projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
stat: Print number of samples in bw and iops stats
[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,
235 precision, time_width);
236 /* fmt will be something like " %5.2fth=[%4llu]%c" */
237 per_line = (80 - 7) / (precision + 10 + time_width);
238
239 for (j = 0; j < len; j++) {
240 /* for formatting */
241 if (j != 0 && (j % per_line) == 0)
242 log_buf(out, " |");
243
244 /* end of the list */
245 is_last = (j == len - 1);
246
247 for (i = 0; i < scale_down; i++)
248 ovals[j] = (ovals[j] + 999) / 1000;
249
250 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
251
252 if (is_last)
253 break;
254
255 if ((j % per_line) == per_line - 1) /* for formatting */
256 log_buf(out, "\n");
257 }
258
259out:
260 if (ovals)
261 free(ovals);
262}
263
264bool calc_lat(struct io_stat *is, unsigned long long *min,
265 unsigned long long *max, double *mean, double *dev)
266{
267 double n = (double) is->samples;
268
269 if (n == 0)
270 return false;
271
272 *min = is->min_val;
273 *max = is->max_val;
274 *mean = is->mean.u.f;
275
276 if (n > 1.0)
277 *dev = sqrt(is->S.u.f / (n - 1.0));
278 else
279 *dev = 0;
280
281 return true;
282}
283
284void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
285{
286 char *io, *agg, *min, *max;
287 char *ioalt, *aggalt, *minalt, *maxalt;
288 const char *str[] = { " READ", " WRITE" , " TRIM"};
289 int i;
290
291 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
292
293 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
294 const int i2p = is_power_of_2(rs->kb_base);
295
296 if (!rs->max_run[i])
297 continue;
298
299 io = num2str(rs->iobytes[i], 4, 1, i2p, N2S_BYTE);
300 ioalt = num2str(rs->iobytes[i], 4, 1, !i2p, N2S_BYTE);
301 agg = num2str(rs->agg[i], 4, 1, i2p, rs->unit_base);
302 aggalt = num2str(rs->agg[i], 4, 1, !i2p, rs->unit_base);
303 min = num2str(rs->min_bw[i], 4, 1, i2p, rs->unit_base);
304 minalt = num2str(rs->min_bw[i], 4, 1, !i2p, rs->unit_base);
305 max = num2str(rs->max_bw[i], 4, 1, i2p, rs->unit_base);
306 maxalt = num2str(rs->max_bw[i], 4, 1, !i2p, rs->unit_base);
307 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
308 rs->unified_rw_rep ? " MIXED" : str[i],
309 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
310 (unsigned long long) rs->min_run[i],
311 (unsigned long long) rs->max_run[i]);
312
313 free(io);
314 free(agg);
315 free(min);
316 free(max);
317 free(ioalt);
318 free(aggalt);
319 free(minalt);
320 free(maxalt);
321 }
322}
323
324void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
325{
326 int i;
327
328 /*
329 * Do depth distribution calculations
330 */
331 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
332 if (total) {
333 io_u_dist[i] = (double) map[i] / (double) total;
334 io_u_dist[i] *= 100.0;
335 if (io_u_dist[i] < 0.1 && map[i])
336 io_u_dist[i] = 0.1;
337 } else
338 io_u_dist[i] = 0.0;
339 }
340}
341
342static void stat_calc_lat(struct thread_stat *ts, double *dst,
343 unsigned int *src, int nr)
344{
345 unsigned long total = ddir_rw_sum(ts->total_io_u);
346 int i;
347
348 /*
349 * Do latency distribution calculations
350 */
351 for (i = 0; i < nr; i++) {
352 if (total) {
353 dst[i] = (double) src[i] / (double) total;
354 dst[i] *= 100.0;
355 if (dst[i] < 0.01 && src[i])
356 dst[i] = 0.01;
357 } else
358 dst[i] = 0.0;
359 }
360}
361
362/*
363 * To keep the terse format unaltered, add all of the ns latency
364 * buckets to the first us latency bucket
365 */
366void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
367{
368 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
369 int i;
370
371 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
372
373 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
374 ntotal += ts->io_u_lat_n[i];
375
376 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
377}
378
379void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
380{
381 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
382}
383
384void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
385{
386 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
387}
388
389void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
390{
391 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
392}
393
394static void display_lat(const char *name, unsigned long long min,
395 unsigned long long max, double mean, double dev,
396 struct buf_output *out)
397{
398 const char *base = "(nsec)";
399 char *minp, *maxp;
400
401 if (nsec_to_msec(&min, &max, &mean, &dev))
402 base = "(msec)";
403 else if (nsec_to_usec(&min, &max, &mean, &dev))
404 base = "(usec)";
405
406 minp = num2str(min, 6, 1, 0, N2S_NONE);
407 maxp = num2str(max, 6, 1, 0, N2S_NONE);
408
409 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
410 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
411
412 free(minp);
413 free(maxp);
414}
415
416static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
417 int ddir, struct buf_output *out)
418{
419 const char *str[] = { " read", "write", " trim" };
420 unsigned long runt;
421 unsigned long long min, max, bw, iops;
422 double mean, dev;
423 char *io_p, *bw_p, *bw_p_alt, *iops_p;
424 int i2p;
425
426 assert(ddir_rw(ddir));
427
428 if (!ts->runtime[ddir])
429 return;
430
431 i2p = is_power_of_2(rs->kb_base);
432 runt = ts->runtime[ddir];
433
434 bw = (1000 * ts->io_bytes[ddir]) / runt;
435 io_p = num2str(ts->io_bytes[ddir], 4, 1, i2p, N2S_BYTE);
436 bw_p = num2str(bw, 4, 1, i2p, ts->unit_base);
437 bw_p_alt = num2str(bw, 4, 1, !i2p, ts->unit_base);
438
439 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
440 iops_p = num2str(iops, 4, 1, 0, N2S_NONE);
441
442 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
443 rs->unified_rw_rep ? "mixed" : str[ddir],
444 iops_p, bw_p, bw_p_alt, io_p,
445 (unsigned long long) ts->runtime[ddir]);
446
447 free(io_p);
448 free(bw_p);
449 free(bw_p_alt);
450 free(iops_p);
451
452 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
453 display_lat("slat", min, max, mean, dev, out);
454 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
455 display_lat("clat", min, max, mean, dev, out);
456 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
457 display_lat(" lat", min, max, mean, dev, out);
458
459 if (ts->clat_percentiles) {
460 show_clat_percentiles(ts->io_u_plat[ddir],
461 ts->clat_stat[ddir].samples,
462 ts->percentile_list,
463 ts->percentile_precision, out);
464 }
465 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
466 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
467 const char *bw_str;
468
469 if ((rs->unit_base == 1) && i2p)
470 bw_str = "Kibit";
471 else if (rs->unit_base == 1)
472 bw_str = "kbit";
473 else if (i2p)
474 bw_str = "KiB";
475 else
476 bw_str = "kB";
477
478 if (rs->unit_base == 1) {
479 min *= 8.0;
480 max *= 8.0;
481 mean *= 8.0;
482 dev *= 8.0;
483 }
484
485 if (rs->agg[ddir]) {
486 p_of_agg = mean * 100 / (double) rs->agg[ddir];
487 if (p_of_agg > 100.0)
488 p_of_agg = 100.0;
489 }
490
491 if (mean > fkb_base * fkb_base) {
492 min /= fkb_base;
493 max /= fkb_base;
494 mean /= fkb_base;
495 dev /= fkb_base;
496 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
497 }
498
499 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
500 "avg=%5.02f, stdev=%5.02f, samples=%5lu\n",
501 bw_str, min, max, p_of_agg, mean, dev,
502 (&ts->bw_stat[ddir])->samples);
503 }
504 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
505 log_buf(out, " iops : min=%5llu, max=%5llu, avg=%5.02f, "
506 "stdev=%5.02f, samples=%5lu\n",
507 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
508 }
509}
510
511static int show_lat(double *io_u_lat, int nr, const char **ranges,
512 const char *msg, struct buf_output *out)
513{
514 int new_line = 1, i, line = 0, shown = 0;
515
516 for (i = 0; i < nr; i++) {
517 if (io_u_lat[i] <= 0.0)
518 continue;
519 shown = 1;
520 if (new_line) {
521 if (line)
522 log_buf(out, "\n");
523 log_buf(out, " lat (%s) : ", msg);
524 new_line = 0;
525 line = 0;
526 }
527 if (line)
528 log_buf(out, ", ");
529 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
530 line++;
531 if (line == 5)
532 new_line = 1;
533 }
534
535 if (shown)
536 log_buf(out, "\n");
537
538 return shown;
539}
540
541static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
542{
543 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
544 "250=", "500=", "750=", "1000=", };
545
546 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
547}
548
549static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
550{
551 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
552 "250=", "500=", "750=", "1000=", };
553
554 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
555}
556
557static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
558{
559 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
560 "250=", "500=", "750=", "1000=", "2000=",
561 ">=2000=", };
562
563 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
564}
565
566static void show_latencies(struct thread_stat *ts, struct buf_output *out)
567{
568 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
569 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
570 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
571
572 stat_calc_lat_n(ts, io_u_lat_n);
573 stat_calc_lat_u(ts, io_u_lat_u);
574 stat_calc_lat_m(ts, io_u_lat_m);
575
576 show_lat_n(io_u_lat_n, out);
577 show_lat_u(io_u_lat_u, out);
578 show_lat_m(io_u_lat_m, out);
579}
580
581static int block_state_category(int block_state)
582{
583 switch (block_state) {
584 case BLOCK_STATE_UNINIT:
585 return 0;
586 case BLOCK_STATE_TRIMMED:
587 case BLOCK_STATE_WRITTEN:
588 return 1;
589 case BLOCK_STATE_WRITE_FAILURE:
590 case BLOCK_STATE_TRIM_FAILURE:
591 return 2;
592 default:
593 /* Silence compile warning on some BSDs and have a return */
594 assert(0);
595 return -1;
596 }
597}
598
599static int compare_block_infos(const void *bs1, const void *bs2)
600{
601 uint32_t block1 = *(uint32_t *)bs1;
602 uint32_t block2 = *(uint32_t *)bs2;
603 int state1 = BLOCK_INFO_STATE(block1);
604 int state2 = BLOCK_INFO_STATE(block2);
605 int bscat1 = block_state_category(state1);
606 int bscat2 = block_state_category(state2);
607 int cycles1 = BLOCK_INFO_TRIMS(block1);
608 int cycles2 = BLOCK_INFO_TRIMS(block2);
609
610 if (bscat1 < bscat2)
611 return -1;
612 if (bscat1 > bscat2)
613 return 1;
614
615 if (cycles1 < cycles2)
616 return -1;
617 if (cycles1 > cycles2)
618 return 1;
619
620 if (state1 < state2)
621 return -1;
622 if (state1 > state2)
623 return 1;
624
625 assert(block1 == block2);
626 return 0;
627}
628
629static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
630 fio_fp64_t *plist, unsigned int **percentiles,
631 unsigned int *types)
632{
633 int len = 0;
634 int i, nr_uninit;
635
636 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
637
638 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
639 len++;
640
641 if (!len)
642 return 0;
643
644 /*
645 * Sort the percentile list. Note that it may already be sorted if
646 * we are using the default values, but since it's a short list this
647 * isn't a worry. Also note that this does not work for NaN values.
648 */
649 if (len > 1)
650 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
651
652 nr_uninit = 0;
653 /* Start only after the uninit entries end */
654 for (nr_uninit = 0;
655 nr_uninit < nr_block_infos
656 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
657 nr_uninit ++)
658 ;
659
660 if (nr_uninit == nr_block_infos)
661 return 0;
662
663 *percentiles = calloc(len, sizeof(**percentiles));
664
665 for (i = 0; i < len; i++) {
666 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
667 + nr_uninit;
668 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
669 }
670
671 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
672 for (i = 0; i < nr_block_infos; i++)
673 types[BLOCK_INFO_STATE(block_infos[i])]++;
674
675 return len;
676}
677
678static const char *block_state_names[] = {
679 [BLOCK_STATE_UNINIT] = "unwritten",
680 [BLOCK_STATE_TRIMMED] = "trimmed",
681 [BLOCK_STATE_WRITTEN] = "written",
682 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
683 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
684};
685
686static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
687 fio_fp64_t *plist, struct buf_output *out)
688{
689 int len, pos, i;
690 unsigned int *percentiles = NULL;
691 unsigned int block_state_counts[BLOCK_STATE_COUNT];
692
693 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
694 &percentiles, block_state_counts);
695
696 log_buf(out, " block lifetime percentiles :\n |");
697 pos = 0;
698 for (i = 0; i < len; i++) {
699 uint32_t block_info = percentiles[i];
700#define LINE_LENGTH 75
701 char str[LINE_LENGTH];
702 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
703 plist[i].u.f, block_info,
704 i == len - 1 ? '\n' : ',');
705 assert(strln < LINE_LENGTH);
706 if (pos + strln > LINE_LENGTH) {
707 pos = 0;
708 log_buf(out, "\n |");
709 }
710 log_buf(out, "%s", str);
711 pos += strln;
712#undef LINE_LENGTH
713 }
714 if (percentiles)
715 free(percentiles);
716
717 log_buf(out, " states :");
718 for (i = 0; i < BLOCK_STATE_COUNT; i++)
719 log_buf(out, " %s=%u%c",
720 block_state_names[i], block_state_counts[i],
721 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
722}
723
724static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
725{
726 char *p1, *p1alt, *p2;
727 unsigned long long bw_mean, iops_mean;
728 const int i2p = is_power_of_2(ts->kb_base);
729
730 if (!ts->ss_dur)
731 return;
732
733 bw_mean = steadystate_bw_mean(ts);
734 iops_mean = steadystate_iops_mean(ts);
735
736 p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
737 p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
738 p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
739
740 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
741 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
742 p1, p1alt, p2,
743 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
744 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
745 ts->ss_criterion.u.f,
746 ts->ss_state & __FIO_SS_PCT ? "%" : "");
747
748 free(p1);
749 free(p1alt);
750 free(p2);
751}
752
753static void show_thread_status_normal(struct thread_stat *ts,
754 struct group_run_stats *rs,
755 struct buf_output *out)
756{
757 double usr_cpu, sys_cpu;
758 unsigned long runtime;
759 double io_u_dist[FIO_IO_U_MAP_NR];
760 time_t time_p;
761 char time_buf[32];
762
763 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
764 return;
765
766 memset(time_buf, 0, sizeof(time_buf));
767
768 time(&time_p);
769 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
770
771 if (!ts->error) {
772 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
773 ts->name, ts->groupid, ts->members,
774 ts->error, (int) ts->pid, time_buf);
775 } else {
776 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
777 ts->name, ts->groupid, ts->members,
778 ts->error, ts->verror, (int) ts->pid,
779 time_buf);
780 }
781
782 if (strlen(ts->description))
783 log_buf(out, " Description : [%s]\n", ts->description);
784
785 if (ts->io_bytes[DDIR_READ])
786 show_ddir_status(rs, ts, DDIR_READ, out);
787 if (ts->io_bytes[DDIR_WRITE])
788 show_ddir_status(rs, ts, DDIR_WRITE, out);
789 if (ts->io_bytes[DDIR_TRIM])
790 show_ddir_status(rs, ts, DDIR_TRIM, out);
791
792 show_latencies(ts, out);
793
794 runtime = ts->total_run_time;
795 if (runtime) {
796 double runt = (double) runtime;
797
798 usr_cpu = (double) ts->usr_time * 100 / runt;
799 sys_cpu = (double) ts->sys_time * 100 / runt;
800 } else {
801 usr_cpu = 0;
802 sys_cpu = 0;
803 }
804
805 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
806 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
807 (unsigned long long) ts->ctx,
808 (unsigned long long) ts->majf,
809 (unsigned long long) ts->minf);
810
811 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
812 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
813 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
814 io_u_dist[1], io_u_dist[2],
815 io_u_dist[3], io_u_dist[4],
816 io_u_dist[5], io_u_dist[6]);
817
818 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
819 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
820 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
821 io_u_dist[1], io_u_dist[2],
822 io_u_dist[3], io_u_dist[4],
823 io_u_dist[5], io_u_dist[6]);
824 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
825 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
826 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
827 io_u_dist[1], io_u_dist[2],
828 io_u_dist[3], io_u_dist[4],
829 io_u_dist[5], io_u_dist[6]);
830 log_buf(out, " issued rwt: total=%llu,%llu,%llu,"
831 " short=%llu,%llu,%llu,"
832 " dropped=%llu,%llu,%llu\n",
833 (unsigned long long) ts->total_io_u[0],
834 (unsigned long long) ts->total_io_u[1],
835 (unsigned long long) ts->total_io_u[2],
836 (unsigned long long) ts->short_io_u[0],
837 (unsigned long long) ts->short_io_u[1],
838 (unsigned long long) ts->short_io_u[2],
839 (unsigned long long) ts->drop_io_u[0],
840 (unsigned long long) ts->drop_io_u[1],
841 (unsigned long long) ts->drop_io_u[2]);
842 if (ts->continue_on_error) {
843 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
844 (unsigned long long)ts->total_err_count,
845 ts->first_error,
846 strerror(ts->first_error));
847 }
848 if (ts->latency_depth) {
849 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
850 (unsigned long long)ts->latency_target,
851 (unsigned long long)ts->latency_window,
852 ts->latency_percentile.u.f,
853 ts->latency_depth);
854 }
855
856 if (ts->nr_block_infos)
857 show_block_infos(ts->nr_block_infos, ts->block_infos,
858 ts->percentile_list, out);
859
860 if (ts->ss_dur)
861 show_ss_normal(ts, out);
862}
863
864static void show_ddir_status_terse(struct thread_stat *ts,
865 struct group_run_stats *rs, int ddir,
866 struct buf_output *out)
867{
868 unsigned long long min, max, minv, maxv, bw, iops;
869 unsigned long long *ovals = NULL;
870 double mean, dev;
871 unsigned int len;
872 int i;
873
874 assert(ddir_rw(ddir));
875
876 iops = bw = 0;
877 if (ts->runtime[ddir]) {
878 uint64_t runt = ts->runtime[ddir];
879
880 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
881 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
882 }
883
884 log_buf(out, ";%llu;%llu;%llu;%llu",
885 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
886 (unsigned long long) ts->runtime[ddir]);
887
888 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
889 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
890 else
891 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
892
893 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
894 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
895 else
896 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
897
898 if (ts->clat_percentiles) {
899 len = calc_clat_percentiles(ts->io_u_plat[ddir],
900 ts->clat_stat[ddir].samples,
901 ts->percentile_list, &ovals, &maxv,
902 &minv);
903 } else
904 len = 0;
905
906 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
907 if (i >= len) {
908 log_buf(out, ";0%%=0");
909 continue;
910 }
911 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
912 }
913
914 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
915 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
916 else
917 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
918
919 if (ovals)
920 free(ovals);
921
922 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
923 double p_of_agg = 100.0;
924
925 if (rs->agg[ddir]) {
926 p_of_agg = mean * 100 / (double) rs->agg[ddir];
927 if (p_of_agg > 100.0)
928 p_of_agg = 100.0;
929 }
930
931 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
932 } else
933 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
934}
935
936static void add_ddir_status_json(struct thread_stat *ts,
937 struct group_run_stats *rs, int ddir, struct json_object *parent)
938{
939 unsigned long long min, max, minv, maxv;
940 unsigned long long bw;
941 unsigned long long *ovals = NULL;
942 double mean, dev, iops;
943 unsigned int len;
944 int i;
945 const char *ddirname[] = {"read", "write", "trim"};
946 struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
947 char buf[120];
948 double p_of_agg = 100.0;
949
950 assert(ddir_rw(ddir));
951
952 if (ts->unified_rw_rep && ddir != DDIR_READ)
953 return;
954
955 dir_object = json_create_object();
956 json_object_add_value_object(parent,
957 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
958
959 bw = 0;
960 iops = 0.0;
961 if (ts->runtime[ddir]) {
962 uint64_t runt = ts->runtime[ddir];
963
964 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
965 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
966 }
967
968 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
969 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
970 json_object_add_value_int(dir_object, "bw", bw);
971 json_object_add_value_float(dir_object, "iops", iops);
972 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
973 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
974 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
975 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
976
977 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
978 min = max = 0;
979 mean = dev = 0.0;
980 }
981 tmp_object = json_create_object();
982 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
983 json_object_add_value_int(tmp_object, "min", min);
984 json_object_add_value_int(tmp_object, "max", max);
985 json_object_add_value_float(tmp_object, "mean", mean);
986 json_object_add_value_float(tmp_object, "stddev", dev);
987
988 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
989 min = max = 0;
990 mean = dev = 0.0;
991 }
992 tmp_object = json_create_object();
993 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
994 json_object_add_value_int(tmp_object, "min", min);
995 json_object_add_value_int(tmp_object, "max", max);
996 json_object_add_value_float(tmp_object, "mean", mean);
997 json_object_add_value_float(tmp_object, "stddev", dev);
998
999 if (ts->clat_percentiles) {
1000 len = calc_clat_percentiles(ts->io_u_plat[ddir],
1001 ts->clat_stat[ddir].samples,
1002 ts->percentile_list, &ovals, &maxv,
1003 &minv);
1004 } else
1005 len = 0;
1006
1007 percentile_object = json_create_object();
1008 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1009 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1010 if (i >= len) {
1011 json_object_add_value_int(percentile_object, "0.00", 0);
1012 continue;
1013 }
1014 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1015 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1016 }
1017
1018 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1019 clat_bins_object = json_create_object();
1020 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1021 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1022 if (ts->io_u_plat[ddir][i]) {
1023 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1024 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1025 }
1026 }
1027 }
1028
1029 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1030 min = max = 0;
1031 mean = dev = 0.0;
1032 }
1033 tmp_object = json_create_object();
1034 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1035 json_object_add_value_int(tmp_object, "min", min);
1036 json_object_add_value_int(tmp_object, "max", max);
1037 json_object_add_value_float(tmp_object, "mean", mean);
1038 json_object_add_value_float(tmp_object, "stddev", dev);
1039 if (ovals)
1040 free(ovals);
1041
1042 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1043 if (rs->agg[ddir]) {
1044 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1045 if (p_of_agg > 100.0)
1046 p_of_agg = 100.0;
1047 }
1048 } else {
1049 min = max = 0;
1050 p_of_agg = mean = dev = 0.0;
1051 }
1052 json_object_add_value_int(dir_object, "bw_min", min);
1053 json_object_add_value_int(dir_object, "bw_max", max);
1054 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1055 json_object_add_value_float(dir_object, "bw_mean", mean);
1056 json_object_add_value_float(dir_object, "bw_dev", dev);
1057 json_object_add_value_int(dir_object, "bw_samples",
1058 (&ts->bw_stat[ddir])->samples);
1059
1060 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1061 min = max = 0;
1062 mean = dev = 0.0;
1063 }
1064 json_object_add_value_int(dir_object, "iops_min", min);
1065 json_object_add_value_int(dir_object, "iops_max", max);
1066 json_object_add_value_float(dir_object, "iops_mean", mean);
1067 json_object_add_value_float(dir_object, "iops_stddev", dev);
1068 json_object_add_value_int(dir_object, "iops_samples",
1069 (&ts->iops_stat[ddir])->samples);
1070}
1071
1072static void show_thread_status_terse_v2(struct thread_stat *ts,
1073 struct group_run_stats *rs,
1074 struct buf_output *out)
1075{
1076 double io_u_dist[FIO_IO_U_MAP_NR];
1077 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1078 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1079 double usr_cpu, sys_cpu;
1080 int i;
1081
1082 /* General Info */
1083 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1084 /* Log Read Status */
1085 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1086 /* Log Write Status */
1087 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1088 /* Log Trim Status */
1089 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1090
1091 /* CPU Usage */
1092 if (ts->total_run_time) {
1093 double runt = (double) ts->total_run_time;
1094
1095 usr_cpu = (double) ts->usr_time * 100 / runt;
1096 sys_cpu = (double) ts->sys_time * 100 / runt;
1097 } else {
1098 usr_cpu = 0;
1099 sys_cpu = 0;
1100 }
1101
1102 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1103 (unsigned long long) ts->ctx,
1104 (unsigned long long) ts->majf,
1105 (unsigned long long) ts->minf);
1106
1107 /* Calc % distribution of IO depths, usecond, msecond latency */
1108 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1109 stat_calc_lat_nu(ts, io_u_lat_u);
1110 stat_calc_lat_m(ts, io_u_lat_m);
1111
1112 /* Only show fixed 7 I/O depth levels*/
1113 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1114 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1115 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1116
1117 /* Microsecond latency */
1118 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1119 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1120 /* Millisecond latency */
1121 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1122 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1123 /* Additional output if continue_on_error set - default off*/
1124 if (ts->continue_on_error)
1125 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1126 log_buf(out, "\n");
1127
1128 /* Additional output if description is set */
1129 if (strlen(ts->description))
1130 log_buf(out, ";%s", ts->description);
1131
1132 log_buf(out, "\n");
1133}
1134
1135static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1136 struct group_run_stats *rs, int ver,
1137 struct buf_output *out)
1138{
1139 double io_u_dist[FIO_IO_U_MAP_NR];
1140 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1141 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1142 double usr_cpu, sys_cpu;
1143 int i;
1144
1145 /* General Info */
1146 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1147 ts->name, ts->groupid, ts->error);
1148 /* Log Read Status */
1149 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1150 /* Log Write Status */
1151 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1152 /* Log Trim Status */
1153 if (ver == 4)
1154 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1155
1156 /* CPU Usage */
1157 if (ts->total_run_time) {
1158 double runt = (double) ts->total_run_time;
1159
1160 usr_cpu = (double) ts->usr_time * 100 / runt;
1161 sys_cpu = (double) ts->sys_time * 100 / runt;
1162 } else {
1163 usr_cpu = 0;
1164 sys_cpu = 0;
1165 }
1166
1167 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1168 (unsigned long long) ts->ctx,
1169 (unsigned long long) ts->majf,
1170 (unsigned long long) ts->minf);
1171
1172 /* Calc % distribution of IO depths, usecond, msecond latency */
1173 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1174 stat_calc_lat_nu(ts, io_u_lat_u);
1175 stat_calc_lat_m(ts, io_u_lat_m);
1176
1177 /* Only show fixed 7 I/O depth levels*/
1178 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1179 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1180 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1181
1182 /* Microsecond latency */
1183 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1184 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1185 /* Millisecond latency */
1186 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1187 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1188
1189 /* disk util stats, if any */
1190 show_disk_util(1, NULL, out);
1191
1192 /* Additional output if continue_on_error set - default off*/
1193 if (ts->continue_on_error)
1194 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1195
1196 /* Additional output if description is set */
1197 if (strlen(ts->description))
1198 log_buf(out, ";%s", ts->description);
1199
1200 log_buf(out, "\n");
1201}
1202
1203static void json_add_job_opts(struct json_object *root, const char *name,
1204 struct flist_head *opt_list, bool num_jobs)
1205{
1206 struct json_object *dir_object;
1207 struct flist_head *entry;
1208 struct print_option *p;
1209
1210 if (flist_empty(opt_list))
1211 return;
1212
1213 dir_object = json_create_object();
1214 json_object_add_value_object(root, name, dir_object);
1215
1216 flist_for_each(entry, opt_list) {
1217 const char *pos = "";
1218
1219 p = flist_entry(entry, struct print_option, list);
1220 if (!num_jobs && !strcmp(p->name, "numjobs"))
1221 continue;
1222 if (p->value)
1223 pos = p->value;
1224 json_object_add_value_string(dir_object, p->name, pos);
1225 }
1226}
1227
1228static struct json_object *show_thread_status_json(struct thread_stat *ts,
1229 struct group_run_stats *rs,
1230 struct flist_head *opt_list)
1231{
1232 struct json_object *root, *tmp;
1233 struct jobs_eta *je;
1234 double io_u_dist[FIO_IO_U_MAP_NR];
1235 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1236 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1237 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1238 double usr_cpu, sys_cpu;
1239 int i;
1240 size_t size;
1241
1242 root = json_create_object();
1243 json_object_add_value_string(root, "jobname", ts->name);
1244 json_object_add_value_int(root, "groupid", ts->groupid);
1245 json_object_add_value_int(root, "error", ts->error);
1246
1247 /* ETA Info */
1248 je = get_jobs_eta(true, &size);
1249 if (je) {
1250 json_object_add_value_int(root, "eta", je->eta_sec);
1251 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1252 }
1253
1254 if (opt_list)
1255 json_add_job_opts(root, "job options", opt_list, true);
1256
1257 add_ddir_status_json(ts, rs, DDIR_READ, root);
1258 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1259 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1260
1261 /* CPU Usage */
1262 if (ts->total_run_time) {
1263 double runt = (double) ts->total_run_time;
1264
1265 usr_cpu = (double) ts->usr_time * 100 / runt;
1266 sys_cpu = (double) ts->sys_time * 100 / runt;
1267 } else {
1268 usr_cpu = 0;
1269 sys_cpu = 0;
1270 }
1271 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1272 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1273 json_object_add_value_int(root, "ctx", ts->ctx);
1274 json_object_add_value_int(root, "majf", ts->majf);
1275 json_object_add_value_int(root, "minf", ts->minf);
1276
1277
1278 /* Calc % distribution of IO depths, usecond, msecond latency */
1279 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1280 stat_calc_lat_n(ts, io_u_lat_n);
1281 stat_calc_lat_u(ts, io_u_lat_u);
1282 stat_calc_lat_m(ts, io_u_lat_m);
1283
1284 tmp = json_create_object();
1285 json_object_add_value_object(root, "iodepth_level", tmp);
1286 /* Only show fixed 7 I/O depth levels*/
1287 for (i = 0; i < 7; i++) {
1288 char name[20];
1289 if (i < 6)
1290 snprintf(name, 20, "%d", 1 << i);
1291 else
1292 snprintf(name, 20, ">=%d", 1 << i);
1293 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1294 }
1295
1296 /* Nanosecond latency */
1297 tmp = json_create_object();
1298 json_object_add_value_object(root, "latency_ns", tmp);
1299 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1300 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1301 "250", "500", "750", "1000", };
1302 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1303 }
1304 /* Microsecond latency */
1305 tmp = json_create_object();
1306 json_object_add_value_object(root, "latency_us", tmp);
1307 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1308 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1309 "250", "500", "750", "1000", };
1310 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1311 }
1312 /* Millisecond latency */
1313 tmp = json_create_object();
1314 json_object_add_value_object(root, "latency_ms", tmp);
1315 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1316 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1317 "250", "500", "750", "1000", "2000",
1318 ">=2000", };
1319 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1320 }
1321
1322 /* Additional output if continue_on_error set - default off*/
1323 if (ts->continue_on_error) {
1324 json_object_add_value_int(root, "total_err", ts->total_err_count);
1325 json_object_add_value_int(root, "first_error", ts->first_error);
1326 }
1327
1328 if (ts->latency_depth) {
1329 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1330 json_object_add_value_int(root, "latency_target", ts->latency_target);
1331 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1332 json_object_add_value_int(root, "latency_window", ts->latency_window);
1333 }
1334
1335 /* Additional output if description is set */
1336 if (strlen(ts->description))
1337 json_object_add_value_string(root, "desc", ts->description);
1338
1339 if (ts->nr_block_infos) {
1340 /* Block error histogram and types */
1341 int len;
1342 unsigned int *percentiles = NULL;
1343 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1344
1345 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1346 ts->percentile_list,
1347 &percentiles, block_state_counts);
1348
1349 if (len) {
1350 struct json_object *block, *percentile_object, *states;
1351 int state;
1352 block = json_create_object();
1353 json_object_add_value_object(root, "block", block);
1354
1355 percentile_object = json_create_object();
1356 json_object_add_value_object(block, "percentiles",
1357 percentile_object);
1358 for (i = 0; i < len; i++) {
1359 char buf[20];
1360 snprintf(buf, sizeof(buf), "%f",
1361 ts->percentile_list[i].u.f);
1362 json_object_add_value_int(percentile_object,
1363 (const char *)buf,
1364 percentiles[i]);
1365 }
1366
1367 states = json_create_object();
1368 json_object_add_value_object(block, "states", states);
1369 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1370 json_object_add_value_int(states,
1371 block_state_names[state],
1372 block_state_counts[state]);
1373 }
1374 free(percentiles);
1375 }
1376 }
1377
1378 if (ts->ss_dur) {
1379 struct json_object *data;
1380 struct json_array *iops, *bw;
1381 int i, j, k;
1382 char ss_buf[64];
1383
1384 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1385 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1386 ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1387 (float) ts->ss_limit.u.f,
1388 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1389
1390 tmp = json_create_object();
1391 json_object_add_value_object(root, "steadystate", tmp);
1392 json_object_add_value_string(tmp, "ss", ss_buf);
1393 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1394 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1395
1396 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1397 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1398 json_object_add_value_string(tmp, "criterion", ss_buf);
1399 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1400 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1401
1402 data = json_create_object();
1403 json_object_add_value_object(tmp, "data", data);
1404 bw = json_create_array();
1405 iops = json_create_array();
1406
1407 /*
1408 ** if ss was attained or the buffer is not full,
1409 ** ss->head points to the first element in the list.
1410 ** otherwise it actually points to the second element
1411 ** in the list
1412 */
1413 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1414 j = ts->ss_head;
1415 else
1416 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1417 for (i = 0; i < ts->ss_dur; i++) {
1418 k = (j + i) % ts->ss_dur;
1419 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1420 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1421 }
1422 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1423 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1424 json_object_add_value_array(data, "iops", iops);
1425 json_object_add_value_array(data, "bw", bw);
1426 }
1427
1428 return root;
1429}
1430
1431static void show_thread_status_terse(struct thread_stat *ts,
1432 struct group_run_stats *rs,
1433 struct buf_output *out)
1434{
1435 if (terse_version == 2)
1436 show_thread_status_terse_v2(ts, rs, out);
1437 else if (terse_version == 3 || terse_version == 4)
1438 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1439 else
1440 log_err("fio: bad terse version!? %d\n", terse_version);
1441}
1442
1443struct json_object *show_thread_status(struct thread_stat *ts,
1444 struct group_run_stats *rs,
1445 struct flist_head *opt_list,
1446 struct buf_output *out)
1447{
1448 struct json_object *ret = NULL;
1449
1450 if (output_format & FIO_OUTPUT_TERSE)
1451 show_thread_status_terse(ts, rs, out);
1452 if (output_format & FIO_OUTPUT_JSON)
1453 ret = show_thread_status_json(ts, rs, opt_list);
1454 if (output_format & FIO_OUTPUT_NORMAL)
1455 show_thread_status_normal(ts, rs, out);
1456
1457 return ret;
1458}
1459
1460static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1461{
1462 double mean, S;
1463
1464 if (src->samples == 0)
1465 return;
1466
1467 dst->min_val = min(dst->min_val, src->min_val);
1468 dst->max_val = max(dst->max_val, src->max_val);
1469
1470 /*
1471 * Compute new mean and S after the merge
1472 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1473 * #Parallel_algorithm>
1474 */
1475 if (first) {
1476 mean = src->mean.u.f;
1477 S = src->S.u.f;
1478 } else {
1479 double delta = src->mean.u.f - dst->mean.u.f;
1480
1481 mean = ((src->mean.u.f * src->samples) +
1482 (dst->mean.u.f * dst->samples)) /
1483 (dst->samples + src->samples);
1484
1485 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1486 (dst->samples * src->samples) /
1487 (dst->samples + src->samples);
1488 }
1489
1490 dst->samples += src->samples;
1491 dst->mean.u.f = mean;
1492 dst->S.u.f = S;
1493}
1494
1495void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1496{
1497 int i;
1498
1499 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1500 if (dst->max_run[i] < src->max_run[i])
1501 dst->max_run[i] = src->max_run[i];
1502 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1503 dst->min_run[i] = src->min_run[i];
1504 if (dst->max_bw[i] < src->max_bw[i])
1505 dst->max_bw[i] = src->max_bw[i];
1506 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1507 dst->min_bw[i] = src->min_bw[i];
1508
1509 dst->iobytes[i] += src->iobytes[i];
1510 dst->agg[i] += src->agg[i];
1511 }
1512
1513 if (!dst->kb_base)
1514 dst->kb_base = src->kb_base;
1515 if (!dst->unit_base)
1516 dst->unit_base = src->unit_base;
1517}
1518
1519void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1520 bool first)
1521{
1522 int l, k;
1523
1524 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1525 if (!dst->unified_rw_rep) {
1526 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1527 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1528 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1529 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1530 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first);
1531
1532 dst->io_bytes[l] += src->io_bytes[l];
1533
1534 if (dst->runtime[l] < src->runtime[l])
1535 dst->runtime[l] = src->runtime[l];
1536 } else {
1537 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1538 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1539 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1540 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1541 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first);
1542
1543 dst->io_bytes[0] += src->io_bytes[l];
1544
1545 if (dst->runtime[0] < src->runtime[l])
1546 dst->runtime[0] = src->runtime[l];
1547
1548 /*
1549 * We're summing to the same destination, so override
1550 * 'first' after the first iteration of the loop
1551 */
1552 first = false;
1553 }
1554 }
1555
1556 dst->usr_time += src->usr_time;
1557 dst->sys_time += src->sys_time;
1558 dst->ctx += src->ctx;
1559 dst->majf += src->majf;
1560 dst->minf += src->minf;
1561
1562 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1563 dst->io_u_map[k] += src->io_u_map[k];
1564 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1565 dst->io_u_submit[k] += src->io_u_submit[k];
1566 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1567 dst->io_u_complete[k] += src->io_u_complete[k];
1568 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1569 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1570 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1571 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1572 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1573 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1574
1575 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1576 if (!dst->unified_rw_rep) {
1577 dst->total_io_u[k] += src->total_io_u[k];
1578 dst->short_io_u[k] += src->short_io_u[k];
1579 dst->drop_io_u[k] += src->drop_io_u[k];
1580 } else {
1581 dst->total_io_u[0] += src->total_io_u[k];
1582 dst->short_io_u[0] += src->short_io_u[k];
1583 dst->drop_io_u[0] += src->drop_io_u[k];
1584 }
1585 }
1586
1587 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1588 int m;
1589
1590 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1591 if (!dst->unified_rw_rep)
1592 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1593 else
1594 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1595 }
1596 }
1597
1598 dst->total_run_time += src->total_run_time;
1599 dst->total_submit += src->total_submit;
1600 dst->total_complete += src->total_complete;
1601}
1602
1603void init_group_run_stat(struct group_run_stats *gs)
1604{
1605 int i;
1606 memset(gs, 0, sizeof(*gs));
1607
1608 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1609 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1610}
1611
1612void init_thread_stat(struct thread_stat *ts)
1613{
1614 int j;
1615
1616 memset(ts, 0, sizeof(*ts));
1617
1618 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1619 ts->lat_stat[j].min_val = -1UL;
1620 ts->clat_stat[j].min_val = -1UL;
1621 ts->slat_stat[j].min_val = -1UL;
1622 ts->bw_stat[j].min_val = -1UL;
1623 ts->iops_stat[j].min_val = -1UL;
1624 }
1625 ts->groupid = -1;
1626}
1627
1628void __show_run_stats(void)
1629{
1630 struct group_run_stats *runstats, *rs;
1631 struct thread_data *td;
1632 struct thread_stat *threadstats, *ts;
1633 int i, j, k, nr_ts, last_ts, idx;
1634 int kb_base_warned = 0;
1635 int unit_base_warned = 0;
1636 struct json_object *root = NULL;
1637 struct json_array *array = NULL;
1638 struct buf_output output[FIO_OUTPUT_NR];
1639 struct flist_head **opt_lists;
1640
1641 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1642
1643 for (i = 0; i < groupid + 1; i++)
1644 init_group_run_stat(&runstats[i]);
1645
1646 /*
1647 * find out how many threads stats we need. if group reporting isn't
1648 * enabled, it's one-per-td.
1649 */
1650 nr_ts = 0;
1651 last_ts = -1;
1652 for_each_td(td, i) {
1653 if (!td->o.group_reporting) {
1654 nr_ts++;
1655 continue;
1656 }
1657 if (last_ts == td->groupid)
1658 continue;
1659 if (!td->o.stats)
1660 continue;
1661
1662 last_ts = td->groupid;
1663 nr_ts++;
1664 }
1665
1666 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1667 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1668
1669 for (i = 0; i < nr_ts; i++) {
1670 init_thread_stat(&threadstats[i]);
1671 opt_lists[i] = NULL;
1672 }
1673
1674 j = 0;
1675 last_ts = -1;
1676 idx = 0;
1677 for_each_td(td, i) {
1678 if (!td->o.stats)
1679 continue;
1680 if (idx && (!td->o.group_reporting ||
1681 (td->o.group_reporting && last_ts != td->groupid))) {
1682 idx = 0;
1683 j++;
1684 }
1685
1686 last_ts = td->groupid;
1687
1688 ts = &threadstats[j];
1689
1690 ts->clat_percentiles = td->o.clat_percentiles;
1691 ts->percentile_precision = td->o.percentile_precision;
1692 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1693 opt_lists[j] = &td->opt_list;
1694
1695 idx++;
1696 ts->members++;
1697
1698 if (ts->groupid == -1) {
1699 /*
1700 * These are per-group shared already
1701 */
1702 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1703 if (td->o.description)
1704 strncpy(ts->description, td->o.description,
1705 FIO_JOBDESC_SIZE - 1);
1706 else
1707 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1708
1709 /*
1710 * If multiple entries in this group, this is
1711 * the first member.
1712 */
1713 ts->thread_number = td->thread_number;
1714 ts->groupid = td->groupid;
1715
1716 /*
1717 * first pid in group, not very useful...
1718 */
1719 ts->pid = td->pid;
1720
1721 ts->kb_base = td->o.kb_base;
1722 ts->unit_base = td->o.unit_base;
1723 ts->unified_rw_rep = td->o.unified_rw_rep;
1724 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1725 log_info("fio: kb_base differs for jobs in group, using"
1726 " %u as the base\n", ts->kb_base);
1727 kb_base_warned = 1;
1728 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1729 log_info("fio: unit_base differs for jobs in group, using"
1730 " %u as the base\n", ts->unit_base);
1731 unit_base_warned = 1;
1732 }
1733
1734 ts->continue_on_error = td->o.continue_on_error;
1735 ts->total_err_count += td->total_err_count;
1736 ts->first_error = td->first_error;
1737 if (!ts->error) {
1738 if (!td->error && td->o.continue_on_error &&
1739 td->first_error) {
1740 ts->error = td->first_error;
1741 ts->verror[sizeof(ts->verror) - 1] = '\0';
1742 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1743 } else if (td->error) {
1744 ts->error = td->error;
1745 ts->verror[sizeof(ts->verror) - 1] = '\0';
1746 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1747 }
1748 }
1749
1750 ts->latency_depth = td->latency_qd;
1751 ts->latency_target = td->o.latency_target;
1752 ts->latency_percentile = td->o.latency_percentile;
1753 ts->latency_window = td->o.latency_window;
1754
1755 ts->nr_block_infos = td->ts.nr_block_infos;
1756 for (k = 0; k < ts->nr_block_infos; k++)
1757 ts->block_infos[k] = td->ts.block_infos[k];
1758
1759 sum_thread_stats(ts, &td->ts, idx == 1);
1760
1761 if (td->o.ss_dur) {
1762 ts->ss_state = td->ss.state;
1763 ts->ss_dur = td->ss.dur;
1764 ts->ss_head = td->ss.head;
1765 ts->ss_bw_data = td->ss.bw_data;
1766 ts->ss_iops_data = td->ss.iops_data;
1767 ts->ss_limit.u.f = td->ss.limit;
1768 ts->ss_slope.u.f = td->ss.slope;
1769 ts->ss_deviation.u.f = td->ss.deviation;
1770 ts->ss_criterion.u.f = td->ss.criterion;
1771 }
1772 else
1773 ts->ss_dur = ts->ss_state = 0;
1774 }
1775
1776 for (i = 0; i < nr_ts; i++) {
1777 unsigned long long bw;
1778
1779 ts = &threadstats[i];
1780 if (ts->groupid == -1)
1781 continue;
1782 rs = &runstats[ts->groupid];
1783 rs->kb_base = ts->kb_base;
1784 rs->unit_base = ts->unit_base;
1785 rs->unified_rw_rep += ts->unified_rw_rep;
1786
1787 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1788 if (!ts->runtime[j])
1789 continue;
1790 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1791 rs->min_run[j] = ts->runtime[j];
1792 if (ts->runtime[j] > rs->max_run[j])
1793 rs->max_run[j] = ts->runtime[j];
1794
1795 bw = 0;
1796 if (ts->runtime[j])
1797 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1798 if (bw < rs->min_bw[j])
1799 rs->min_bw[j] = bw;
1800 if (bw > rs->max_bw[j])
1801 rs->max_bw[j] = bw;
1802
1803 rs->iobytes[j] += ts->io_bytes[j];
1804 }
1805 }
1806
1807 for (i = 0; i < groupid + 1; i++) {
1808 int ddir;
1809
1810 rs = &runstats[i];
1811
1812 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1813 if (rs->max_run[ddir])
1814 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1815 rs->max_run[ddir];
1816 }
1817 }
1818
1819 for (i = 0; i < FIO_OUTPUT_NR; i++)
1820 buf_output_init(&output[i]);
1821
1822 /*
1823 * don't overwrite last signal output
1824 */
1825 if (output_format & FIO_OUTPUT_NORMAL)
1826 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1827 if (output_format & FIO_OUTPUT_JSON) {
1828 struct thread_data *global;
1829 char time_buf[32];
1830 struct timeval now;
1831 unsigned long long ms_since_epoch;
1832
1833 gettimeofday(&now, NULL);
1834 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1835 (unsigned long long)(now.tv_usec) / 1000;
1836
1837 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1838 sizeof(time_buf));
1839 if (time_buf[strlen(time_buf) - 1] == '\n')
1840 time_buf[strlen(time_buf) - 1] = '\0';
1841
1842 root = json_create_object();
1843 json_object_add_value_string(root, "fio version", fio_version_string);
1844 json_object_add_value_int(root, "timestamp", now.tv_sec);
1845 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1846 json_object_add_value_string(root, "time", time_buf);
1847 global = get_global_options();
1848 json_add_job_opts(root, "global options", &global->opt_list, false);
1849 array = json_create_array();
1850 json_object_add_value_array(root, "jobs", array);
1851 }
1852
1853 if (is_backend)
1854 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1855
1856 for (i = 0; i < nr_ts; i++) {
1857 ts = &threadstats[i];
1858 rs = &runstats[ts->groupid];
1859
1860 if (is_backend) {
1861 fio_server_send_job_options(opt_lists[i], i);
1862 fio_server_send_ts(ts, rs);
1863 } else {
1864 if (output_format & FIO_OUTPUT_TERSE)
1865 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1866 if (output_format & FIO_OUTPUT_JSON) {
1867 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1868 json_array_add_value_object(array, tmp);
1869 }
1870 if (output_format & FIO_OUTPUT_NORMAL)
1871 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1872 }
1873 }
1874 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1875 /* disk util stats, if any */
1876 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1877
1878 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1879
1880 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1881 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1882 json_free_object(root);
1883 }
1884
1885 for (i = 0; i < groupid + 1; i++) {
1886 rs = &runstats[i];
1887
1888 rs->groupid = i;
1889 if (is_backend)
1890 fio_server_send_gs(rs);
1891 else if (output_format & FIO_OUTPUT_NORMAL)
1892 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1893 }
1894
1895 if (is_backend)
1896 fio_server_send_du();
1897 else if (output_format & FIO_OUTPUT_NORMAL) {
1898 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1899 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1900 }
1901
1902 for (i = 0; i < FIO_OUTPUT_NR; i++) {
1903 struct buf_output *out = &output[i];
1904
1905 log_info_buf(out->buf, out->buflen);
1906 buf_output_free(out);
1907 }
1908
1909 log_info_flush();
1910 free(runstats);
1911 free(threadstats);
1912 free(opt_lists);
1913}
1914
1915void show_run_stats(void)
1916{
1917 fio_mutex_down(stat_mutex);
1918 __show_run_stats();
1919 fio_mutex_up(stat_mutex);
1920}
1921
1922void __show_running_run_stats(void)
1923{
1924 struct thread_data *td;
1925 unsigned long long *rt;
1926 struct timespec ts;
1927 int i;
1928
1929 fio_mutex_down(stat_mutex);
1930
1931 rt = malloc(thread_number * sizeof(unsigned long long));
1932 fio_gettime(&ts, NULL);
1933
1934 for_each_td(td, i) {
1935 td->update_rusage = 1;
1936 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1937 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1938 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1939 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1940
1941 rt[i] = mtime_since(&td->start, &ts);
1942 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1943 td->ts.runtime[DDIR_READ] += rt[i];
1944 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1945 td->ts.runtime[DDIR_WRITE] += rt[i];
1946 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1947 td->ts.runtime[DDIR_TRIM] += rt[i];
1948 }
1949
1950 for_each_td(td, i) {
1951 if (td->runstate >= TD_EXITED)
1952 continue;
1953 if (td->rusage_sem) {
1954 td->update_rusage = 1;
1955 fio_mutex_down(td->rusage_sem);
1956 }
1957 td->update_rusage = 0;
1958 }
1959
1960 __show_run_stats();
1961
1962 for_each_td(td, i) {
1963 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1964 td->ts.runtime[DDIR_READ] -= rt[i];
1965 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1966 td->ts.runtime[DDIR_WRITE] -= rt[i];
1967 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1968 td->ts.runtime[DDIR_TRIM] -= rt[i];
1969 }
1970
1971 free(rt);
1972 fio_mutex_up(stat_mutex);
1973}
1974
1975static int status_interval_init;
1976static struct timespec status_time;
1977static int status_file_disabled;
1978
1979#define FIO_STATUS_FILE "fio-dump-status"
1980
1981static int check_status_file(void)
1982{
1983 struct stat sb;
1984 const char *temp_dir;
1985 char fio_status_file_path[PATH_MAX];
1986
1987 if (status_file_disabled)
1988 return 0;
1989
1990 temp_dir = getenv("TMPDIR");
1991 if (temp_dir == NULL) {
1992 temp_dir = getenv("TEMP");
1993 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1994 temp_dir = NULL;
1995 }
1996 if (temp_dir == NULL)
1997 temp_dir = "/tmp";
1998
1999 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2000
2001 if (stat(fio_status_file_path, &sb))
2002 return 0;
2003
2004 if (unlink(fio_status_file_path) < 0) {
2005 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2006 strerror(errno));
2007 log_err("fio: disabling status file updates\n");
2008 status_file_disabled = 1;
2009 }
2010
2011 return 1;
2012}
2013
2014void check_for_running_stats(void)
2015{
2016 if (status_interval) {
2017 if (!status_interval_init) {
2018 fio_gettime(&status_time, NULL);
2019 status_interval_init = 1;
2020 } else if (mtime_since_now(&status_time) >= status_interval) {
2021 show_running_run_stats();
2022 fio_gettime(&status_time, NULL);
2023 return;
2024 }
2025 }
2026 if (check_status_file()) {
2027 show_running_run_stats();
2028 return;
2029 }
2030}
2031
2032static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2033{
2034 double val = data;
2035 double delta;
2036
2037 if (data > is->max_val)
2038 is->max_val = data;
2039 if (data < is->min_val)
2040 is->min_val = data;
2041
2042 delta = val - is->mean.u.f;
2043 if (delta) {
2044 is->mean.u.f += delta / (is->samples + 1.0);
2045 is->S.u.f += delta * (val - is->mean.u.f);
2046 }
2047
2048 is->samples++;
2049}
2050
2051/*
2052 * Return a struct io_logs, which is added to the tail of the log
2053 * list for 'iolog'.
2054 */
2055static struct io_logs *get_new_log(struct io_log *iolog)
2056{
2057 size_t new_size, new_samples;
2058 struct io_logs *cur_log;
2059
2060 /*
2061 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2062 * forever
2063 */
2064 if (!iolog->cur_log_max)
2065 new_samples = DEF_LOG_ENTRIES;
2066 else {
2067 new_samples = iolog->cur_log_max * 2;
2068 if (new_samples > MAX_LOG_ENTRIES)
2069 new_samples = MAX_LOG_ENTRIES;
2070 }
2071
2072 new_size = new_samples * log_entry_sz(iolog);
2073
2074 cur_log = smalloc(sizeof(*cur_log));
2075 if (cur_log) {
2076 INIT_FLIST_HEAD(&cur_log->list);
2077 cur_log->log = malloc(new_size);
2078 if (cur_log->log) {
2079 cur_log->nr_samples = 0;
2080 cur_log->max_samples = new_samples;
2081 flist_add_tail(&cur_log->list, &iolog->io_logs);
2082 iolog->cur_log_max = new_samples;
2083 return cur_log;
2084 }
2085 sfree(cur_log);
2086 }
2087
2088 return NULL;
2089}
2090
2091/*
2092 * Add and return a new log chunk, or return current log if big enough
2093 */
2094static struct io_logs *regrow_log(struct io_log *iolog)
2095{
2096 struct io_logs *cur_log;
2097 int i;
2098
2099 if (!iolog || iolog->disabled)
2100 goto disable;
2101
2102 cur_log = iolog_cur_log(iolog);
2103 if (!cur_log) {
2104 cur_log = get_new_log(iolog);
2105 if (!cur_log)
2106 return NULL;
2107 }
2108
2109 if (cur_log->nr_samples < cur_log->max_samples)
2110 return cur_log;
2111
2112 /*
2113 * No room for a new sample. If we're compressing on the fly, flush
2114 * out the current chunk
2115 */
2116 if (iolog->log_gz) {
2117 if (iolog_cur_flush(iolog, cur_log)) {
2118 log_err("fio: failed flushing iolog! Will stop logging.\n");
2119 return NULL;
2120 }
2121 }
2122
2123 /*
2124 * Get a new log array, and add to our list
2125 */
2126 cur_log = get_new_log(iolog);
2127 if (!cur_log) {
2128 log_err("fio: failed extending iolog! Will stop logging.\n");
2129 return NULL;
2130 }
2131
2132 if (!iolog->pending || !iolog->pending->nr_samples)
2133 return cur_log;
2134
2135 /*
2136 * Flush pending items to new log
2137 */
2138 for (i = 0; i < iolog->pending->nr_samples; i++) {
2139 struct io_sample *src, *dst;
2140
2141 src = get_sample(iolog, iolog->pending, i);
2142 dst = get_sample(iolog, cur_log, i);
2143 memcpy(dst, src, log_entry_sz(iolog));
2144 }
2145 cur_log->nr_samples = iolog->pending->nr_samples;
2146
2147 iolog->pending->nr_samples = 0;
2148 return cur_log;
2149disable:
2150 if (iolog)
2151 iolog->disabled = true;
2152 return NULL;
2153}
2154
2155void regrow_logs(struct thread_data *td)
2156{
2157 regrow_log(td->slat_log);
2158 regrow_log(td->clat_log);
2159 regrow_log(td->clat_hist_log);
2160 regrow_log(td->lat_log);
2161 regrow_log(td->bw_log);
2162 regrow_log(td->iops_log);
2163 td->flags &= ~TD_F_REGROW_LOGS;
2164}
2165
2166static struct io_logs *get_cur_log(struct io_log *iolog)
2167{
2168 struct io_logs *cur_log;
2169
2170 cur_log = iolog_cur_log(iolog);
2171 if (!cur_log) {
2172 cur_log = get_new_log(iolog);
2173 if (!cur_log)
2174 return NULL;
2175 }
2176
2177 if (cur_log->nr_samples < cur_log->max_samples)
2178 return cur_log;
2179
2180 /*
2181 * Out of space. If we're in IO offload mode, or we're not doing
2182 * per unit logging (hence logging happens outside of the IO thread
2183 * as well), add a new log chunk inline. If we're doing inline
2184 * submissions, flag 'td' as needing a log regrow and we'll take
2185 * care of it on the submission side.
2186 */
2187 if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2188 !per_unit_log(iolog))
2189 return regrow_log(iolog);
2190
2191 iolog->td->flags |= TD_F_REGROW_LOGS;
2192 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2193 return iolog->pending;
2194}
2195
2196static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2197 enum fio_ddir ddir, unsigned int bs,
2198 unsigned long t, uint64_t offset)
2199{
2200 struct io_logs *cur_log;
2201
2202 if (iolog->disabled)
2203 return;
2204 if (flist_empty(&iolog->io_logs))
2205 iolog->avg_last = t;
2206
2207 cur_log = get_cur_log(iolog);
2208 if (cur_log) {
2209 struct io_sample *s;
2210
2211 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2212
2213 s->data = data;
2214 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2215 io_sample_set_ddir(iolog, s, ddir);
2216 s->bs = bs;
2217
2218 if (iolog->log_offset) {
2219 struct io_sample_offset *so = (void *) s;
2220
2221 so->offset = offset;
2222 }
2223
2224 cur_log->nr_samples++;
2225 return;
2226 }
2227
2228 iolog->disabled = true;
2229}
2230
2231static inline void reset_io_stat(struct io_stat *ios)
2232{
2233 ios->max_val = ios->min_val = ios->samples = 0;
2234 ios->mean.u.f = ios->S.u.f = 0;
2235}
2236
2237void reset_io_stats(struct thread_data *td)
2238{
2239 struct thread_stat *ts = &td->ts;
2240 int i, j;
2241
2242 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2243 reset_io_stat(&ts->clat_stat[i]);
2244 reset_io_stat(&ts->slat_stat[i]);
2245 reset_io_stat(&ts->lat_stat[i]);
2246 reset_io_stat(&ts->bw_stat[i]);
2247 reset_io_stat(&ts->iops_stat[i]);
2248
2249 ts->io_bytes[i] = 0;
2250 ts->runtime[i] = 0;
2251 ts->total_io_u[i] = 0;
2252 ts->short_io_u[i] = 0;
2253 ts->drop_io_u[i] = 0;
2254
2255 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2256 ts->io_u_plat[i][j] = 0;
2257 }
2258
2259 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2260 ts->io_u_map[i] = 0;
2261 ts->io_u_submit[i] = 0;
2262 ts->io_u_complete[i] = 0;
2263 }
2264
2265 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2266 ts->io_u_lat_n[i] = 0;
2267 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2268 ts->io_u_lat_u[i] = 0;
2269 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2270 ts->io_u_lat_m[i] = 0;
2271
2272 ts->total_submit = 0;
2273 ts->total_complete = 0;
2274}
2275
2276static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2277 unsigned long elapsed, bool log_max)
2278{
2279 /*
2280 * Note an entry in the log. Use the mean from the logged samples,
2281 * making sure to properly round up. Only write a log entry if we
2282 * had actual samples done.
2283 */
2284 if (iolog->avg_window[ddir].samples) {
2285 union io_sample_data data;
2286
2287 if (log_max)
2288 data.val = iolog->avg_window[ddir].max_val;
2289 else
2290 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2291
2292 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2293 }
2294
2295 reset_io_stat(&iolog->avg_window[ddir]);
2296}
2297
2298static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2299 bool log_max)
2300{
2301 int ddir;
2302
2303 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2304 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2305}
2306
2307static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2308 union io_sample_data data, enum fio_ddir ddir,
2309 unsigned int bs, uint64_t offset)
2310{
2311 unsigned long elapsed, this_window;
2312
2313 if (!ddir_rw(ddir))
2314 return 0;
2315
2316 elapsed = mtime_since_now(&td->epoch);
2317
2318 /*
2319 * If no time averaging, just add the log sample.
2320 */
2321 if (!iolog->avg_msec) {
2322 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2323 return 0;
2324 }
2325
2326 /*
2327 * Add the sample. If the time period has passed, then
2328 * add that entry to the log and clear.
2329 */
2330 add_stat_sample(&iolog->avg_window[ddir], data.val);
2331
2332 /*
2333 * If period hasn't passed, adding the above sample is all we
2334 * need to do.
2335 */
2336 this_window = elapsed - iolog->avg_last;
2337 if (elapsed < iolog->avg_last)
2338 return iolog->avg_last - elapsed;
2339 else if (this_window < iolog->avg_msec) {
2340 int diff = iolog->avg_msec - this_window;
2341
2342 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2343 return diff;
2344 }
2345
2346 _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2347
2348 iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2349 return iolog->avg_msec;
2350}
2351
2352void finalize_logs(struct thread_data *td, bool unit_logs)
2353{
2354 unsigned long elapsed;
2355
2356 elapsed = mtime_since_now(&td->epoch);
2357
2358 if (td->clat_log && unit_logs)
2359 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2360 if (td->slat_log && unit_logs)
2361 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2362 if (td->lat_log && unit_logs)
2363 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2364 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2365 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2366 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2367 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2368}
2369
2370void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2371{
2372 struct io_log *iolog;
2373
2374 if (!ddir_rw(ddir))
2375 return;
2376
2377 iolog = agg_io_log[ddir];
2378 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2379}
2380
2381static void add_clat_percentile_sample(struct thread_stat *ts,
2382 unsigned long long nsec, enum fio_ddir ddir)
2383{
2384 unsigned int idx = plat_val_to_idx(nsec);
2385 assert(idx < FIO_IO_U_PLAT_NR);
2386
2387 ts->io_u_plat[ddir][idx]++;
2388}
2389
2390void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2391 unsigned long long nsec, unsigned int bs, uint64_t offset)
2392{
2393 unsigned long elapsed, this_window;
2394 struct thread_stat *ts = &td->ts;
2395 struct io_log *iolog = td->clat_hist_log;
2396
2397 td_io_u_lock(td);
2398
2399 add_stat_sample(&ts->clat_stat[ddir], nsec);
2400
2401 if (td->clat_log)
2402 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2403 offset);
2404
2405 if (ts->clat_percentiles)
2406 add_clat_percentile_sample(ts, nsec, ddir);
2407
2408 if (iolog && iolog->hist_msec) {
2409 struct io_hist *hw = &iolog->hist_window[ddir];
2410
2411 hw->samples++;
2412 elapsed = mtime_since_now(&td->epoch);
2413 if (!hw->hist_last)
2414 hw->hist_last = elapsed;
2415 this_window = elapsed - hw->hist_last;
2416
2417 if (this_window >= iolog->hist_msec) {
2418 unsigned int *io_u_plat;
2419 struct io_u_plat_entry *dst;
2420
2421 /*
2422 * Make a byte-for-byte copy of the latency histogram
2423 * stored in td->ts.io_u_plat[ddir], recording it in a
2424 * log sample. Note that the matching call to free() is
2425 * located in iolog.c after printing this sample to the
2426 * log file.
2427 */
2428 io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2429 dst = malloc(sizeof(struct io_u_plat_entry));
2430 memcpy(&(dst->io_u_plat), io_u_plat,
2431 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2432 flist_add(&dst->list, &hw->list);
2433 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2434 elapsed, offset);
2435
2436 /*
2437 * Update the last time we recorded as being now, minus
2438 * any drift in time we encountered before actually
2439 * making the record.
2440 */
2441 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2442 hw->samples = 0;
2443 }
2444 }
2445
2446 td_io_u_unlock(td);
2447}
2448
2449void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2450 unsigned long usec, unsigned int bs, uint64_t offset)
2451{
2452 struct thread_stat *ts = &td->ts;
2453
2454 if (!ddir_rw(ddir))
2455 return;
2456
2457 td_io_u_lock(td);
2458
2459 add_stat_sample(&ts->slat_stat[ddir], usec);
2460
2461 if (td->slat_log)
2462 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2463
2464 td_io_u_unlock(td);
2465}
2466
2467void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2468 unsigned long long nsec, unsigned int bs, uint64_t offset)
2469{
2470 struct thread_stat *ts = &td->ts;
2471
2472 if (!ddir_rw(ddir))
2473 return;
2474
2475 td_io_u_lock(td);
2476
2477 add_stat_sample(&ts->lat_stat[ddir], nsec);
2478
2479 if (td->lat_log)
2480 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2481 offset);
2482
2483 td_io_u_unlock(td);
2484}
2485
2486void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2487 unsigned int bytes, unsigned long long spent)
2488{
2489 struct thread_stat *ts = &td->ts;
2490 unsigned long rate;
2491
2492 if (spent)
2493 rate = (unsigned long) (bytes * 1000000ULL / spent);
2494 else
2495 rate = 0;
2496
2497 td_io_u_lock(td);
2498
2499 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2500
2501 if (td->bw_log)
2502 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2503 bytes, io_u->offset);
2504
2505 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2506 td_io_u_unlock(td);
2507}
2508
2509static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2510 struct timespec *t, unsigned int avg_time,
2511 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2512 struct io_stat *stat, struct io_log *log,
2513 bool is_kb)
2514{
2515 unsigned long spent, rate;
2516 enum fio_ddir ddir;
2517 unsigned int next, next_log;
2518
2519 next_log = avg_time;
2520
2521 spent = mtime_since(parent_tv, t);
2522 if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2523 return avg_time - spent;
2524
2525 td_io_u_lock(td);
2526
2527 /*
2528 * Compute both read and write rates for the interval.
2529 */
2530 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2531 uint64_t delta;
2532
2533 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2534 if (!delta)
2535 continue; /* No entries for interval */
2536
2537 if (spent) {
2538 if (is_kb)
2539 rate = delta * 1000 / spent / 1024; /* KiB/s */
2540 else
2541 rate = (delta * 1000) / spent;
2542 } else
2543 rate = 0;
2544
2545 add_stat_sample(&stat[ddir], rate);
2546
2547 if (log) {
2548 unsigned int bs = 0;
2549
2550 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2551 bs = td->o.min_bs[ddir];
2552
2553 next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2554 next_log = min(next_log, next);
2555 }
2556
2557 stat_io_bytes[ddir] = this_io_bytes[ddir];
2558 }
2559
2560 timespec_add_msec(parent_tv, avg_time);
2561
2562 td_io_u_unlock(td);
2563
2564 if (spent <= avg_time)
2565 next = avg_time;
2566 else
2567 next = avg_time - (1 + spent - avg_time);
2568
2569 return min(next, next_log);
2570}
2571
2572static int add_bw_samples(struct thread_data *td, struct timespec *t)
2573{
2574 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2575 td->this_io_bytes, td->stat_io_bytes,
2576 td->ts.bw_stat, td->bw_log, true);
2577}
2578
2579void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2580 unsigned int bytes)
2581{
2582 struct thread_stat *ts = &td->ts;
2583
2584 td_io_u_lock(td);
2585
2586 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2587
2588 if (td->iops_log)
2589 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2590 bytes, io_u->offset);
2591
2592 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2593 td_io_u_unlock(td);
2594}
2595
2596static int add_iops_samples(struct thread_data *td, struct timespec *t)
2597{
2598 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2599 td->this_io_blocks, td->stat_io_blocks,
2600 td->ts.iops_stat, td->iops_log, false);
2601}
2602
2603/*
2604 * Returns msecs to next event
2605 */
2606int calc_log_samples(void)
2607{
2608 struct thread_data *td;
2609 unsigned int next = ~0U, tmp;
2610 struct timespec now;
2611 int i;
2612
2613 fio_gettime(&now, NULL);
2614
2615 for_each_td(td, i) {
2616 if (!td->o.stats)
2617 continue;
2618 if (in_ramp_time(td) ||
2619 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2620 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2621 continue;
2622 }
2623 if (!td->bw_log ||
2624 (td->bw_log && !per_unit_log(td->bw_log))) {
2625 tmp = add_bw_samples(td, &now);
2626 if (tmp < next)
2627 next = tmp;
2628 }
2629 if (!td->iops_log ||
2630 (td->iops_log && !per_unit_log(td->iops_log))) {
2631 tmp = add_iops_samples(td, &now);
2632 if (tmp < next)
2633 next = tmp;
2634 }
2635 }
2636
2637 return next == ~0U ? 0 : next;
2638}
2639
2640void stat_init(void)
2641{
2642 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2643}
2644
2645void stat_exit(void)
2646{
2647 /*
2648 * When we have the mutex, we know out-of-band access to it
2649 * have ended.
2650 */
2651 fio_mutex_down(stat_mutex);
2652 fio_mutex_remove(stat_mutex);
2653}
2654
2655/*
2656 * Called from signal handler. Wake up status thread.
2657 */
2658void show_running_run_stats(void)
2659{
2660 helper_do_stat();
2661}
2662
2663uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2664{
2665 /* Ignore io_u's which span multiple blocks--they will just get
2666 * inaccurate counts. */
2667 int idx = (io_u->offset - io_u->file->file_offset)
2668 / td->o.bs[DDIR_TRIM];
2669 uint32_t *info = &td->ts.block_infos[idx];
2670 assert(idx < td->ts.nr_block_infos);
2671 return info;
2672}
fio - Flexible IO Tester