;
2169 }
2170
2171 memcpy(dst->clat_prio[dst_ddir], src->clat_prio[src_ddir], sz);
2172 dst->nr_clat_prio[dst_ddir] = src->nr_clat_prio[src_ddir];
2173
2174 return 0;
2175}
2176
2177static int clat_prio_stat_add_samples(struct thread_stat *dst,
2178 enum fio_ddir dst_ddir, uint32_t ioprio,
2179 struct io_stat *io_stat,
2180 uint64_t *io_u_plat)
2181{
2182 int i, dst_index;
2183
2184 if (!io_stat->samples)
2185 return 0;
2186
2187 if (alloc_or_get_clat_prio_index(dst, dst_ddir, ioprio, &dst_index))
2188 return 1;
2189
2190 sum_stat(&dst->clat_prio[dst_ddir][dst_index].clat_stat, io_stat,
2191 false);
2192
2193 for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
2194 dst->clat_prio[dst_ddir][dst_index].io_u_plat[i] += io_u_plat[i];
2195
2196 return 0;
2197}
2198
2199static int sum_clat_prio_stats_src_single_prio(struct thread_stat *dst,
2200 struct thread_stat *src,
2201 enum fio_ddir dst_ddir,
2202 enum fio_ddir src_ddir)
2203{
2204 struct io_stat *io_stat;
2205 uint64_t *io_u_plat;
2206
2207 /*
2208 * If src ts has no clat_prio_stat array, then all I/Os were submitted
2209 * using src->ioprio. Thus, the global samples in src->clat_stat (or
2210 * src->lat_stat) can be used as the 'per prio' samples for src->ioprio.
2211 */
2212 assert(!src->clat_prio[src_ddir]);
2213 assert(src->nr_clat_prio[src_ddir] == 0);
2214
2215 if (src->lat_percentiles) {
2216 io_u_plat = src->io_u_plat[FIO_LAT][src_ddir];
2217 io_stat = &src->lat_stat[src_ddir];
2218 } else {
2219 io_u_plat = src->io_u_plat[FIO_CLAT][src_ddir];
2220 io_stat = &src->clat_stat[src_ddir];
2221 }
2222
2223 return clat_prio_stat_add_samples(dst, dst_ddir, src->ioprio, io_stat,
2224 io_u_plat);
2225}
2226
2227static int sum_clat_prio_stats_src_multi_prio(struct thread_stat *dst,
2228 struct thread_stat *src,
2229 enum fio_ddir dst_ddir,
2230 enum fio_ddir src_ddir)
2231{
2232 int i;
2233
2234 /*
2235 * If src ts has a clat_prio_stat array, then there are multiple prios
2236 * in use (i.e. src ts had cmdprio_percentage or cmdprio_bssplit set).
2237 * The samples for the default prio will exist in the src->clat_prio
2238 * array, just like the samples for any other prio.
2239 */
2240 assert(src->clat_prio[src_ddir]);
2241 assert(src->nr_clat_prio[src_ddir]);
2242
2243 /* If the dst ts doesn't yet have a clat_prio array, simply memcpy. */
2244 if (!dst->clat_prio[dst_ddir])
2245 return clat_prio_stats_copy(dst, src, dst_ddir, src_ddir);
2246
2247 /* The dst ts already has a clat_prio_array, add src stats into it. */
2248 for (i = 0; i < src->nr_clat_prio[src_ddir]; i++) {
2249 struct io_stat *io_stat = &src->clat_prio[src_ddir][i].clat_stat;
2250 uint64_t *io_u_plat = src->clat_prio[src_ddir][i].io_u_plat;
2251 uint32_t ioprio = src->clat_prio[src_ddir][i].ioprio;
2252
2253 if (clat_prio_stat_add_samples(dst, dst_ddir, ioprio, io_stat, io_u_plat))
2254 return 1;
2255 }
2256
2257 return 0;
2258}
2259
2260static int sum_clat_prio_stats(struct thread_stat *dst, struct thread_stat *src,
2261 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2262{
2263 if (dst->disable_prio_stat)
2264 return 0;
2265
2266 if (!src->clat_prio[src_ddir])
2267 return sum_clat_prio_stats_src_single_prio(dst, src, dst_ddir,
2268 src_ddir);
2269
2270 return sum_clat_prio_stats_src_multi_prio(dst, src, dst_ddir, src_ddir);
2271}
2272
2273void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
2274{
2275 int k, l, m;
2276
2277 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2278 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2279 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2280 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2281 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2282 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2283 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2284 sum_clat_prio_stats(dst, src, l, l);
2285
2286 dst->io_bytes[l] += src->io_bytes[l];
2287
2288 if (dst->runtime[l] < src->runtime[l])
2289 dst->runtime[l] = src->runtime[l];
2290 } else {
2291 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2292 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2293 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2294 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2295 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2296 sum_clat_prio_stats(dst, src, 0, l);
2297
2298 dst->io_bytes[0] += src->io_bytes[l];
2299
2300 if (dst->runtime[0] < src->runtime[l])
2301 dst->runtime[0] = src->runtime[l];
2302 }
2303 }
2304
2305 sum_stat(&dst->sync_stat, &src->sync_stat, false);
2306 dst->usr_time += src->usr_time;
2307 dst->sys_time += src->sys_time;
2308 dst->ctx += src->ctx;
2309 dst->majf += src->majf;
2310 dst->minf += src->minf;
2311
2312 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2313 dst->io_u_map[k] += src->io_u_map[k];
2314 dst->io_u_submit[k] += src->io_u_submit[k];
2315 dst->io_u_complete[k] += src->io_u_complete[k];
2316 }
2317
2318 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2319 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2320 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2321 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2322 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2323 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2324
2325 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2326 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2327 dst->total_io_u[k] += src->total_io_u[k];
2328 dst->short_io_u[k] += src->short_io_u[k];
2329 dst->drop_io_u[k] += src->drop_io_u[k];
2330 } else {
2331 dst->total_io_u[0] += src->total_io_u[k];
2332 dst->short_io_u[0] += src->short_io_u[k];
2333 dst->drop_io_u[0] += src->drop_io_u[k];
2334 }
2335 }
2336
2337 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2338
2339 for (k = 0; k < FIO_LAT_CNT; k++)
2340 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2341 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2342 if (dst->unified_rw_rep != UNIFIED_MIXED)
2343 dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2344 else
2345 dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2346
2347 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2348 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2349
2350 dst->total_run_time += src->total_run_time;
2351 dst->total_submit += src->total_submit;
2352 dst->total_complete += src->total_complete;
2353 dst->nr_zone_resets += src->nr_zone_resets;
2354 dst->cachehit += src->cachehit;
2355 dst->cachemiss += src->cachemiss;
2356}
2357
2358void init_group_run_stat(struct group_run_stats *gs)
2359{
2360 int i;
2361 memset(gs, 0, sizeof(*gs));
2362
2363 for (i = 0; i < DDIR_RWDIR_CNT; i++)
2364 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2365}
2366
2367void init_thread_stat_min_vals(struct thread_stat *ts)
2368{
2369 int i;
2370
2371 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2372 ts->clat_stat[i].min_val = ULONG_MAX;
2373 ts->slat_stat[i].min_val = ULONG_MAX;
2374 ts->lat_stat[i].min_val = ULONG_MAX;
2375 ts->bw_stat[i].min_val = ULONG_MAX;
2376 ts->iops_stat[i].min_val = ULONG_MAX;
2377 }
2378 ts->sync_stat.min_val = ULONG_MAX;
2379}
2380
2381void init_thread_stat(struct thread_stat *ts)
2382{
2383 memset(ts, 0, sizeof(*ts));
2384
2385 init_thread_stat_min_vals(ts);
2386 ts->groupid = -1;
2387}
2388
2389static void init_per_prio_stats(struct thread_stat *threadstats, int nr_ts)
2390{
2391 struct thread_stat *ts;
2392 int i, j, last_ts, idx;
2393 enum fio_ddir ddir;
2394
2395 j = 0;
2396 last_ts = -1;
2397 idx = 0;
2398
2399 /*
2400 * Loop through all tds, if a td requires per prio stats, temporarily
2401 * store a 1 in ts->disable_prio_stat, and then do an additional
2402 * loop at the end where we invert the ts->disable_prio_stat values.
2403 */
2404 for_each_td(td) {
2405 if (!td->o.stats)
2406 continue;
2407 if (idx &&
2408 (!td->o.group_reporting ||
2409 (td->o.group_reporting && last_ts != td->groupid))) {
2410 idx = 0;
2411 j++;
2412 }
2413
2414 last_ts = td->groupid;
2415 ts = &threadstats[j];
2416
2417 /* idx == 0 means first td in group, or td is not in a group. */
2418 if (idx == 0)
2419 ts->ioprio = td->ioprio;
2420 else if (td->ioprio != ts->ioprio)
2421 ts->disable_prio_stat = 1;
2422
2423 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2424 if (td->ts.clat_prio[ddir]) {
2425 ts->disable_prio_stat = 1;
2426 break;
2427 }
2428 }
2429
2430 idx++;
2431 } end_for_each();
2432
2433 /* Loop through all dst threadstats and fixup the values. */
2434 for (i = 0; i < nr_ts; i++) {
2435 ts = &threadstats[i];
2436 ts->disable_prio_stat = !ts->disable_prio_stat;
2437 }
2438}
2439
2440void __show_run_stats(void)
2441{
2442 struct group_run_stats *runstats, *rs;
2443 struct thread_stat *threadstats, *ts;
2444 int i, j, k, nr_ts, last_ts, idx;
2445 bool kb_base_warned = false;
2446 bool unit_base_warned = false;
2447 struct json_object *root = NULL;
2448 struct json_array *array = NULL;
2449 struct buf_output output[FIO_OUTPUT_NR];
2450 struct flist_head **opt_lists;
2451
2452 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2453
2454 for (i = 0; i < groupid + 1; i++)
2455 init_group_run_stat(&runstats[i]);
2456
2457 /*
2458 * find out how many threads stats we need. if group reporting isn't
2459 * enabled, it's one-per-td.
2460 */
2461 nr_ts = 0;
2462 last_ts = -1;
2463 for_each_td(td) {
2464 if (!td->o.group_reporting) {
2465 nr_ts++;
2466 continue;
2467 }
2468 if (last_ts == td->groupid)
2469 continue;
2470 if (!td->o.stats)
2471 continue;
2472
2473 last_ts = td->groupid;
2474 nr_ts++;
2475 } end_for_each();
2476
2477 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2478 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2479
2480 for (i = 0; i < nr_ts; i++) {
2481 init_thread_stat(&threadstats[i]);
2482 opt_lists[i] = NULL;
2483 }
2484
2485 init_per_prio_stats(threadstats, nr_ts);
2486
2487 j = 0;
2488 last_ts = -1;
2489 idx = 0;
2490 for_each_td(td) {
2491 if (!td->o.stats)
2492 continue;
2493 if (idx && (!td->o.group_reporting ||
2494 (td->o.group_reporting && last_ts != td->groupid))) {
2495 idx = 0;
2496 j++;
2497 }
2498
2499 last_ts = td->groupid;
2500
2501 ts = &threadstats[j];
2502
2503 ts->clat_percentiles = td->o.clat_percentiles;
2504 ts->lat_percentiles = td->o.lat_percentiles;
2505 ts->slat_percentiles = td->o.slat_percentiles;
2506 ts->percentile_precision = td->o.percentile_precision;
2507 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2508 opt_lists[j] = &td->opt_list;
2509
2510 idx++;
2511
2512 if (ts->groupid == -1) {
2513 /*
2514 * These are per-group shared already
2515 */
2516 snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2517 if (td->o.description)
2518 snprintf(ts->description,
2519 sizeof(ts->description), "%s",
2520 td->o.description);
2521 else
2522 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2523
2524 /*
2525 * If multiple entries in this group, this is
2526 * the first member.
2527 */
2528 ts->thread_number = td->thread_number;
2529 ts->groupid = td->groupid;
2530 ts->job_start = td->job_start;
2531
2532 /*
2533 * first pid in group, not very useful...
2534 */
2535 ts->pid = td->pid;
2536
2537 ts->kb_base = td->o.kb_base;
2538 ts->unit_base = td->o.unit_base;
2539 ts->sig_figs = td->o.sig_figs;
2540 ts->unified_rw_rep = td->o.unified_rw_rep;
2541 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2542 log_info("fio: kb_base differs for jobs in group, using"
2543 " %u as the base\n", ts->kb_base);
2544 kb_base_warned = true;
2545 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2546 log_info("fio: unit_base differs for jobs in group, using"
2547 " %u as the base\n", ts->unit_base);
2548 unit_base_warned = true;
2549 }
2550
2551 ts->continue_on_error = td->o.continue_on_error;
2552 ts->total_err_count += td->total_err_count;
2553 ts->first_error = td->first_error;
2554 if (!ts->error) {
2555 if (!td->error && td->o.continue_on_error &&
2556 td->first_error) {
2557 ts->error = td->first_error;
2558 snprintf(ts->verror, sizeof(ts->verror), "%s",
2559 td->verror);
2560 } else if (td->error) {
2561 ts->error = td->error;
2562 snprintf(ts->verror, sizeof(ts->verror), "%s",
2563 td->verror);
2564 }
2565 }
2566
2567 ts->latency_depth = td->latency_qd;
2568 ts->latency_target = td->o.latency_target;
2569 ts->latency_percentile = td->o.latency_percentile;
2570 ts->latency_window = td->o.latency_window;
2571
2572 ts->nr_block_infos = td->ts.nr_block_infos;
2573 for (k = 0; k < ts->nr_block_infos; k++)
2574 ts->block_infos[k] = td->ts.block_infos[k];
2575
2576 sum_thread_stats(ts, &td->ts);
2577
2578 ts->members++;
2579
2580 if (td->o.ss_dur) {
2581 ts->ss_state = td->ss.state;
2582 ts->ss_dur = td->ss.dur;
2583 ts->ss_head = td->ss.head;
2584 ts->ss_bw_data = td->ss.bw_data;
2585 ts->ss_iops_data = td->ss.iops_data;
2586 ts->ss_limit.u.f = td->ss.limit;
2587 ts->ss_slope.u.f = td->ss.slope;
2588 ts->ss_deviation.u.f = td->ss.deviation;
2589 ts->ss_criterion.u.f = td->ss.criterion;
2590 }
2591 else
2592 ts->ss_dur = ts->ss_state = 0;
2593 } end_for_each();
2594
2595 for (i = 0; i < nr_ts; i++) {
2596 unsigned long long bw;
2597
2598 ts = &threadstats[i];
2599 if (ts->groupid == -1)
2600 continue;
2601 rs = &runstats[ts->groupid];
2602 rs->kb_base = ts->kb_base;
2603 rs->unit_base = ts->unit_base;
2604 rs->sig_figs = ts->sig_figs;
2605 rs->unified_rw_rep |= ts->unified_rw_rep;
2606
2607 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2608 if (!ts->runtime[j])
2609 continue;
2610 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2611 rs->min_run[j] = ts->runtime[j];
2612 if (ts->runtime[j] > rs->max_run[j])
2613 rs->max_run[j] = ts->runtime[j];
2614
2615 bw = 0;
2616 if (ts->runtime[j])
2617 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2618 if (bw < rs->min_bw[j])
2619 rs->min_bw[j] = bw;
2620 if (bw > rs->max_bw[j])
2621 rs->max_bw[j] = bw;
2622
2623 rs->iobytes[j] += ts->io_bytes[j];
2624 }
2625 }
2626
2627 for (i = 0; i < groupid + 1; i++) {
2628 enum fio_ddir ddir;
2629
2630 rs = &runstats[i];
2631
2632 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2633 if (rs->max_run[ddir])
2634 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2635 rs->max_run[ddir];
2636 }
2637 }
2638
2639 for (i = 0; i < FIO_OUTPUT_NR; i++)
2640 buf_output_init(&output[i]);
2641
2642 /*
2643 * don't overwrite last signal output
2644 */
2645 if (output_format & FIO_OUTPUT_NORMAL)
2646 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2647 if (output_format & FIO_OUTPUT_JSON) {
2648 struct thread_data *global;
2649 char time_buf[32];
2650 struct timeval now;
2651 unsigned long long ms_since_epoch;
2652 time_t tv_sec;
2653
2654 gettimeofday(&now, NULL);
2655 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2656 (unsigned long long)(now.tv_usec) / 1000;
2657
2658 tv_sec = now.tv_sec;
2659 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2660 if (time_buf[strlen(time_buf) - 1] == '\n')
2661 time_buf[strlen(time_buf) - 1] = '\0';
2662
2663 root = json_create_object();
2664 json_object_add_value_string(root, "fio version", fio_version_string);
2665 json_object_add_value_int(root, "timestamp", now.tv_sec);
2666 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2667 json_object_add_value_string(root, "time", time_buf);
2668 global = get_global_options();
2669 json_add_job_opts(root, "global options", &global->opt_list);
2670 array = json_create_array();
2671 json_object_add_value_array(root, "jobs", array);
2672 }
2673
2674 if (is_backend)
2675 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2676
2677 for (i = 0; i < nr_ts; i++) {
2678 ts = &threadstats[i];
2679 rs = &runstats[ts->groupid];
2680
2681 if (is_backend) {
2682 fio_server_send_job_options(opt_lists[i], i);
2683 fio_server_send_ts(ts, rs);
2684 } else {
2685 if (output_format & FIO_OUTPUT_TERSE)
2686 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2687 if (output_format & FIO_OUTPUT_JSON) {
2688 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2689 json_array_add_value_object(array, tmp);
2690 }
2691 if (output_format & FIO_OUTPUT_NORMAL)
2692 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2693 }
2694 }
2695 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2696 /* disk util stats, if any */
2697 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2698
2699 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2700
2701 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2702 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2703 json_free_object(root);
2704 }
2705
2706 for (i = 0; i < groupid + 1; i++) {
2707 rs = &runstats[i];
2708
2709 rs->groupid = i;
2710 if (is_backend)
2711 fio_server_send_gs(rs);
2712 else if (output_format & FIO_OUTPUT_NORMAL)
2713 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2714 }
2715
2716 if (is_backend)
2717 fio_server_send_du();
2718 else if (output_format & FIO_OUTPUT_NORMAL) {
2719 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2720 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2721 }
2722
2723 for (i = 0; i < FIO_OUTPUT_NR; i++) {
2724 struct buf_output *out = &output[i];
2725
2726 log_info_buf(out->buf, out->buflen);
2727 buf_output_free(out);
2728 }
2729
2730 fio_idle_prof_cleanup();
2731
2732 log_info_flush();
2733 free(runstats);
2734
2735 /* free arrays allocated by sum_thread_stats(), if any */
2736 for (i = 0; i < nr_ts; i++) {
2737 ts = &threadstats[i];
2738 free_clat_prio_stats(ts);
2739 }
2740 free(threadstats);
2741 free(opt_lists);
2742}
2743
2744int __show_running_run_stats(void)
2745{
2746 unsigned long long *rt;
2747 struct timespec ts;
2748
2749 fio_sem_down(stat_sem);
2750
2751 rt = malloc(thread_number * sizeof(unsigned long long));
2752 fio_gettime(&ts, NULL);
2753
2754 for_each_td(td) {
2755 if (td->runstate >= TD_EXITED)
2756 continue;
2757
2758 td->update_rusage = 1;
2759 for_each_rw_ddir(ddir) {
2760 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2761 }
2762 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2763
2764 rt[__td_index] = mtime_since(&td->start, &ts);
2765 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2766 td->ts.runtime[DDIR_READ] += rt[__td_index];
2767 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2768 td->ts.runtime[DDIR_WRITE] += rt[__td_index];
2769 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2770 td->ts.runtime[DDIR_TRIM] += rt[__td_index];
2771 } end_for_each();
2772
2773 for_each_td(td) {
2774 if (td->runstate >= TD_EXITED)
2775 continue;
2776 if (td->rusage_sem) {
2777 td->update_rusage = 1;
2778 fio_sem_down(td->rusage_sem);
2779 }
2780 td->update_rusage = 0;
2781 } end_for_each();
2782
2783 __show_run_stats();
2784
2785 for_each_td(td) {
2786 if (td->runstate >= TD_EXITED)
2787 continue;
2788
2789 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2790 td->ts.runtime[DDIR_READ] -= rt[__td_index];
2791 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2792 td->ts.runtime[DDIR_WRITE] -= rt[__td_index];
2793 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2794 td->ts.runtime[DDIR_TRIM] -= rt[__td_index];
2795 } end_for_each();
2796
2797 free(rt);
2798 fio_sem_up(stat_sem);
2799
2800 return 0;
2801}
2802
2803static bool status_file_disabled;
2804
2805#define FIO_STATUS_FILE "fio-dump-status"
2806
2807static int check_status_file(void)
2808{
2809 struct stat sb;
2810 const char *temp_dir;
2811 char fio_status_file_path[PATH_MAX];
2812
2813 if (status_file_disabled)
2814 return 0;
2815
2816 temp_dir = getenv("TMPDIR");
2817 if (temp_dir == NULL) {
2818 temp_dir = getenv("TEMP");
2819 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2820 temp_dir = NULL;
2821 }
2822 if (temp_dir == NULL)
2823 temp_dir = "/tmp";
2824#ifdef __COVERITY__
2825 __coverity_tainted_data_sanitize__(temp_dir);
2826#endif
2827
2828 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2829
2830 if (stat(fio_status_file_path, &sb))
2831 return 0;
2832
2833 if (unlink(fio_status_file_path) < 0) {
2834 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2835 strerror(errno));
2836 log_err("fio: disabling status file updates\n");
2837 status_file_disabled = true;
2838 }
2839
2840 return 1;
2841}
2842
2843void check_for_running_stats(void)
2844{
2845 if (check_status_file()) {
2846 show_running_run_stats();
2847 return;
2848 }
2849}
2850
2851static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2852{
2853 double val = data;
2854 double delta;
2855
2856 if (data > is->max_val)
2857 is->max_val = data;
2858 if (data < is->min_val)
2859 is->min_val = data;
2860
2861 delta = val - is->mean.u.f;
2862 if (delta) {
2863 is->mean.u.f += delta / (is->samples + 1.0);
2864 is->S.u.f += delta * (val - is->mean.u.f);
2865 }
2866
2867 is->samples++;
2868}
2869
2870static inline void add_stat_prio_sample(struct clat_prio_stat *clat_prio,
2871 unsigned short clat_prio_index,
2872 unsigned long long nsec)
2873{
2874 if (clat_prio)
2875 add_stat_sample(&clat_prio[clat_prio_index].clat_stat, nsec);
2876}
2877
2878/*
2879 * Return a struct io_logs, which is added to the tail of the log
2880 * list for 'iolog'.
2881 */
2882static struct io_logs *get_new_log(struct io_log *iolog)
2883{
2884 size_t new_samples;
2885 struct io_logs *cur_log;
2886
2887 /*
2888 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2889 * forever
2890 */
2891 if (!iolog->cur_log_max) {
2892 if (iolog->td)
2893 new_samples = iolog->td->o.log_entries;
2894 else
2895 new_samples = DEF_LOG_ENTRIES;
2896 } else {
2897 new_samples = iolog->cur_log_max * 2;
2898 if (new_samples > MAX_LOG_ENTRIES)
2899 new_samples = MAX_LOG_ENTRIES;
2900 }
2901
2902 cur_log = smalloc(sizeof(*cur_log));
2903 if (cur_log) {
2904 INIT_FLIST_HEAD(&cur_log->list);
2905 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2906 if (cur_log->log) {
2907 cur_log->nr_samples = 0;
2908 cur_log->max_samples = new_samples;
2909 flist_add_tail(&cur_log->list, &iolog->io_logs);
2910 iolog->cur_log_max = new_samples;
2911 return cur_log;
2912 }
2913 sfree(cur_log);
2914 }
2915
2916 return NULL;
2917}
2918
2919/*
2920 * Add and return a new log chunk, or return current log if big enough
2921 */
2922static struct io_logs *regrow_log(struct io_log *iolog)
2923{
2924 struct io_logs *cur_log;
2925 int i;
2926
2927 if (!iolog || iolog->disabled)
2928 goto disable;
2929
2930 cur_log = iolog_cur_log(iolog);
2931 if (!cur_log) {
2932 cur_log = get_new_log(iolog);
2933 if (!cur_log)
2934 return NULL;
2935 }
2936
2937 if (cur_log->nr_samples < cur_log->max_samples)
2938 return cur_log;
2939
2940 /*
2941 * No room for a new sample. If we're compressing on the fly, flush
2942 * out the current chunk
2943 */
2944 if (iolog->log_gz) {
2945 if (iolog_cur_flush(iolog, cur_log)) {
2946 log_err("fio: failed flushing iolog! Will stop logging.\n");
2947 return NULL;
2948 }
2949 }
2950
2951 /*
2952 * Get a new log array, and add to our list
2953 */
2954 cur_log = get_new_log(iolog);
2955 if (!cur_log) {
2956 log_err("fio: failed extending iolog! Will stop logging.\n");
2957 return NULL;
2958 }
2959
2960 if (!iolog->pending || !iolog->pending->nr_samples)
2961 return cur_log;
2962
2963 /*
2964 * Flush pending items to new log
2965 */
2966 for (i = 0; i < iolog->pending->nr_samples; i++) {
2967 struct io_sample *src, *dst;
2968
2969 src = get_sample(iolog, iolog->pending, i);
2970 dst = get_sample(iolog, cur_log, i);
2971 memcpy(dst, src, log_entry_sz(iolog));
2972 }
2973 cur_log->nr_samples = iolog->pending->nr_samples;
2974
2975 iolog->pending->nr_samples = 0;
2976 return cur_log;
2977disable:
2978 if (iolog)
2979 iolog->disabled = true;
2980 return NULL;
2981}
2982
2983void regrow_logs(struct thread_data *td)
2984{
2985 regrow_log(td->slat_log);
2986 regrow_log(td->clat_log);
2987 regrow_log(td->clat_hist_log);
2988 regrow_log(td->lat_log);
2989 regrow_log(td->bw_log);
2990 regrow_log(td->iops_log);
2991 td->flags &= ~TD_F_REGROW_LOGS;
2992}
2993
2994void regrow_agg_logs(void)
2995{
2996 enum fio_ddir ddir;
2997
2998 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2999 regrow_log(agg_io_log[ddir]);
3000}
3001
3002static struct io_logs *get_cur_log(struct io_log *iolog)
3003{
3004 struct io_logs *cur_log;
3005
3006 cur_log = iolog_cur_log(iolog);
3007 if (!cur_log) {
3008 cur_log = get_new_log(iolog);
3009 if (!cur_log)
3010 return NULL;
3011 }
3012
3013 if (cur_log->nr_samples < cur_log->max_samples)
3014 return cur_log;
3015
3016 /*
3017 * Out of space. If we're in IO offload mode, or we're not doing
3018 * per unit logging (hence logging happens outside of the IO thread
3019 * as well), add a new log chunk inline. If we're doing inline
3020 * submissions, flag 'td' as needing a log regrow and we'll take
3021 * care of it on the submission side.
3022 */
3023 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
3024 !per_unit_log(iolog))
3025 return regrow_log(iolog);
3026
3027 if (iolog->td)
3028 iolog->td->flags |= TD_F_REGROW_LOGS;
3029 if (iolog->pending)
3030 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
3031 return iolog->pending;
3032}
3033
3034static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
3035 enum fio_ddir ddir, unsigned long long bs,
3036 unsigned long t, uint64_t offset,
3037 unsigned int priority)
3038{
3039 struct io_logs *cur_log;
3040
3041 if (iolog->disabled)
3042 return;
3043 if (flist_empty(&iolog->io_logs))
3044 iolog->avg_last[ddir] = t;
3045
3046 cur_log = get_cur_log(iolog);
3047 if (cur_log) {
3048 struct io_sample *s;
3049
3050 s = get_sample(iolog, cur_log, cur_log->nr_samples);
3051
3052 s->data = data;
3053 s->time = t;
3054 if (iolog->td && iolog->td->o.log_alternate_epoch)
3055 s->time += iolog->td->alternate_epoch;
3056 io_sample_set_ddir(iolog, s, ddir);
3057 s->bs = bs;
3058 s->priority = priority;
3059
3060 if (iolog->log_offset) {
3061 struct io_sample_offset *so = (void *) s;
3062
3063 so->offset = offset;
3064 }
3065
3066 cur_log->nr_samples++;
3067 return;
3068 }
3069
3070 iolog->disabled = true;
3071}
3072
3073static inline void reset_io_stat(struct io_stat *ios)
3074{
3075 ios->min_val = -1ULL;
3076 ios->max_val = ios->samples = 0;
3077 ios->mean.u.f = ios->S.u.f = 0;
3078}
3079
3080static inline void reset_io_u_plat(uint64_t *io_u_plat)
3081{
3082 int i;
3083
3084 for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
3085 io_u_plat[i] = 0;
3086}
3087
3088static inline void reset_clat_prio_stats(struct thread_stat *ts)
3089{
3090 enum fio_ddir ddir;
3091 int i;
3092
3093 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3094 if (!ts->clat_prio[ddir])
3095 continue;
3096
3097 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
3098 reset_io_stat(&ts->clat_prio[ddir][i].clat_stat);
3099 reset_io_u_plat(ts->clat_prio[ddir][i].io_u_plat);
3100 }
3101 }
3102}
3103
3104void reset_io_stats(struct thread_data *td)
3105{
3106 struct thread_stat *ts = &td->ts;
3107 int i, j;
3108
3109 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
3110 reset_io_stat(&ts->clat_stat[i]);
3111 reset_io_stat(&ts->slat_stat[i]);
3112 reset_io_stat(&ts->lat_stat[i]);
3113 reset_io_stat(&ts->bw_stat[i]);
3114 reset_io_stat(&ts->iops_stat[i]);
3115
3116 ts->io_bytes[i] = 0;
3117 ts->runtime[i] = 0;
3118 ts->total_io_u[i] = 0;
3119 ts->short_io_u[i] = 0;
3120 ts->drop_io_u[i] = 0;
3121 }
3122
3123 for (i = 0; i < FIO_LAT_CNT; i++)
3124 for (j = 0; j < DDIR_RWDIR_CNT; j++)
3125 reset_io_u_plat(ts->io_u_plat[i][j]);
3126
3127 reset_clat_prio_stats(ts);
3128
3129 ts->total_io_u[DDIR_SYNC] = 0;
3130 reset_io_u_plat(ts->io_u_sync_plat);
3131
3132 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
3133 ts->io_u_map[i] = 0;
3134 ts->io_u_submit[i] = 0;
3135 ts->io_u_complete[i] = 0;
3136 }
3137
3138 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
3139 ts->io_u_lat_n[i] = 0;
3140 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
3141 ts->io_u_lat_u[i] = 0;
3142 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
3143 ts->io_u_lat_m[i] = 0;
3144
3145 ts->total_submit = 0;
3146 ts->total_complete = 0;
3147 ts->nr_zone_resets = 0;
3148 ts->cachehit = ts->cachemiss = 0;
3149}
3150
3151static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
3152 unsigned long elapsed, int log_max)
3153{
3154 /*
3155 * Note an entry in the log. Use the mean from the logged samples,
3156 * making sure to properly round up. Only write a log entry if we
3157 * had actual samples done.
3158 */
3159 if (iolog->avg_window[ddir].samples) {
3160 union io_sample_data data;
3161
3162 if (log_max == IO_LOG_SAMPLE_AVG) {
3163 data.val.val0 = iolog->avg_window[ddir].mean.u.f + 0.50;
3164 data.val.val1 = 0;
3165 } else if (log_max == IO_LOG_SAMPLE_MAX) {
3166 data.val.val0 = iolog->avg_window[ddir].max_val;
3167 data.val.val1 = 0;
3168 } else {
3169 data.val.val0 = iolog->avg_window[ddir].mean.u.f + 0.50;
3170 data.val.val1 = iolog->avg_window[ddir].max_val;
3171 }
3172
3173 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
3174 }
3175
3176 reset_io_stat(&iolog->avg_window[ddir]);
3177}
3178
3179static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
3180 int log_max)
3181{
3182 enum fio_ddir ddir;
3183
3184 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
3185 __add_stat_to_log(iolog, ddir, elapsed, log_max);
3186}
3187
3188static unsigned long add_log_sample(struct thread_data *td,
3189 struct io_log *iolog,
3190 union io_sample_data data,
3191 enum fio_ddir ddir, unsigned long long bs,
3192 uint64_t offset, unsigned int ioprio)
3193{
3194 unsigned long elapsed, this_window;
3195
3196 if (!ddir_rw(ddir))
3197 return 0;
3198
3199 elapsed = mtime_since_now(&td->epoch);
3200
3201 /*
3202 * If no time averaging, just add the log sample.
3203 */
3204 if (!iolog->avg_msec) {
3205 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
3206 ioprio);
3207 return 0;
3208 }
3209
3210 /*
3211 * Add the sample. If the time period has passed, then
3212 * add that entry to the log and clear.
3213 */
3214 add_stat_sample(&iolog->avg_window[ddir], data.val.val0);
3215
3216 /*
3217 * If period hasn't passed, adding the above sample is all we
3218 * need to do.
3219 */
3220 this_window = elapsed - iolog->avg_last[ddir];
3221 if (elapsed < iolog->avg_last[ddir])
3222 return iolog->avg_last[ddir] - elapsed;
3223 else if (this_window < iolog->avg_msec) {
3224 unsigned long diff = iolog->avg_msec - this_window;
3225
3226 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
3227 return diff;
3228 }
3229
3230 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max);
3231
3232 iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
3233
3234 return iolog->avg_msec;
3235}
3236
3237void finalize_logs(struct thread_data *td, bool unit_logs)
3238{
3239 unsigned long elapsed;
3240
3241 elapsed = mtime_since_now(&td->epoch);
3242
3243 if (td->clat_log && unit_logs)
3244 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max);
3245 if (td->slat_log && unit_logs)
3246 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max);
3247 if (td->lat_log && unit_logs)
3248 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max);
3249 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
3250 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max);
3251 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
3252 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max);
3253}
3254
3255void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
3256 unsigned long long bs)
3257{
3258 struct io_log *iolog;
3259
3260 if (!ddir_rw(ddir))
3261 return;
3262
3263 iolog = agg_io_log[ddir];
3264 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
3265}
3266
3267void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
3268{
3269 unsigned int idx = plat_val_to_idx(nsec);
3270 assert(idx < FIO_IO_U_PLAT_NR);
3271
3272 ts->io_u_sync_plat[idx]++;
3273 add_stat_sample(&ts->sync_stat, nsec);
3274}
3275
3276static inline void add_lat_percentile_sample(struct thread_stat *ts,
3277 unsigned long long nsec,
3278 enum fio_ddir ddir,
3279 enum fio_lat lat)
3280{
3281 unsigned int idx = plat_val_to_idx(nsec);
3282 assert(idx < FIO_IO_U_PLAT_NR);
3283
3284 ts->io_u_plat[lat][ddir][idx]++;
3285}
3286
3287static inline void
3288add_lat_percentile_prio_sample(struct thread_stat *ts, unsigned long long nsec,
3289 enum fio_ddir ddir,
3290 unsigned short clat_prio_index)
3291{
3292 unsigned int idx = plat_val_to_idx(nsec);
3293
3294 if (ts->clat_prio[ddir])
3295 ts->clat_prio[ddir][clat_prio_index].io_u_plat[idx]++;
3296}
3297
3298void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
3299 unsigned long long nsec, unsigned long long bs,
3300 uint64_t offset, unsigned int ioprio,
3301 unsigned short clat_prio_index)
3302{
3303 const bool needs_lock = td_async_processing(td);
3304 unsigned long elapsed, this_window;
3305 struct thread_stat *ts = &td->ts;
3306 struct io_log *iolog = td->clat_hist_log;
3307
3308 if (needs_lock)
3309 __td_io_u_lock(td);
3310
3311 add_stat_sample(&ts->clat_stat[ddir], nsec);
3312
3313 /*
3314 * When lat_percentiles=1 (default 0), the reported per priority
3315 * percentiles and stats are used for describing total latency values,
3316 * even though the variable names themselves start with clat_.
3317 *
3318 * Because of the above definition, add a prio stat sample only when
3319 * lat_percentiles=0. add_lat_sample() will add the prio stat sample
3320 * when lat_percentiles=1.
3321 */
3322 if (!ts->lat_percentiles)
3323 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3324 nsec);
3325
3326 if (td->clat_log)
3327 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
3328 offset, ioprio);
3329
3330 if (ts->clat_percentiles) {
3331 /*
3332 * Because of the above definition, add a prio lat percentile
3333 * sample only when lat_percentiles=0. add_lat_sample() will add
3334 * the prio lat percentile sample when lat_percentiles=1.
3335 */
3336 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3337 if (!ts->lat_percentiles)
3338 add_lat_percentile_prio_sample(ts, nsec, ddir,
3339 clat_prio_index);
3340 }
3341
3342 if (iolog && iolog->hist_msec) {
3343 struct io_hist *hw = &iolog->hist_window[ddir];
3344
3345 hw->samples++;
3346 elapsed = mtime_since_now(&td->epoch);
3347 if (!hw->hist_last)
3348 hw->hist_last = elapsed;
3349 this_window = elapsed - hw->hist_last;
3350
3351 if (this_window >= iolog->hist_msec) {
3352 uint64_t *io_u_plat;
3353 struct io_u_plat_entry *dst;
3354
3355 /*
3356 * Make a byte-for-byte copy of the latency histogram
3357 * stored in td->ts.io_u_plat[ddir], recording it in a
3358 * log sample. Note that the matching call to free() is
3359 * located in iolog.c after printing this sample to the
3360 * log file.
3361 */
3362 io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3363 dst = malloc(sizeof(struct io_u_plat_entry));
3364 memcpy(&(dst->io_u_plat), io_u_plat,
3365 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3366 flist_add(&dst->list, &hw->list);
3367 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3368 elapsed, offset, ioprio);
3369
3370 /*
3371 * Update the last time we recorded as being now, minus
3372 * any drift in time we encountered before actually
3373 * making the record.
3374 */
3375 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3376 hw->samples = 0;
3377 }
3378 }
3379
3380 if (needs_lock)
3381 __td_io_u_unlock(td);
3382}
3383
3384void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3385 unsigned long long nsec, unsigned long long bs,
3386 uint64_t offset, unsigned int ioprio)
3387{
3388 const bool needs_lock = td_async_processing(td);
3389 struct thread_stat *ts = &td->ts;
3390
3391 if (!ddir_rw(ddir))
3392 return;
3393
3394 if (needs_lock)
3395 __td_io_u_lock(td);
3396
3397 add_stat_sample(&ts->slat_stat[ddir], nsec);
3398
3399 if (td->slat_log)
3400 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3401 offset, ioprio);
3402
3403 if (ts->slat_percentiles)
3404 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3405
3406 if (needs_lock)
3407 __td_io_u_unlock(td);
3408}
3409
3410void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3411 unsigned long long nsec, unsigned long long bs,
3412 uint64_t offset, unsigned int ioprio,
3413 unsigned short clat_prio_index)
3414{
3415 const bool needs_lock = td_async_processing(td);
3416 struct thread_stat *ts = &td->ts;
3417
3418 if (!ddir_rw(ddir))
3419 return;
3420
3421 if (needs_lock)
3422 __td_io_u_lock(td);
3423
3424 add_stat_sample(&ts->lat_stat[ddir], nsec);
3425
3426 if (td->lat_log)
3427 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3428 offset, ioprio);
3429
3430 /*
3431 * When lat_percentiles=1 (default 0), the reported per priority
3432 * percentiles and stats are used for describing total latency values,
3433 * even though the variable names themselves start with clat_.
3434 *
3435 * Because of the above definition, add a prio stat and prio lat
3436 * percentile sample only when lat_percentiles=1. add_clat_sample() will
3437 * add the prio stat and prio lat percentile sample when
3438 * lat_percentiles=0.
3439 */
3440 if (ts->lat_percentiles) {
3441 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3442 add_lat_percentile_prio_sample(ts, nsec, ddir, clat_prio_index);
3443 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3444 nsec);
3445 }
3446 if (needs_lock)
3447 __td_io_u_unlock(td);
3448}
3449
3450void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3451 unsigned int bytes, unsigned long long spent)
3452{
3453 const bool needs_lock = td_async_processing(td);
3454 struct thread_stat *ts = &td->ts;
3455 unsigned long rate;
3456
3457 if (spent)
3458 rate = (unsigned long) (bytes * 1000000ULL / spent);
3459 else
3460 rate = 0;
3461
3462 if (needs_lock)
3463 __td_io_u_lock(td);
3464
3465 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3466
3467 if (td->bw_log)
3468 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3469 bytes, io_u->offset, io_u->ioprio);
3470
3471 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3472
3473 if (needs_lock)
3474 __td_io_u_unlock(td);
3475}
3476
3477static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3478 struct timespec *t, unsigned int avg_time,
3479 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3480 struct io_stat *stat, struct io_log *log,
3481 bool is_kb)
3482{
3483 const bool needs_lock = td_async_processing(td);
3484 unsigned long spent, rate;
3485 enum fio_ddir ddir;
3486 unsigned long next, next_log;
3487
3488 next_log = avg_time;
3489
3490 spent = mtime_since(parent_tv, t);
3491 if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3492 return avg_time - spent;
3493
3494 if (needs_lock)
3495 __td_io_u_lock(td);
3496
3497 /*
3498 * Compute both read and write rates for the interval.
3499 */
3500 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3501 uint64_t delta;
3502
3503 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3504 if (!delta)
3505 continue; /* No entries for interval */
3506
3507 if (spent) {
3508 if (is_kb)
3509 rate = delta * 1000 / spent / 1024; /* KiB/s */
3510 else
3511 rate = (delta * 1000) / spent;
3512 } else
3513 rate = 0;
3514
3515 add_stat_sample(&stat[ddir], rate);
3516
3517 if (log) {
3518 unsigned long long bs = 0;
3519
3520 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3521 bs = td->o.min_bs[ddir];
3522
3523 next = add_log_sample(td, log, sample_val(rate), ddir,
3524 bs, 0, 0);
3525 next_log = min(next_log, next);
3526 }
3527
3528 stat_io_bytes[ddir] = this_io_bytes[ddir];
3529 }
3530
3531 *parent_tv = *t;
3532
3533 if (needs_lock)
3534 __td_io_u_unlock(td);
3535
3536 if (spent <= avg_time)
3537 next = avg_time;
3538 else
3539 next = avg_time - (1 + spent - avg_time);
3540
3541 return min(next, next_log);
3542}
3543
3544static int add_bw_samples(struct thread_data *td, struct timespec *t)
3545{
3546 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3547 td->this_io_bytes, td->stat_io_bytes,
3548 td->ts.bw_stat, td->bw_log, true);
3549}
3550
3551void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3552 unsigned int bytes)
3553{
3554 const bool needs_lock = td_async_processing(td);
3555 struct thread_stat *ts = &td->ts;
3556
3557 if (needs_lock)
3558 __td_io_u_lock(td);
3559
3560 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3561
3562 if (td->iops_log)
3563 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3564 bytes, io_u->offset, io_u->ioprio);
3565
3566 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3567
3568 if (needs_lock)
3569 __td_io_u_unlock(td);
3570}
3571
3572static int add_iops_samples(struct thread_data *td, struct timespec *t)
3573{
3574 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3575 td->this_io_blocks, td->stat_io_blocks,
3576 td->ts.iops_stat, td->iops_log, false);
3577}
3578
3579static bool td_in_logging_state(struct thread_data *td)
3580{
3581 if (in_ramp_time(td))
3582 return false;
3583
3584 switch(td->runstate) {
3585 case TD_RUNNING:
3586 case TD_VERIFYING:
3587 case TD_FINISHING:
3588 case TD_EXITED:
3589 return true;
3590 default:
3591 return false;
3592 }
3593}
3594
3595/*
3596 * Returns msecs to next event
3597 */
3598int calc_log_samples(void)
3599{
3600 unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3601 struct timespec now;
3602 long elapsed_time = 0;
3603
3604 for_each_td(td) {
3605 fio_gettime(&now, NULL);
3606 elapsed_time = mtime_since(&td->epoch, &now);
3607
3608 if (!td->o.stats)
3609 continue;
3610 if (!td_in_logging_state(td)) {
3611 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3612 continue;
3613 }
3614 if (!td->bw_log ||
3615 (td->bw_log && !per_unit_log(td->bw_log))) {
3616 tmp = add_bw_samples(td, &now);
3617
3618 if (td->bw_log)
3619 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3620 }
3621 if (!td->iops_log ||
3622 (td->iops_log && !per_unit_log(td->iops_log))) {
3623 tmp = add_iops_samples(td, &now);
3624
3625 if (td->iops_log)
3626 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3627 }
3628
3629 if (tmp < next)
3630 next = tmp;
3631 } end_for_each();
3632
3633 /* if log_avg_msec_min has not been changed, set it to 0 */
3634 if (log_avg_msec_min == -1U)
3635 log_avg_msec_min = 0;
3636
3637 if (log_avg_msec_min == 0)
3638 next_mod = elapsed_time;
3639 else
3640 next_mod = elapsed_time % log_avg_msec_min;
3641
3642 /* correction to keep the time on the log avg msec boundary */
3643 next = min(next, (log_avg_msec_min - next_mod));
3644
3645 return next == ~0U ? 0 : next;
3646}
3647
3648void stat_init(void)
3649{
3650 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3651}
3652
3653void stat_exit(void)
3654{
3655 /*
3656 * When we have the mutex, we know out-of-band access to it
3657 * have ended.
3658 */
3659 fio_sem_down(stat_sem);
3660 fio_sem_remove(stat_sem);
3661}
3662
3663/*
3664 * Called from signal handler. Wake up status thread.
3665 */
3666void show_running_run_stats(void)
3667{
3668 helper_do_stat();
3669}
3670
3671uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3672{
3673 /* Ignore io_u's which span multiple blocks--they will just get
3674 * inaccurate counts. */
3675 int idx = (io_u->offset - io_u->file->file_offset)
3676 / td->o.bs[DDIR_TRIM];
3677 uint32_t *info = &td->ts.block_infos[idx];
3678 assert(idx < td->ts.nr_block_infos);
3679 return info;
3680}
3681
fio - Flexible IO Tester