projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
fio: report percentiles for slat, clat, lat
[fio.git] / stat.c
... / ...
CommitLineData
1#include <stdio.h>
2#include <string.h>
3#include <sys/time.h>
4#include <sys/stat.h>
5#include <math.h>
6
7#include "fio.h"
8#include "diskutil.h"
9#include "lib/ieee754.h"
10#include "json.h"
11#include "lib/getrusage.h"
12#include "idletime.h"
13#include "lib/pow2.h"
14#include "lib/output_buffer.h"
15#include "helper_thread.h"
16#include "smalloc.h"
17#include "zbd.h"
18#include "oslib/asprintf.h"
19
20#define LOG_MSEC_SLACK 1
21
22struct fio_sem *stat_sem;
23
24void clear_rusage_stat(struct thread_data *td)
25{
26 struct thread_stat *ts = &td->ts;
27
28 fio_getrusage(&td->ru_start);
29 ts->usr_time = ts->sys_time = 0;
30 ts->ctx = 0;
31 ts->minf = ts->majf = 0;
32}
33
34void update_rusage_stat(struct thread_data *td)
35{
36 struct thread_stat *ts = &td->ts;
37
38 fio_getrusage(&td->ru_end);
39 ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
40 &td->ru_end.ru_utime);
41 ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
42 &td->ru_end.ru_stime);
43 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
44 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
45 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
46 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
47
48 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
49}
50
51/*
52 * Given a latency, return the index of the corresponding bucket in
53 * the structure tracking percentiles.
54 *
55 * (1) find the group (and error bits) that the value (latency)
56 * belongs to by looking at its MSB. (2) find the bucket number in the
57 * group by looking at the index bits.
58 *
59 */
60static unsigned int plat_val_to_idx(unsigned long long val)
61{
62 unsigned int msb, error_bits, base, offset, idx;
63
64 /* Find MSB starting from bit 0 */
65 if (val == 0)
66 msb = 0;
67 else
68 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
69
70 /*
71 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
72 * all bits of the sample as index
73 */
74 if (msb <= FIO_IO_U_PLAT_BITS)
75 return val;
76
77 /* Compute the number of error bits to discard*/
78 error_bits = msb - FIO_IO_U_PLAT_BITS;
79
80 /* Compute the number of buckets before the group */
81 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
82
83 /*
84 * Discard the error bits and apply the mask to find the
85 * index for the buckets in the group
86 */
87 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
88
89 /* Make sure the index does not exceed (array size - 1) */
90 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
91 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
92
93 return idx;
94}
95
96/*
97 * Convert the given index of the bucket array to the value
98 * represented by the bucket
99 */
100static unsigned long long plat_idx_to_val(unsigned int idx)
101{
102 unsigned int error_bits;
103 unsigned long long 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 = ((unsigned long long) 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(uint64_t *io_u_plat, unsigned long long nr,
138 fio_fp64_t *plist, unsigned long long **output,
139 unsigned long long *maxv, unsigned long long *minv)
140{
141 unsigned long long sum = 0;
142 unsigned int len, i, j = 0;
143 unsigned long long *ovals = NULL;
144 bool is_last;
145
146 *minv = -1ULL;
147 *maxv = 0;
148
149 len = 0;
150 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
151 len++;
152
153 if (!len)
154 return 0;
155
156 /*
157 * Sort the percentile list. Note that it may already be sorted if
158 * we are using the default values, but since it's a short list this
159 * isn't a worry. Also note that this does not work for NaN values.
160 */
161 if (len > 1)
162 qsort(plist, len, sizeof(plist[0]), double_cmp);
163
164 ovals = malloc(len * sizeof(*ovals));
165 if (!ovals)
166 return 0;
167
168 /*
169 * Calculate bucket values, note down max and min values
170 */
171 is_last = false;
172 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
173 sum += io_u_plat[i];
174 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
175 assert(plist[j].u.f <= 100.0);
176
177 ovals[j] = plat_idx_to_val(i);
178 if (ovals[j] < *minv)
179 *minv = ovals[j];
180 if (ovals[j] > *maxv)
181 *maxv = ovals[j];
182
183 is_last = (j == len - 1) != 0;
184 if (is_last)
185 break;
186
187 j++;
188 }
189 }
190
191 if (!is_last)
192 log_err("fio: error calculating latency percentiles\n");
193
194 *output = ovals;
195 return len;
196}
197
198/*
199 * Find and display the p-th percentile of clat
200 */
201static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
202 fio_fp64_t *plist, unsigned int precision,
203 const char *pre, struct buf_output *out)
204{
205 unsigned int divisor, len, i, j = 0;
206 unsigned long long minv, maxv;
207 unsigned long long *ovals;
208 int per_line, scale_down, time_width;
209 bool is_last;
210 char fmt[32];
211
212 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
213 if (!len || !ovals)
214 goto out;
215
216 /*
217 * We default to nsecs, but if the value range is such that we
218 * should scale down to usecs or msecs, do that.
219 */
220 if (minv > 2000000 && maxv > 99999999ULL) {
221 scale_down = 2;
222 divisor = 1000000;
223 log_buf(out, " %s percentiles (msec):\n |", pre);
224 } else if (minv > 2000 && maxv > 99999) {
225 scale_down = 1;
226 divisor = 1000;
227 log_buf(out, " %s percentiles (usec):\n |", pre);
228 } else {
229 scale_down = 0;
230 divisor = 1;
231 log_buf(out, " %s percentiles (nsec):\n |", pre);
232 }
233
234
235 time_width = max(5, (int) (log10(maxv / divisor) + 1));
236 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
237 precision, time_width);
238 /* fmt will be something like " %5.2fth=[%4llu]%c" */
239 per_line = (80 - 7) / (precision + 10 + time_width);
240
241 for (j = 0; j < len; j++) {
242 /* for formatting */
243 if (j != 0 && (j % per_line) == 0)
244 log_buf(out, " |");
245
246 /* end of the list */
247 is_last = (j == len - 1) != 0;
248
249 for (i = 0; i < scale_down; i++)
250 ovals[j] = (ovals[j] + 999) / 1000;
251
252 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
253
254 if (is_last)
255 break;
256
257 if ((j % per_line) == per_line - 1) /* for formatting */
258 log_buf(out, "\n");
259 }
260
261out:
262 free(ovals);
263}
264
265bool calc_lat(struct io_stat *is, unsigned long long *min,
266 unsigned long long *max, double *mean, double *dev)
267{
268 double n = (double) is->samples;
269
270 if (n == 0)
271 return false;
272
273 *min = is->min_val;
274 *max = is->max_val;
275 *mean = is->mean.u.f;
276
277 if (n > 1.0)
278 *dev = sqrt(is->S.u.f / (n - 1.0));
279 else
280 *dev = 0;
281
282 return true;
283}
284
285void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
286{
287 char *io, *agg, *min, *max;
288 char *ioalt, *aggalt, *minalt, *maxalt;
289 const char *str[] = { " READ", " WRITE" , " TRIM"};
290 int i;
291
292 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
293
294 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
295 const int i2p = is_power_of_2(rs->kb_base);
296
297 if (!rs->max_run[i])
298 continue;
299
300 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
301 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
302 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
303 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
304 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
305 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
306 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
307 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
308 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
309 rs->unified_rw_rep ? " MIXED" : str[i],
310 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
311 (unsigned long long) rs->min_run[i],
312 (unsigned long long) rs->max_run[i]);
313
314 free(io);
315 free(agg);
316 free(min);
317 free(max);
318 free(ioalt);
319 free(aggalt);
320 free(minalt);
321 free(maxalt);
322 }
323}
324
325void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
326{
327 int i;
328
329 /*
330 * Do depth distribution calculations
331 */
332 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
333 if (total) {
334 io_u_dist[i] = (double) map[i] / (double) total;
335 io_u_dist[i] *= 100.0;
336 if (io_u_dist[i] < 0.1 && map[i])
337 io_u_dist[i] = 0.1;
338 } else
339 io_u_dist[i] = 0.0;
340 }
341}
342
343static void stat_calc_lat(struct thread_stat *ts, double *dst,
344 uint64_t *src, int nr)
345{
346 unsigned long total = ddir_rw_sum(ts->total_io_u);
347 int i;
348
349 /*
350 * Do latency distribution calculations
351 */
352 for (i = 0; i < nr; i++) {
353 if (total) {
354 dst[i] = (double) src[i] / (double) total;
355 dst[i] *= 100.0;
356 if (dst[i] < 0.01 && src[i])
357 dst[i] = 0.01;
358 } else
359 dst[i] = 0.0;
360 }
361}
362
363/*
364 * To keep the terse format unaltered, add all of the ns latency
365 * buckets to the first us latency bucket
366 */
367static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
368{
369 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
370 int i;
371
372 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
373
374 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
375 ntotal += ts->io_u_lat_n[i];
376
377 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
378}
379
380void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
381{
382 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
383}
384
385void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
386{
387 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
388}
389
390void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
391{
392 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
393}
394
395static void display_lat(const char *name, unsigned long long min,
396 unsigned long long max, double mean, double dev,
397 struct buf_output *out)
398{
399 const char *base = "(nsec)";
400 char *minp, *maxp;
401
402 if (nsec_to_msec(&min, &max, &mean, &dev))
403 base = "(msec)";
404 else if (nsec_to_usec(&min, &max, &mean, &dev))
405 base = "(usec)";
406
407 minp = num2str(min, 6, 1, 0, N2S_NONE);
408 maxp = num2str(max, 6, 1, 0, N2S_NONE);
409
410 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
411 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
412
413 free(minp);
414 free(maxp);
415}
416
417static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
418 int ddir, struct buf_output *out)
419{
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, *post_st = NULL;
424 int i2p;
425
426 if (ddir_sync(ddir)) {
427 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
428 log_buf(out, " %s:\n", "fsync/fdatasync/sync_file_range");
429 display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
430 show_clat_percentiles(ts->io_u_sync_plat,
431 ts->sync_stat.samples,
432 ts->percentile_list,
433 ts->percentile_precision,
434 io_ddir_name(ddir), out);
435 }
436 return;
437 }
438
439 assert(ddir_rw(ddir));
440
441 if (!ts->runtime[ddir])
442 return;
443
444 i2p = is_power_of_2(rs->kb_base);
445 runt = ts->runtime[ddir];
446
447 bw = (1000 * ts->io_bytes[ddir]) / runt;
448 io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
449 bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
450 bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
451
452 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
453 iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
454 if (ddir == DDIR_WRITE)
455 post_st = zbd_write_status(ts);
456 else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
457 uint64_t total;
458 double hit;
459
460 total = ts->cachehit + ts->cachemiss;
461 hit = (double) ts->cachehit / (double) total;
462 hit *= 100.0;
463 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
464 post_st = NULL;
465 }
466
467 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
468 rs->unified_rw_rep ? "mixed" : io_ddir_name(ddir),
469 iops_p, bw_p, bw_p_alt, io_p,
470 (unsigned long long) ts->runtime[ddir],
471 post_st ? : "");
472
473 free(post_st);
474 free(io_p);
475 free(bw_p);
476 free(bw_p_alt);
477 free(iops_p);
478
479 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
480 display_lat("slat", min, max, mean, dev, out);
481 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
482 display_lat("clat", min, max, mean, dev, out);
483 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
484 display_lat(" lat", min, max, mean, dev, out);
485 if (calc_lat(&ts->clat_high_prio_stat[ddir], &min, &max, &mean, &dev))
486 display_lat("prio_clat", min, max, mean, dev, out);
487
488 if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
489 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
490 ts->slat_stat[ddir].samples,
491 ts->percentile_list,
492 ts->percentile_precision, "slat", out);
493 if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
494 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
495 ts->clat_stat[ddir].samples,
496 ts->percentile_list,
497 ts->percentile_precision, "clat", out);
498 if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
499 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
500 ts->lat_stat[ddir].samples,
501 ts->percentile_list,
502 ts->percentile_precision, "lat", out);
503
504 if (ts->clat_percentiles || ts->lat_percentiles) {
505 const char *name = ts->lat_percentiles ? " lat" : "clat";
506 char prio_name[32];
507 uint64_t samples;
508
509 if (ts->lat_percentiles)
510 samples = ts->lat_stat[ddir].samples;
511 else
512 samples = ts->clat_stat[ddir].samples;
513
514 /* Only print this if some high and low priority stats were collected */
515 if (ts->clat_high_prio_stat[ddir].samples > 0 &&
516 ts->clat_prio_stat[ddir].samples > 0)
517 {
518 sprintf(prio_name, "high prio (%.2f%%) %s",
519 100. * (double) ts->clat_high_prio_stat[ddir].samples / (double) samples,
520 name);
521 show_clat_percentiles(ts->io_u_plat_high_prio[ddir],
522 ts->clat_high_prio_stat[ddir].samples,
523 ts->percentile_list,
524 ts->percentile_precision, prio_name, out);
525
526 sprintf(prio_name, "low prio (%.2f%%) %s",
527 100. * (double) ts->clat_prio_stat[ddir].samples / (double) samples,
528 name);
529 show_clat_percentiles(ts->io_u_plat_prio[ddir],
530 ts->clat_prio_stat[ddir].samples,
531 ts->percentile_list,
532 ts->percentile_precision, prio_name, out);
533 }
534 }
535
536 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
537 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
538 const char *bw_str;
539
540 if ((rs->unit_base == 1) && i2p)
541 bw_str = "Kibit";
542 else if (rs->unit_base == 1)
543 bw_str = "kbit";
544 else if (i2p)
545 bw_str = "KiB";
546 else
547 bw_str = "kB";
548
549 if (rs->agg[ddir]) {
550 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
551 if (p_of_agg > 100.0)
552 p_of_agg = 100.0;
553 }
554
555 if (rs->unit_base == 1) {
556 min *= 8.0;
557 max *= 8.0;
558 mean *= 8.0;
559 dev *= 8.0;
560 }
561
562 if (mean > fkb_base * fkb_base) {
563 min /= fkb_base;
564 max /= fkb_base;
565 mean /= fkb_base;
566 dev /= fkb_base;
567 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
568 }
569
570 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
571 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
572 bw_str, min, max, p_of_agg, mean, dev,
573 (&ts->bw_stat[ddir])->samples);
574 }
575 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
576 log_buf(out, " iops : min=%5llu, max=%5llu, "
577 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
578 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
579 }
580}
581
582static bool show_lat(double *io_u_lat, int nr, const char **ranges,
583 const char *msg, struct buf_output *out)
584{
585 bool new_line = true, shown = false;
586 int i, line = 0;
587
588 for (i = 0; i < nr; i++) {
589 if (io_u_lat[i] <= 0.0)
590 continue;
591 shown = true;
592 if (new_line) {
593 if (line)
594 log_buf(out, "\n");
595 log_buf(out, " lat (%s) : ", msg);
596 new_line = false;
597 line = 0;
598 }
599 if (line)
600 log_buf(out, ", ");
601 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
602 line++;
603 if (line == 5)
604 new_line = true;
605 }
606
607 if (shown)
608 log_buf(out, "\n");
609
610 return true;
611}
612
613static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
614{
615 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
616 "250=", "500=", "750=", "1000=", };
617
618 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
619}
620
621static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
622{
623 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
624 "250=", "500=", "750=", "1000=", };
625
626 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
627}
628
629static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
630{
631 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
632 "250=", "500=", "750=", "1000=", "2000=",
633 ">=2000=", };
634
635 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
636}
637
638static void show_latencies(struct thread_stat *ts, struct buf_output *out)
639{
640 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
641 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
642 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
643
644 stat_calc_lat_n(ts, io_u_lat_n);
645 stat_calc_lat_u(ts, io_u_lat_u);
646 stat_calc_lat_m(ts, io_u_lat_m);
647
648 show_lat_n(io_u_lat_n, out);
649 show_lat_u(io_u_lat_u, out);
650 show_lat_m(io_u_lat_m, out);
651}
652
653static int block_state_category(int block_state)
654{
655 switch (block_state) {
656 case BLOCK_STATE_UNINIT:
657 return 0;
658 case BLOCK_STATE_TRIMMED:
659 case BLOCK_STATE_WRITTEN:
660 return 1;
661 case BLOCK_STATE_WRITE_FAILURE:
662 case BLOCK_STATE_TRIM_FAILURE:
663 return 2;
664 default:
665 /* Silence compile warning on some BSDs and have a return */
666 assert(0);
667 return -1;
668 }
669}
670
671static int compare_block_infos(const void *bs1, const void *bs2)
672{
673 uint64_t block1 = *(uint64_t *)bs1;
674 uint64_t block2 = *(uint64_t *)bs2;
675 int state1 = BLOCK_INFO_STATE(block1);
676 int state2 = BLOCK_INFO_STATE(block2);
677 int bscat1 = block_state_category(state1);
678 int bscat2 = block_state_category(state2);
679 int cycles1 = BLOCK_INFO_TRIMS(block1);
680 int cycles2 = BLOCK_INFO_TRIMS(block2);
681
682 if (bscat1 < bscat2)
683 return -1;
684 if (bscat1 > bscat2)
685 return 1;
686
687 if (cycles1 < cycles2)
688 return -1;
689 if (cycles1 > cycles2)
690 return 1;
691
692 if (state1 < state2)
693 return -1;
694 if (state1 > state2)
695 return 1;
696
697 assert(block1 == block2);
698 return 0;
699}
700
701static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
702 fio_fp64_t *plist, unsigned int **percentiles,
703 unsigned int *types)
704{
705 int len = 0;
706 int i, nr_uninit;
707
708 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
709
710 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
711 len++;
712
713 if (!len)
714 return 0;
715
716 /*
717 * Sort the percentile list. Note that it may already be sorted if
718 * we are using the default values, but since it's a short list this
719 * isn't a worry. Also note that this does not work for NaN values.
720 */
721 if (len > 1)
722 qsort(plist, len, sizeof(plist[0]), double_cmp);
723
724 /* Start only after the uninit entries end */
725 for (nr_uninit = 0;
726 nr_uninit < nr_block_infos
727 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
728 nr_uninit ++)
729 ;
730
731 if (nr_uninit == nr_block_infos)
732 return 0;
733
734 *percentiles = calloc(len, sizeof(**percentiles));
735
736 for (i = 0; i < len; i++) {
737 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
738 + nr_uninit;
739 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
740 }
741
742 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
743 for (i = 0; i < nr_block_infos; i++)
744 types[BLOCK_INFO_STATE(block_infos[i])]++;
745
746 return len;
747}
748
749static const char *block_state_names[] = {
750 [BLOCK_STATE_UNINIT] = "unwritten",
751 [BLOCK_STATE_TRIMMED] = "trimmed",
752 [BLOCK_STATE_WRITTEN] = "written",
753 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
754 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
755};
756
757static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
758 fio_fp64_t *plist, struct buf_output *out)
759{
760 int len, pos, i;
761 unsigned int *percentiles = NULL;
762 unsigned int block_state_counts[BLOCK_STATE_COUNT];
763
764 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
765 &percentiles, block_state_counts);
766
767 log_buf(out, " block lifetime percentiles :\n |");
768 pos = 0;
769 for (i = 0; i < len; i++) {
770 uint32_t block_info = percentiles[i];
771#define LINE_LENGTH 75
772 char str[LINE_LENGTH];
773 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
774 plist[i].u.f, block_info,
775 i == len - 1 ? '\n' : ',');
776 assert(strln < LINE_LENGTH);
777 if (pos + strln > LINE_LENGTH) {
778 pos = 0;
779 log_buf(out, "\n |");
780 }
781 log_buf(out, "%s", str);
782 pos += strln;
783#undef LINE_LENGTH
784 }
785 if (percentiles)
786 free(percentiles);
787
788 log_buf(out, " states :");
789 for (i = 0; i < BLOCK_STATE_COUNT; i++)
790 log_buf(out, " %s=%u%c",
791 block_state_names[i], block_state_counts[i],
792 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
793}
794
795static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
796{
797 char *p1, *p1alt, *p2;
798 unsigned long long bw_mean, iops_mean;
799 const int i2p = is_power_of_2(ts->kb_base);
800
801 if (!ts->ss_dur)
802 return;
803
804 bw_mean = steadystate_bw_mean(ts);
805 iops_mean = steadystate_iops_mean(ts);
806
807 p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
808 p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
809 p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
810
811 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
812 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
813 p1, p1alt, p2,
814 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
815 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
816 ts->ss_criterion.u.f,
817 ts->ss_state & FIO_SS_PCT ? "%" : "");
818
819 free(p1);
820 free(p1alt);
821 free(p2);
822}
823
824static void show_agg_stats(struct disk_util_agg *agg, int terse,
825 struct buf_output *out)
826{
827 if (!agg->slavecount)
828 return;
829
830 if (!terse) {
831 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
832 "aggrticks=%llu/%llu, aggrin_queue=%llu, "
833 "aggrutil=%3.2f%%",
834 (unsigned long long) agg->ios[0] / agg->slavecount,
835 (unsigned long long) agg->ios[1] / agg->slavecount,
836 (unsigned long long) agg->merges[0] / agg->slavecount,
837 (unsigned long long) agg->merges[1] / agg->slavecount,
838 (unsigned long long) agg->ticks[0] / agg->slavecount,
839 (unsigned long long) agg->ticks[1] / agg->slavecount,
840 (unsigned long long) agg->time_in_queue / agg->slavecount,
841 agg->max_util.u.f);
842 } else {
843 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
844 (unsigned long long) agg->ios[0] / agg->slavecount,
845 (unsigned long long) agg->ios[1] / agg->slavecount,
846 (unsigned long long) agg->merges[0] / agg->slavecount,
847 (unsigned long long) agg->merges[1] / agg->slavecount,
848 (unsigned long long) agg->ticks[0] / agg->slavecount,
849 (unsigned long long) agg->ticks[1] / agg->slavecount,
850 (unsigned long long) agg->time_in_queue / agg->slavecount,
851 agg->max_util.u.f);
852 }
853}
854
855static void aggregate_slaves_stats(struct disk_util *masterdu)
856{
857 struct disk_util_agg *agg = &masterdu->agg;
858 struct disk_util_stat *dus;
859 struct flist_head *entry;
860 struct disk_util *slavedu;
861 double util;
862
863 flist_for_each(entry, &masterdu->slaves) {
864 slavedu = flist_entry(entry, struct disk_util, slavelist);
865 dus = &slavedu->dus;
866 agg->ios[0] += dus->s.ios[0];
867 agg->ios[1] += dus->s.ios[1];
868 agg->merges[0] += dus->s.merges[0];
869 agg->merges[1] += dus->s.merges[1];
870 agg->sectors[0] += dus->s.sectors[0];
871 agg->sectors[1] += dus->s.sectors[1];
872 agg->ticks[0] += dus->s.ticks[0];
873 agg->ticks[1] += dus->s.ticks[1];
874 agg->time_in_queue += dus->s.time_in_queue;
875 agg->slavecount++;
876
877 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
878 /* System utilization is the utilization of the
879 * component with the highest utilization.
880 */
881 if (util > agg->max_util.u.f)
882 agg->max_util.u.f = util;
883
884 }
885
886 if (agg->max_util.u.f > 100.0)
887 agg->max_util.u.f = 100.0;
888}
889
890void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
891 int terse, struct buf_output *out)
892{
893 double util = 0;
894
895 if (dus->s.msec)
896 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
897 if (util > 100.0)
898 util = 100.0;
899
900 if (!terse) {
901 if (agg->slavecount)
902 log_buf(out, " ");
903
904 log_buf(out, " %s: ios=%llu/%llu, merge=%llu/%llu, "
905 "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
906 dus->name,
907 (unsigned long long) dus->s.ios[0],
908 (unsigned long long) dus->s.ios[1],
909 (unsigned long long) dus->s.merges[0],
910 (unsigned long long) dus->s.merges[1],
911 (unsigned long long) dus->s.ticks[0],
912 (unsigned long long) dus->s.ticks[1],
913 (unsigned long long) dus->s.time_in_queue,
914 util);
915 } else {
916 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
917 dus->name,
918 (unsigned long long) dus->s.ios[0],
919 (unsigned long long) dus->s.ios[1],
920 (unsigned long long) dus->s.merges[0],
921 (unsigned long long) dus->s.merges[1],
922 (unsigned long long) dus->s.ticks[0],
923 (unsigned long long) dus->s.ticks[1],
924 (unsigned long long) dus->s.time_in_queue,
925 util);
926 }
927
928 /*
929 * If the device has slaves, aggregate the stats for
930 * those slave devices also.
931 */
932 show_agg_stats(agg, terse, out);
933
934 if (!terse)
935 log_buf(out, "\n");
936}
937
938void json_array_add_disk_util(struct disk_util_stat *dus,
939 struct disk_util_agg *agg, struct json_array *array)
940{
941 struct json_object *obj;
942 double util = 0;
943
944 if (dus->s.msec)
945 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
946 if (util > 100.0)
947 util = 100.0;
948
949 obj = json_create_object();
950 json_array_add_value_object(array, obj);
951
952 json_object_add_value_string(obj, "name", dus->name);
953 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
954 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
955 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
956 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
957 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
958 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
959 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
960 json_object_add_value_float(obj, "util", util);
961
962 /*
963 * If the device has slaves, aggregate the stats for
964 * those slave devices also.
965 */
966 if (!agg->slavecount)
967 return;
968 json_object_add_value_int(obj, "aggr_read_ios",
969 agg->ios[0] / agg->slavecount);
970 json_object_add_value_int(obj, "aggr_write_ios",
971 agg->ios[1] / agg->slavecount);
972 json_object_add_value_int(obj, "aggr_read_merges",
973 agg->merges[0] / agg->slavecount);
974 json_object_add_value_int(obj, "aggr_write_merge",
975 agg->merges[1] / agg->slavecount);
976 json_object_add_value_int(obj, "aggr_read_ticks",
977 agg->ticks[0] / agg->slavecount);
978 json_object_add_value_int(obj, "aggr_write_ticks",
979 agg->ticks[1] / agg->slavecount);
980 json_object_add_value_int(obj, "aggr_in_queue",
981 agg->time_in_queue / agg->slavecount);
982 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
983}
984
985static void json_object_add_disk_utils(struct json_object *obj,
986 struct flist_head *head)
987{
988 struct json_array *array = json_create_array();
989 struct flist_head *entry;
990 struct disk_util *du;
991
992 json_object_add_value_array(obj, "disk_util", array);
993
994 flist_for_each(entry, head) {
995 du = flist_entry(entry, struct disk_util, list);
996
997 aggregate_slaves_stats(du);
998 json_array_add_disk_util(&du->dus, &du->agg, array);
999 }
1000}
1001
1002void show_disk_util(int terse, struct json_object *parent,
1003 struct buf_output *out)
1004{
1005 struct flist_head *entry;
1006 struct disk_util *du;
1007 bool do_json;
1008
1009 if (!is_running_backend())
1010 return;
1011
1012 if (flist_empty(&disk_list)) {
1013 return;
1014 }
1015
1016 if ((output_format & FIO_OUTPUT_JSON) && parent)
1017 do_json = true;
1018 else
1019 do_json = false;
1020
1021 if (!terse && !do_json)
1022 log_buf(out, "\nDisk stats (read/write):\n");
1023
1024 if (do_json)
1025 json_object_add_disk_utils(parent, &disk_list);
1026 else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1027 flist_for_each(entry, &disk_list) {
1028 du = flist_entry(entry, struct disk_util, list);
1029
1030 aggregate_slaves_stats(du);
1031 print_disk_util(&du->dus, &du->agg, terse, out);
1032 }
1033 }
1034}
1035
1036static void show_thread_status_normal(struct thread_stat *ts,
1037 struct group_run_stats *rs,
1038 struct buf_output *out)
1039{
1040 double usr_cpu, sys_cpu;
1041 unsigned long runtime;
1042 double io_u_dist[FIO_IO_U_MAP_NR];
1043 time_t time_p;
1044 char time_buf[32];
1045
1046 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1047 return;
1048
1049 memset(time_buf, 0, sizeof(time_buf));
1050
1051 time(&time_p);
1052 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1053
1054 if (!ts->error) {
1055 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1056 ts->name, ts->groupid, ts->members,
1057 ts->error, (int) ts->pid, time_buf);
1058 } else {
1059 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1060 ts->name, ts->groupid, ts->members,
1061 ts->error, ts->verror, (int) ts->pid,
1062 time_buf);
1063 }
1064
1065 if (strlen(ts->description))
1066 log_buf(out, " Description : [%s]\n", ts->description);
1067
1068 if (ts->io_bytes[DDIR_READ])
1069 show_ddir_status(rs, ts, DDIR_READ, out);
1070 if (ts->io_bytes[DDIR_WRITE])
1071 show_ddir_status(rs, ts, DDIR_WRITE, out);
1072 if (ts->io_bytes[DDIR_TRIM])
1073 show_ddir_status(rs, ts, DDIR_TRIM, out);
1074
1075 show_latencies(ts, out);
1076
1077 if (ts->sync_stat.samples)
1078 show_ddir_status(rs, ts, DDIR_SYNC, out);
1079
1080 runtime = ts->total_run_time;
1081 if (runtime) {
1082 double runt = (double) runtime;
1083
1084 usr_cpu = (double) ts->usr_time * 100 / runt;
1085 sys_cpu = (double) ts->sys_time * 100 / runt;
1086 } else {
1087 usr_cpu = 0;
1088 sys_cpu = 0;
1089 }
1090
1091 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1092 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1093 (unsigned long long) ts->ctx,
1094 (unsigned long long) ts->majf,
1095 (unsigned long long) ts->minf);
1096
1097 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1098 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1099 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1100 io_u_dist[1], io_u_dist[2],
1101 io_u_dist[3], io_u_dist[4],
1102 io_u_dist[5], io_u_dist[6]);
1103
1104 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1105 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1106 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1107 io_u_dist[1], io_u_dist[2],
1108 io_u_dist[3], io_u_dist[4],
1109 io_u_dist[5], io_u_dist[6]);
1110 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1111 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1112 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1113 io_u_dist[1], io_u_dist[2],
1114 io_u_dist[3], io_u_dist[4],
1115 io_u_dist[5], io_u_dist[6]);
1116 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
1117 " short=%llu,%llu,%llu,0"
1118 " dropped=%llu,%llu,%llu,0\n",
1119 (unsigned long long) ts->total_io_u[0],
1120 (unsigned long long) ts->total_io_u[1],
1121 (unsigned long long) ts->total_io_u[2],
1122 (unsigned long long) ts->total_io_u[3],
1123 (unsigned long long) ts->short_io_u[0],
1124 (unsigned long long) ts->short_io_u[1],
1125 (unsigned long long) ts->short_io_u[2],
1126 (unsigned long long) ts->drop_io_u[0],
1127 (unsigned long long) ts->drop_io_u[1],
1128 (unsigned long long) ts->drop_io_u[2]);
1129 if (ts->continue_on_error) {
1130 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
1131 (unsigned long long)ts->total_err_count,
1132 ts->first_error,
1133 strerror(ts->first_error));
1134 }
1135 if (ts->latency_depth) {
1136 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1137 (unsigned long long)ts->latency_target,
1138 (unsigned long long)ts->latency_window,
1139 ts->latency_percentile.u.f,
1140 ts->latency_depth);
1141 }
1142
1143 if (ts->nr_block_infos)
1144 show_block_infos(ts->nr_block_infos, ts->block_infos,
1145 ts->percentile_list, out);
1146
1147 if (ts->ss_dur)
1148 show_ss_normal(ts, out);
1149}
1150
1151static void show_ddir_status_terse(struct thread_stat *ts,
1152 struct group_run_stats *rs, int ddir,
1153 int ver, struct buf_output *out)
1154{
1155 unsigned long long min, max, minv, maxv, bw, iops;
1156 unsigned long long *ovals = NULL;
1157 double mean, dev;
1158 unsigned int len;
1159 int i, bw_stat;
1160
1161 assert(ddir_rw(ddir));
1162
1163 iops = bw = 0;
1164 if (ts->runtime[ddir]) {
1165 uint64_t runt = ts->runtime[ddir];
1166
1167 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1168 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1169 }
1170
1171 log_buf(out, ";%llu;%llu;%llu;%llu",
1172 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1173 (unsigned long long) ts->runtime[ddir]);
1174
1175 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1176 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1177 else
1178 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1179
1180 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1181 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1182 else
1183 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1184
1185 if (ts->lat_percentiles)
1186 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1187 ts->lat_stat[ddir].samples,
1188 ts->percentile_list, &ovals, &maxv,
1189 &minv);
1190 else if (ts->clat_percentiles)
1191 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1192 ts->clat_stat[ddir].samples,
1193 ts->percentile_list, &ovals, &maxv,
1194 &minv);
1195 else
1196 len = 0;
1197
1198 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1199 if (i >= len) {
1200 log_buf(out, ";0%%=0");
1201 continue;
1202 }
1203 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1204 }
1205
1206 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1207 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1208 else
1209 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1210
1211 free(ovals);
1212
1213 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1214 if (bw_stat) {
1215 double p_of_agg = 100.0;
1216
1217 if (rs->agg[ddir]) {
1218 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1219 if (p_of_agg > 100.0)
1220 p_of_agg = 100.0;
1221 }
1222
1223 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1224 } else
1225 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1226
1227 if (ver == 5) {
1228 if (bw_stat)
1229 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1230 else
1231 log_buf(out, ";%lu", 0UL);
1232
1233 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1234 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1235 mean, dev, (&ts->iops_stat[ddir])->samples);
1236 else
1237 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1238 }
1239}
1240
1241static struct json_object *add_ddir_lat_json(struct thread_stat *ts, uint32_t percentiles,
1242 struct io_stat *lat_stat, uint64_t *io_u_plat)
1243{
1244 char buf[120];
1245 double mean, dev;
1246 unsigned int i, len;
1247 struct json_object *lat_object, *percentile_object, *clat_bins_object;
1248 unsigned long long min, max, maxv, minv, *ovals = NULL;
1249
1250 if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1251 min = max = 0;
1252 mean = dev = 0.0;
1253 }
1254 lat_object = json_create_object();
1255 json_object_add_value_int(lat_object, "min", min);
1256 json_object_add_value_int(lat_object, "max", max);
1257 json_object_add_value_float(lat_object, "mean", mean);
1258 json_object_add_value_float(lat_object, "stddev", dev);
1259 json_object_add_value_int(lat_object, "N", lat_stat->samples);
1260
1261 if (percentiles && lat_stat->samples) {
1262 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1263 ts->percentile_list, &ovals, &maxv, &minv);
1264
1265 if (len > FIO_IO_U_LIST_MAX_LEN)
1266 len = FIO_IO_U_LIST_MAX_LEN;
1267
1268 percentile_object = json_create_object();
1269 json_object_add_value_object(lat_object, "percentile", percentile_object);
1270 for (i = 0; i < len; i++) {
1271 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1272 json_object_add_value_int(percentile_object, buf, ovals[i]);
1273 }
1274 free(ovals);
1275
1276 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1277 clat_bins_object = json_create_object();
1278 json_object_add_value_object(lat_object, "bins", clat_bins_object);
1279
1280 for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1281 if (io_u_plat[i]) {
1282 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1283 json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1284 }
1285 }
1286 }
1287
1288 return lat_object;
1289}
1290
1291static void add_ddir_status_json(struct thread_stat *ts,
1292 struct group_run_stats *rs, int ddir, struct json_object *parent)
1293{
1294 unsigned long long min, max;
1295 unsigned long long bw_bytes, bw;
1296 double mean, dev, iops;
1297 struct json_object *dir_object, *tmp_object;
1298 double p_of_agg = 100.0;
1299
1300 assert(ddir_rw(ddir) || ddir_sync(ddir));
1301
1302 if (ts->unified_rw_rep && ddir != DDIR_READ)
1303 return;
1304
1305 dir_object = json_create_object();
1306 json_object_add_value_object(parent,
1307 ts->unified_rw_rep ? "mixed" : io_ddir_name(ddir), dir_object);
1308
1309 if (ddir_rw(ddir)) {
1310 bw_bytes = 0;
1311 bw = 0;
1312 iops = 0.0;
1313 if (ts->runtime[ddir]) {
1314 uint64_t runt = ts->runtime[ddir];
1315
1316 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1317 bw = bw_bytes / 1024; /* KiB/s */
1318 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1319 }
1320
1321 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1322 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1323 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1324 json_object_add_value_int(dir_object, "bw", bw);
1325 json_object_add_value_float(dir_object, "iops", iops);
1326 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1327 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1328 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1329 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1330
1331 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1332 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1333 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1334
1335 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1336 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1337 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1338
1339 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1340 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1341 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1342 } else {
1343 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1344 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1345 &ts->sync_stat, ts->io_u_sync_plat);
1346 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1347 }
1348
1349 if (!ddir_rw(ddir))
1350 return;
1351
1352 /* Only print PRIO latencies if some high priority samples were gathered */
1353 if (ts->clat_high_prio_stat[ddir].samples > 0) {
1354 const char *high, *low;
1355
1356 if (ts->lat_percentiles) {
1357 high = "lat_prio";
1358 low = "lat_low_prio";
1359 } else {
1360 high = "clat_prio";
1361 low = "clat_low_prio";
1362 }
1363
1364 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1365 &ts->clat_high_prio_stat[ddir], ts->io_u_plat_high_prio[ddir]);
1366 json_object_add_value_object(dir_object, high, tmp_object);
1367
1368 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1369 &ts->clat_prio_stat[ddir], ts->io_u_plat_prio[ddir]);
1370 json_object_add_value_object(dir_object, low, tmp_object);
1371 }
1372
1373 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1374 if (rs->agg[ddir]) {
1375 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1376 if (p_of_agg > 100.0)
1377 p_of_agg = 100.0;
1378 }
1379 } else {
1380 min = max = 0;
1381 p_of_agg = mean = dev = 0.0;
1382 }
1383
1384 json_object_add_value_int(dir_object, "bw_min", min);
1385 json_object_add_value_int(dir_object, "bw_max", max);
1386 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1387 json_object_add_value_float(dir_object, "bw_mean", mean);
1388 json_object_add_value_float(dir_object, "bw_dev", dev);
1389 json_object_add_value_int(dir_object, "bw_samples",
1390 (&ts->bw_stat[ddir])->samples);
1391
1392 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1393 min = max = 0;
1394 mean = dev = 0.0;
1395 }
1396 json_object_add_value_int(dir_object, "iops_min", min);
1397 json_object_add_value_int(dir_object, "iops_max", max);
1398 json_object_add_value_float(dir_object, "iops_mean", mean);
1399 json_object_add_value_float(dir_object, "iops_stddev", dev);
1400 json_object_add_value_int(dir_object, "iops_samples",
1401 (&ts->iops_stat[ddir])->samples);
1402
1403 if (ts->cachehit + ts->cachemiss) {
1404 uint64_t total;
1405 double hit;
1406
1407 total = ts->cachehit + ts->cachemiss;
1408 hit = (double) ts->cachehit / (double) total;
1409 hit *= 100.0;
1410 json_object_add_value_float(dir_object, "cachehit", hit);
1411 }
1412}
1413
1414static void show_thread_status_terse_all(struct thread_stat *ts,
1415 struct group_run_stats *rs, int ver,
1416 struct buf_output *out)
1417{
1418 double io_u_dist[FIO_IO_U_MAP_NR];
1419 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1420 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1421 double usr_cpu, sys_cpu;
1422 int i;
1423
1424 /* General Info */
1425 if (ver == 2)
1426 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1427 else
1428 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1429 ts->name, ts->groupid, ts->error);
1430
1431 /* Log Read Status */
1432 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1433 /* Log Write Status */
1434 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1435 /* Log Trim Status */
1436 if (ver == 2 || ver == 4 || ver == 5)
1437 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1438
1439 /* CPU Usage */
1440 if (ts->total_run_time) {
1441 double runt = (double) ts->total_run_time;
1442
1443 usr_cpu = (double) ts->usr_time * 100 / runt;
1444 sys_cpu = (double) ts->sys_time * 100 / runt;
1445 } else {
1446 usr_cpu = 0;
1447 sys_cpu = 0;
1448 }
1449
1450 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1451 (unsigned long long) ts->ctx,
1452 (unsigned long long) ts->majf,
1453 (unsigned long long) ts->minf);
1454
1455 /* Calc % distribution of IO depths, usecond, msecond latency */
1456 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1457 stat_calc_lat_nu(ts, io_u_lat_u);
1458 stat_calc_lat_m(ts, io_u_lat_m);
1459
1460 /* Only show fixed 7 I/O depth levels*/
1461 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1462 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1463 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1464
1465 /* Microsecond latency */
1466 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1467 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1468 /* Millisecond latency */
1469 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1470 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1471
1472 /* disk util stats, if any */
1473 if (ver >= 3 && is_running_backend())
1474 show_disk_util(1, NULL, out);
1475
1476 /* Additional output if continue_on_error set - default off*/
1477 if (ts->continue_on_error)
1478 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1479
1480 /* Additional output if description is set */
1481 if (strlen(ts->description)) {
1482 if (ver == 2)
1483 log_buf(out, "\n");
1484 log_buf(out, ";%s", ts->description);
1485 }
1486
1487 log_buf(out, "\n");
1488}
1489
1490static void json_add_job_opts(struct json_object *root, const char *name,
1491 struct flist_head *opt_list)
1492{
1493 struct json_object *dir_object;
1494 struct flist_head *entry;
1495 struct print_option *p;
1496
1497 if (flist_empty(opt_list))
1498 return;
1499
1500 dir_object = json_create_object();
1501 json_object_add_value_object(root, name, dir_object);
1502
1503 flist_for_each(entry, opt_list) {
1504 const char *pos = "";
1505
1506 p = flist_entry(entry, struct print_option, list);
1507 if (p->value)
1508 pos = p->value;
1509 json_object_add_value_string(dir_object, p->name, pos);
1510 }
1511}
1512
1513static struct json_object *show_thread_status_json(struct thread_stat *ts,
1514 struct group_run_stats *rs,
1515 struct flist_head *opt_list)
1516{
1517 struct json_object *root, *tmp;
1518 struct jobs_eta *je;
1519 double io_u_dist[FIO_IO_U_MAP_NR];
1520 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1521 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1522 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1523 double usr_cpu, sys_cpu;
1524 int i;
1525 size_t size;
1526
1527 root = json_create_object();
1528 json_object_add_value_string(root, "jobname", ts->name);
1529 json_object_add_value_int(root, "groupid", ts->groupid);
1530 json_object_add_value_int(root, "error", ts->error);
1531
1532 /* ETA Info */
1533 je = get_jobs_eta(true, &size);
1534 if (je) {
1535 json_object_add_value_int(root, "eta", je->eta_sec);
1536 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1537 }
1538
1539 if (opt_list)
1540 json_add_job_opts(root, "job options", opt_list);
1541
1542 add_ddir_status_json(ts, rs, DDIR_READ, root);
1543 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1544 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1545 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1546
1547 /* CPU Usage */
1548 if (ts->total_run_time) {
1549 double runt = (double) ts->total_run_time;
1550
1551 usr_cpu = (double) ts->usr_time * 100 / runt;
1552 sys_cpu = (double) ts->sys_time * 100 / runt;
1553 } else {
1554 usr_cpu = 0;
1555 sys_cpu = 0;
1556 }
1557 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1558 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1559 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1560 json_object_add_value_int(root, "ctx", ts->ctx);
1561 json_object_add_value_int(root, "majf", ts->majf);
1562 json_object_add_value_int(root, "minf", ts->minf);
1563
1564 /* Calc % distribution of IO depths */
1565 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1566 tmp = json_create_object();
1567 json_object_add_value_object(root, "iodepth_level", tmp);
1568 /* Only show fixed 7 I/O depth levels*/
1569 for (i = 0; i < 7; i++) {
1570 char name[20];
1571 if (i < 6)
1572 snprintf(name, 20, "%d", 1 << i);
1573 else
1574 snprintf(name, 20, ">=%d", 1 << i);
1575 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1576 }
1577
1578 /* Calc % distribution of submit IO depths */
1579 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1580 tmp = json_create_object();
1581 json_object_add_value_object(root, "iodepth_submit", tmp);
1582 /* Only show fixed 7 I/O depth levels*/
1583 for (i = 0; i < 7; i++) {
1584 char name[20];
1585 if (i == 0)
1586 snprintf(name, 20, "0");
1587 else if (i < 6)
1588 snprintf(name, 20, "%d", 1 << (i+1));
1589 else
1590 snprintf(name, 20, ">=%d", 1 << i);
1591 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1592 }
1593
1594 /* Calc % distribution of completion IO depths */
1595 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1596 tmp = json_create_object();
1597 json_object_add_value_object(root, "iodepth_complete", tmp);
1598 /* Only show fixed 7 I/O depth levels*/
1599 for (i = 0; i < 7; i++) {
1600 char name[20];
1601 if (i == 0)
1602 snprintf(name, 20, "0");
1603 else if (i < 6)
1604 snprintf(name, 20, "%d", 1 << (i+1));
1605 else
1606 snprintf(name, 20, ">=%d", 1 << i);
1607 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1608 }
1609
1610 /* Calc % distribution of nsecond, usecond, msecond latency */
1611 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1612 stat_calc_lat_n(ts, io_u_lat_n);
1613 stat_calc_lat_u(ts, io_u_lat_u);
1614 stat_calc_lat_m(ts, io_u_lat_m);
1615
1616 /* Nanosecond latency */
1617 tmp = json_create_object();
1618 json_object_add_value_object(root, "latency_ns", tmp);
1619 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1620 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1621 "250", "500", "750", "1000", };
1622 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1623 }
1624 /* Microsecond latency */
1625 tmp = json_create_object();
1626 json_object_add_value_object(root, "latency_us", tmp);
1627 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1628 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1629 "250", "500", "750", "1000", };
1630 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1631 }
1632 /* Millisecond latency */
1633 tmp = json_create_object();
1634 json_object_add_value_object(root, "latency_ms", tmp);
1635 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1636 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1637 "250", "500", "750", "1000", "2000",
1638 ">=2000", };
1639 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1640 }
1641
1642 /* Additional output if continue_on_error set - default off*/
1643 if (ts->continue_on_error) {
1644 json_object_add_value_int(root, "total_err", ts->total_err_count);
1645 json_object_add_value_int(root, "first_error", ts->first_error);
1646 }
1647
1648 if (ts->latency_depth) {
1649 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1650 json_object_add_value_int(root, "latency_target", ts->latency_target);
1651 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1652 json_object_add_value_int(root, "latency_window", ts->latency_window);
1653 }
1654
1655 /* Additional output if description is set */
1656 if (strlen(ts->description))
1657 json_object_add_value_string(root, "desc", ts->description);
1658
1659 if (ts->nr_block_infos) {
1660 /* Block error histogram and types */
1661 int len;
1662 unsigned int *percentiles = NULL;
1663 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1664
1665 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1666 ts->percentile_list,
1667 &percentiles, block_state_counts);
1668
1669 if (len) {
1670 struct json_object *block, *percentile_object, *states;
1671 int state;
1672 block = json_create_object();
1673 json_object_add_value_object(root, "block", block);
1674
1675 percentile_object = json_create_object();
1676 json_object_add_value_object(block, "percentiles",
1677 percentile_object);
1678 for (i = 0; i < len; i++) {
1679 char buf[20];
1680 snprintf(buf, sizeof(buf), "%f",
1681 ts->percentile_list[i].u.f);
1682 json_object_add_value_int(percentile_object,
1683 buf,
1684 percentiles[i]);
1685 }
1686
1687 states = json_create_object();
1688 json_object_add_value_object(block, "states", states);
1689 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1690 json_object_add_value_int(states,
1691 block_state_names[state],
1692 block_state_counts[state]);
1693 }
1694 free(percentiles);
1695 }
1696 }
1697
1698 if (ts->ss_dur) {
1699 struct json_object *data;
1700 struct json_array *iops, *bw;
1701 int j, k, l;
1702 char ss_buf[64];
1703
1704 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1705 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1706 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1707 (float) ts->ss_limit.u.f,
1708 ts->ss_state & FIO_SS_PCT ? "%" : "");
1709
1710 tmp = json_create_object();
1711 json_object_add_value_object(root, "steadystate", tmp);
1712 json_object_add_value_string(tmp, "ss", ss_buf);
1713 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1714 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1715
1716 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1717 ts->ss_state & FIO_SS_PCT ? "%" : "");
1718 json_object_add_value_string(tmp, "criterion", ss_buf);
1719 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1720 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1721
1722 data = json_create_object();
1723 json_object_add_value_object(tmp, "data", data);
1724 bw = json_create_array();
1725 iops = json_create_array();
1726
1727 /*
1728 ** if ss was attained or the buffer is not full,
1729 ** ss->head points to the first element in the list.
1730 ** otherwise it actually points to the second element
1731 ** in the list
1732 */
1733 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1734 j = ts->ss_head;
1735 else
1736 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1737 for (l = 0; l < ts->ss_dur; l++) {
1738 k = (j + l) % ts->ss_dur;
1739 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1740 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1741 }
1742 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1743 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1744 json_object_add_value_array(data, "iops", iops);
1745 json_object_add_value_array(data, "bw", bw);
1746 }
1747
1748 return root;
1749}
1750
1751static void show_thread_status_terse(struct thread_stat *ts,
1752 struct group_run_stats *rs,
1753 struct buf_output *out)
1754{
1755 if (terse_version >= 2 && terse_version <= 5)
1756 show_thread_status_terse_all(ts, rs, terse_version, out);
1757 else
1758 log_err("fio: bad terse version!? %d\n", terse_version);
1759}
1760
1761struct json_object *show_thread_status(struct thread_stat *ts,
1762 struct group_run_stats *rs,
1763 struct flist_head *opt_list,
1764 struct buf_output *out)
1765{
1766 struct json_object *ret = NULL;
1767
1768 if (output_format & FIO_OUTPUT_TERSE)
1769 show_thread_status_terse(ts, rs, out);
1770 if (output_format & FIO_OUTPUT_JSON)
1771 ret = show_thread_status_json(ts, rs, opt_list);
1772 if (output_format & FIO_OUTPUT_NORMAL)
1773 show_thread_status_normal(ts, rs, out);
1774
1775 return ret;
1776}
1777
1778static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1779{
1780 double mean, S;
1781
1782 dst->min_val = min(dst->min_val, src->min_val);
1783 dst->max_val = max(dst->max_val, src->max_val);
1784
1785 /*
1786 * Compute new mean and S after the merge
1787 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1788 * #Parallel_algorithm>
1789 */
1790 if (first) {
1791 mean = src->mean.u.f;
1792 S = src->S.u.f;
1793 } else {
1794 double delta = src->mean.u.f - dst->mean.u.f;
1795
1796 mean = ((src->mean.u.f * src->samples) +
1797 (dst->mean.u.f * dst->samples)) /
1798 (dst->samples + src->samples);
1799
1800 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1801 (dst->samples * src->samples) /
1802 (dst->samples + src->samples);
1803 }
1804
1805 dst->samples += src->samples;
1806 dst->mean.u.f = mean;
1807 dst->S.u.f = S;
1808
1809}
1810
1811/*
1812 * We sum two kinds of stats - one that is time based, in which case we
1813 * apply the proper summing technique, and then one that is iops/bw
1814 * numbers. For group_reporting, we should just add those up, not make
1815 * them the mean of everything.
1816 */
1817static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first,
1818 bool pure_sum)
1819{
1820 if (src->samples == 0)
1821 return;
1822
1823 if (!pure_sum) {
1824 __sum_stat(dst, src, first);
1825 return;
1826 }
1827
1828 if (first) {
1829 dst->min_val = src->min_val;
1830 dst->max_val = src->max_val;
1831 dst->samples = src->samples;
1832 dst->mean.u.f = src->mean.u.f;
1833 dst->S.u.f = src->S.u.f;
1834 } else {
1835 dst->min_val += src->min_val;
1836 dst->max_val += src->max_val;
1837 dst->samples += src->samples;
1838 dst->mean.u.f += src->mean.u.f;
1839 dst->S.u.f += src->S.u.f;
1840 }
1841}
1842
1843void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1844{
1845 int i;
1846
1847 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1848 if (dst->max_run[i] < src->max_run[i])
1849 dst->max_run[i] = src->max_run[i];
1850 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1851 dst->min_run[i] = src->min_run[i];
1852 if (dst->max_bw[i] < src->max_bw[i])
1853 dst->max_bw[i] = src->max_bw[i];
1854 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1855 dst->min_bw[i] = src->min_bw[i];
1856
1857 dst->iobytes[i] += src->iobytes[i];
1858 dst->agg[i] += src->agg[i];
1859 }
1860
1861 if (!dst->kb_base)
1862 dst->kb_base = src->kb_base;
1863 if (!dst->unit_base)
1864 dst->unit_base = src->unit_base;
1865 if (!dst->sig_figs)
1866 dst->sig_figs = src->sig_figs;
1867}
1868
1869void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1870 bool first)
1871{
1872 int k, l, m;
1873
1874 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1875 if (!dst->unified_rw_rep) {
1876 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first, false);
1877 sum_stat(&dst->clat_high_prio_stat[l], &src->clat_high_prio_stat[l], first, false);
1878 sum_stat(&dst->clat_prio_stat[l], &src->clat_prio_stat[l], first, false);
1879 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first, false);
1880 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first, false);
1881 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first, true);
1882 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first, true);
1883
1884 dst->io_bytes[l] += src->io_bytes[l];
1885
1886 if (dst->runtime[l] < src->runtime[l])
1887 dst->runtime[l] = src->runtime[l];
1888 } else {
1889 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first, false);
1890 sum_stat(&dst->clat_high_prio_stat[l], &src->clat_high_prio_stat[l], first, false);
1891 sum_stat(&dst->clat_prio_stat[l], &src->clat_prio_stat[l], first, false);
1892 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first, false);
1893 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first, false);
1894 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first, true);
1895 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first, true);
1896
1897 dst->io_bytes[0] += src->io_bytes[l];
1898
1899 if (dst->runtime[0] < src->runtime[l])
1900 dst->runtime[0] = src->runtime[l];
1901
1902 /*
1903 * We're summing to the same destination, so override
1904 * 'first' after the first iteration of the loop
1905 */
1906 first = false;
1907 }
1908 }
1909
1910 sum_stat(&dst->sync_stat, &src->sync_stat, first, false);
1911 dst->usr_time += src->usr_time;
1912 dst->sys_time += src->sys_time;
1913 dst->ctx += src->ctx;
1914 dst->majf += src->majf;
1915 dst->minf += src->minf;
1916
1917 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
1918 dst->io_u_map[k] += src->io_u_map[k];
1919 dst->io_u_submit[k] += src->io_u_submit[k];
1920 dst->io_u_complete[k] += src->io_u_complete[k];
1921 }
1922
1923 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1924 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1925 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1926 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1927 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1928 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1929
1930 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1931 if (!dst->unified_rw_rep) {
1932 dst->total_io_u[k] += src->total_io_u[k];
1933 dst->short_io_u[k] += src->short_io_u[k];
1934 dst->drop_io_u[k] += src->drop_io_u[k];
1935 } else {
1936 dst->total_io_u[0] += src->total_io_u[k];
1937 dst->short_io_u[0] += src->short_io_u[k];
1938 dst->drop_io_u[0] += src->drop_io_u[k];
1939 }
1940 }
1941
1942 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
1943
1944 for (k = 0; k < FIO_LAT_CNT; k++)
1945 for (l = 0; l < DDIR_RWDIR_CNT; l++)
1946 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
1947 if (!dst->unified_rw_rep)
1948 dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
1949 else
1950 dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
1951
1952 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1953 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
1954
1955 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1956 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1957 if (!dst->unified_rw_rep) {
1958 dst->io_u_plat_high_prio[k][m] += src->io_u_plat_high_prio[k][m];
1959 dst->io_u_plat_prio[k][m] += src->io_u_plat_prio[k][m];
1960 } else {
1961 dst->io_u_plat_high_prio[0][m] += src->io_u_plat_high_prio[k][m];
1962 dst->io_u_plat_prio[0][m] += src->io_u_plat_prio[k][m];
1963 }
1964
1965 }
1966 }
1967
1968 dst->total_run_time += src->total_run_time;
1969 dst->total_submit += src->total_submit;
1970 dst->total_complete += src->total_complete;
1971 dst->nr_zone_resets += src->nr_zone_resets;
1972 dst->cachehit += src->cachehit;
1973 dst->cachemiss += src->cachemiss;
1974}
1975
1976void init_group_run_stat(struct group_run_stats *gs)
1977{
1978 int i;
1979 memset(gs, 0, sizeof(*gs));
1980
1981 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1982 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1983}
1984
1985void init_thread_stat(struct thread_stat *ts)
1986{
1987 int j;
1988
1989 memset(ts, 0, sizeof(*ts));
1990
1991 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1992 ts->lat_stat[j].min_val = -1UL;
1993 ts->clat_stat[j].min_val = -1UL;
1994 ts->slat_stat[j].min_val = -1UL;
1995 ts->bw_stat[j].min_val = -1UL;
1996 ts->iops_stat[j].min_val = -1UL;
1997 ts->clat_high_prio_stat[j].min_val = -1UL;
1998 ts->clat_prio_stat[j].min_val = -1UL;
1999 }
2000 ts->sync_stat.min_val = -1UL;
2001 ts->groupid = -1;
2002}
2003
2004void __show_run_stats(void)
2005{
2006 struct group_run_stats *runstats, *rs;
2007 struct thread_data *td;
2008 struct thread_stat *threadstats, *ts;
2009 int i, j, k, nr_ts, last_ts, idx;
2010 bool kb_base_warned = false;
2011 bool unit_base_warned = false;
2012 struct json_object *root = NULL;
2013 struct json_array *array = NULL;
2014 struct buf_output output[FIO_OUTPUT_NR];
2015 struct flist_head **opt_lists;
2016
2017 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2018
2019 for (i = 0; i < groupid + 1; i++)
2020 init_group_run_stat(&runstats[i]);
2021
2022 /*
2023 * find out how many threads stats we need. if group reporting isn't
2024 * enabled, it's one-per-td.
2025 */
2026 nr_ts = 0;
2027 last_ts = -1;
2028 for_each_td(td, i) {
2029 if (!td->o.group_reporting) {
2030 nr_ts++;
2031 continue;
2032 }
2033 if (last_ts == td->groupid)
2034 continue;
2035 if (!td->o.stats)
2036 continue;
2037
2038 last_ts = td->groupid;
2039 nr_ts++;
2040 }
2041
2042 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2043 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2044
2045 for (i = 0; i < nr_ts; i++) {
2046 init_thread_stat(&threadstats[i]);
2047 opt_lists[i] = NULL;
2048 }
2049
2050 j = 0;
2051 last_ts = -1;
2052 idx = 0;
2053 for_each_td(td, i) {
2054 if (!td->o.stats)
2055 continue;
2056 if (idx && (!td->o.group_reporting ||
2057 (td->o.group_reporting && last_ts != td->groupid))) {
2058 idx = 0;
2059 j++;
2060 }
2061
2062 last_ts = td->groupid;
2063
2064 ts = &threadstats[j];
2065
2066 ts->clat_percentiles = td->o.clat_percentiles;
2067 ts->lat_percentiles = td->o.lat_percentiles;
2068 ts->slat_percentiles = td->o.slat_percentiles;
2069 ts->percentile_precision = td->o.percentile_precision;
2070 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2071 opt_lists[j] = &td->opt_list;
2072
2073 idx++;
2074 ts->members++;
2075
2076 if (ts->groupid == -1) {
2077 /*
2078 * These are per-group shared already
2079 */
2080 snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2081 if (td->o.description)
2082 snprintf(ts->description,
2083 sizeof(ts->description), "%s",
2084 td->o.description);
2085 else
2086 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2087
2088 /*
2089 * If multiple entries in this group, this is
2090 * the first member.
2091 */
2092 ts->thread_number = td->thread_number;
2093 ts->groupid = td->groupid;
2094
2095 /*
2096 * first pid in group, not very useful...
2097 */
2098 ts->pid = td->pid;
2099
2100 ts->kb_base = td->o.kb_base;
2101 ts->unit_base = td->o.unit_base;
2102 ts->sig_figs = td->o.sig_figs;
2103 ts->unified_rw_rep = td->o.unified_rw_rep;
2104 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2105 log_info("fio: kb_base differs for jobs in group, using"
2106 " %u as the base\n", ts->kb_base);
2107 kb_base_warned = true;
2108 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2109 log_info("fio: unit_base differs for jobs in group, using"
2110 " %u as the base\n", ts->unit_base);
2111 unit_base_warned = true;
2112 }
2113
2114 ts->continue_on_error = td->o.continue_on_error;
2115 ts->total_err_count += td->total_err_count;
2116 ts->first_error = td->first_error;
2117 if (!ts->error) {
2118 if (!td->error && td->o.continue_on_error &&
2119 td->first_error) {
2120 ts->error = td->first_error;
2121 snprintf(ts->verror, sizeof(ts->verror), "%s",
2122 td->verror);
2123 } else if (td->error) {
2124 ts->error = td->error;
2125 snprintf(ts->verror, sizeof(ts->verror), "%s",
2126 td->verror);
2127 }
2128 }
2129
2130 ts->latency_depth = td->latency_qd;
2131 ts->latency_target = td->o.latency_target;
2132 ts->latency_percentile = td->o.latency_percentile;
2133 ts->latency_window = td->o.latency_window;
2134
2135 ts->nr_block_infos = td->ts.nr_block_infos;
2136 for (k = 0; k < ts->nr_block_infos; k++)
2137 ts->block_infos[k] = td->ts.block_infos[k];
2138
2139 sum_thread_stats(ts, &td->ts, idx == 1);
2140
2141 if (td->o.ss_dur) {
2142 ts->ss_state = td->ss.state;
2143 ts->ss_dur = td->ss.dur;
2144 ts->ss_head = td->ss.head;
2145 ts->ss_bw_data = td->ss.bw_data;
2146 ts->ss_iops_data = td->ss.iops_data;
2147 ts->ss_limit.u.f = td->ss.limit;
2148 ts->ss_slope.u.f = td->ss.slope;
2149 ts->ss_deviation.u.f = td->ss.deviation;
2150 ts->ss_criterion.u.f = td->ss.criterion;
2151 }
2152 else
2153 ts->ss_dur = ts->ss_state = 0;
2154 }
2155
2156 for (i = 0; i < nr_ts; i++) {
2157 unsigned long long bw;
2158
2159 ts = &threadstats[i];
2160 if (ts->groupid == -1)
2161 continue;
2162 rs = &runstats[ts->groupid];
2163 rs->kb_base = ts->kb_base;
2164 rs->unit_base = ts->unit_base;
2165 rs->sig_figs = ts->sig_figs;
2166 rs->unified_rw_rep += ts->unified_rw_rep;
2167
2168 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2169 if (!ts->runtime[j])
2170 continue;
2171 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2172 rs->min_run[j] = ts->runtime[j];
2173 if (ts->runtime[j] > rs->max_run[j])
2174 rs->max_run[j] = ts->runtime[j];
2175
2176 bw = 0;
2177 if (ts->runtime[j])
2178 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2179 if (bw < rs->min_bw[j])
2180 rs->min_bw[j] = bw;
2181 if (bw > rs->max_bw[j])
2182 rs->max_bw[j] = bw;
2183
2184 rs->iobytes[j] += ts->io_bytes[j];
2185 }
2186 }
2187
2188 for (i = 0; i < groupid + 1; i++) {
2189 int ddir;
2190
2191 rs = &runstats[i];
2192
2193 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2194 if (rs->max_run[ddir])
2195 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2196 rs->max_run[ddir];
2197 }
2198 }
2199
2200 for (i = 0; i < FIO_OUTPUT_NR; i++)
2201 buf_output_init(&output[i]);
2202
2203 /*
2204 * don't overwrite last signal output
2205 */
2206 if (output_format & FIO_OUTPUT_NORMAL)
2207 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2208 if (output_format & FIO_OUTPUT_JSON) {
2209 struct thread_data *global;
2210 char time_buf[32];
2211 struct timeval now;
2212 unsigned long long ms_since_epoch;
2213 time_t tv_sec;
2214
2215 gettimeofday(&now, NULL);
2216 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2217 (unsigned long long)(now.tv_usec) / 1000;
2218
2219 tv_sec = now.tv_sec;
2220 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2221 if (time_buf[strlen(time_buf) - 1] == '\n')
2222 time_buf[strlen(time_buf) - 1] = '\0';
2223
2224 root = json_create_object();
2225 json_object_add_value_string(root, "fio version", fio_version_string);
2226 json_object_add_value_int(root, "timestamp", now.tv_sec);
2227 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2228 json_object_add_value_string(root, "time", time_buf);
2229 global = get_global_options();
2230 json_add_job_opts(root, "global options", &global->opt_list);
2231 array = json_create_array();
2232 json_object_add_value_array(root, "jobs", array);
2233 }
2234
2235 if (is_backend)
2236 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2237
2238 for (i = 0; i < nr_ts; i++) {
2239 ts = &threadstats[i];
2240 rs = &runstats[ts->groupid];
2241
2242 if (is_backend) {
2243 fio_server_send_job_options(opt_lists[i], i);
2244 fio_server_send_ts(ts, rs);
2245 } else {
2246 if (output_format & FIO_OUTPUT_TERSE)
2247 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2248 if (output_format & FIO_OUTPUT_JSON) {
2249 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2250 json_array_add_value_object(array, tmp);
2251 }
2252 if (output_format & FIO_OUTPUT_NORMAL)
2253 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2254 }
2255 }
2256 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2257 /* disk util stats, if any */
2258 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2259
2260 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2261
2262 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2263 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2264 json_free_object(root);
2265 }
2266
2267 for (i = 0; i < groupid + 1; i++) {
2268 rs = &runstats[i];
2269
2270 rs->groupid = i;
2271 if (is_backend)
2272 fio_server_send_gs(rs);
2273 else if (output_format & FIO_OUTPUT_NORMAL)
2274 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2275 }
2276
2277 if (is_backend)
2278 fio_server_send_du();
2279 else if (output_format & FIO_OUTPUT_NORMAL) {
2280 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2281 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2282 }
2283
2284 for (i = 0; i < FIO_OUTPUT_NR; i++) {
2285 struct buf_output *out = &output[i];
2286
2287 log_info_buf(out->buf, out->buflen);
2288 buf_output_free(out);
2289 }
2290
2291 fio_idle_prof_cleanup();
2292
2293 log_info_flush();
2294 free(runstats);
2295 free(threadstats);
2296 free(opt_lists);
2297}
2298
2299void __show_running_run_stats(void)
2300{
2301 struct thread_data *td;
2302 unsigned long long *rt;
2303 struct timespec ts;
2304 int i;
2305
2306 fio_sem_down(stat_sem);
2307
2308 rt = malloc(thread_number * sizeof(unsigned long long));
2309 fio_gettime(&ts, NULL);
2310
2311 for_each_td(td, i) {
2312 td->update_rusage = 1;
2313 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
2314 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
2315 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2316 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2317
2318 rt[i] = mtime_since(&td->start, &ts);
2319 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2320 td->ts.runtime[DDIR_READ] += rt[i];
2321 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2322 td->ts.runtime[DDIR_WRITE] += rt[i];
2323 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2324 td->ts.runtime[DDIR_TRIM] += rt[i];
2325 }
2326
2327 for_each_td(td, i) {
2328 if (td->runstate >= TD_EXITED)
2329 continue;
2330 if (td->rusage_sem) {
2331 td->update_rusage = 1;
2332 fio_sem_down(td->rusage_sem);
2333 }
2334 td->update_rusage = 0;
2335 }
2336
2337 __show_run_stats();
2338
2339 for_each_td(td, i) {
2340 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2341 td->ts.runtime[DDIR_READ] -= rt[i];
2342 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2343 td->ts.runtime[DDIR_WRITE] -= rt[i];
2344 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2345 td->ts.runtime[DDIR_TRIM] -= rt[i];
2346 }
2347
2348 free(rt);
2349 fio_sem_up(stat_sem);
2350}
2351
2352static bool status_interval_init;
2353static struct timespec status_time;
2354static bool status_file_disabled;
2355
2356#define FIO_STATUS_FILE "fio-dump-status"
2357
2358static int check_status_file(void)
2359{
2360 struct stat sb;
2361 const char *temp_dir;
2362 char fio_status_file_path[PATH_MAX];
2363
2364 if (status_file_disabled)
2365 return 0;
2366
2367 temp_dir = getenv("TMPDIR");
2368 if (temp_dir == NULL) {
2369 temp_dir = getenv("TEMP");
2370 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2371 temp_dir = NULL;
2372 }
2373 if (temp_dir == NULL)
2374 temp_dir = "/tmp";
2375#ifdef __COVERITY__
2376 __coverity_tainted_data_sanitize__(temp_dir);
2377#endif
2378
2379 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2380
2381 if (stat(fio_status_file_path, &sb))
2382 return 0;
2383
2384 if (unlink(fio_status_file_path) < 0) {
2385 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2386 strerror(errno));
2387 log_err("fio: disabling status file updates\n");
2388 status_file_disabled = true;
2389 }
2390
2391 return 1;
2392}
2393
2394void check_for_running_stats(void)
2395{
2396 if (status_interval) {
2397 if (!status_interval_init) {
2398 fio_gettime(&status_time, NULL);
2399 status_interval_init = true;
2400 } else if (mtime_since_now(&status_time) >= status_interval) {
2401 show_running_run_stats();
2402 fio_gettime(&status_time, NULL);
2403 return;
2404 }
2405 }
2406 if (check_status_file()) {
2407 show_running_run_stats();
2408 return;
2409 }
2410}
2411
2412static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2413{
2414 double val = data;
2415 double delta;
2416
2417 if (data > is->max_val)
2418 is->max_val = data;
2419 if (data < is->min_val)
2420 is->min_val = data;
2421
2422 delta = val - is->mean.u.f;
2423 if (delta) {
2424 is->mean.u.f += delta / (is->samples + 1.0);
2425 is->S.u.f += delta * (val - is->mean.u.f);
2426 }
2427
2428 is->samples++;
2429}
2430
2431/*
2432 * Return a struct io_logs, which is added to the tail of the log
2433 * list for 'iolog'.
2434 */
2435static struct io_logs *get_new_log(struct io_log *iolog)
2436{
2437 size_t new_size, new_samples;
2438 struct io_logs *cur_log;
2439
2440 /*
2441 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2442 * forever
2443 */
2444 if (!iolog->cur_log_max)
2445 new_samples = DEF_LOG_ENTRIES;
2446 else {
2447 new_samples = iolog->cur_log_max * 2;
2448 if (new_samples > MAX_LOG_ENTRIES)
2449 new_samples = MAX_LOG_ENTRIES;
2450 }
2451
2452 new_size = new_samples * log_entry_sz(iolog);
2453
2454 cur_log = smalloc(sizeof(*cur_log));
2455 if (cur_log) {
2456 INIT_FLIST_HEAD(&cur_log->list);
2457 cur_log->log = malloc(new_size);
2458 if (cur_log->log) {
2459 cur_log->nr_samples = 0;
2460 cur_log->max_samples = new_samples;
2461 flist_add_tail(&cur_log->list, &iolog->io_logs);
2462 iolog->cur_log_max = new_samples;
2463 return cur_log;
2464 }
2465 sfree(cur_log);
2466 }
2467
2468 return NULL;
2469}
2470
2471/*
2472 * Add and return a new log chunk, or return current log if big enough
2473 */
2474static struct io_logs *regrow_log(struct io_log *iolog)
2475{
2476 struct io_logs *cur_log;
2477 int i;
2478
2479 if (!iolog || iolog->disabled)
2480 goto disable;
2481
2482 cur_log = iolog_cur_log(iolog);
2483 if (!cur_log) {
2484 cur_log = get_new_log(iolog);
2485 if (!cur_log)
2486 return NULL;
2487 }
2488
2489 if (cur_log->nr_samples < cur_log->max_samples)
2490 return cur_log;
2491
2492 /*
2493 * No room for a new sample. If we're compressing on the fly, flush
2494 * out the current chunk
2495 */
2496 if (iolog->log_gz) {
2497 if (iolog_cur_flush(iolog, cur_log)) {
2498 log_err("fio: failed flushing iolog! Will stop logging.\n");
2499 return NULL;
2500 }
2501 }
2502
2503 /*
2504 * Get a new log array, and add to our list
2505 */
2506 cur_log = get_new_log(iolog);
2507 if (!cur_log) {
2508 log_err("fio: failed extending iolog! Will stop logging.\n");
2509 return NULL;
2510 }
2511
2512 if (!iolog->pending || !iolog->pending->nr_samples)
2513 return cur_log;
2514
2515 /*
2516 * Flush pending items to new log
2517 */
2518 for (i = 0; i < iolog->pending->nr_samples; i++) {
2519 struct io_sample *src, *dst;
2520
2521 src = get_sample(iolog, iolog->pending, i);
2522 dst = get_sample(iolog, cur_log, i);
2523 memcpy(dst, src, log_entry_sz(iolog));
2524 }
2525 cur_log->nr_samples = iolog->pending->nr_samples;
2526
2527 iolog->pending->nr_samples = 0;
2528 return cur_log;
2529disable:
2530 if (iolog)
2531 iolog->disabled = true;
2532 return NULL;
2533}
2534
2535void regrow_logs(struct thread_data *td)
2536{
2537 regrow_log(td->slat_log);
2538 regrow_log(td->clat_log);
2539 regrow_log(td->clat_hist_log);
2540 regrow_log(td->lat_log);
2541 regrow_log(td->bw_log);
2542 regrow_log(td->iops_log);
2543 td->flags &= ~TD_F_REGROW_LOGS;
2544}
2545
2546static struct io_logs *get_cur_log(struct io_log *iolog)
2547{
2548 struct io_logs *cur_log;
2549
2550 cur_log = iolog_cur_log(iolog);
2551 if (!cur_log) {
2552 cur_log = get_new_log(iolog);
2553 if (!cur_log)
2554 return NULL;
2555 }
2556
2557 if (cur_log->nr_samples < cur_log->max_samples)
2558 return cur_log;
2559
2560 /*
2561 * Out of space. If we're in IO offload mode, or we're not doing
2562 * per unit logging (hence logging happens outside of the IO thread
2563 * as well), add a new log chunk inline. If we're doing inline
2564 * submissions, flag 'td' as needing a log regrow and we'll take
2565 * care of it on the submission side.
2566 */
2567 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2568 !per_unit_log(iolog))
2569 return regrow_log(iolog);
2570
2571 if (iolog->td)
2572 iolog->td->flags |= TD_F_REGROW_LOGS;
2573 if (iolog->pending)
2574 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2575 return iolog->pending;
2576}
2577
2578static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2579 enum fio_ddir ddir, unsigned long long bs,
2580 unsigned long t, uint64_t offset, uint8_t priority_bit)
2581{
2582 struct io_logs *cur_log;
2583
2584 if (iolog->disabled)
2585 return;
2586 if (flist_empty(&iolog->io_logs))
2587 iolog->avg_last[ddir] = t;
2588
2589 cur_log = get_cur_log(iolog);
2590 if (cur_log) {
2591 struct io_sample *s;
2592
2593 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2594
2595 s->data = data;
2596 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2597 io_sample_set_ddir(iolog, s, ddir);
2598 s->bs = bs;
2599 s->priority_bit = priority_bit;
2600
2601 if (iolog->log_offset) {
2602 struct io_sample_offset *so = (void *) s;
2603
2604 so->offset = offset;
2605 }
2606
2607 cur_log->nr_samples++;
2608 return;
2609 }
2610
2611 iolog->disabled = true;
2612}
2613
2614static inline void reset_io_stat(struct io_stat *ios)
2615{
2616 ios->min_val = -1ULL;
2617 ios->max_val = ios->samples = 0;
2618 ios->mean.u.f = ios->S.u.f = 0;
2619}
2620
2621void reset_io_stats(struct thread_data *td)
2622{
2623 struct thread_stat *ts = &td->ts;
2624 int i, j, k;
2625
2626 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2627 reset_io_stat(&ts->clat_high_prio_stat[i]);
2628 reset_io_stat(&ts->clat_prio_stat[i]);
2629 reset_io_stat(&ts->clat_stat[i]);
2630 reset_io_stat(&ts->slat_stat[i]);
2631 reset_io_stat(&ts->lat_stat[i]);
2632 reset_io_stat(&ts->bw_stat[i]);
2633 reset_io_stat(&ts->iops_stat[i]);
2634
2635 ts->io_bytes[i] = 0;
2636 ts->runtime[i] = 0;
2637 ts->total_io_u[i] = 0;
2638 ts->short_io_u[i] = 0;
2639 ts->drop_io_u[i] = 0;
2640
2641 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2642 ts->io_u_plat_high_prio[i][j] = 0;
2643 ts->io_u_plat_prio[i][j] = 0;
2644 if (!i)
2645 ts->io_u_sync_plat[j] = 0;
2646 }
2647 }
2648
2649 for (i = 0; i < FIO_LAT_CNT; i++)
2650 for (j = 0; j < DDIR_RWDIR_CNT; j++)
2651 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2652 ts->io_u_plat[i][j][k] = 0;
2653
2654 ts->total_io_u[DDIR_SYNC] = 0;
2655
2656 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2657 ts->io_u_map[i] = 0;
2658 ts->io_u_submit[i] = 0;
2659 ts->io_u_complete[i] = 0;
2660 }
2661
2662 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2663 ts->io_u_lat_n[i] = 0;
2664 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2665 ts->io_u_lat_u[i] = 0;
2666 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2667 ts->io_u_lat_m[i] = 0;
2668
2669 ts->total_submit = 0;
2670 ts->total_complete = 0;
2671 ts->nr_zone_resets = 0;
2672 ts->cachehit = ts->cachemiss = 0;
2673}
2674
2675static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2676 unsigned long elapsed, bool log_max, uint8_t priority_bit)
2677{
2678 /*
2679 * Note an entry in the log. Use the mean from the logged samples,
2680 * making sure to properly round up. Only write a log entry if we
2681 * had actual samples done.
2682 */
2683 if (iolog->avg_window[ddir].samples) {
2684 union io_sample_data data;
2685
2686 if (log_max)
2687 data.val = iolog->avg_window[ddir].max_val;
2688 else
2689 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2690
2691 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, priority_bit);
2692 }
2693
2694 reset_io_stat(&iolog->avg_window[ddir]);
2695}
2696
2697static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2698 bool log_max, uint8_t priority_bit)
2699{
2700 int ddir;
2701
2702 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2703 __add_stat_to_log(iolog, ddir, elapsed, log_max, priority_bit);
2704}
2705
2706static unsigned long add_log_sample(struct thread_data *td,
2707 struct io_log *iolog,
2708 union io_sample_data data,
2709 enum fio_ddir ddir, unsigned long long bs,
2710 uint64_t offset, uint8_t priority_bit)
2711{
2712 unsigned long elapsed, this_window;
2713
2714 if (!ddir_rw(ddir))
2715 return 0;
2716
2717 elapsed = mtime_since_now(&td->epoch);
2718
2719 /*
2720 * If no time averaging, just add the log sample.
2721 */
2722 if (!iolog->avg_msec) {
2723 __add_log_sample(iolog, data, ddir, bs, elapsed, offset, priority_bit);
2724 return 0;
2725 }
2726
2727 /*
2728 * Add the sample. If the time period has passed, then
2729 * add that entry to the log and clear.
2730 */
2731 add_stat_sample(&iolog->avg_window[ddir], data.val);
2732
2733 /*
2734 * If period hasn't passed, adding the above sample is all we
2735 * need to do.
2736 */
2737 this_window = elapsed - iolog->avg_last[ddir];
2738 if (elapsed < iolog->avg_last[ddir])
2739 return iolog->avg_last[ddir] - elapsed;
2740 else if (this_window < iolog->avg_msec) {
2741 unsigned long diff = iolog->avg_msec - this_window;
2742
2743 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2744 return diff;
2745 }
2746
2747 _add_stat_to_log(iolog, elapsed, td->o.log_max != 0, priority_bit);
2748
2749 iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2750 return iolog->avg_msec;
2751}
2752
2753void finalize_logs(struct thread_data *td, bool unit_logs)
2754{
2755 unsigned long elapsed;
2756
2757 elapsed = mtime_since_now(&td->epoch);
2758
2759 if (td->clat_log && unit_logs)
2760 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0, 0);
2761 if (td->slat_log && unit_logs)
2762 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0, 0);
2763 if (td->lat_log && unit_logs)
2764 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0, 0);
2765 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2766 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0, 0);
2767 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2768 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0, 0);
2769}
2770
2771void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned long long bs,
2772 uint8_t priority_bit)
2773{
2774 struct io_log *iolog;
2775
2776 if (!ddir_rw(ddir))
2777 return;
2778
2779 iolog = agg_io_log[ddir];
2780 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, priority_bit);
2781}
2782
2783void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2784{
2785 unsigned int idx = plat_val_to_idx(nsec);
2786 assert(idx < FIO_IO_U_PLAT_NR);
2787
2788 ts->io_u_sync_plat[idx]++;
2789 add_stat_sample(&ts->sync_stat, nsec);
2790}
2791
2792static void add_lat_percentile_sample_noprio(struct thread_stat *ts,
2793 unsigned long long nsec, enum fio_ddir ddir, enum fio_lat lat)
2794{
2795 unsigned int idx = plat_val_to_idx(nsec);
2796 assert(idx < FIO_IO_U_PLAT_NR);
2797
2798 ts->io_u_plat[lat][ddir][idx]++;
2799}
2800
2801static void add_lat_percentile_sample(struct thread_stat *ts,
2802 unsigned long long nsec, enum fio_ddir ddir, uint8_t priority_bit,
2803 enum fio_lat lat)
2804{
2805 unsigned int idx = plat_val_to_idx(nsec);
2806
2807 add_lat_percentile_sample_noprio(ts, nsec, ddir, lat);
2808
2809 if (!priority_bit)
2810 ts->io_u_plat_prio[ddir][idx]++;
2811 else
2812 ts->io_u_plat_high_prio[ddir][idx]++;
2813}
2814
2815void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2816 unsigned long long nsec, unsigned long long bs,
2817 uint64_t offset, uint8_t priority_bit)
2818{
2819 const bool needs_lock = td_async_processing(td);
2820 unsigned long elapsed, this_window;
2821 struct thread_stat *ts = &td->ts;
2822 struct io_log *iolog = td->clat_hist_log;
2823
2824 if (needs_lock)
2825 __td_io_u_lock(td);
2826
2827 add_stat_sample(&ts->clat_stat[ddir], nsec);
2828
2829 if (priority_bit) {
2830 add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
2831 } else {
2832 add_stat_sample(&ts->clat_prio_stat[ddir], nsec);
2833 }
2834
2835 if (td->clat_log)
2836 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2837 offset, priority_bit);
2838
2839 if (ts->clat_percentiles) {
2840 if (ts->lat_percentiles)
2841 add_lat_percentile_sample_noprio(ts, nsec, ddir, FIO_CLAT);
2842 else
2843 add_lat_percentile_sample(ts, nsec, ddir, priority_bit, FIO_CLAT);
2844 }
2845
2846 if (iolog && iolog->hist_msec) {
2847 struct io_hist *hw = &iolog->hist_window[ddir];
2848
2849 hw->samples++;
2850 elapsed = mtime_since_now(&td->epoch);
2851 if (!hw->hist_last)
2852 hw->hist_last = elapsed;
2853 this_window = elapsed - hw->hist_last;
2854
2855 if (this_window >= iolog->hist_msec) {
2856 uint64_t *io_u_plat;
2857 struct io_u_plat_entry *dst;
2858
2859 /*
2860 * Make a byte-for-byte copy of the latency histogram
2861 * stored in td->ts.io_u_plat[ddir], recording it in a
2862 * log sample. Note that the matching call to free() is
2863 * located in iolog.c after printing this sample to the
2864 * log file.
2865 */
2866 io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
2867 dst = malloc(sizeof(struct io_u_plat_entry));
2868 memcpy(&(dst->io_u_plat), io_u_plat,
2869 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
2870 flist_add(&dst->list, &hw->list);
2871 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2872 elapsed, offset, priority_bit);
2873
2874 /*
2875 * Update the last time we recorded as being now, minus
2876 * any drift in time we encountered before actually
2877 * making the record.
2878 */
2879 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2880 hw->samples = 0;
2881 }
2882 }
2883
2884 if (needs_lock)
2885 __td_io_u_unlock(td);
2886}
2887
2888void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2889 unsigned long long nsec, unsigned long long bs, uint64_t offset,
2890 uint8_t priority_bit)
2891{
2892 const bool needs_lock = td_async_processing(td);
2893 struct thread_stat *ts = &td->ts;
2894
2895 if (!ddir_rw(ddir))
2896 return;
2897
2898 if (needs_lock)
2899 __td_io_u_lock(td);
2900
2901 add_stat_sample(&ts->slat_stat[ddir], nsec);
2902
2903 if (td->slat_log)
2904 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs, offset,
2905 priority_bit);
2906
2907 if (ts->slat_percentiles)
2908 add_lat_percentile_sample_noprio(ts, nsec, ddir, FIO_SLAT);
2909
2910 if (needs_lock)
2911 __td_io_u_unlock(td);
2912}
2913
2914void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2915 unsigned long long nsec, unsigned long long bs,
2916 uint64_t offset, uint8_t priority_bit)
2917{
2918 const bool needs_lock = td_async_processing(td);
2919 struct thread_stat *ts = &td->ts;
2920
2921 if (!ddir_rw(ddir))
2922 return;
2923
2924 if (needs_lock)
2925 __td_io_u_lock(td);
2926
2927 add_stat_sample(&ts->lat_stat[ddir], nsec);
2928
2929 if (td->lat_log)
2930 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2931 offset, priority_bit);
2932
2933 if (ts->lat_percentiles)
2934 add_lat_percentile_sample(ts, nsec, ddir, priority_bit, FIO_LAT);
2935
2936 if (needs_lock)
2937 __td_io_u_unlock(td);
2938}
2939
2940void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2941 unsigned int bytes, unsigned long long spent)
2942{
2943 const bool needs_lock = td_async_processing(td);
2944 struct thread_stat *ts = &td->ts;
2945 unsigned long rate;
2946
2947 if (spent)
2948 rate = (unsigned long) (bytes * 1000000ULL / spent);
2949 else
2950 rate = 0;
2951
2952 if (needs_lock)
2953 __td_io_u_lock(td);
2954
2955 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2956
2957 if (td->bw_log)
2958 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2959 bytes, io_u->offset, io_u_is_prio(io_u));
2960
2961 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2962
2963 if (needs_lock)
2964 __td_io_u_unlock(td);
2965}
2966
2967static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2968 struct timespec *t, unsigned int avg_time,
2969 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2970 struct io_stat *stat, struct io_log *log,
2971 bool is_kb)
2972{
2973 const bool needs_lock = td_async_processing(td);
2974 unsigned long spent, rate;
2975 enum fio_ddir ddir;
2976 unsigned long next, next_log;
2977
2978 next_log = avg_time;
2979
2980 spent = mtime_since(parent_tv, t);
2981 if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2982 return avg_time - spent;
2983
2984 if (needs_lock)
2985 __td_io_u_lock(td);
2986
2987 /*
2988 * Compute both read and write rates for the interval.
2989 */
2990 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2991 uint64_t delta;
2992
2993 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2994 if (!delta)
2995 continue; /* No entries for interval */
2996
2997 if (spent) {
2998 if (is_kb)
2999 rate = delta * 1000 / spent / 1024; /* KiB/s */
3000 else
3001 rate = (delta * 1000) / spent;
3002 } else
3003 rate = 0;
3004
3005 add_stat_sample(&stat[ddir], rate);
3006
3007 if (log) {
3008 unsigned long long bs = 0;
3009
3010 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3011 bs = td->o.min_bs[ddir];
3012
3013 next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0, 0);
3014 next_log = min(next_log, next);
3015 }
3016
3017 stat_io_bytes[ddir] = this_io_bytes[ddir];
3018 }
3019
3020 *parent_tv = *t;
3021
3022 if (needs_lock)
3023 __td_io_u_unlock(td);
3024
3025 if (spent <= avg_time)
3026 next = avg_time;
3027 else
3028 next = avg_time - (1 + spent - avg_time);
3029
3030 return min(next, next_log);
3031}
3032
3033static int add_bw_samples(struct thread_data *td, struct timespec *t)
3034{
3035 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3036 td->this_io_bytes, td->stat_io_bytes,
3037 td->ts.bw_stat, td->bw_log, true);
3038}
3039
3040void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3041 unsigned int bytes)
3042{
3043 const bool needs_lock = td_async_processing(td);
3044 struct thread_stat *ts = &td->ts;
3045
3046 if (needs_lock)
3047 __td_io_u_lock(td);
3048
3049 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3050
3051 if (td->iops_log)
3052 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3053 bytes, io_u->offset, io_u_is_prio(io_u));
3054
3055 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3056
3057 if (needs_lock)
3058 __td_io_u_unlock(td);
3059}
3060
3061static int add_iops_samples(struct thread_data *td, struct timespec *t)
3062{
3063 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3064 td->this_io_blocks, td->stat_io_blocks,
3065 td->ts.iops_stat, td->iops_log, false);
3066}
3067
3068/*
3069 * Returns msecs to next event
3070 */
3071int calc_log_samples(void)
3072{
3073 struct thread_data *td;
3074 unsigned int next = ~0U, tmp;
3075 struct timespec now;
3076 int i;
3077
3078 fio_gettime(&now, NULL);
3079
3080 for_each_td(td, i) {
3081 if (!td->o.stats)
3082 continue;
3083 if (in_ramp_time(td) ||
3084 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3085 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3086 continue;
3087 }
3088 if (!td->bw_log ||
3089 (td->bw_log && !per_unit_log(td->bw_log))) {
3090 tmp = add_bw_samples(td, &now);
3091 if (tmp < next)
3092 next = tmp;
3093 }
3094 if (!td->iops_log ||
3095 (td->iops_log && !per_unit_log(td->iops_log))) {
3096 tmp = add_iops_samples(td, &now);
3097 if (tmp < next)
3098 next = tmp;
3099 }
3100 }
3101
3102 return next == ~0U ? 0 : next;
3103}
3104
3105void stat_init(void)
3106{
3107 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3108}
3109
3110void stat_exit(void)
3111{
3112 /*
3113 * When we have the mutex, we know out-of-band access to it
3114 * have ended.
3115 */
3116 fio_sem_down(stat_sem);
3117 fio_sem_remove(stat_sem);
3118}
3119
3120/*
3121 * Called from signal handler. Wake up status thread.
3122 */
3123void show_running_run_stats(void)
3124{
3125 helper_do_stat();
3126}
3127
3128uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3129{
3130 /* Ignore io_u's which span multiple blocks--they will just get
3131 * inaccurate counts. */
3132 int idx = (io_u->offset - io_u->file->file_offset)
3133 / td->o.bs[DDIR_TRIM];
3134 uint32_t *info = &td->ts.block_infos[idx];
3135 assert(idx < td->ts.nr_block_infos);
3136 return info;
3137}
fio - Flexible IO Tester