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