projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
nanosecond: reconcile terse output with nanosecond timing for latencies
[fio.git] / stat.c
... / ...
CommitLineData
1#include <stdio.h>
2#include <string.h>
3#include <sys/time.h>
4#include <sys/types.h>
5#include <sys/stat.h>
6#include <dirent.h>
7#include <libgen.h>
8#include <math.h>
9
10#include "fio.h"
11#include "diskutil.h"
12#include "lib/ieee754.h"
13#include "json.h"
14#include "lib/getrusage.h"
15#include "idletime.h"
16#include "lib/pow2.h"
17#include "lib/output_buffer.h"
18#include "helper_thread.h"
19#include "smalloc.h"
20
21#define LOG_MSEC_SLACK 10
22
23struct fio_mutex *stat_mutex;
24
25void clear_rusage_stat(struct thread_data *td)
26{
27 struct thread_stat *ts = &td->ts;
28
29 fio_getrusage(&td->ru_start);
30 ts->usr_time = ts->sys_time = 0;
31 ts->ctx = 0;
32 ts->minf = ts->majf = 0;
33}
34
35void update_rusage_stat(struct thread_data *td)
36{
37 struct thread_stat *ts = &td->ts;
38
39 fio_getrusage(&td->ru_end);
40 ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
41 &td->ru_end.ru_utime);
42 ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
43 &td->ru_end.ru_stime);
44 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
45 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
46 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
47 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
48
49 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
50}
51
52/*
53 * Given a latency, return the index of the corresponding bucket in
54 * the structure tracking percentiles.
55 *
56 * (1) find the group (and error bits) that the value (latency)
57 * belongs to by looking at its MSB. (2) find the bucket number in the
58 * group by looking at the index bits.
59 *
60 */
61static unsigned int plat_val_to_idx(unsigned long long val)
62{
63 unsigned int msb, error_bits, base, offset, idx;
64
65 /* Find MSB starting from bit 0 */
66 if (val == 0)
67 msb = 0;
68 else
69 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
70
71 /*
72 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
73 * all bits of the sample as index
74 */
75 if (msb <= FIO_IO_U_PLAT_BITS)
76 return val;
77
78 /* Compute the number of error bits to discard*/
79 error_bits = msb - FIO_IO_U_PLAT_BITS;
80
81 /* Compute the number of buckets before the group */
82 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
83
84 /*
85 * Discard the error bits and apply the mask to find the
86 * index for the buckets in the group
87 */
88 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
89
90 /* Make sure the index does not exceed (array size - 1) */
91 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
92 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
93
94 return idx;
95}
96
97/*
98 * Convert the given index of the bucket array to the value
99 * represented by the bucket
100 */
101static unsigned long long plat_idx_to_val(unsigned int idx)
102{
103 unsigned int error_bits, k, base;
104
105 assert(idx < FIO_IO_U_PLAT_NR);
106
107 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
108 * all bits of the sample as index */
109 if (idx < (FIO_IO_U_PLAT_VAL << 1))
110 return idx;
111
112 /* Find the group and compute the minimum value of that group */
113 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
114 base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
115
116 /* Find its bucket number of the group */
117 k = idx % FIO_IO_U_PLAT_VAL;
118
119 /* Return the mean of the range of the bucket */
120 return base + ((k + 0.5) * (1 << error_bits));
121}
122
123static int double_cmp(const void *a, const void *b)
124{
125 const fio_fp64_t fa = *(const fio_fp64_t *) a;
126 const fio_fp64_t fb = *(const fio_fp64_t *) b;
127 int cmp = 0;
128
129 if (fa.u.f > fb.u.f)
130 cmp = 1;
131 else if (fa.u.f < fb.u.f)
132 cmp = -1;
133
134 return cmp;
135}
136
137unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
138 fio_fp64_t *plist, unsigned long long **output,
139 unsigned long long *maxv, unsigned long long *minv)
140{
141 unsigned long sum = 0;
142 unsigned int len, i, j = 0;
143 unsigned int oval_len = 0;
144 unsigned long long *ovals = NULL;
145 int is_last;
146
147 *minv = -1ULL;
148 *maxv = 0;
149
150 len = 0;
151 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
152 len++;
153
154 if (!len)
155 return 0;
156
157 /*
158 * Sort the percentile list. Note that it may already be sorted if
159 * we are using the default values, but since it's a short list this
160 * isn't a worry. Also note that this does not work for NaN values.
161 */
162 if (len > 1)
163 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
164
165 /*
166 * Calculate bucket values, note down max and min values
167 */
168 is_last = 0;
169 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
170 sum += io_u_plat[i];
171 while (sum >= (plist[j].u.f / 100.0 * nr)) {
172 assert(plist[j].u.f <= 100.0);
173
174 if (j == oval_len) {
175 oval_len += 100;
176 ovals = realloc(ovals, oval_len * sizeof(*ovals));
177 }
178
179 ovals[j] = plat_idx_to_val(i);
180 if (ovals[j] < *minv)
181 *minv = ovals[j];
182 if (ovals[j] > *maxv)
183 *maxv = ovals[j];
184
185 is_last = (j == len - 1);
186 if (is_last)
187 break;
188
189 j++;
190 }
191 }
192
193 *output = ovals;
194 return len;
195}
196
197/*
198 * Find and display the p-th percentile of clat
199 */
200static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
201 fio_fp64_t *plist, unsigned int precision,
202 struct buf_output *out)
203{
204 unsigned int divisor, len, i, j = 0;
205 unsigned long long minv, maxv;
206 unsigned long long *ovals;
207 int is_last, per_line, scale_down, time_width;
208 char fmt[32];
209
210 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
211 if (!len)
212 goto out;
213
214 /*
215 * We default to nsecs, but if the value range is such that we
216 * should scale down to usecs or msecs, do that.
217 */
218 if (minv > 2000000 && maxv > 99999999ULL) {
219 scale_down = 2;
220 divisor = 1000000;
221 log_buf(out, " clat percentiles (msec):\n |");
222 } else if (minv > 2000 && maxv > 99999) {
223 scale_down = 1;
224 divisor = 1000;
225 log_buf(out, " clat percentiles (usec):\n |");
226 } else {
227 scale_down = 0;
228 divisor = 1;
229 log_buf(out, " clat percentiles (nsec):\n |");
230 }
231
232
233 time_width = max(5, (int) (log10(maxv / divisor) + 1));
234 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision+3, precision, time_width);
235 // fmt will be something like " %5.2fth=[%4llu]%c"
236 per_line = (80 - 7) / (precision + 10 + time_width);
237
238 for (j = 0; j < len; j++) {
239 /* for formatting */
240 if (j != 0 && (j % per_line) == 0)
241 log_buf(out, " |");
242
243 /* end of the list */
244 is_last = (j == len - 1);
245
246 for (i = 0; i < scale_down; i++)
247 ovals[j] = (ovals[j] + 999) / 1000;
248
249 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
250
251 if (is_last)
252 break;
253
254 if ((j % per_line) == per_line - 1) /* for formatting */
255 log_buf(out, "\n");
256 }
257
258out:
259 if (ovals)
260 free(ovals);
261}
262
263bool calc_lat(struct io_stat *is, unsigned long long *min, unsigned long long *max,
264 double *mean, double *dev)
265{
266 double n = (double) is->samples;
267
268 if (n == 0)
269 return false;
270
271 *min = is->min_val;
272 *max = is->max_val;
273 *mean = is->mean.u.f;
274
275 if (n > 1.0)
276 *dev = sqrt(is->S.u.f / (n - 1.0));
277 else
278 *dev = 0;
279
280 return true;
281}
282
283void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
284{
285 char *io, *agg, *min, *max;
286 char *ioalt, *aggalt, *minalt, *maxalt;
287 const char *str[] = { " READ", " WRITE" , " TRIM"};
288 int i;
289
290 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
291
292 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
293 const int i2p = is_power_of_2(rs->kb_base);
294
295 if (!rs->max_run[i])
296 continue;
297
298 io = num2str(rs->iobytes[i], 4, 1, i2p, N2S_BYTE);
299 ioalt = num2str(rs->iobytes[i], 4, 1, !i2p, N2S_BYTE);
300 agg = num2str(rs->agg[i], 4, 1, i2p, rs->unit_base);
301 aggalt = num2str(rs->agg[i], 4, 1, !i2p, rs->unit_base);
302 min = num2str(rs->min_bw[i], 4, 1, i2p, rs->unit_base);
303 minalt = num2str(rs->min_bw[i], 4, 1, !i2p, rs->unit_base);
304 max = num2str(rs->max_bw[i], 4, 1, i2p, rs->unit_base);
305 maxalt = num2str(rs->max_bw[i], 4, 1, !i2p, rs->unit_base);
306 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
307 rs->unified_rw_rep ? " MIXED" : str[i],
308 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
309 (unsigned long long) rs->min_run[i],
310 (unsigned long long) rs->max_run[i]);
311
312 free(io);
313 free(agg);
314 free(min);
315 free(max);
316 free(ioalt);
317 free(aggalt);
318 free(minalt);
319 free(maxalt);
320 }
321}
322
323void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
324{
325 int i;
326
327 /*
328 * Do depth distribution calculations
329 */
330 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
331 if (total) {
332 io_u_dist[i] = (double) map[i] / (double) total;
333 io_u_dist[i] *= 100.0;
334 if (io_u_dist[i] < 0.1 && map[i])
335 io_u_dist[i] = 0.1;
336 } else
337 io_u_dist[i] = 0.0;
338 }
339}
340
341static void stat_calc_lat(struct thread_stat *ts, double *dst,
342 unsigned int *src, int nr)
343{
344 unsigned long total = ddir_rw_sum(ts->total_io_u);
345 int i;
346
347 /*
348 * Do latency distribution calculations
349 */
350 for (i = 0; i < nr; i++) {
351 if (total) {
352 dst[i] = (double) src[i] / (double) total;
353 dst[i] *= 100.0;
354 if (dst[i] < 0.01 && src[i])
355 dst[i] = 0.01;
356 } else
357 dst[i] = 0.0;
358 }
359}
360
361/*
362 * To keep the terse format unaltered, add all of the ns latency
363 * buckets to the first us latency bucket
364 */
365void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
366{
367 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
368 int i;
369
370 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
371
372 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
373 ntotal += ts->io_u_lat_n[i];
374
375 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
376}
377
378void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
379{
380 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
381}
382
383void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
384{
385 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
386}
387
388void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
389{
390 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
391}
392
393static void display_lat(const char *name, unsigned long long min, unsigned long long max,
394 double mean, double dev, struct buf_output *out)
395{
396 const char *base = "(nsec)";
397 char *minp, *maxp;
398
399 if (nsec_to_msec(&min, &max, &mean, &dev))
400 base = "(msec)";
401 else if (nsec_to_usec(&min, &max, &mean, &dev))
402 base = "(usec)";
403
404 minp = num2str(min, 6, 1, 0, N2S_NONE);
405 maxp = num2str(max, 6, 1, 0, N2S_NONE);
406
407 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
408 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
409
410 free(minp);
411 free(maxp);
412}
413
414static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
415 int ddir, struct buf_output *out)
416{
417 const char *str[] = { " read", "write", " trim" };
418 unsigned long runt;
419 unsigned long long min, max, bw, iops;
420 double mean, dev;
421 char *io_p, *bw_p, *bw_p_alt, *iops_p;
422 int i2p;
423
424 assert(ddir_rw(ddir));
425
426 if (!ts->runtime[ddir])
427 return;
428
429 i2p = is_power_of_2(rs->kb_base);
430 runt = ts->runtime[ddir];
431
432 bw = (1000 * ts->io_bytes[ddir]) / runt;
433 io_p = num2str(ts->io_bytes[ddir], 4, 1, i2p, N2S_BYTE);
434 bw_p = num2str(bw, 4, 1, i2p, ts->unit_base);
435 bw_p_alt = num2str(bw, 4, 1, !i2p, ts->unit_base);
436
437 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
438 iops_p = num2str(iops, 4, 1, 0, N2S_NONE);
439
440 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
441 rs->unified_rw_rep ? "mixed" : str[ddir],
442 iops_p, bw_p, bw_p_alt, io_p,
443 (unsigned long long) ts->runtime[ddir]);
444
445 free(io_p);
446 free(bw_p);
447 free(bw_p_alt);
448 free(iops_p);
449
450 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
451 display_lat("slat", min, max, mean, dev, out);
452 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
453 display_lat("clat", min, max, mean, dev, out);
454 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
455 display_lat(" lat", min, max, mean, dev, out);
456
457 if (ts->clat_percentiles) {
458 show_clat_percentiles(ts->io_u_plat[ddir],
459 ts->clat_stat[ddir].samples,
460 ts->percentile_list,
461 ts->percentile_precision, out);
462 }
463 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
464 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
465 const char *bw_str;
466
467 if ((rs->unit_base == 1) && i2p)
468 bw_str = "Kibit";
469 else if (rs->unit_base == 1)
470 bw_str = "kbit";
471 else if (i2p)
472 bw_str = "KiB";
473 else
474 bw_str = "kB";
475
476 if (rs->unit_base == 1) {
477 min *= 8.0;
478 max *= 8.0;
479 mean *= 8.0;
480 dev *= 8.0;
481 }
482
483 if (rs->agg[ddir]) {
484 p_of_agg = mean * 100 / (double) rs->agg[ddir];
485 if (p_of_agg > 100.0)
486 p_of_agg = 100.0;
487 }
488
489 if (mean > fkb_base * fkb_base) {
490 min /= fkb_base;
491 max /= fkb_base;
492 mean /= fkb_base;
493 dev /= fkb_base;
494 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
495 }
496
497 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, avg=%5.02f, stdev=%5.02f\n",
498 bw_str, min, max, p_of_agg, mean, dev);
499 }
500}
501
502static int show_lat(double *io_u_lat, int nr, const char **ranges,
503 const char *msg, struct buf_output *out)
504{
505 int new_line = 1, i, line = 0, shown = 0;
506
507 for (i = 0; i < nr; i++) {
508 if (io_u_lat[i] <= 0.0)
509 continue;
510 shown = 1;
511 if (new_line) {
512 if (line)
513 log_buf(out, "\n");
514 log_buf(out, " lat (%s) : ", msg);
515 new_line = 0;
516 line = 0;
517 }
518 if (line)
519 log_buf(out, ", ");
520 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
521 line++;
522 if (line == 5)
523 new_line = 1;
524 }
525
526 if (shown)
527 log_buf(out, "\n");
528
529 return shown;
530}
531
532static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
533{
534 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
535 "250=", "500=", "750=", "1000=", };
536
537 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
538}
539
540static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
541{
542 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
543 "250=", "500=", "750=", "1000=", };
544
545 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
546}
547
548static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
549{
550 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
551 "250=", "500=", "750=", "1000=", "2000=",
552 ">=2000=", };
553
554 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
555}
556
557static void show_latencies(struct thread_stat *ts, struct buf_output *out)
558{
559 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
560 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
561 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
562
563 stat_calc_lat_n(ts, io_u_lat_n);
564 stat_calc_lat_u(ts, io_u_lat_u);
565 stat_calc_lat_m(ts, io_u_lat_m);
566
567 show_lat_n(io_u_lat_n, out);
568 show_lat_u(io_u_lat_u, out);
569 show_lat_m(io_u_lat_m, out);
570}
571
572static int block_state_category(int block_state)
573{
574 switch (block_state) {
575 case BLOCK_STATE_UNINIT:
576 return 0;
577 case BLOCK_STATE_TRIMMED:
578 case BLOCK_STATE_WRITTEN:
579 return 1;
580 case BLOCK_STATE_WRITE_FAILURE:
581 case BLOCK_STATE_TRIM_FAILURE:
582 return 2;
583 default:
584 /* Silence compile warning on some BSDs and have a return */
585 assert(0);
586 return -1;
587 }
588}
589
590static int compare_block_infos(const void *bs1, const void *bs2)
591{
592 uint32_t block1 = *(uint32_t *)bs1;
593 uint32_t block2 = *(uint32_t *)bs2;
594 int state1 = BLOCK_INFO_STATE(block1);
595 int state2 = BLOCK_INFO_STATE(block2);
596 int bscat1 = block_state_category(state1);
597 int bscat2 = block_state_category(state2);
598 int cycles1 = BLOCK_INFO_TRIMS(block1);
599 int cycles2 = BLOCK_INFO_TRIMS(block2);
600
601 if (bscat1 < bscat2)
602 return -1;
603 if (bscat1 > bscat2)
604 return 1;
605
606 if (cycles1 < cycles2)
607 return -1;
608 if (cycles1 > cycles2)
609 return 1;
610
611 if (state1 < state2)
612 return -1;
613 if (state1 > state2)
614 return 1;
615
616 assert(block1 == block2);
617 return 0;
618}
619
620static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
621 fio_fp64_t *plist, unsigned int **percentiles,
622 unsigned int *types)
623{
624 int len = 0;
625 int i, nr_uninit;
626
627 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
628
629 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
630 len++;
631
632 if (!len)
633 return 0;
634
635 /*
636 * Sort the percentile list. Note that it may already be sorted if
637 * we are using the default values, but since it's a short list this
638 * isn't a worry. Also note that this does not work for NaN values.
639 */
640 if (len > 1)
641 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
642
643 nr_uninit = 0;
644 /* Start only after the uninit entries end */
645 for (nr_uninit = 0;
646 nr_uninit < nr_block_infos
647 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
648 nr_uninit ++)
649 ;
650
651 if (nr_uninit == nr_block_infos)
652 return 0;
653
654 *percentiles = calloc(len, sizeof(**percentiles));
655
656 for (i = 0; i < len; i++) {
657 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
658 + nr_uninit;
659 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
660 }
661
662 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
663 for (i = 0; i < nr_block_infos; i++)
664 types[BLOCK_INFO_STATE(block_infos[i])]++;
665
666 return len;
667}
668
669static const char *block_state_names[] = {
670 [BLOCK_STATE_UNINIT] = "unwritten",
671 [BLOCK_STATE_TRIMMED] = "trimmed",
672 [BLOCK_STATE_WRITTEN] = "written",
673 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
674 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
675};
676
677static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
678 fio_fp64_t *plist, struct buf_output *out)
679{
680 int len, pos, i;
681 unsigned int *percentiles = NULL;
682 unsigned int block_state_counts[BLOCK_STATE_COUNT];
683
684 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
685 &percentiles, block_state_counts);
686
687 log_buf(out, " block lifetime percentiles :\n |");
688 pos = 0;
689 for (i = 0; i < len; i++) {
690 uint32_t block_info = percentiles[i];
691#define LINE_LENGTH 75
692 char str[LINE_LENGTH];
693 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
694 plist[i].u.f, block_info,
695 i == len - 1 ? '\n' : ',');
696 assert(strln < LINE_LENGTH);
697 if (pos + strln > LINE_LENGTH) {
698 pos = 0;
699 log_buf(out, "\n |");
700 }
701 log_buf(out, "%s", str);
702 pos += strln;
703#undef LINE_LENGTH
704 }
705 if (percentiles)
706 free(percentiles);
707
708 log_buf(out, " states :");
709 for (i = 0; i < BLOCK_STATE_COUNT; i++)
710 log_buf(out, " %s=%u%c",
711 block_state_names[i], block_state_counts[i],
712 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
713}
714
715static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
716{
717 char *p1, *p1alt, *p2;
718 unsigned long long bw_mean, iops_mean;
719 const int i2p = is_power_of_2(ts->kb_base);
720
721 if (!ts->ss_dur)
722 return;
723
724 bw_mean = steadystate_bw_mean(ts);
725 iops_mean = steadystate_iops_mean(ts);
726
727 p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
728 p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
729 p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
730
731 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
732 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
733 p1, p1alt, p2,
734 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
735 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
736 ts->ss_criterion.u.f,
737 ts->ss_state & __FIO_SS_PCT ? "%" : "");
738
739 free(p1);
740 free(p1alt);
741 free(p2);
742}
743
744static void show_thread_status_normal(struct thread_stat *ts,
745 struct group_run_stats *rs,
746 struct buf_output *out)
747{
748 double usr_cpu, sys_cpu;
749 unsigned long runtime;
750 double io_u_dist[FIO_IO_U_MAP_NR];
751 time_t time_p;
752 char time_buf[32];
753
754 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
755 return;
756
757 memset(time_buf, 0, sizeof(time_buf));
758
759 time(&time_p);
760 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
761
762 if (!ts->error) {
763 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
764 ts->name, ts->groupid, ts->members,
765 ts->error, (int) ts->pid, time_buf);
766 } else {
767 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
768 ts->name, ts->groupid, ts->members,
769 ts->error, ts->verror, (int) ts->pid,
770 time_buf);
771 }
772
773 if (strlen(ts->description))
774 log_buf(out, " Description : [%s]\n", ts->description);
775
776 if (ts->io_bytes[DDIR_READ])
777 show_ddir_status(rs, ts, DDIR_READ, out);
778 if (ts->io_bytes[DDIR_WRITE])
779 show_ddir_status(rs, ts, DDIR_WRITE, out);
780 if (ts->io_bytes[DDIR_TRIM])
781 show_ddir_status(rs, ts, DDIR_TRIM, out);
782
783 show_latencies(ts, out);
784
785 runtime = ts->total_run_time;
786 if (runtime) {
787 double runt = (double) runtime;
788
789 usr_cpu = (double) ts->usr_time * 100 / runt;
790 sys_cpu = (double) ts->sys_time * 100 / runt;
791 } else {
792 usr_cpu = 0;
793 sys_cpu = 0;
794 }
795
796 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
797 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
798 (unsigned long long) ts->ctx,
799 (unsigned long long) ts->majf,
800 (unsigned long long) ts->minf);
801
802 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
803 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
804 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
805 io_u_dist[1], io_u_dist[2],
806 io_u_dist[3], io_u_dist[4],
807 io_u_dist[5], io_u_dist[6]);
808
809 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
810 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
811 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
812 io_u_dist[1], io_u_dist[2],
813 io_u_dist[3], io_u_dist[4],
814 io_u_dist[5], io_u_dist[6]);
815 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
816 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
817 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
818 io_u_dist[1], io_u_dist[2],
819 io_u_dist[3], io_u_dist[4],
820 io_u_dist[5], io_u_dist[6]);
821 log_buf(out, " issued rwt: total=%llu,%llu,%llu,"
822 " short=%llu,%llu,%llu,"
823 " dropped=%llu,%llu,%llu\n",
824 (unsigned long long) ts->total_io_u[0],
825 (unsigned long long) ts->total_io_u[1],
826 (unsigned long long) ts->total_io_u[2],
827 (unsigned long long) ts->short_io_u[0],
828 (unsigned long long) ts->short_io_u[1],
829 (unsigned long long) ts->short_io_u[2],
830 (unsigned long long) ts->drop_io_u[0],
831 (unsigned long long) ts->drop_io_u[1],
832 (unsigned long long) ts->drop_io_u[2]);
833 if (ts->continue_on_error) {
834 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
835 (unsigned long long)ts->total_err_count,
836 ts->first_error,
837 strerror(ts->first_error));
838 }
839 if (ts->latency_depth) {
840 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
841 (unsigned long long)ts->latency_target,
842 (unsigned long long)ts->latency_window,
843 ts->latency_percentile.u.f,
844 ts->latency_depth);
845 }
846
847 if (ts->nr_block_infos)
848 show_block_infos(ts->nr_block_infos, ts->block_infos,
849 ts->percentile_list, out);
850
851 if (ts->ss_dur)
852 show_ss_normal(ts, out);
853}
854
855static void show_ddir_status_terse(struct thread_stat *ts,
856 struct group_run_stats *rs, int ddir,
857 struct buf_output *out)
858{
859 unsigned long long min, max, minv, maxv, bw, iops;
860 unsigned long long *ovals = NULL;
861 double mean, dev;
862 unsigned int len;
863 int i;
864
865 assert(ddir_rw(ddir));
866
867 iops = bw = 0;
868 if (ts->runtime[ddir]) {
869 uint64_t runt = ts->runtime[ddir];
870
871 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
872 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
873 }
874
875 log_buf(out, ";%llu;%llu;%llu;%llu",
876 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
877 (unsigned long long) ts->runtime[ddir]);
878
879 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
880 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
881 else
882 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
883
884 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
885 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
886 else
887 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
888
889 if (ts->clat_percentiles) {
890 len = calc_clat_percentiles(ts->io_u_plat[ddir],
891 ts->clat_stat[ddir].samples,
892 ts->percentile_list, &ovals, &maxv,
893 &minv);
894 } else
895 len = 0;
896
897 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
898 if (i >= len) {
899 log_buf(out, ";0%%=0");
900 continue;
901 }
902 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
903 }
904
905 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
906 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
907 else
908 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
909
910 if (ovals)
911 free(ovals);
912
913 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
914 double p_of_agg = 100.0;
915
916 if (rs->agg[ddir]) {
917 p_of_agg = mean * 100 / (double) rs->agg[ddir];
918 if (p_of_agg > 100.0)
919 p_of_agg = 100.0;
920 }
921
922 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
923 } else
924 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
925}
926
927static void add_ddir_status_json(struct thread_stat *ts,
928 struct group_run_stats *rs, int ddir, struct json_object *parent)
929{
930 unsigned long long min, max, minv, maxv;
931 unsigned long long bw;
932 unsigned long long *ovals = NULL;
933 double mean, dev, iops;
934 unsigned int len;
935 int i;
936 const char *ddirname[] = {"read", "write", "trim"};
937 struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
938 char buf[120];
939 double p_of_agg = 100.0;
940
941 assert(ddir_rw(ddir));
942
943 if (ts->unified_rw_rep && ddir != DDIR_READ)
944 return;
945
946 dir_object = json_create_object();
947 json_object_add_value_object(parent,
948 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
949
950 bw = 0;
951 iops = 0.0;
952 if (ts->runtime[ddir]) {
953 uint64_t runt = ts->runtime[ddir];
954
955 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
956 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
957 }
958
959 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
960 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
961 json_object_add_value_int(dir_object, "bw", bw);
962 json_object_add_value_float(dir_object, "iops", iops);
963 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
964 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
965 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
966 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
967
968 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
969 min = max = 0;
970 mean = dev = 0.0;
971 }
972 tmp_object = json_create_object();
973 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
974 json_object_add_value_int(tmp_object, "min", min);
975 json_object_add_value_int(tmp_object, "max", max);
976 json_object_add_value_float(tmp_object, "mean", mean);
977 json_object_add_value_float(tmp_object, "stddev", dev);
978
979 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
980 min = max = 0;
981 mean = dev = 0.0;
982 }
983 tmp_object = json_create_object();
984 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
985 json_object_add_value_int(tmp_object, "min", min);
986 json_object_add_value_int(tmp_object, "max", max);
987 json_object_add_value_float(tmp_object, "mean", mean);
988 json_object_add_value_float(tmp_object, "stddev", dev);
989
990 if (ts->clat_percentiles) {
991 len = calc_clat_percentiles(ts->io_u_plat[ddir],
992 ts->clat_stat[ddir].samples,
993 ts->percentile_list, &ovals, &maxv,
994 &minv);
995 } else
996 len = 0;
997
998 percentile_object = json_create_object();
999 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1000 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1001 if (i >= len) {
1002 json_object_add_value_int(percentile_object, "0.00", 0);
1003 continue;
1004 }
1005 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1006 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1007 }
1008
1009 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1010 clat_bins_object = json_create_object();
1011 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1012 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1013 if (ts->io_u_plat[ddir][i]) {
1014 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1015 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1016 }
1017 }
1018 }
1019
1020 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1021 min = max = 0;
1022 mean = dev = 0.0;
1023 }
1024 tmp_object = json_create_object();
1025 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1026 json_object_add_value_int(tmp_object, "min", min);
1027 json_object_add_value_int(tmp_object, "max", max);
1028 json_object_add_value_float(tmp_object, "mean", mean);
1029 json_object_add_value_float(tmp_object, "stddev", dev);
1030 if (ovals)
1031 free(ovals);
1032
1033 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1034 if (rs->agg[ddir]) {
1035 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1036 if (p_of_agg > 100.0)
1037 p_of_agg = 100.0;
1038 }
1039 } else {
1040 min = max = 0;
1041 p_of_agg = mean = dev = 0.0;
1042 }
1043 json_object_add_value_int(dir_object, "bw_min", min);
1044 json_object_add_value_int(dir_object, "bw_max", max);
1045 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1046 json_object_add_value_float(dir_object, "bw_mean", mean);
1047 json_object_add_value_float(dir_object, "bw_dev", dev);
1048}
1049
1050static void show_thread_status_terse_v2(struct thread_stat *ts,
1051 struct group_run_stats *rs,
1052 struct buf_output *out)
1053{
1054 double io_u_dist[FIO_IO_U_MAP_NR];
1055 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1056 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1057 double usr_cpu, sys_cpu;
1058 int i;
1059
1060 /* General Info */
1061 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1062 /* Log Read Status */
1063 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1064 /* Log Write Status */
1065 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1066 /* Log Trim Status */
1067 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1068
1069 /* CPU Usage */
1070 if (ts->total_run_time) {
1071 double runt = (double) ts->total_run_time;
1072
1073 usr_cpu = (double) ts->usr_time * 100 / runt;
1074 sys_cpu = (double) ts->sys_time * 100 / runt;
1075 } else {
1076 usr_cpu = 0;
1077 sys_cpu = 0;
1078 }
1079
1080 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1081 (unsigned long long) ts->ctx,
1082 (unsigned long long) ts->majf,
1083 (unsigned long long) ts->minf);
1084
1085 /* Calc % distribution of IO depths, usecond, msecond latency */
1086 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1087 stat_calc_lat_nu(ts, io_u_lat_u);
1088 stat_calc_lat_m(ts, io_u_lat_m);
1089
1090 /* Only show fixed 7 I/O depth levels*/
1091 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1092 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1093 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1094
1095 /* Microsecond latency */
1096 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1097 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1098 /* Millisecond latency */
1099 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1100 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1101 /* Additional output if continue_on_error set - default off*/
1102 if (ts->continue_on_error)
1103 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1104 log_buf(out, "\n");
1105
1106 /* Additional output if description is set */
1107 if (strlen(ts->description))
1108 log_buf(out, ";%s", ts->description);
1109
1110 log_buf(out, "\n");
1111}
1112
1113static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1114 struct group_run_stats *rs, int ver,
1115 struct buf_output *out)
1116{
1117 double io_u_dist[FIO_IO_U_MAP_NR];
1118 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1119 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1120 double usr_cpu, sys_cpu;
1121 int i;
1122
1123 /* General Info */
1124 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1125 ts->name, ts->groupid, ts->error);
1126 /* Log Read Status */
1127 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1128 /* Log Write Status */
1129 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1130 /* Log Trim Status */
1131 if (ver == 4)
1132 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1133
1134 /* CPU Usage */
1135 if (ts->total_run_time) {
1136 double runt = (double) ts->total_run_time;
1137
1138 usr_cpu = (double) ts->usr_time * 100 / runt;
1139 sys_cpu = (double) ts->sys_time * 100 / runt;
1140 } else {
1141 usr_cpu = 0;
1142 sys_cpu = 0;
1143 }
1144
1145 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1146 (unsigned long long) ts->ctx,
1147 (unsigned long long) ts->majf,
1148 (unsigned long long) ts->minf);
1149
1150 /* Calc % distribution of IO depths, usecond, msecond latency */
1151 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1152 stat_calc_lat_nu(ts, io_u_lat_u);
1153 stat_calc_lat_m(ts, io_u_lat_m);
1154
1155 /* Only show fixed 7 I/O depth levels*/
1156 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1157 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1158 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1159
1160 /* Microsecond latency */
1161 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1162 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1163 /* Millisecond latency */
1164 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1165 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1166
1167 /* disk util stats, if any */
1168 show_disk_util(1, NULL, out);
1169
1170 /* Additional output if continue_on_error set - default off*/
1171 if (ts->continue_on_error)
1172 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1173
1174 /* Additional output if description is set */
1175 if (strlen(ts->description))
1176 log_buf(out, ";%s", ts->description);
1177
1178 log_buf(out, "\n");
1179}
1180
1181static void json_add_job_opts(struct json_object *root, const char *name,
1182 struct flist_head *opt_list, bool num_jobs)
1183{
1184 struct json_object *dir_object;
1185 struct flist_head *entry;
1186 struct print_option *p;
1187
1188 if (flist_empty(opt_list))
1189 return;
1190
1191 dir_object = json_create_object();
1192 json_object_add_value_object(root, name, dir_object);
1193
1194 flist_for_each(entry, opt_list) {
1195 const char *pos = "";
1196
1197 p = flist_entry(entry, struct print_option, list);
1198 if (!num_jobs && !strcmp(p->name, "numjobs"))
1199 continue;
1200 if (p->value)
1201 pos = p->value;
1202 json_object_add_value_string(dir_object, p->name, pos);
1203 }
1204}
1205
1206static struct json_object *show_thread_status_json(struct thread_stat *ts,
1207 struct group_run_stats *rs,
1208 struct flist_head *opt_list)
1209{
1210 struct json_object *root, *tmp;
1211 struct jobs_eta *je;
1212 double io_u_dist[FIO_IO_U_MAP_NR];
1213 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1214 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1215 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1216 double usr_cpu, sys_cpu;
1217 int i;
1218 size_t size;
1219
1220 root = json_create_object();
1221 json_object_add_value_string(root, "jobname", ts->name);
1222 json_object_add_value_int(root, "groupid", ts->groupid);
1223 json_object_add_value_int(root, "error", ts->error);
1224
1225 /* ETA Info */
1226 je = get_jobs_eta(true, &size);
1227 if (je) {
1228 json_object_add_value_int(root, "eta", je->eta_sec);
1229 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1230 }
1231
1232 if (opt_list)
1233 json_add_job_opts(root, "job options", opt_list, true);
1234
1235 add_ddir_status_json(ts, rs, DDIR_READ, root);
1236 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1237 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1238
1239 /* CPU Usage */
1240 if (ts->total_run_time) {
1241 double runt = (double) ts->total_run_time;
1242
1243 usr_cpu = (double) ts->usr_time * 100 / runt;
1244 sys_cpu = (double) ts->sys_time * 100 / runt;
1245 } else {
1246 usr_cpu = 0;
1247 sys_cpu = 0;
1248 }
1249 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1250 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1251 json_object_add_value_int(root, "ctx", ts->ctx);
1252 json_object_add_value_int(root, "majf", ts->majf);
1253 json_object_add_value_int(root, "minf", ts->minf);
1254
1255
1256 /* Calc % distribution of IO depths, usecond, msecond latency */
1257 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1258 stat_calc_lat_n(ts, io_u_lat_n);
1259 stat_calc_lat_u(ts, io_u_lat_u);
1260 stat_calc_lat_m(ts, io_u_lat_m);
1261
1262 tmp = json_create_object();
1263 json_object_add_value_object(root, "iodepth_level", tmp);
1264 /* Only show fixed 7 I/O depth levels*/
1265 for (i = 0; i < 7; i++) {
1266 char name[20];
1267 if (i < 6)
1268 snprintf(name, 20, "%d", 1 << i);
1269 else
1270 snprintf(name, 20, ">=%d", 1 << i);
1271 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1272 }
1273
1274 /* Nanosecond latency */
1275 tmp = json_create_object();
1276 json_object_add_value_object(root, "latency_ns", tmp);
1277 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1278 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1279 "250", "500", "750", "1000", };
1280 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1281 }
1282 /* Microsecond latency */
1283 tmp = json_create_object();
1284 json_object_add_value_object(root, "latency_us", tmp);
1285 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1286 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1287 "250", "500", "750", "1000", };
1288 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1289 }
1290 /* Millisecond latency */
1291 tmp = json_create_object();
1292 json_object_add_value_object(root, "latency_ms", tmp);
1293 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1294 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1295 "250", "500", "750", "1000", "2000",
1296 ">=2000", };
1297 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1298 }
1299
1300 /* Additional output if continue_on_error set - default off*/
1301 if (ts->continue_on_error) {
1302 json_object_add_value_int(root, "total_err", ts->total_err_count);
1303 json_object_add_value_int(root, "first_error", ts->first_error);
1304 }
1305
1306 if (ts->latency_depth) {
1307 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1308 json_object_add_value_int(root, "latency_target", ts->latency_target);
1309 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1310 json_object_add_value_int(root, "latency_window", ts->latency_window);
1311 }
1312
1313 /* Additional output if description is set */
1314 if (strlen(ts->description))
1315 json_object_add_value_string(root, "desc", ts->description);
1316
1317 if (ts->nr_block_infos) {
1318 /* Block error histogram and types */
1319 int len;
1320 unsigned int *percentiles = NULL;
1321 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1322
1323 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1324 ts->percentile_list,
1325 &percentiles, block_state_counts);
1326
1327 if (len) {
1328 struct json_object *block, *percentile_object, *states;
1329 int state;
1330 block = json_create_object();
1331 json_object_add_value_object(root, "block", block);
1332
1333 percentile_object = json_create_object();
1334 json_object_add_value_object(block, "percentiles",
1335 percentile_object);
1336 for (i = 0; i < len; i++) {
1337 char buf[20];
1338 snprintf(buf, sizeof(buf), "%f",
1339 ts->percentile_list[i].u.f);
1340 json_object_add_value_int(percentile_object,
1341 (const char *)buf,
1342 percentiles[i]);
1343 }
1344
1345 states = json_create_object();
1346 json_object_add_value_object(block, "states", states);
1347 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1348 json_object_add_value_int(states,
1349 block_state_names[state],
1350 block_state_counts[state]);
1351 }
1352 free(percentiles);
1353 }
1354 }
1355
1356 if (ts->ss_dur) {
1357 struct json_object *data;
1358 struct json_array *iops, *bw;
1359 int i, j, k;
1360 char ss_buf[64];
1361
1362 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1363 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1364 ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1365 (float) ts->ss_limit.u.f,
1366 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1367
1368 tmp = json_create_object();
1369 json_object_add_value_object(root, "steadystate", tmp);
1370 json_object_add_value_string(tmp, "ss", ss_buf);
1371 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1372 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1373
1374 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1375 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1376 json_object_add_value_string(tmp, "criterion", ss_buf);
1377 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1378 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1379
1380 data = json_create_object();
1381 json_object_add_value_object(tmp, "data", data);
1382 bw = json_create_array();
1383 iops = json_create_array();
1384
1385 /*
1386 ** if ss was attained or the buffer is not full,
1387 ** ss->head points to the first element in the list.
1388 ** otherwise it actually points to the second element
1389 ** in the list
1390 */
1391 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1392 j = ts->ss_head;
1393 else
1394 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1395 for (i = 0; i < ts->ss_dur; i++) {
1396 k = (j + i) % ts->ss_dur;
1397 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1398 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1399 }
1400 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1401 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1402 json_object_add_value_array(data, "iops", iops);
1403 json_object_add_value_array(data, "bw", bw);
1404 }
1405
1406 return root;
1407}
1408
1409static void show_thread_status_terse(struct thread_stat *ts,
1410 struct group_run_stats *rs,
1411 struct buf_output *out)
1412{
1413 if (terse_version == 2)
1414 show_thread_status_terse_v2(ts, rs, out);
1415 else if (terse_version == 3 || terse_version == 4)
1416 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1417 else
1418 log_err("fio: bad terse version!? %d\n", terse_version);
1419}
1420
1421struct json_object *show_thread_status(struct thread_stat *ts,
1422 struct group_run_stats *rs,
1423 struct flist_head *opt_list,
1424 struct buf_output *out)
1425{
1426 struct json_object *ret = NULL;
1427
1428 if (output_format & FIO_OUTPUT_TERSE)
1429 show_thread_status_terse(ts, rs, out);
1430 if (output_format & FIO_OUTPUT_JSON)
1431 ret = show_thread_status_json(ts, rs, opt_list);
1432 if (output_format & FIO_OUTPUT_NORMAL)
1433 show_thread_status_normal(ts, rs, out);
1434
1435 return ret;
1436}
1437
1438static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1439{
1440 double mean, S;
1441
1442 if (src->samples == 0)
1443 return;
1444
1445 dst->min_val = min(dst->min_val, src->min_val);
1446 dst->max_val = max(dst->max_val, src->max_val);
1447
1448 /*
1449 * Compute new mean and S after the merge
1450 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1451 * #Parallel_algorithm>
1452 */
1453 if (first) {
1454 mean = src->mean.u.f;
1455 S = src->S.u.f;
1456 } else {
1457 double delta = src->mean.u.f - dst->mean.u.f;
1458
1459 mean = ((src->mean.u.f * src->samples) +
1460 (dst->mean.u.f * dst->samples)) /
1461 (dst->samples + src->samples);
1462
1463 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1464 (dst->samples * src->samples) /
1465 (dst->samples + src->samples);
1466 }
1467
1468 dst->samples += src->samples;
1469 dst->mean.u.f = mean;
1470 dst->S.u.f = S;
1471}
1472
1473void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1474{
1475 int i;
1476
1477 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1478 if (dst->max_run[i] < src->max_run[i])
1479 dst->max_run[i] = src->max_run[i];
1480 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1481 dst->min_run[i] = src->min_run[i];
1482 if (dst->max_bw[i] < src->max_bw[i])
1483 dst->max_bw[i] = src->max_bw[i];
1484 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1485 dst->min_bw[i] = src->min_bw[i];
1486
1487 dst->iobytes[i] += src->iobytes[i];
1488 dst->agg[i] += src->agg[i];
1489 }
1490
1491 if (!dst->kb_base)
1492 dst->kb_base = src->kb_base;
1493 if (!dst->unit_base)
1494 dst->unit_base = src->unit_base;
1495}
1496
1497void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1498 bool first)
1499{
1500 int l, k;
1501
1502 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1503 if (!dst->unified_rw_rep) {
1504 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1505 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1506 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1507 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1508
1509 dst->io_bytes[l] += src->io_bytes[l];
1510
1511 if (dst->runtime[l] < src->runtime[l])
1512 dst->runtime[l] = src->runtime[l];
1513 } else {
1514 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1515 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1516 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1517 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1518
1519 dst->io_bytes[0] += src->io_bytes[l];
1520
1521 if (dst->runtime[0] < src->runtime[l])
1522 dst->runtime[0] = src->runtime[l];
1523
1524 /*
1525 * We're summing to the same destination, so override
1526 * 'first' after the first iteration of the loop
1527 */
1528 first = false;
1529 }
1530 }
1531
1532 dst->usr_time += src->usr_time;
1533 dst->sys_time += src->sys_time;
1534 dst->ctx += src->ctx;
1535 dst->majf += src->majf;
1536 dst->minf += src->minf;
1537
1538 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1539 dst->io_u_map[k] += src->io_u_map[k];
1540 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1541 dst->io_u_submit[k] += src->io_u_submit[k];
1542 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1543 dst->io_u_complete[k] += src->io_u_complete[k];
1544 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1545 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1546 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1547 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1548 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1549 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1550
1551 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1552 if (!dst->unified_rw_rep) {
1553 dst->total_io_u[k] += src->total_io_u[k];
1554 dst->short_io_u[k] += src->short_io_u[k];
1555 dst->drop_io_u[k] += src->drop_io_u[k];
1556 } else {
1557 dst->total_io_u[0] += src->total_io_u[k];
1558 dst->short_io_u[0] += src->short_io_u[k];
1559 dst->drop_io_u[0] += src->drop_io_u[k];
1560 }
1561 }
1562
1563 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1564 int m;
1565
1566 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1567 if (!dst->unified_rw_rep)
1568 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1569 else
1570 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1571 }
1572 }
1573
1574 dst->total_run_time += src->total_run_time;
1575 dst->total_submit += src->total_submit;
1576 dst->total_complete += src->total_complete;
1577}
1578
1579void init_group_run_stat(struct group_run_stats *gs)
1580{
1581 int i;
1582 memset(gs, 0, sizeof(*gs));
1583
1584 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1585 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1586}
1587
1588void init_thread_stat(struct thread_stat *ts)
1589{
1590 int j;
1591
1592 memset(ts, 0, sizeof(*ts));
1593
1594 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1595 ts->lat_stat[j].min_val = -1UL;
1596 ts->clat_stat[j].min_val = -1UL;
1597 ts->slat_stat[j].min_val = -1UL;
1598 ts->bw_stat[j].min_val = -1UL;
1599 }
1600 ts->groupid = -1;
1601}
1602
1603void __show_run_stats(void)
1604{
1605 struct group_run_stats *runstats, *rs;
1606 struct thread_data *td;
1607 struct thread_stat *threadstats, *ts;
1608 int i, j, k, nr_ts, last_ts, idx;
1609 int kb_base_warned = 0;
1610 int unit_base_warned = 0;
1611 struct json_object *root = NULL;
1612 struct json_array *array = NULL;
1613 struct buf_output output[FIO_OUTPUT_NR];
1614 struct flist_head **opt_lists;
1615
1616 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1617
1618 for (i = 0; i < groupid + 1; i++)
1619 init_group_run_stat(&runstats[i]);
1620
1621 /*
1622 * find out how many threads stats we need. if group reporting isn't
1623 * enabled, it's one-per-td.
1624 */
1625 nr_ts = 0;
1626 last_ts = -1;
1627 for_each_td(td, i) {
1628 if (!td->o.group_reporting) {
1629 nr_ts++;
1630 continue;
1631 }
1632 if (last_ts == td->groupid)
1633 continue;
1634 if (!td->o.stats)
1635 continue;
1636
1637 last_ts = td->groupid;
1638 nr_ts++;
1639 }
1640
1641 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1642 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1643
1644 for (i = 0; i < nr_ts; i++) {
1645 init_thread_stat(&threadstats[i]);
1646 opt_lists[i] = NULL;
1647 }
1648
1649 j = 0;
1650 last_ts = -1;
1651 idx = 0;
1652 for_each_td(td, i) {
1653 if (!td->o.stats)
1654 continue;
1655 if (idx && (!td->o.group_reporting ||
1656 (td->o.group_reporting && last_ts != td->groupid))) {
1657 idx = 0;
1658 j++;
1659 }
1660
1661 last_ts = td->groupid;
1662
1663 ts = &threadstats[j];
1664
1665 ts->clat_percentiles = td->o.clat_percentiles;
1666 ts->percentile_precision = td->o.percentile_precision;
1667 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1668 opt_lists[j] = &td->opt_list;
1669
1670 idx++;
1671 ts->members++;
1672
1673 if (ts->groupid == -1) {
1674 /*
1675 * These are per-group shared already
1676 */
1677 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1678 if (td->o.description)
1679 strncpy(ts->description, td->o.description,
1680 FIO_JOBDESC_SIZE - 1);
1681 else
1682 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1683
1684 /*
1685 * If multiple entries in this group, this is
1686 * the first member.
1687 */
1688 ts->thread_number = td->thread_number;
1689 ts->groupid = td->groupid;
1690
1691 /*
1692 * first pid in group, not very useful...
1693 */
1694 ts->pid = td->pid;
1695
1696 ts->kb_base = td->o.kb_base;
1697 ts->unit_base = td->o.unit_base;
1698 ts->unified_rw_rep = td->o.unified_rw_rep;
1699 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1700 log_info("fio: kb_base differs for jobs in group, using"
1701 " %u as the base\n", ts->kb_base);
1702 kb_base_warned = 1;
1703 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1704 log_info("fio: unit_base differs for jobs in group, using"
1705 " %u as the base\n", ts->unit_base);
1706 unit_base_warned = 1;
1707 }
1708
1709 ts->continue_on_error = td->o.continue_on_error;
1710 ts->total_err_count += td->total_err_count;
1711 ts->first_error = td->first_error;
1712 if (!ts->error) {
1713 if (!td->error && td->o.continue_on_error &&
1714 td->first_error) {
1715 ts->error = td->first_error;
1716 ts->verror[sizeof(ts->verror) - 1] = '\0';
1717 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1718 } else if (td->error) {
1719 ts->error = td->error;
1720 ts->verror[sizeof(ts->verror) - 1] = '\0';
1721 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1722 }
1723 }
1724
1725 ts->latency_depth = td->latency_qd;
1726 ts->latency_target = td->o.latency_target;
1727 ts->latency_percentile = td->o.latency_percentile;
1728 ts->latency_window = td->o.latency_window;
1729
1730 ts->nr_block_infos = td->ts.nr_block_infos;
1731 for (k = 0; k < ts->nr_block_infos; k++)
1732 ts->block_infos[k] = td->ts.block_infos[k];
1733
1734 sum_thread_stats(ts, &td->ts, idx == 1);
1735
1736 if (td->o.ss_dur) {
1737 ts->ss_state = td->ss.state;
1738 ts->ss_dur = td->ss.dur;
1739 ts->ss_head = td->ss.head;
1740 ts->ss_bw_data = td->ss.bw_data;
1741 ts->ss_iops_data = td->ss.iops_data;
1742 ts->ss_limit.u.f = td->ss.limit;
1743 ts->ss_slope.u.f = td->ss.slope;
1744 ts->ss_deviation.u.f = td->ss.deviation;
1745 ts->ss_criterion.u.f = td->ss.criterion;
1746 }
1747 else
1748 ts->ss_dur = ts->ss_state = 0;
1749 }
1750
1751 for (i = 0; i < nr_ts; i++) {
1752 unsigned long long bw;
1753
1754 ts = &threadstats[i];
1755 if (ts->groupid == -1)
1756 continue;
1757 rs = &runstats[ts->groupid];
1758 rs->kb_base = ts->kb_base;
1759 rs->unit_base = ts->unit_base;
1760 rs->unified_rw_rep += ts->unified_rw_rep;
1761
1762 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1763 if (!ts->runtime[j])
1764 continue;
1765 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1766 rs->min_run[j] = ts->runtime[j];
1767 if (ts->runtime[j] > rs->max_run[j])
1768 rs->max_run[j] = ts->runtime[j];
1769
1770 bw = 0;
1771 if (ts->runtime[j])
1772 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1773 if (bw < rs->min_bw[j])
1774 rs->min_bw[j] = bw;
1775 if (bw > rs->max_bw[j])
1776 rs->max_bw[j] = bw;
1777
1778 rs->iobytes[j] += ts->io_bytes[j];
1779 }
1780 }
1781
1782 for (i = 0; i < groupid + 1; i++) {
1783 int ddir;
1784
1785 rs = &runstats[i];
1786
1787 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1788 if (rs->max_run[ddir])
1789 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1790 rs->max_run[ddir];
1791 }
1792 }
1793
1794 for (i = 0; i < FIO_OUTPUT_NR; i++)
1795 buf_output_init(&output[i]);
1796
1797 /*
1798 * don't overwrite last signal output
1799 */
1800 if (output_format & FIO_OUTPUT_NORMAL)
1801 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1802 if (output_format & FIO_OUTPUT_JSON) {
1803 struct thread_data *global;
1804 char time_buf[32];
1805 struct timeval now;
1806 unsigned long long ms_since_epoch;
1807
1808 gettimeofday(&now, NULL);
1809 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1810 (unsigned long long)(now.tv_usec) / 1000;
1811
1812 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1813 sizeof(time_buf));
1814 if (time_buf[strlen(time_buf) - 1] == '\n')
1815 time_buf[strlen(time_buf) - 1] = '\0';
1816
1817 root = json_create_object();
1818 json_object_add_value_string(root, "fio version", fio_version_string);
1819 json_object_add_value_int(root, "timestamp", now.tv_sec);
1820 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1821 json_object_add_value_string(root, "time", time_buf);
1822 global = get_global_options();
1823 json_add_job_opts(root, "global options", &global->opt_list, false);
1824 array = json_create_array();
1825 json_object_add_value_array(root, "jobs", array);
1826 }
1827
1828 if (is_backend)
1829 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1830
1831 for (i = 0; i < nr_ts; i++) {
1832 ts = &threadstats[i];
1833 rs = &runstats[ts->groupid];
1834
1835 if (is_backend) {
1836 fio_server_send_job_options(opt_lists[i], i);
1837 fio_server_send_ts(ts, rs);
1838 } else {
1839 if (output_format & FIO_OUTPUT_TERSE)
1840 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1841 if (output_format & FIO_OUTPUT_JSON) {
1842 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1843 json_array_add_value_object(array, tmp);
1844 }
1845 if (output_format & FIO_OUTPUT_NORMAL)
1846 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1847 }
1848 }
1849 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1850 /* disk util stats, if any */
1851 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1852
1853 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1854
1855 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1856 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1857 json_free_object(root);
1858 }
1859
1860 for (i = 0; i < groupid + 1; i++) {
1861 rs = &runstats[i];
1862
1863 rs->groupid = i;
1864 if (is_backend)
1865 fio_server_send_gs(rs);
1866 else if (output_format & FIO_OUTPUT_NORMAL)
1867 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1868 }
1869
1870 if (is_backend)
1871 fio_server_send_du();
1872 else if (output_format & FIO_OUTPUT_NORMAL) {
1873 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1874 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1875 }
1876
1877 for (i = 0; i < FIO_OUTPUT_NR; i++) {
1878 struct buf_output *out = &output[i];
1879
1880 log_info_buf(out->buf, out->buflen);
1881 buf_output_free(out);
1882 }
1883
1884 log_info_flush();
1885 free(runstats);
1886 free(threadstats);
1887 free(opt_lists);
1888}
1889
1890void show_run_stats(void)
1891{
1892 fio_mutex_down(stat_mutex);
1893 __show_run_stats();
1894 fio_mutex_up(stat_mutex);
1895}
1896
1897void __show_running_run_stats(void)
1898{
1899 struct thread_data *td;
1900 unsigned long long *rt;
1901 struct timespec ts;
1902 int i;
1903
1904 fio_mutex_down(stat_mutex);
1905
1906 rt = malloc(thread_number * sizeof(unsigned long long));
1907 fio_gettime(&ts, NULL);
1908
1909 for_each_td(td, i) {
1910 td->update_rusage = 1;
1911 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1912 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1913 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1914 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1915
1916 rt[i] = mtime_since(&td->start, &ts);
1917 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1918 td->ts.runtime[DDIR_READ] += rt[i];
1919 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1920 td->ts.runtime[DDIR_WRITE] += rt[i];
1921 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1922 td->ts.runtime[DDIR_TRIM] += rt[i];
1923 }
1924
1925 for_each_td(td, i) {
1926 if (td->runstate >= TD_EXITED)
1927 continue;
1928 if (td->rusage_sem) {
1929 td->update_rusage = 1;
1930 fio_mutex_down(td->rusage_sem);
1931 }
1932 td->update_rusage = 0;
1933 }
1934
1935 __show_run_stats();
1936
1937 for_each_td(td, i) {
1938 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1939 td->ts.runtime[DDIR_READ] -= rt[i];
1940 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1941 td->ts.runtime[DDIR_WRITE] -= rt[i];
1942 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1943 td->ts.runtime[DDIR_TRIM] -= rt[i];
1944 }
1945
1946 free(rt);
1947 fio_mutex_up(stat_mutex);
1948}
1949
1950static int status_interval_init;
1951static struct timespec status_time;
1952static int status_file_disabled;
1953
1954#define FIO_STATUS_FILE "fio-dump-status"
1955
1956static int check_status_file(void)
1957{
1958 struct stat sb;
1959 const char *temp_dir;
1960 char fio_status_file_path[PATH_MAX];
1961
1962 if (status_file_disabled)
1963 return 0;
1964
1965 temp_dir = getenv("TMPDIR");
1966 if (temp_dir == NULL) {
1967 temp_dir = getenv("TEMP");
1968 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1969 temp_dir = NULL;
1970 }
1971 if (temp_dir == NULL)
1972 temp_dir = "/tmp";
1973
1974 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1975
1976 if (stat(fio_status_file_path, &sb))
1977 return 0;
1978
1979 if (unlink(fio_status_file_path) < 0) {
1980 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1981 strerror(errno));
1982 log_err("fio: disabling status file updates\n");
1983 status_file_disabled = 1;
1984 }
1985
1986 return 1;
1987}
1988
1989void check_for_running_stats(void)
1990{
1991 if (status_interval) {
1992 if (!status_interval_init) {
1993 fio_gettime(&status_time, NULL);
1994 status_interval_init = 1;
1995 } else if (mtime_since_now(&status_time) >= status_interval) {
1996 show_running_run_stats();
1997 fio_gettime(&status_time, NULL);
1998 return;
1999 }
2000 }
2001 if (check_status_file()) {
2002 show_running_run_stats();
2003 return;
2004 }
2005}
2006
2007static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2008{
2009 double val = data;
2010 double delta;
2011
2012 if (data > is->max_val)
2013 is->max_val = data;
2014 if (data < is->min_val)
2015 is->min_val = data;
2016
2017 delta = val - is->mean.u.f;
2018 if (delta) {
2019 is->mean.u.f += delta / (is->samples + 1.0);
2020 is->S.u.f += delta * (val - is->mean.u.f);
2021 }
2022
2023 is->samples++;
2024}
2025
2026/*
2027 * Return a struct io_logs, which is added to the tail of the log
2028 * list for 'iolog'.
2029 */
2030static struct io_logs *get_new_log(struct io_log *iolog)
2031{
2032 size_t new_size, new_samples;
2033 struct io_logs *cur_log;
2034
2035 /*
2036 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2037 * forever
2038 */
2039 if (!iolog->cur_log_max)
2040 new_samples = DEF_LOG_ENTRIES;
2041 else {
2042 new_samples = iolog->cur_log_max * 2;
2043 if (new_samples > MAX_LOG_ENTRIES)
2044 new_samples = MAX_LOG_ENTRIES;
2045 }
2046
2047 new_size = new_samples * log_entry_sz(iolog);
2048
2049 cur_log = smalloc(sizeof(*cur_log));
2050 if (cur_log) {
2051 INIT_FLIST_HEAD(&cur_log->list);
2052 cur_log->log = malloc(new_size);
2053 if (cur_log->log) {
2054 cur_log->nr_samples = 0;
2055 cur_log->max_samples = new_samples;
2056 flist_add_tail(&cur_log->list, &iolog->io_logs);
2057 iolog->cur_log_max = new_samples;
2058 return cur_log;
2059 }
2060 sfree(cur_log);
2061 }
2062
2063 return NULL;
2064}
2065
2066/*
2067 * Add and return a new log chunk, or return current log if big enough
2068 */
2069static struct io_logs *regrow_log(struct io_log *iolog)
2070{
2071 struct io_logs *cur_log;
2072 int i;
2073
2074 if (!iolog || iolog->disabled)
2075 goto disable;
2076
2077 cur_log = iolog_cur_log(iolog);
2078 if (!cur_log) {
2079 cur_log = get_new_log(iolog);
2080 if (!cur_log)
2081 return NULL;
2082 }
2083
2084 if (cur_log->nr_samples < cur_log->max_samples)
2085 return cur_log;
2086
2087 /*
2088 * No room for a new sample. If we're compressing on the fly, flush
2089 * out the current chunk
2090 */
2091 if (iolog->log_gz) {
2092 if (iolog_cur_flush(iolog, cur_log)) {
2093 log_err("fio: failed flushing iolog! Will stop logging.\n");
2094 return NULL;
2095 }
2096 }
2097
2098 /*
2099 * Get a new log array, and add to our list
2100 */
2101 cur_log = get_new_log(iolog);
2102 if (!cur_log) {
2103 log_err("fio: failed extending iolog! Will stop logging.\n");
2104 return NULL;
2105 }
2106
2107 if (!iolog->pending || !iolog->pending->nr_samples)
2108 return cur_log;
2109
2110 /*
2111 * Flush pending items to new log
2112 */
2113 for (i = 0; i < iolog->pending->nr_samples; i++) {
2114 struct io_sample *src, *dst;
2115
2116 src = get_sample(iolog, iolog->pending, i);
2117 dst = get_sample(iolog, cur_log, i);
2118 memcpy(dst, src, log_entry_sz(iolog));
2119 }
2120 cur_log->nr_samples = iolog->pending->nr_samples;
2121
2122 iolog->pending->nr_samples = 0;
2123 return cur_log;
2124disable:
2125 if (iolog)
2126 iolog->disabled = true;
2127 return NULL;
2128}
2129
2130void regrow_logs(struct thread_data *td)
2131{
2132 regrow_log(td->slat_log);
2133 regrow_log(td->clat_log);
2134 regrow_log(td->clat_hist_log);
2135 regrow_log(td->lat_log);
2136 regrow_log(td->bw_log);
2137 regrow_log(td->iops_log);
2138 td->flags &= ~TD_F_REGROW_LOGS;
2139}
2140
2141static struct io_logs *get_cur_log(struct io_log *iolog)
2142{
2143 struct io_logs *cur_log;
2144
2145 cur_log = iolog_cur_log(iolog);
2146 if (!cur_log) {
2147 cur_log = get_new_log(iolog);
2148 if (!cur_log)
2149 return NULL;
2150 }
2151
2152 if (cur_log->nr_samples < cur_log->max_samples)
2153 return cur_log;
2154
2155 /*
2156 * Out of space. If we're in IO offload mode, or we're not doing
2157 * per unit logging (hence logging happens outside of the IO thread
2158 * as well), add a new log chunk inline. If we're doing inline
2159 * submissions, flag 'td' as needing a log regrow and we'll take
2160 * care of it on the submission side.
2161 */
2162 if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2163 !per_unit_log(iolog))
2164 return regrow_log(iolog);
2165
2166 iolog->td->flags |= TD_F_REGROW_LOGS;
2167 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2168 return iolog->pending;
2169}
2170
2171static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2172 enum fio_ddir ddir, unsigned int bs,
2173 unsigned long t, uint64_t offset)
2174{
2175 struct io_logs *cur_log;
2176
2177 if (iolog->disabled)
2178 return;
2179 if (flist_empty(&iolog->io_logs))
2180 iolog->avg_last = t;
2181
2182 cur_log = get_cur_log(iolog);
2183 if (cur_log) {
2184 struct io_sample *s;
2185
2186 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2187
2188 s->data = data;
2189 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2190 io_sample_set_ddir(iolog, s, ddir);
2191 s->bs = bs;
2192
2193 if (iolog->log_offset) {
2194 struct io_sample_offset *so = (void *) s;
2195
2196 so->offset = offset;
2197 }
2198
2199 cur_log->nr_samples++;
2200 return;
2201 }
2202
2203 iolog->disabled = true;
2204}
2205
2206static inline void reset_io_stat(struct io_stat *ios)
2207{
2208 ios->max_val = ios->min_val = ios->samples = 0;
2209 ios->mean.u.f = ios->S.u.f = 0;
2210}
2211
2212void reset_io_stats(struct thread_data *td)
2213{
2214 struct thread_stat *ts = &td->ts;
2215 int i, j;
2216
2217 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2218 reset_io_stat(&ts->clat_stat[i]);
2219 reset_io_stat(&ts->slat_stat[i]);
2220 reset_io_stat(&ts->lat_stat[i]);
2221 reset_io_stat(&ts->bw_stat[i]);
2222 reset_io_stat(&ts->iops_stat[i]);
2223
2224 ts->io_bytes[i] = 0;
2225 ts->runtime[i] = 0;
2226 ts->total_io_u[i] = 0;
2227 ts->short_io_u[i] = 0;
2228 ts->drop_io_u[i] = 0;
2229
2230 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2231 ts->io_u_plat[i][j] = 0;
2232 }
2233
2234 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2235 ts->io_u_map[i] = 0;
2236 ts->io_u_submit[i] = 0;
2237 ts->io_u_complete[i] = 0;
2238 }
2239
2240 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2241 ts->io_u_lat_n[i] = 0;
2242 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2243 ts->io_u_lat_u[i] = 0;
2244 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2245 ts->io_u_lat_m[i] = 0;
2246
2247 ts->total_submit = 0;
2248 ts->total_complete = 0;
2249}
2250
2251static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2252 unsigned long elapsed, bool log_max)
2253{
2254 /*
2255 * Note an entry in the log. Use the mean from the logged samples,
2256 * making sure to properly round up. Only write a log entry if we
2257 * had actual samples done.
2258 */
2259 if (iolog->avg_window[ddir].samples) {
2260 union io_sample_data data;
2261
2262 if (log_max)
2263 data.val = iolog->avg_window[ddir].max_val;
2264 else
2265 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2266
2267 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2268 }
2269
2270 reset_io_stat(&iolog->avg_window[ddir]);
2271}
2272
2273static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2274 bool log_max)
2275{
2276 int ddir;
2277
2278 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2279 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2280}
2281
2282static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2283 union io_sample_data data, enum fio_ddir ddir,
2284 unsigned int bs, uint64_t offset)
2285{
2286 unsigned long elapsed, this_window;
2287
2288 if (!ddir_rw(ddir))
2289 return 0;
2290
2291 elapsed = mtime_since_now(&td->epoch);
2292
2293 /*
2294 * If no time averaging, just add the log sample.
2295 */
2296 if (!iolog->avg_msec) {
2297 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2298 return 0;
2299 }
2300
2301 /*
2302 * Add the sample. If the time period has passed, then
2303 * add that entry to the log and clear.
2304 */
2305 add_stat_sample(&iolog->avg_window[ddir], data.val);
2306
2307 /*
2308 * If period hasn't passed, adding the above sample is all we
2309 * need to do.
2310 */
2311 this_window = elapsed - iolog->avg_last;
2312 if (elapsed < iolog->avg_last)
2313 return iolog->avg_last - elapsed;
2314 else if (this_window < iolog->avg_msec) {
2315 int diff = iolog->avg_msec - this_window;
2316
2317 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2318 return diff;
2319 }
2320
2321 _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2322
2323 iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2324 return iolog->avg_msec;
2325}
2326
2327void finalize_logs(struct thread_data *td, bool unit_logs)
2328{
2329 unsigned long elapsed;
2330
2331 elapsed = mtime_since_now(&td->epoch);
2332
2333 if (td->clat_log && unit_logs)
2334 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2335 if (td->slat_log && unit_logs)
2336 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2337 if (td->lat_log && unit_logs)
2338 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2339 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2340 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2341 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2342 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2343}
2344
2345void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2346{
2347 struct io_log *iolog;
2348
2349 if (!ddir_rw(ddir))
2350 return;
2351
2352 iolog = agg_io_log[ddir];
2353 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2354}
2355
2356static void add_clat_percentile_sample(struct thread_stat *ts,
2357 unsigned long long nsec, enum fio_ddir ddir)
2358{
2359 unsigned int idx = plat_val_to_idx(nsec);
2360 assert(idx < FIO_IO_U_PLAT_NR);
2361
2362 ts->io_u_plat[ddir][idx]++;
2363}
2364
2365void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2366 unsigned long long nsec, unsigned int bs, uint64_t offset)
2367{
2368 unsigned long elapsed, this_window;
2369 struct thread_stat *ts = &td->ts;
2370 struct io_log *iolog = td->clat_hist_log;
2371
2372 td_io_u_lock(td);
2373
2374 add_stat_sample(&ts->clat_stat[ddir], nsec);
2375
2376 if (td->clat_log)
2377 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2378 offset);
2379
2380 if (ts->clat_percentiles)
2381 add_clat_percentile_sample(ts, nsec, ddir);
2382
2383 if (iolog && iolog->hist_msec) {
2384 struct io_hist *hw = &iolog->hist_window[ddir];
2385
2386 hw->samples++;
2387 elapsed = mtime_since_now(&td->epoch);
2388 if (!hw->hist_last)
2389 hw->hist_last = elapsed;
2390 this_window = elapsed - hw->hist_last;
2391
2392 if (this_window >= iolog->hist_msec) {
2393 unsigned int *io_u_plat;
2394 struct io_u_plat_entry *dst;
2395
2396 /*
2397 * Make a byte-for-byte copy of the latency histogram
2398 * stored in td->ts.io_u_plat[ddir], recording it in a
2399 * log sample. Note that the matching call to free() is
2400 * located in iolog.c after printing this sample to the
2401 * log file.
2402 */
2403 io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2404 dst = malloc(sizeof(struct io_u_plat_entry));
2405 memcpy(&(dst->io_u_plat), io_u_plat,
2406 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2407 flist_add(&dst->list, &hw->list);
2408 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2409 elapsed, offset);
2410
2411 /*
2412 * Update the last time we recorded as being now, minus
2413 * any drift in time we encountered before actually
2414 * making the record.
2415 */
2416 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2417 hw->samples = 0;
2418 }
2419 }
2420
2421 td_io_u_unlock(td);
2422}
2423
2424void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2425 unsigned long usec, unsigned int bs, uint64_t offset)
2426{
2427 struct thread_stat *ts = &td->ts;
2428
2429 if (!ddir_rw(ddir))
2430 return;
2431
2432 td_io_u_lock(td);
2433
2434 add_stat_sample(&ts->slat_stat[ddir], usec);
2435
2436 if (td->slat_log)
2437 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2438
2439 td_io_u_unlock(td);
2440}
2441
2442void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2443 unsigned long long nsec, unsigned int bs, uint64_t offset)
2444{
2445 struct thread_stat *ts = &td->ts;
2446
2447 if (!ddir_rw(ddir))
2448 return;
2449
2450 td_io_u_lock(td);
2451
2452 add_stat_sample(&ts->lat_stat[ddir], nsec);
2453
2454 if (td->lat_log)
2455 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2456 offset);
2457
2458 td_io_u_unlock(td);
2459}
2460
2461void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2462 unsigned int bytes, unsigned long long spent)
2463{
2464 struct thread_stat *ts = &td->ts;
2465 unsigned long rate;
2466
2467 if (spent)
2468 rate = (unsigned long) (bytes * 1000000ULL / spent);
2469 else
2470 rate = 0;
2471
2472 td_io_u_lock(td);
2473
2474 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2475
2476 if (td->bw_log)
2477 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2478 bytes, io_u->offset);
2479
2480 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2481 td_io_u_unlock(td);
2482}
2483
2484static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2485 struct timespec *t, unsigned int avg_time,
2486 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2487 struct io_stat *stat, struct io_log *log,
2488 bool is_kb)
2489{
2490 unsigned long spent, rate;
2491 enum fio_ddir ddir;
2492 unsigned int next, next_log;
2493
2494 next_log = avg_time;
2495
2496 spent = mtime_since(parent_tv, t);
2497 if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2498 return avg_time - spent;
2499
2500 td_io_u_lock(td);
2501
2502 /*
2503 * Compute both read and write rates for the interval.
2504 */
2505 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2506 uint64_t delta;
2507
2508 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2509 if (!delta)
2510 continue; /* No entries for interval */
2511
2512 if (spent) {
2513 if (is_kb)
2514 rate = delta * 1000 / spent / 1024; /* KiB/s */
2515 else
2516 rate = (delta * 1000) / spent;
2517 } else
2518 rate = 0;
2519
2520 add_stat_sample(&stat[ddir], rate);
2521
2522 if (log) {
2523 unsigned int bs = 0;
2524
2525 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2526 bs = td->o.min_bs[ddir];
2527
2528 next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2529 next_log = min(next_log, next);
2530 }
2531
2532 stat_io_bytes[ddir] = this_io_bytes[ddir];
2533 }
2534
2535 timespec_add_msec(parent_tv, avg_time);
2536
2537 td_io_u_unlock(td);
2538
2539 if (spent <= avg_time)
2540 next = avg_time;
2541 else
2542 next = avg_time - (1 + spent - avg_time);
2543
2544 return min(next, next_log);
2545}
2546
2547static int add_bw_samples(struct thread_data *td, struct timespec *t)
2548{
2549 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2550 td->this_io_bytes, td->stat_io_bytes,
2551 td->ts.bw_stat, td->bw_log, true);
2552}
2553
2554void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2555 unsigned int bytes)
2556{
2557 struct thread_stat *ts = &td->ts;
2558
2559 td_io_u_lock(td);
2560
2561 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2562
2563 if (td->iops_log)
2564 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2565 bytes, io_u->offset);
2566
2567 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2568 td_io_u_unlock(td);
2569}
2570
2571static int add_iops_samples(struct thread_data *td, struct timespec *t)
2572{
2573 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2574 td->this_io_blocks, td->stat_io_blocks,
2575 td->ts.iops_stat, td->iops_log, false);
2576}
2577
2578/*
2579 * Returns msecs to next event
2580 */
2581int calc_log_samples(void)
2582{
2583 struct thread_data *td;
2584 unsigned int next = ~0U, tmp;
2585 struct timespec now;
2586 int i;
2587
2588 fio_gettime(&now, NULL);
2589
2590 for_each_td(td, i) {
2591 if (!td->o.stats)
2592 continue;
2593 if (in_ramp_time(td) ||
2594 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2595 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2596 continue;
2597 }
2598 if (!td->bw_log ||
2599 (td->bw_log && !per_unit_log(td->bw_log))) {
2600 tmp = add_bw_samples(td, &now);
2601 if (tmp < next)
2602 next = tmp;
2603 }
2604 if (!td->iops_log ||
2605 (td->iops_log && !per_unit_log(td->iops_log))) {
2606 tmp = add_iops_samples(td, &now);
2607 if (tmp < next)
2608 next = tmp;
2609 }
2610 }
2611
2612 return next == ~0U ? 0 : next;
2613}
2614
2615void stat_init(void)
2616{
2617 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2618}
2619
2620void stat_exit(void)
2621{
2622 /*
2623 * When we have the mutex, we know out-of-band access to it
2624 * have ended.
2625 */
2626 fio_mutex_down(stat_mutex);
2627 fio_mutex_remove(stat_mutex);
2628}
2629
2630/*
2631 * Called from signal handler. Wake up status thread.
2632 */
2633void show_running_run_stats(void)
2634{
2635 helper_do_stat();
2636}
2637
2638uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2639{
2640 /* Ignore io_u's which span multiple blocks--they will just get
2641 * inaccurate counts. */
2642 int idx = (io_u->offset - io_u->file->file_offset)
2643 / td->o.bs[DDIR_TRIM];
2644 uint32_t *info = &td->ts.block_infos[idx];
2645 assert(idx < td->ts.nr_block_infos);
2646 return info;
2647}
fio - Flexible IO Tester