projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
client: switch to per-client buffer
[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
19#define LOG_MSEC_SLACK 1
20
21struct fio_sem *stat_sem;
22
23void clear_rusage_stat(struct thread_data *td)
24{
25 struct thread_stat *ts = &td->ts;
26
27 fio_getrusage(&td->ru_start);
28 ts->usr_time = ts->sys_time = 0;
29 ts->ctx = 0;
30 ts->minf = ts->majf = 0;
31}
32
33void update_rusage_stat(struct thread_data *td)
34{
35 struct thread_stat *ts = &td->ts;
36
37 fio_getrusage(&td->ru_end);
38 ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
39 &td->ru_end.ru_utime);
40 ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
41 &td->ru_end.ru_stime);
42 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
43 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
44 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
45 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
46
47 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
48}
49
50/*
51 * Given a latency, return the index of the corresponding bucket in
52 * the structure tracking percentiles.
53 *
54 * (1) find the group (and error bits) that the value (latency)
55 * belongs to by looking at its MSB. (2) find the bucket number in the
56 * group by looking at the index bits.
57 *
58 */
59static unsigned int plat_val_to_idx(unsigned long long val)
60{
61 unsigned int msb, error_bits, base, offset, idx;
62
63 /* Find MSB starting from bit 0 */
64 if (val == 0)
65 msb = 0;
66 else
67 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
68
69 /*
70 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
71 * all bits of the sample as index
72 */
73 if (msb <= FIO_IO_U_PLAT_BITS)
74 return val;
75
76 /* Compute the number of error bits to discard*/
77 error_bits = msb - FIO_IO_U_PLAT_BITS;
78
79 /* Compute the number of buckets before the group */
80 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
81
82 /*
83 * Discard the error bits and apply the mask to find the
84 * index for the buckets in the group
85 */
86 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
87
88 /* Make sure the index does not exceed (array size - 1) */
89 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
90 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
91
92 return idx;
93}
94
95/*
96 * Convert the given index of the bucket array to the value
97 * represented by the bucket
98 */
99static unsigned long long plat_idx_to_val(unsigned int idx)
100{
101 unsigned int error_bits;
102 unsigned long long k, base;
103
104 assert(idx < FIO_IO_U_PLAT_NR);
105
106 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
107 * all bits of the sample as index */
108 if (idx < (FIO_IO_U_PLAT_VAL << 1))
109 return idx;
110
111 /* Find the group and compute the minimum value of that group */
112 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
113 base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
114
115 /* Find its bucket number of the group */
116 k = idx % FIO_IO_U_PLAT_VAL;
117
118 /* Return the mean of the range of the bucket */
119 return base + ((k + 0.5) * (1 << error_bits));
120}
121
122static int double_cmp(const void *a, const void *b)
123{
124 const fio_fp64_t fa = *(const fio_fp64_t *) a;
125 const fio_fp64_t fb = *(const fio_fp64_t *) b;
126 int cmp = 0;
127
128 if (fa.u.f > fb.u.f)
129 cmp = 1;
130 else if (fa.u.f < fb.u.f)
131 cmp = -1;
132
133 return cmp;
134}
135
136unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
137 fio_fp64_t *plist, unsigned long long **output,
138 unsigned long long *maxv, unsigned long long *minv)
139{
140 unsigned long long sum = 0;
141 unsigned int len, i, j = 0;
142 unsigned int oval_len = 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((void *)plist, len, sizeof(plist[0]), double_cmp);
163
164 /*
165 * Calculate bucket values, note down max and min values
166 */
167 is_last = false;
168 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
169 sum += io_u_plat[i];
170 while (sum >= (plist[j].u.f / 100.0 * nr)) {
171 assert(plist[j].u.f <= 100.0);
172
173 if (j == oval_len) {
174 oval_len += 100;
175 ovals = realloc(ovals, oval_len * sizeof(*ovals));
176 }
177
178 ovals[j] = plat_idx_to_val(i);
179 if (ovals[j] < *minv)
180 *minv = ovals[j];
181 if (ovals[j] > *maxv)
182 *maxv = ovals[j];
183
184 is_last = (j == len - 1) != 0;
185 if (is_last)
186 break;
187
188 j++;
189 }
190 }
191
192 *output = ovals;
193 return len;
194}
195
196/*
197 * Find and display the p-th percentile of clat
198 */
199static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
200 fio_fp64_t *plist, unsigned int precision,
201 const char *pre, struct buf_output *out)
202{
203 unsigned int divisor, len, i, j = 0;
204 unsigned long long minv, maxv;
205 unsigned long long *ovals;
206 int per_line, scale_down, time_width;
207 bool is_last;
208 char fmt[32];
209
210 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
211 if (!len || !ovals)
212 goto out;
213
214 /*
215 * We default to nsecs, but if the value range is such that we
216 * should scale down to usecs or msecs, do that.
217 */
218 if (minv > 2000000 && maxv > 99999999ULL) {
219 scale_down = 2;
220 divisor = 1000000;
221 log_buf(out, " %s percentiles (msec):\n |", pre);
222 } else if (minv > 2000 && maxv > 99999) {
223 scale_down = 1;
224 divisor = 1000;
225 log_buf(out, " %s percentiles (usec):\n |", pre);
226 } else {
227 scale_down = 0;
228 divisor = 1;
229 log_buf(out, " %s percentiles (nsec):\n |", pre);
230 }
231
232
233 time_width = max(5, (int) (log10(maxv / divisor) + 1));
234 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
235 precision, time_width);
236 /* fmt will be something like " %5.2fth=[%4llu]%c" */
237 per_line = (80 - 7) / (precision + 10 + time_width);
238
239 for (j = 0; j < len; j++) {
240 /* for formatting */
241 if (j != 0 && (j % per_line) == 0)
242 log_buf(out, " |");
243
244 /* end of the list */
245 is_last = (j == len - 1) != 0;
246
247 for (i = 0; i < scale_down; i++)
248 ovals[j] = (ovals[j] + 999) / 1000;
249
250 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
251
252 if (is_last)
253 break;
254
255 if ((j % per_line) == per_line - 1) /* for formatting */
256 log_buf(out, "\n");
257 }
258
259out:
260 if (ovals)
261 free(ovals);
262}
263
264bool calc_lat(struct io_stat *is, unsigned long long *min,
265 unsigned long long *max, double *mean, double *dev)
266{
267 double n = (double) is->samples;
268
269 if (n == 0)
270 return false;
271
272 *min = is->min_val;
273 *max = is->max_val;
274 *mean = is->mean.u.f;
275
276 if (n > 1.0)
277 *dev = sqrt(is->S.u.f / (n - 1.0));
278 else
279 *dev = 0;
280
281 return true;
282}
283
284void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
285{
286 char *io, *agg, *min, *max;
287 char *ioalt, *aggalt, *minalt, *maxalt;
288 const char *str[] = { " READ", " WRITE" , " TRIM"};
289 int i;
290
291 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
292
293 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
294 const int i2p = is_power_of_2(rs->kb_base);
295
296 if (!rs->max_run[i])
297 continue;
298
299 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
300 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
301 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
302 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
303 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
304 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
305 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
306 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
307 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
308 rs->unified_rw_rep ? " MIXED" : str[i],
309 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
310 (unsigned long long) rs->min_run[i],
311 (unsigned long long) rs->max_run[i]);
312
313 free(io);
314 free(agg);
315 free(min);
316 free(max);
317 free(ioalt);
318 free(aggalt);
319 free(minalt);
320 free(maxalt);
321 }
322}
323
324void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
325{
326 int i;
327
328 /*
329 * Do depth distribution calculations
330 */
331 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
332 if (total) {
333 io_u_dist[i] = (double) map[i] / (double) total;
334 io_u_dist[i] *= 100.0;
335 if (io_u_dist[i] < 0.1 && map[i])
336 io_u_dist[i] = 0.1;
337 } else
338 io_u_dist[i] = 0.0;
339 }
340}
341
342static void stat_calc_lat(struct thread_stat *ts, double *dst,
343 uint64_t *src, int nr)
344{
345 unsigned long total = ddir_rw_sum(ts->total_io_u);
346 int i;
347
348 /*
349 * Do latency distribution calculations
350 */
351 for (i = 0; i < nr; i++) {
352 if (total) {
353 dst[i] = (double) src[i] / (double) total;
354 dst[i] *= 100.0;
355 if (dst[i] < 0.01 && src[i])
356 dst[i] = 0.01;
357 } else
358 dst[i] = 0.0;
359 }
360}
361
362/*
363 * To keep the terse format unaltered, add all of the ns latency
364 * buckets to the first us latency bucket
365 */
366static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
367{
368 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
369 int i;
370
371 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
372
373 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
374 ntotal += ts->io_u_lat_n[i];
375
376 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
377}
378
379void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
380{
381 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
382}
383
384void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
385{
386 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
387}
388
389void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
390{
391 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
392}
393
394static void display_lat(const char *name, unsigned long long min,
395 unsigned long long max, double mean, double dev,
396 struct buf_output *out)
397{
398 const char *base = "(nsec)";
399 char *minp, *maxp;
400
401 if (nsec_to_msec(&min, &max, &mean, &dev))
402 base = "(msec)";
403 else if (nsec_to_usec(&min, &max, &mean, &dev))
404 base = "(usec)";
405
406 minp = num2str(min, 6, 1, 0, N2S_NONE);
407 maxp = num2str(max, 6, 1, 0, N2S_NONE);
408
409 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
410 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
411
412 free(minp);
413 free(maxp);
414}
415
416static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
417 int ddir, struct buf_output *out)
418{
419 const char *str[] = { " read", "write", " trim", "sync" };
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, *zbd_w_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(str[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 str[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 zbd_w_st = zbd_write_status(ts);
456
457 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
458 rs->unified_rw_rep ? "mixed" : str[ddir],
459 iops_p, bw_p, bw_p_alt, io_p,
460 (unsigned long long) ts->runtime[ddir],
461 zbd_w_st ? : "");
462
463 free(zbd_w_st);
464 free(io_p);
465 free(bw_p);
466 free(bw_p_alt);
467 free(iops_p);
468
469 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
470 display_lat("slat", min, max, mean, dev, out);
471 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
472 display_lat("clat", min, max, mean, dev, out);
473 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
474 display_lat(" lat", min, max, mean, dev, out);
475
476 if (ts->clat_percentiles || ts->lat_percentiles) {
477 const char *name = ts->clat_percentiles ? "clat" : " lat";
478 uint64_t samples;
479
480 if (ts->clat_percentiles)
481 samples = ts->clat_stat[ddir].samples;
482 else
483 samples = ts->lat_stat[ddir].samples;
484
485 show_clat_percentiles(ts->io_u_plat[ddir],
486 samples,
487 ts->percentile_list,
488 ts->percentile_precision, name, out);
489 }
490 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
491 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
492 const char *bw_str;
493
494 if ((rs->unit_base == 1) && i2p)
495 bw_str = "Kibit";
496 else if (rs->unit_base == 1)
497 bw_str = "kbit";
498 else if (i2p)
499 bw_str = "KiB";
500 else
501 bw_str = "kB";
502
503 if (rs->agg[ddir]) {
504 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
505 if (p_of_agg > 100.0)
506 p_of_agg = 100.0;
507 }
508
509 if (rs->unit_base == 1) {
510 min *= 8.0;
511 max *= 8.0;
512 mean *= 8.0;
513 dev *= 8.0;
514 }
515
516 if (mean > fkb_base * fkb_base) {
517 min /= fkb_base;
518 max /= fkb_base;
519 mean /= fkb_base;
520 dev /= fkb_base;
521 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
522 }
523
524 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
525 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
526 bw_str, min, max, p_of_agg, mean, dev,
527 (&ts->bw_stat[ddir])->samples);
528 }
529 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
530 log_buf(out, " iops : min=%5llu, max=%5llu, "
531 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
532 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
533 }
534}
535
536static bool show_lat(double *io_u_lat, int nr, const char **ranges,
537 const char *msg, struct buf_output *out)
538{
539 bool new_line = true, shown = false;
540 int i, line = 0;
541
542 for (i = 0; i < nr; i++) {
543 if (io_u_lat[i] <= 0.0)
544 continue;
545 shown = true;
546 if (new_line) {
547 if (line)
548 log_buf(out, "\n");
549 log_buf(out, " lat (%s) : ", msg);
550 new_line = false;
551 line = 0;
552 }
553 if (line)
554 log_buf(out, ", ");
555 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
556 line++;
557 if (line == 5)
558 new_line = true;
559 }
560
561 if (shown)
562 log_buf(out, "\n");
563
564 return true;
565}
566
567static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
568{
569 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
570 "250=", "500=", "750=", "1000=", };
571
572 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
573}
574
575static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
576{
577 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
578 "250=", "500=", "750=", "1000=", };
579
580 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
581}
582
583static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
584{
585 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
586 "250=", "500=", "750=", "1000=", "2000=",
587 ">=2000=", };
588
589 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
590}
591
592static void show_latencies(struct thread_stat *ts, struct buf_output *out)
593{
594 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
595 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
596 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
597
598 stat_calc_lat_n(ts, io_u_lat_n);
599 stat_calc_lat_u(ts, io_u_lat_u);
600 stat_calc_lat_m(ts, io_u_lat_m);
601
602 show_lat_n(io_u_lat_n, out);
603 show_lat_u(io_u_lat_u, out);
604 show_lat_m(io_u_lat_m, out);
605}
606
607static int block_state_category(int block_state)
608{
609 switch (block_state) {
610 case BLOCK_STATE_UNINIT:
611 return 0;
612 case BLOCK_STATE_TRIMMED:
613 case BLOCK_STATE_WRITTEN:
614 return 1;
615 case BLOCK_STATE_WRITE_FAILURE:
616 case BLOCK_STATE_TRIM_FAILURE:
617 return 2;
618 default:
619 /* Silence compile warning on some BSDs and have a return */
620 assert(0);
621 return -1;
622 }
623}
624
625static int compare_block_infos(const void *bs1, const void *bs2)
626{
627 uint64_t block1 = *(uint64_t *)bs1;
628 uint64_t block2 = *(uint64_t *)bs2;
629 int state1 = BLOCK_INFO_STATE(block1);
630 int state2 = BLOCK_INFO_STATE(block2);
631 int bscat1 = block_state_category(state1);
632 int bscat2 = block_state_category(state2);
633 int cycles1 = BLOCK_INFO_TRIMS(block1);
634 int cycles2 = BLOCK_INFO_TRIMS(block2);
635
636 if (bscat1 < bscat2)
637 return -1;
638 if (bscat1 > bscat2)
639 return 1;
640
641 if (cycles1 < cycles2)
642 return -1;
643 if (cycles1 > cycles2)
644 return 1;
645
646 if (state1 < state2)
647 return -1;
648 if (state1 > state2)
649 return 1;
650
651 assert(block1 == block2);
652 return 0;
653}
654
655static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
656 fio_fp64_t *plist, unsigned int **percentiles,
657 unsigned int *types)
658{
659 int len = 0;
660 int i, nr_uninit;
661
662 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
663
664 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
665 len++;
666
667 if (!len)
668 return 0;
669
670 /*
671 * Sort the percentile list. Note that it may already be sorted if
672 * we are using the default values, but since it's a short list this
673 * isn't a worry. Also note that this does not work for NaN values.
674 */
675 if (len > 1)
676 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
677
678 /* Start only after the uninit entries end */
679 for (nr_uninit = 0;
680 nr_uninit < nr_block_infos
681 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
682 nr_uninit ++)
683 ;
684
685 if (nr_uninit == nr_block_infos)
686 return 0;
687
688 *percentiles = calloc(len, sizeof(**percentiles));
689
690 for (i = 0; i < len; i++) {
691 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
692 + nr_uninit;
693 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
694 }
695
696 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
697 for (i = 0; i < nr_block_infos; i++)
698 types[BLOCK_INFO_STATE(block_infos[i])]++;
699
700 return len;
701}
702
703static const char *block_state_names[] = {
704 [BLOCK_STATE_UNINIT] = "unwritten",
705 [BLOCK_STATE_TRIMMED] = "trimmed",
706 [BLOCK_STATE_WRITTEN] = "written",
707 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
708 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
709};
710
711static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
712 fio_fp64_t *plist, struct buf_output *out)
713{
714 int len, pos, i;
715 unsigned int *percentiles = NULL;
716 unsigned int block_state_counts[BLOCK_STATE_COUNT];
717
718 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
719 &percentiles, block_state_counts);
720
721 log_buf(out, " block lifetime percentiles :\n |");
722 pos = 0;
723 for (i = 0; i < len; i++) {
724 uint32_t block_info = percentiles[i];
725#define LINE_LENGTH 75
726 char str[LINE_LENGTH];
727 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
728 plist[i].u.f, block_info,
729 i == len - 1 ? '\n' : ',');
730 assert(strln < LINE_LENGTH);
731 if (pos + strln > LINE_LENGTH) {
732 pos = 0;
733 log_buf(out, "\n |");
734 }
735 log_buf(out, "%s", str);
736 pos += strln;
737#undef LINE_LENGTH
738 }
739 if (percentiles)
740 free(percentiles);
741
742 log_buf(out, " states :");
743 for (i = 0; i < BLOCK_STATE_COUNT; i++)
744 log_buf(out, " %s=%u%c",
745 block_state_names[i], block_state_counts[i],
746 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
747}
748
749static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
750{
751 char *p1, *p1alt, *p2;
752 unsigned long long bw_mean, iops_mean;
753 const int i2p = is_power_of_2(ts->kb_base);
754
755 if (!ts->ss_dur)
756 return;
757
758 bw_mean = steadystate_bw_mean(ts);
759 iops_mean = steadystate_iops_mean(ts);
760
761 p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
762 p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
763 p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
764
765 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
766 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
767 p1, p1alt, p2,
768 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
769 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
770 ts->ss_criterion.u.f,
771 ts->ss_state & FIO_SS_PCT ? "%" : "");
772
773 free(p1);
774 free(p1alt);
775 free(p2);
776}
777
778static void show_thread_status_normal(struct thread_stat *ts,
779 struct group_run_stats *rs,
780 struct buf_output *out)
781{
782 double usr_cpu, sys_cpu;
783 unsigned long runtime;
784 double io_u_dist[FIO_IO_U_MAP_NR];
785 time_t time_p;
786 char time_buf[32];
787
788 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
789 return;
790
791 memset(time_buf, 0, sizeof(time_buf));
792
793 time(&time_p);
794 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
795
796 if (!ts->error) {
797 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
798 ts->name, ts->groupid, ts->members,
799 ts->error, (int) ts->pid, time_buf);
800 } else {
801 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
802 ts->name, ts->groupid, ts->members,
803 ts->error, ts->verror, (int) ts->pid,
804 time_buf);
805 }
806
807 if (strlen(ts->description))
808 log_buf(out, " Description : [%s]\n", ts->description);
809
810 if (ts->io_bytes[DDIR_READ])
811 show_ddir_status(rs, ts, DDIR_READ, out);
812 if (ts->io_bytes[DDIR_WRITE])
813 show_ddir_status(rs, ts, DDIR_WRITE, out);
814 if (ts->io_bytes[DDIR_TRIM])
815 show_ddir_status(rs, ts, DDIR_TRIM, out);
816
817 show_latencies(ts, out);
818
819 if (ts->sync_stat.samples)
820 show_ddir_status(rs, ts, DDIR_SYNC, out);
821
822 runtime = ts->total_run_time;
823 if (runtime) {
824 double runt = (double) runtime;
825
826 usr_cpu = (double) ts->usr_time * 100 / runt;
827 sys_cpu = (double) ts->sys_time * 100 / runt;
828 } else {
829 usr_cpu = 0;
830 sys_cpu = 0;
831 }
832
833 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
834 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
835 (unsigned long long) ts->ctx,
836 (unsigned long long) ts->majf,
837 (unsigned long long) ts->minf);
838
839 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
840 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
841 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
842 io_u_dist[1], io_u_dist[2],
843 io_u_dist[3], io_u_dist[4],
844 io_u_dist[5], io_u_dist[6]);
845
846 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
847 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
848 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
849 io_u_dist[1], io_u_dist[2],
850 io_u_dist[3], io_u_dist[4],
851 io_u_dist[5], io_u_dist[6]);
852 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
853 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
854 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
855 io_u_dist[1], io_u_dist[2],
856 io_u_dist[3], io_u_dist[4],
857 io_u_dist[5], io_u_dist[6]);
858 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
859 " short=%llu,%llu,%llu,0"
860 " dropped=%llu,%llu,%llu,0\n",
861 (unsigned long long) ts->total_io_u[0],
862 (unsigned long long) ts->total_io_u[1],
863 (unsigned long long) ts->total_io_u[2],
864 (unsigned long long) ts->total_io_u[3],
865 (unsigned long long) ts->short_io_u[0],
866 (unsigned long long) ts->short_io_u[1],
867 (unsigned long long) ts->short_io_u[2],
868 (unsigned long long) ts->drop_io_u[0],
869 (unsigned long long) ts->drop_io_u[1],
870 (unsigned long long) ts->drop_io_u[2]);
871 if (ts->continue_on_error) {
872 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
873 (unsigned long long)ts->total_err_count,
874 ts->first_error,
875 strerror(ts->first_error));
876 }
877 if (ts->latency_depth) {
878 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
879 (unsigned long long)ts->latency_target,
880 (unsigned long long)ts->latency_window,
881 ts->latency_percentile.u.f,
882 ts->latency_depth);
883 }
884
885 if (ts->nr_block_infos)
886 show_block_infos(ts->nr_block_infos, ts->block_infos,
887 ts->percentile_list, out);
888
889 if (ts->ss_dur)
890 show_ss_normal(ts, out);
891}
892
893static void show_ddir_status_terse(struct thread_stat *ts,
894 struct group_run_stats *rs, int ddir,
895 int ver, struct buf_output *out)
896{
897 unsigned long long min, max, minv, maxv, bw, iops;
898 unsigned long long *ovals = NULL;
899 double mean, dev;
900 unsigned int len;
901 int i, bw_stat;
902
903 assert(ddir_rw(ddir));
904
905 iops = bw = 0;
906 if (ts->runtime[ddir]) {
907 uint64_t runt = ts->runtime[ddir];
908
909 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
910 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
911 }
912
913 log_buf(out, ";%llu;%llu;%llu;%llu",
914 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
915 (unsigned long long) ts->runtime[ddir]);
916
917 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
918 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
919 else
920 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
921
922 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
923 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
924 else
925 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
926
927 if (ts->clat_percentiles || ts->lat_percentiles) {
928 len = calc_clat_percentiles(ts->io_u_plat[ddir],
929 ts->clat_stat[ddir].samples,
930 ts->percentile_list, &ovals, &maxv,
931 &minv);
932 } else
933 len = 0;
934
935 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
936 if (i >= len) {
937 log_buf(out, ";0%%=0");
938 continue;
939 }
940 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
941 }
942
943 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
944 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
945 else
946 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
947
948 if (ovals)
949 free(ovals);
950
951 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
952 if (bw_stat) {
953 double p_of_agg = 100.0;
954
955 if (rs->agg[ddir]) {
956 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
957 if (p_of_agg > 100.0)
958 p_of_agg = 100.0;
959 }
960
961 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
962 } else
963 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
964
965 if (ver == 5) {
966 if (bw_stat)
967 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
968 else
969 log_buf(out, ";%lu", 0UL);
970
971 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
972 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
973 mean, dev, (&ts->iops_stat[ddir])->samples);
974 else
975 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
976 }
977}
978
979static void add_ddir_status_json(struct thread_stat *ts,
980 struct group_run_stats *rs, int ddir, struct json_object *parent)
981{
982 unsigned long long min, max, minv, maxv;
983 unsigned long long bw_bytes, bw;
984 unsigned long long *ovals = NULL;
985 double mean, dev, iops;
986 unsigned int len;
987 int i;
988 const char *ddirname[] = { "read", "write", "trim", "sync" };
989 struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object = NULL;
990 char buf[120];
991 double p_of_agg = 100.0;
992
993 assert(ddir_rw(ddir) || ddir_sync(ddir));
994
995 if (ts->unified_rw_rep && ddir != DDIR_READ)
996 return;
997
998 dir_object = json_create_object();
999 json_object_add_value_object(parent,
1000 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
1001
1002 if (ddir_rw(ddir)) {
1003 bw_bytes = 0;
1004 bw = 0;
1005 iops = 0.0;
1006 if (ts->runtime[ddir]) {
1007 uint64_t runt = ts->runtime[ddir];
1008
1009 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1010 bw = bw_bytes / 1024; /* KiB/s */
1011 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1012 }
1013
1014 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1015 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1016 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1017 json_object_add_value_int(dir_object, "bw", bw);
1018 json_object_add_value_float(dir_object, "iops", iops);
1019 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1020 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1021 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1022 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1023
1024 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
1025 min = max = 0;
1026 mean = dev = 0.0;
1027 }
1028 tmp_object = json_create_object();
1029 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1030 json_object_add_value_int(tmp_object, "min", min);
1031 json_object_add_value_int(tmp_object, "max", max);
1032 json_object_add_value_float(tmp_object, "mean", mean);
1033 json_object_add_value_float(tmp_object, "stddev", dev);
1034
1035 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1036 min = max = 0;
1037 mean = dev = 0.0;
1038 }
1039 tmp_object = json_create_object();
1040 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1041 json_object_add_value_int(tmp_object, "min", min);
1042 json_object_add_value_int(tmp_object, "max", max);
1043 json_object_add_value_float(tmp_object, "mean", mean);
1044 json_object_add_value_float(tmp_object, "stddev", dev);
1045 } else {
1046 if (!calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
1047 min = max = 0;
1048 mean = dev = 0.0;
1049 }
1050
1051 tmp_object = json_create_object();
1052 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1053 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1054 json_object_add_value_int(tmp_object, "min", min);
1055 json_object_add_value_int(tmp_object, "max", max);
1056 json_object_add_value_float(tmp_object, "mean", mean);
1057 json_object_add_value_float(tmp_object, "stddev", dev);
1058 }
1059
1060 if (ts->clat_percentiles || ts->lat_percentiles) {
1061 if (ddir_rw(ddir)) {
1062 len = calc_clat_percentiles(ts->io_u_plat[ddir],
1063 ts->clat_stat[ddir].samples,
1064 ts->percentile_list, &ovals, &maxv,
1065 &minv);
1066 } else {
1067 len = calc_clat_percentiles(ts->io_u_sync_plat,
1068 ts->sync_stat.samples,
1069 ts->percentile_list, &ovals, &maxv,
1070 &minv);
1071 }
1072
1073 if (len > FIO_IO_U_LIST_MAX_LEN)
1074 len = FIO_IO_U_LIST_MAX_LEN;
1075 } else
1076 len = 0;
1077
1078 percentile_object = json_create_object();
1079 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1080 for (i = 0; i < len; i++) {
1081 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1082 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1083 }
1084
1085 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1086 clat_bins_object = json_create_object();
1087 if (ts->clat_percentiles)
1088 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1089
1090 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1091 if (ddir_rw(ddir)) {
1092 if (ts->io_u_plat[ddir][i]) {
1093 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1094 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1095 }
1096 } else {
1097 if (ts->io_u_sync_plat[i]) {
1098 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1099 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_sync_plat[i]);
1100 }
1101 }
1102 }
1103 }
1104
1105 if (!ddir_rw(ddir))
1106 return;
1107
1108 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1109 min = max = 0;
1110 mean = dev = 0.0;
1111 }
1112 tmp_object = json_create_object();
1113 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1114 json_object_add_value_int(tmp_object, "min", min);
1115 json_object_add_value_int(tmp_object, "max", max);
1116 json_object_add_value_float(tmp_object, "mean", mean);
1117 json_object_add_value_float(tmp_object, "stddev", dev);
1118 if (output_format & FIO_OUTPUT_JSON_PLUS && ts->lat_percentiles)
1119 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1120
1121 if (ovals)
1122 free(ovals);
1123
1124 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1125 if (rs->agg[ddir]) {
1126 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1127 if (p_of_agg > 100.0)
1128 p_of_agg = 100.0;
1129 }
1130 } else {
1131 min = max = 0;
1132 p_of_agg = mean = dev = 0.0;
1133 }
1134 json_object_add_value_int(dir_object, "bw_min", min);
1135 json_object_add_value_int(dir_object, "bw_max", max);
1136 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1137 json_object_add_value_float(dir_object, "bw_mean", mean);
1138 json_object_add_value_float(dir_object, "bw_dev", dev);
1139 json_object_add_value_int(dir_object, "bw_samples",
1140 (&ts->bw_stat[ddir])->samples);
1141
1142 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1143 min = max = 0;
1144 mean = dev = 0.0;
1145 }
1146 json_object_add_value_int(dir_object, "iops_min", min);
1147 json_object_add_value_int(dir_object, "iops_max", max);
1148 json_object_add_value_float(dir_object, "iops_mean", mean);
1149 json_object_add_value_float(dir_object, "iops_stddev", dev);
1150 json_object_add_value_int(dir_object, "iops_samples",
1151 (&ts->iops_stat[ddir])->samples);
1152}
1153
1154static void show_thread_status_terse_all(struct thread_stat *ts,
1155 struct group_run_stats *rs, int ver,
1156 struct buf_output *out)
1157{
1158 double io_u_dist[FIO_IO_U_MAP_NR];
1159 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1160 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1161 double usr_cpu, sys_cpu;
1162 int i;
1163
1164 /* General Info */
1165 if (ver == 2)
1166 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1167 else
1168 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1169 ts->name, ts->groupid, ts->error);
1170
1171 /* Log Read Status */
1172 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1173 /* Log Write Status */
1174 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1175 /* Log Trim Status */
1176 if (ver == 2 || ver == 4 || ver == 5)
1177 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1178
1179 /* CPU Usage */
1180 if (ts->total_run_time) {
1181 double runt = (double) ts->total_run_time;
1182
1183 usr_cpu = (double) ts->usr_time * 100 / runt;
1184 sys_cpu = (double) ts->sys_time * 100 / runt;
1185 } else {
1186 usr_cpu = 0;
1187 sys_cpu = 0;
1188 }
1189
1190 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1191 (unsigned long long) ts->ctx,
1192 (unsigned long long) ts->majf,
1193 (unsigned long long) ts->minf);
1194
1195 /* Calc % distribution of IO depths, usecond, msecond latency */
1196 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1197 stat_calc_lat_nu(ts, io_u_lat_u);
1198 stat_calc_lat_m(ts, io_u_lat_m);
1199
1200 /* Only show fixed 7 I/O depth levels*/
1201 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1202 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1203 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1204
1205 /* Microsecond latency */
1206 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1207 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1208 /* Millisecond latency */
1209 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1210 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1211
1212 /* disk util stats, if any */
1213 if (ver >= 3 && is_running_backend())
1214 show_disk_util(1, NULL, out);
1215
1216 /* Additional output if continue_on_error set - default off*/
1217 if (ts->continue_on_error)
1218 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1219 if (ver == 2)
1220 log_buf(out, "\n");
1221
1222 /* Additional output if description is set */
1223 if (strlen(ts->description))
1224 log_buf(out, ";%s", ts->description);
1225
1226 log_buf(out, "\n");
1227}
1228
1229static void json_add_job_opts(struct json_object *root, const char *name,
1230 struct flist_head *opt_list)
1231{
1232 struct json_object *dir_object;
1233 struct flist_head *entry;
1234 struct print_option *p;
1235
1236 if (flist_empty(opt_list))
1237 return;
1238
1239 dir_object = json_create_object();
1240 json_object_add_value_object(root, name, dir_object);
1241
1242 flist_for_each(entry, opt_list) {
1243 const char *pos = "";
1244
1245 p = flist_entry(entry, struct print_option, list);
1246 if (p->value)
1247 pos = p->value;
1248 json_object_add_value_string(dir_object, p->name, pos);
1249 }
1250}
1251
1252static struct json_object *show_thread_status_json(struct thread_stat *ts,
1253 struct group_run_stats *rs,
1254 struct flist_head *opt_list)
1255{
1256 struct json_object *root, *tmp;
1257 struct jobs_eta *je;
1258 double io_u_dist[FIO_IO_U_MAP_NR];
1259 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1260 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1261 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1262 double usr_cpu, sys_cpu;
1263 int i;
1264 size_t size;
1265
1266 root = json_create_object();
1267 json_object_add_value_string(root, "jobname", ts->name);
1268 json_object_add_value_int(root, "groupid", ts->groupid);
1269 json_object_add_value_int(root, "error", ts->error);
1270
1271 /* ETA Info */
1272 je = get_jobs_eta(true, &size);
1273 if (je) {
1274 json_object_add_value_int(root, "eta", je->eta_sec);
1275 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1276 }
1277
1278 if (opt_list)
1279 json_add_job_opts(root, "job options", opt_list);
1280
1281 add_ddir_status_json(ts, rs, DDIR_READ, root);
1282 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1283 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1284 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1285
1286 /* CPU Usage */
1287 if (ts->total_run_time) {
1288 double runt = (double) ts->total_run_time;
1289
1290 usr_cpu = (double) ts->usr_time * 100 / runt;
1291 sys_cpu = (double) ts->sys_time * 100 / runt;
1292 } else {
1293 usr_cpu = 0;
1294 sys_cpu = 0;
1295 }
1296 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1297 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1298 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1299 json_object_add_value_int(root, "ctx", ts->ctx);
1300 json_object_add_value_int(root, "majf", ts->majf);
1301 json_object_add_value_int(root, "minf", ts->minf);
1302
1303 /* Calc % distribution of IO depths */
1304 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1305 tmp = json_create_object();
1306 json_object_add_value_object(root, "iodepth_level", tmp);
1307 /* Only show fixed 7 I/O depth levels*/
1308 for (i = 0; i < 7; i++) {
1309 char name[20];
1310 if (i < 6)
1311 snprintf(name, 20, "%d", 1 << i);
1312 else
1313 snprintf(name, 20, ">=%d", 1 << i);
1314 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1315 }
1316
1317 /* Calc % distribution of submit IO depths */
1318 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1319 tmp = json_create_object();
1320 json_object_add_value_object(root, "iodepth_submit", tmp);
1321 /* Only show fixed 7 I/O depth levels*/
1322 for (i = 0; i < 7; i++) {
1323 char name[20];
1324 if (i == 0)
1325 snprintf(name, 20, "0");
1326 else if (i < 6)
1327 snprintf(name, 20, "%d", 1 << (i+1));
1328 else
1329 snprintf(name, 20, ">=%d", 1 << i);
1330 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1331 }
1332
1333 /* Calc % distribution of completion IO depths */
1334 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1335 tmp = json_create_object();
1336 json_object_add_value_object(root, "iodepth_complete", tmp);
1337 /* Only show fixed 7 I/O depth levels*/
1338 for (i = 0; i < 7; i++) {
1339 char name[20];
1340 if (i == 0)
1341 snprintf(name, 20, "0");
1342 else if (i < 6)
1343 snprintf(name, 20, "%d", 1 << (i+1));
1344 else
1345 snprintf(name, 20, ">=%d", 1 << i);
1346 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1347 }
1348
1349 /* Calc % distribution of nsecond, usecond, msecond latency */
1350 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1351 stat_calc_lat_n(ts, io_u_lat_n);
1352 stat_calc_lat_u(ts, io_u_lat_u);
1353 stat_calc_lat_m(ts, io_u_lat_m);
1354
1355 /* Nanosecond latency */
1356 tmp = json_create_object();
1357 json_object_add_value_object(root, "latency_ns", tmp);
1358 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1359 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1360 "250", "500", "750", "1000", };
1361 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1362 }
1363 /* Microsecond latency */
1364 tmp = json_create_object();
1365 json_object_add_value_object(root, "latency_us", tmp);
1366 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1367 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1368 "250", "500", "750", "1000", };
1369 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1370 }
1371 /* Millisecond latency */
1372 tmp = json_create_object();
1373 json_object_add_value_object(root, "latency_ms", tmp);
1374 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1375 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1376 "250", "500", "750", "1000", "2000",
1377 ">=2000", };
1378 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1379 }
1380
1381 /* Additional output if continue_on_error set - default off*/
1382 if (ts->continue_on_error) {
1383 json_object_add_value_int(root, "total_err", ts->total_err_count);
1384 json_object_add_value_int(root, "first_error", ts->first_error);
1385 }
1386
1387 if (ts->latency_depth) {
1388 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1389 json_object_add_value_int(root, "latency_target", ts->latency_target);
1390 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1391 json_object_add_value_int(root, "latency_window", ts->latency_window);
1392 }
1393
1394 /* Additional output if description is set */
1395 if (strlen(ts->description))
1396 json_object_add_value_string(root, "desc", ts->description);
1397
1398 if (ts->nr_block_infos) {
1399 /* Block error histogram and types */
1400 int len;
1401 unsigned int *percentiles = NULL;
1402 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1403
1404 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1405 ts->percentile_list,
1406 &percentiles, block_state_counts);
1407
1408 if (len) {
1409 struct json_object *block, *percentile_object, *states;
1410 int state;
1411 block = json_create_object();
1412 json_object_add_value_object(root, "block", block);
1413
1414 percentile_object = json_create_object();
1415 json_object_add_value_object(block, "percentiles",
1416 percentile_object);
1417 for (i = 0; i < len; i++) {
1418 char buf[20];
1419 snprintf(buf, sizeof(buf), "%f",
1420 ts->percentile_list[i].u.f);
1421 json_object_add_value_int(percentile_object,
1422 (const char *)buf,
1423 percentiles[i]);
1424 }
1425
1426 states = json_create_object();
1427 json_object_add_value_object(block, "states", states);
1428 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1429 json_object_add_value_int(states,
1430 block_state_names[state],
1431 block_state_counts[state]);
1432 }
1433 free(percentiles);
1434 }
1435 }
1436
1437 if (ts->ss_dur) {
1438 struct json_object *data;
1439 struct json_array *iops, *bw;
1440 int j, k, l;
1441 char ss_buf[64];
1442
1443 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1444 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1445 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1446 (float) ts->ss_limit.u.f,
1447 ts->ss_state & FIO_SS_PCT ? "%" : "");
1448
1449 tmp = json_create_object();
1450 json_object_add_value_object(root, "steadystate", tmp);
1451 json_object_add_value_string(tmp, "ss", ss_buf);
1452 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1453 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1454
1455 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1456 ts->ss_state & FIO_SS_PCT ? "%" : "");
1457 json_object_add_value_string(tmp, "criterion", ss_buf);
1458 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1459 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1460
1461 data = json_create_object();
1462 json_object_add_value_object(tmp, "data", data);
1463 bw = json_create_array();
1464 iops = json_create_array();
1465
1466 /*
1467 ** if ss was attained or the buffer is not full,
1468 ** ss->head points to the first element in the list.
1469 ** otherwise it actually points to the second element
1470 ** in the list
1471 */
1472 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1473 j = ts->ss_head;
1474 else
1475 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1476 for (l = 0; l < ts->ss_dur; l++) {
1477 k = (j + l) % ts->ss_dur;
1478 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1479 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1480 }
1481 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1482 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1483 json_object_add_value_array(data, "iops", iops);
1484 json_object_add_value_array(data, "bw", bw);
1485 }
1486
1487 return root;
1488}
1489
1490static void show_thread_status_terse(struct thread_stat *ts,
1491 struct group_run_stats *rs,
1492 struct buf_output *out)
1493{
1494 if (terse_version >= 2 && terse_version <= 5)
1495 show_thread_status_terse_all(ts, rs, terse_version, out);
1496 else
1497 log_err("fio: bad terse version!? %d\n", terse_version);
1498}
1499
1500struct json_object *show_thread_status(struct thread_stat *ts,
1501 struct group_run_stats *rs,
1502 struct flist_head *opt_list,
1503 struct buf_output *out)
1504{
1505 struct json_object *ret = NULL;
1506
1507 if (output_format & FIO_OUTPUT_TERSE)
1508 show_thread_status_terse(ts, rs, out);
1509 if (output_format & FIO_OUTPUT_JSON)
1510 ret = show_thread_status_json(ts, rs, opt_list);
1511 if (output_format & FIO_OUTPUT_NORMAL)
1512 show_thread_status_normal(ts, rs, out);
1513
1514 return ret;
1515}
1516
1517static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1518{
1519 double mean, S;
1520
1521 if (src->samples == 0)
1522 return;
1523
1524 dst->min_val = min(dst->min_val, src->min_val);
1525 dst->max_val = max(dst->max_val, src->max_val);
1526
1527 /*
1528 * Compute new mean and S after the merge
1529 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1530 * #Parallel_algorithm>
1531 */
1532 if (first) {
1533 mean = src->mean.u.f;
1534 S = src->S.u.f;
1535 } else {
1536 double delta = src->mean.u.f - dst->mean.u.f;
1537
1538 mean = ((src->mean.u.f * src->samples) +
1539 (dst->mean.u.f * dst->samples)) /
1540 (dst->samples + src->samples);
1541
1542 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1543 (dst->samples * src->samples) /
1544 (dst->samples + src->samples);
1545 }
1546
1547 dst->samples += src->samples;
1548 dst->mean.u.f = mean;
1549 dst->S.u.f = S;
1550}
1551
1552void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1553{
1554 int i;
1555
1556 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1557 if (dst->max_run[i] < src->max_run[i])
1558 dst->max_run[i] = src->max_run[i];
1559 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1560 dst->min_run[i] = src->min_run[i];
1561 if (dst->max_bw[i] < src->max_bw[i])
1562 dst->max_bw[i] = src->max_bw[i];
1563 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1564 dst->min_bw[i] = src->min_bw[i];
1565
1566 dst->iobytes[i] += src->iobytes[i];
1567 dst->agg[i] += src->agg[i];
1568 }
1569
1570 if (!dst->kb_base)
1571 dst->kb_base = src->kb_base;
1572 if (!dst->unit_base)
1573 dst->unit_base = src->unit_base;
1574 if (!dst->sig_figs)
1575 dst->sig_figs = src->sig_figs;
1576}
1577
1578void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1579 bool first)
1580{
1581 int l, k;
1582
1583 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1584 if (!dst->unified_rw_rep) {
1585 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1586 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1587 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1588 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1589 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first);
1590
1591 dst->io_bytes[l] += src->io_bytes[l];
1592
1593 if (dst->runtime[l] < src->runtime[l])
1594 dst->runtime[l] = src->runtime[l];
1595 } else {
1596 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1597 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1598 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1599 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1600 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first);
1601
1602 dst->io_bytes[0] += src->io_bytes[l];
1603
1604 if (dst->runtime[0] < src->runtime[l])
1605 dst->runtime[0] = src->runtime[l];
1606
1607 /*
1608 * We're summing to the same destination, so override
1609 * 'first' after the first iteration of the loop
1610 */
1611 first = false;
1612 }
1613 }
1614
1615 sum_stat(&dst->sync_stat, &src->sync_stat, first);
1616 dst->usr_time += src->usr_time;
1617 dst->sys_time += src->sys_time;
1618 dst->ctx += src->ctx;
1619 dst->majf += src->majf;
1620 dst->minf += src->minf;
1621
1622 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
1623 dst->io_u_map[k] += src->io_u_map[k];
1624 dst->io_u_submit[k] += src->io_u_submit[k];
1625 dst->io_u_complete[k] += src->io_u_complete[k];
1626 }
1627 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++) {
1628 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1629 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1630 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1631 }
1632 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1633 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
1634
1635 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1636 if (!dst->unified_rw_rep) {
1637 dst->total_io_u[k] += src->total_io_u[k];
1638 dst->short_io_u[k] += src->short_io_u[k];
1639 dst->drop_io_u[k] += src->drop_io_u[k];
1640 } else {
1641 dst->total_io_u[0] += src->total_io_u[k];
1642 dst->short_io_u[0] += src->short_io_u[k];
1643 dst->drop_io_u[0] += src->drop_io_u[k];
1644 }
1645 }
1646
1647 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
1648
1649 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1650 int m;
1651
1652 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1653 if (!dst->unified_rw_rep)
1654 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1655 else
1656 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1657 }
1658 }
1659
1660 dst->total_run_time += src->total_run_time;
1661 dst->total_submit += src->total_submit;
1662 dst->total_complete += src->total_complete;
1663 dst->nr_zone_resets += src->nr_zone_resets;
1664}
1665
1666void init_group_run_stat(struct group_run_stats *gs)
1667{
1668 int i;
1669 memset(gs, 0, sizeof(*gs));
1670
1671 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1672 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1673}
1674
1675void init_thread_stat(struct thread_stat *ts)
1676{
1677 int j;
1678
1679 memset(ts, 0, sizeof(*ts));
1680
1681 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1682 ts->lat_stat[j].min_val = -1UL;
1683 ts->clat_stat[j].min_val = -1UL;
1684 ts->slat_stat[j].min_val = -1UL;
1685 ts->bw_stat[j].min_val = -1UL;
1686 ts->iops_stat[j].min_val = -1UL;
1687 }
1688 ts->sync_stat.min_val = -1UL;
1689 ts->groupid = -1;
1690}
1691
1692void __show_run_stats(void)
1693{
1694 struct group_run_stats *runstats, *rs;
1695 struct thread_data *td;
1696 struct thread_stat *threadstats, *ts;
1697 int i, j, k, nr_ts, last_ts, idx;
1698 bool kb_base_warned = false;
1699 bool unit_base_warned = false;
1700 struct json_object *root = NULL;
1701 struct json_array *array = NULL;
1702 struct buf_output output[FIO_OUTPUT_NR];
1703 struct flist_head **opt_lists;
1704
1705 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1706
1707 for (i = 0; i < groupid + 1; i++)
1708 init_group_run_stat(&runstats[i]);
1709
1710 /*
1711 * find out how many threads stats we need. if group reporting isn't
1712 * enabled, it's one-per-td.
1713 */
1714 nr_ts = 0;
1715 last_ts = -1;
1716 for_each_td(td, i) {
1717 if (!td->o.group_reporting) {
1718 nr_ts++;
1719 continue;
1720 }
1721 if (last_ts == td->groupid)
1722 continue;
1723 if (!td->o.stats)
1724 continue;
1725
1726 last_ts = td->groupid;
1727 nr_ts++;
1728 }
1729
1730 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1731 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1732
1733 for (i = 0; i < nr_ts; i++) {
1734 init_thread_stat(&threadstats[i]);
1735 opt_lists[i] = NULL;
1736 }
1737
1738 j = 0;
1739 last_ts = -1;
1740 idx = 0;
1741 for_each_td(td, i) {
1742 if (!td->o.stats)
1743 continue;
1744 if (idx && (!td->o.group_reporting ||
1745 (td->o.group_reporting && last_ts != td->groupid))) {
1746 idx = 0;
1747 j++;
1748 }
1749
1750 last_ts = td->groupid;
1751
1752 ts = &threadstats[j];
1753
1754 ts->clat_percentiles = td->o.clat_percentiles;
1755 ts->lat_percentiles = td->o.lat_percentiles;
1756 ts->percentile_precision = td->o.percentile_precision;
1757 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1758 opt_lists[j] = &td->opt_list;
1759
1760 idx++;
1761 ts->members++;
1762
1763 if (ts->groupid == -1) {
1764 /*
1765 * These are per-group shared already
1766 */
1767 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1768 if (td->o.description)
1769 strncpy(ts->description, td->o.description,
1770 FIO_JOBDESC_SIZE - 1);
1771 else
1772 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1773
1774 /*
1775 * If multiple entries in this group, this is
1776 * the first member.
1777 */
1778 ts->thread_number = td->thread_number;
1779 ts->groupid = td->groupid;
1780
1781 /*
1782 * first pid in group, not very useful...
1783 */
1784 ts->pid = td->pid;
1785
1786 ts->kb_base = td->o.kb_base;
1787 ts->unit_base = td->o.unit_base;
1788 ts->sig_figs = td->o.sig_figs;
1789 ts->unified_rw_rep = td->o.unified_rw_rep;
1790 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1791 log_info("fio: kb_base differs for jobs in group, using"
1792 " %u as the base\n", ts->kb_base);
1793 kb_base_warned = true;
1794 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1795 log_info("fio: unit_base differs for jobs in group, using"
1796 " %u as the base\n", ts->unit_base);
1797 unit_base_warned = true;
1798 }
1799
1800 ts->continue_on_error = td->o.continue_on_error;
1801 ts->total_err_count += td->total_err_count;
1802 ts->first_error = td->first_error;
1803 if (!ts->error) {
1804 if (!td->error && td->o.continue_on_error &&
1805 td->first_error) {
1806 ts->error = td->first_error;
1807 ts->verror[sizeof(ts->verror) - 1] = '\0';
1808 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1809 } else if (td->error) {
1810 ts->error = td->error;
1811 ts->verror[sizeof(ts->verror) - 1] = '\0';
1812 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1813 }
1814 }
1815
1816 ts->latency_depth = td->latency_qd;
1817 ts->latency_target = td->o.latency_target;
1818 ts->latency_percentile = td->o.latency_percentile;
1819 ts->latency_window = td->o.latency_window;
1820
1821 ts->nr_block_infos = td->ts.nr_block_infos;
1822 for (k = 0; k < ts->nr_block_infos; k++)
1823 ts->block_infos[k] = td->ts.block_infos[k];
1824
1825 sum_thread_stats(ts, &td->ts, idx == 1);
1826
1827 if (td->o.ss_dur) {
1828 ts->ss_state = td->ss.state;
1829 ts->ss_dur = td->ss.dur;
1830 ts->ss_head = td->ss.head;
1831 ts->ss_bw_data = td->ss.bw_data;
1832 ts->ss_iops_data = td->ss.iops_data;
1833 ts->ss_limit.u.f = td->ss.limit;
1834 ts->ss_slope.u.f = td->ss.slope;
1835 ts->ss_deviation.u.f = td->ss.deviation;
1836 ts->ss_criterion.u.f = td->ss.criterion;
1837 }
1838 else
1839 ts->ss_dur = ts->ss_state = 0;
1840 }
1841
1842 for (i = 0; i < nr_ts; i++) {
1843 unsigned long long bw;
1844
1845 ts = &threadstats[i];
1846 if (ts->groupid == -1)
1847 continue;
1848 rs = &runstats[ts->groupid];
1849 rs->kb_base = ts->kb_base;
1850 rs->unit_base = ts->unit_base;
1851 rs->sig_figs = ts->sig_figs;
1852 rs->unified_rw_rep += ts->unified_rw_rep;
1853
1854 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1855 if (!ts->runtime[j])
1856 continue;
1857 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1858 rs->min_run[j] = ts->runtime[j];
1859 if (ts->runtime[j] > rs->max_run[j])
1860 rs->max_run[j] = ts->runtime[j];
1861
1862 bw = 0;
1863 if (ts->runtime[j])
1864 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1865 if (bw < rs->min_bw[j])
1866 rs->min_bw[j] = bw;
1867 if (bw > rs->max_bw[j])
1868 rs->max_bw[j] = bw;
1869
1870 rs->iobytes[j] += ts->io_bytes[j];
1871 }
1872 }
1873
1874 for (i = 0; i < groupid + 1; i++) {
1875 int ddir;
1876
1877 rs = &runstats[i];
1878
1879 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1880 if (rs->max_run[ddir])
1881 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1882 rs->max_run[ddir];
1883 }
1884 }
1885
1886 for (i = 0; i < FIO_OUTPUT_NR; i++)
1887 buf_output_init(&output[i]);
1888
1889 /*
1890 * don't overwrite last signal output
1891 */
1892 if (output_format & FIO_OUTPUT_NORMAL)
1893 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1894 if (output_format & FIO_OUTPUT_JSON) {
1895 struct thread_data *global;
1896 char time_buf[32];
1897 struct timeval now;
1898 unsigned long long ms_since_epoch;
1899 time_t tv_sec;
1900
1901 gettimeofday(&now, NULL);
1902 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1903 (unsigned long long)(now.tv_usec) / 1000;
1904
1905 tv_sec = now.tv_sec;
1906 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
1907 if (time_buf[strlen(time_buf) - 1] == '\n')
1908 time_buf[strlen(time_buf) - 1] = '\0';
1909
1910 root = json_create_object();
1911 json_object_add_value_string(root, "fio version", fio_version_string);
1912 json_object_add_value_int(root, "timestamp", now.tv_sec);
1913 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1914 json_object_add_value_string(root, "time", time_buf);
1915 global = get_global_options();
1916 json_add_job_opts(root, "global options", &global->opt_list);
1917 array = json_create_array();
1918 json_object_add_value_array(root, "jobs", array);
1919 }
1920
1921 if (is_backend)
1922 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1923
1924 for (i = 0; i < nr_ts; i++) {
1925 ts = &threadstats[i];
1926 rs = &runstats[ts->groupid];
1927
1928 if (is_backend) {
1929 fio_server_send_job_options(opt_lists[i], i);
1930 fio_server_send_ts(ts, rs);
1931 } else {
1932 if (output_format & FIO_OUTPUT_TERSE)
1933 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1934 if (output_format & FIO_OUTPUT_JSON) {
1935 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1936 json_array_add_value_object(array, tmp);
1937 }
1938 if (output_format & FIO_OUTPUT_NORMAL)
1939 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1940 }
1941 }
1942 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1943 /* disk util stats, if any */
1944 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1945
1946 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1947
1948 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1949 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1950 json_free_object(root);
1951 }
1952
1953 for (i = 0; i < groupid + 1; i++) {
1954 rs = &runstats[i];
1955
1956 rs->groupid = i;
1957 if (is_backend)
1958 fio_server_send_gs(rs);
1959 else if (output_format & FIO_OUTPUT_NORMAL)
1960 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1961 }
1962
1963 if (is_backend)
1964 fio_server_send_du();
1965 else if (output_format & FIO_OUTPUT_NORMAL) {
1966 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1967 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1968 }
1969
1970 for (i = 0; i < FIO_OUTPUT_NR; i++) {
1971 struct buf_output *out = &output[i];
1972
1973 log_info_buf(out->buf, out->buflen);
1974 buf_output_free(out);
1975 }
1976
1977 fio_idle_prof_cleanup();
1978
1979 log_info_flush();
1980 free(runstats);
1981 free(threadstats);
1982 free(opt_lists);
1983}
1984
1985void __show_running_run_stats(void)
1986{
1987 struct thread_data *td;
1988 unsigned long long *rt;
1989 struct timespec ts;
1990 int i;
1991
1992 fio_sem_down(stat_sem);
1993
1994 rt = malloc(thread_number * sizeof(unsigned long long));
1995 fio_gettime(&ts, NULL);
1996
1997 for_each_td(td, i) {
1998 td->update_rusage = 1;
1999 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
2000 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
2001 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2002 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2003
2004 rt[i] = mtime_since(&td->start, &ts);
2005 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2006 td->ts.runtime[DDIR_READ] += rt[i];
2007 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2008 td->ts.runtime[DDIR_WRITE] += rt[i];
2009 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2010 td->ts.runtime[DDIR_TRIM] += rt[i];
2011 }
2012
2013 for_each_td(td, i) {
2014 if (td->runstate >= TD_EXITED)
2015 continue;
2016 if (td->rusage_sem) {
2017 td->update_rusage = 1;
2018 fio_sem_down(td->rusage_sem);
2019 }
2020 td->update_rusage = 0;
2021 }
2022
2023 __show_run_stats();
2024
2025 for_each_td(td, i) {
2026 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2027 td->ts.runtime[DDIR_READ] -= rt[i];
2028 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2029 td->ts.runtime[DDIR_WRITE] -= rt[i];
2030 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2031 td->ts.runtime[DDIR_TRIM] -= rt[i];
2032 }
2033
2034 free(rt);
2035 fio_sem_up(stat_sem);
2036}
2037
2038static bool status_interval_init;
2039static struct timespec status_time;
2040static bool status_file_disabled;
2041
2042#define FIO_STATUS_FILE "fio-dump-status"
2043
2044static int check_status_file(void)
2045{
2046 struct stat sb;
2047 const char *temp_dir;
2048 char fio_status_file_path[PATH_MAX];
2049
2050 if (status_file_disabled)
2051 return 0;
2052
2053 temp_dir = getenv("TMPDIR");
2054 if (temp_dir == NULL) {
2055 temp_dir = getenv("TEMP");
2056 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2057 temp_dir = NULL;
2058 }
2059 if (temp_dir == NULL)
2060 temp_dir = "/tmp";
2061
2062 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2063
2064 if (stat(fio_status_file_path, &sb))
2065 return 0;
2066
2067 if (unlink(fio_status_file_path) < 0) {
2068 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2069 strerror(errno));
2070 log_err("fio: disabling status file updates\n");
2071 status_file_disabled = true;
2072 }
2073
2074 return 1;
2075}
2076
2077void check_for_running_stats(void)
2078{
2079 if (status_interval) {
2080 if (!status_interval_init) {
2081 fio_gettime(&status_time, NULL);
2082 status_interval_init = true;
2083 } else if (mtime_since_now(&status_time) >= status_interval) {
2084 show_running_run_stats();
2085 fio_gettime(&status_time, NULL);
2086 return;
2087 }
2088 }
2089 if (check_status_file()) {
2090 show_running_run_stats();
2091 return;
2092 }
2093}
2094
2095static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2096{
2097 double val = data;
2098 double delta;
2099
2100 if (data > is->max_val)
2101 is->max_val = data;
2102 if (data < is->min_val)
2103 is->min_val = data;
2104
2105 delta = val - is->mean.u.f;
2106 if (delta) {
2107 is->mean.u.f += delta / (is->samples + 1.0);
2108 is->S.u.f += delta * (val - is->mean.u.f);
2109 }
2110
2111 is->samples++;
2112}
2113
2114/*
2115 * Return a struct io_logs, which is added to the tail of the log
2116 * list for 'iolog'.
2117 */
2118static struct io_logs *get_new_log(struct io_log *iolog)
2119{
2120 size_t new_size, new_samples;
2121 struct io_logs *cur_log;
2122
2123 /*
2124 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2125 * forever
2126 */
2127 if (!iolog->cur_log_max)
2128 new_samples = DEF_LOG_ENTRIES;
2129 else {
2130 new_samples = iolog->cur_log_max * 2;
2131 if (new_samples > MAX_LOG_ENTRIES)
2132 new_samples = MAX_LOG_ENTRIES;
2133 }
2134
2135 new_size = new_samples * log_entry_sz(iolog);
2136
2137 cur_log = smalloc(sizeof(*cur_log));
2138 if (cur_log) {
2139 INIT_FLIST_HEAD(&cur_log->list);
2140 cur_log->log = malloc(new_size);
2141 if (cur_log->log) {
2142 cur_log->nr_samples = 0;
2143 cur_log->max_samples = new_samples;
2144 flist_add_tail(&cur_log->list, &iolog->io_logs);
2145 iolog->cur_log_max = new_samples;
2146 return cur_log;
2147 }
2148 sfree(cur_log);
2149 }
2150
2151 return NULL;
2152}
2153
2154/*
2155 * Add and return a new log chunk, or return current log if big enough
2156 */
2157static struct io_logs *regrow_log(struct io_log *iolog)
2158{
2159 struct io_logs *cur_log;
2160 int i;
2161
2162 if (!iolog || iolog->disabled)
2163 goto disable;
2164
2165 cur_log = iolog_cur_log(iolog);
2166 if (!cur_log) {
2167 cur_log = get_new_log(iolog);
2168 if (!cur_log)
2169 return NULL;
2170 }
2171
2172 if (cur_log->nr_samples < cur_log->max_samples)
2173 return cur_log;
2174
2175 /*
2176 * No room for a new sample. If we're compressing on the fly, flush
2177 * out the current chunk
2178 */
2179 if (iolog->log_gz) {
2180 if (iolog_cur_flush(iolog, cur_log)) {
2181 log_err("fio: failed flushing iolog! Will stop logging.\n");
2182 return NULL;
2183 }
2184 }
2185
2186 /*
2187 * Get a new log array, and add to our list
2188 */
2189 cur_log = get_new_log(iolog);
2190 if (!cur_log) {
2191 log_err("fio: failed extending iolog! Will stop logging.\n");
2192 return NULL;
2193 }
2194
2195 if (!iolog->pending || !iolog->pending->nr_samples)
2196 return cur_log;
2197
2198 /*
2199 * Flush pending items to new log
2200 */
2201 for (i = 0; i < iolog->pending->nr_samples; i++) {
2202 struct io_sample *src, *dst;
2203
2204 src = get_sample(iolog, iolog->pending, i);
2205 dst = get_sample(iolog, cur_log, i);
2206 memcpy(dst, src, log_entry_sz(iolog));
2207 }
2208 cur_log->nr_samples = iolog->pending->nr_samples;
2209
2210 iolog->pending->nr_samples = 0;
2211 return cur_log;
2212disable:
2213 if (iolog)
2214 iolog->disabled = true;
2215 return NULL;
2216}
2217
2218void regrow_logs(struct thread_data *td)
2219{
2220 regrow_log(td->slat_log);
2221 regrow_log(td->clat_log);
2222 regrow_log(td->clat_hist_log);
2223 regrow_log(td->lat_log);
2224 regrow_log(td->bw_log);
2225 regrow_log(td->iops_log);
2226 td->flags &= ~TD_F_REGROW_LOGS;
2227}
2228
2229static struct io_logs *get_cur_log(struct io_log *iolog)
2230{
2231 struct io_logs *cur_log;
2232
2233 cur_log = iolog_cur_log(iolog);
2234 if (!cur_log) {
2235 cur_log = get_new_log(iolog);
2236 if (!cur_log)
2237 return NULL;
2238 }
2239
2240 if (cur_log->nr_samples < cur_log->max_samples)
2241 return cur_log;
2242
2243 /*
2244 * Out of space. If we're in IO offload mode, or we're not doing
2245 * per unit logging (hence logging happens outside of the IO thread
2246 * as well), add a new log chunk inline. If we're doing inline
2247 * submissions, flag 'td' as needing a log regrow and we'll take
2248 * care of it on the submission side.
2249 */
2250 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2251 !per_unit_log(iolog))
2252 return regrow_log(iolog);
2253
2254 if (iolog->td)
2255 iolog->td->flags |= TD_F_REGROW_LOGS;
2256 if (iolog->pending)
2257 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2258 return iolog->pending;
2259}
2260
2261static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2262 enum fio_ddir ddir, unsigned long long bs,
2263 unsigned long t, uint64_t offset)
2264{
2265 struct io_logs *cur_log;
2266
2267 if (iolog->disabled)
2268 return;
2269 if (flist_empty(&iolog->io_logs))
2270 iolog->avg_last[ddir] = t;
2271
2272 cur_log = get_cur_log(iolog);
2273 if (cur_log) {
2274 struct io_sample *s;
2275
2276 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2277
2278 s->data = data;
2279 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2280 io_sample_set_ddir(iolog, s, ddir);
2281 s->bs = bs;
2282
2283 if (iolog->log_offset) {
2284 struct io_sample_offset *so = (void *) s;
2285
2286 so->offset = offset;
2287 }
2288
2289 cur_log->nr_samples++;
2290 return;
2291 }
2292
2293 iolog->disabled = true;
2294}
2295
2296static inline void reset_io_stat(struct io_stat *ios)
2297{
2298 ios->max_val = ios->min_val = ios->samples = 0;
2299 ios->mean.u.f = ios->S.u.f = 0;
2300}
2301
2302void reset_io_stats(struct thread_data *td)
2303{
2304 struct thread_stat *ts = &td->ts;
2305 int i, j;
2306
2307 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2308 reset_io_stat(&ts->clat_stat[i]);
2309 reset_io_stat(&ts->slat_stat[i]);
2310 reset_io_stat(&ts->lat_stat[i]);
2311 reset_io_stat(&ts->bw_stat[i]);
2312 reset_io_stat(&ts->iops_stat[i]);
2313
2314 ts->io_bytes[i] = 0;
2315 ts->runtime[i] = 0;
2316 ts->total_io_u[i] = 0;
2317 ts->short_io_u[i] = 0;
2318 ts->drop_io_u[i] = 0;
2319
2320 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2321 ts->io_u_plat[i][j] = 0;
2322 if (!i)
2323 ts->io_u_sync_plat[j] = 0;
2324 }
2325 }
2326
2327 ts->total_io_u[DDIR_SYNC] = 0;
2328
2329 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2330 ts->io_u_map[i] = 0;
2331 ts->io_u_submit[i] = 0;
2332 ts->io_u_complete[i] = 0;
2333 }
2334
2335 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2336 ts->io_u_lat_n[i] = 0;
2337 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2338 ts->io_u_lat_u[i] = 0;
2339 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2340 ts->io_u_lat_m[i] = 0;
2341
2342 ts->total_submit = 0;
2343 ts->total_complete = 0;
2344 ts->nr_zone_resets = 0;
2345}
2346
2347static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2348 unsigned long elapsed, bool log_max)
2349{
2350 /*
2351 * Note an entry in the log. Use the mean from the logged samples,
2352 * making sure to properly round up. Only write a log entry if we
2353 * had actual samples done.
2354 */
2355 if (iolog->avg_window[ddir].samples) {
2356 union io_sample_data data;
2357
2358 if (log_max)
2359 data.val = iolog->avg_window[ddir].max_val;
2360 else
2361 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2362
2363 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2364 }
2365
2366 reset_io_stat(&iolog->avg_window[ddir]);
2367}
2368
2369static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2370 bool log_max)
2371{
2372 int ddir;
2373
2374 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2375 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2376}
2377
2378static unsigned long add_log_sample(struct thread_data *td,
2379 struct io_log *iolog,
2380 union io_sample_data data,
2381 enum fio_ddir ddir, unsigned long long bs,
2382 uint64_t offset)
2383{
2384 unsigned long elapsed, this_window;
2385
2386 if (!ddir_rw(ddir))
2387 return 0;
2388
2389 elapsed = mtime_since_now(&td->epoch);
2390
2391 /*
2392 * If no time averaging, just add the log sample.
2393 */
2394 if (!iolog->avg_msec) {
2395 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2396 return 0;
2397 }
2398
2399 /*
2400 * Add the sample. If the time period has passed, then
2401 * add that entry to the log and clear.
2402 */
2403 add_stat_sample(&iolog->avg_window[ddir], data.val);
2404
2405 /*
2406 * If period hasn't passed, adding the above sample is all we
2407 * need to do.
2408 */
2409 this_window = elapsed - iolog->avg_last[ddir];
2410 if (elapsed < iolog->avg_last[ddir])
2411 return iolog->avg_last[ddir] - elapsed;
2412 else if (this_window < iolog->avg_msec) {
2413 unsigned long diff = iolog->avg_msec - this_window;
2414
2415 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2416 return diff;
2417 }
2418
2419 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2420
2421 iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2422 return iolog->avg_msec;
2423}
2424
2425void finalize_logs(struct thread_data *td, bool unit_logs)
2426{
2427 unsigned long elapsed;
2428
2429 elapsed = mtime_since_now(&td->epoch);
2430
2431 if (td->clat_log && unit_logs)
2432 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2433 if (td->slat_log && unit_logs)
2434 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2435 if (td->lat_log && unit_logs)
2436 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2437 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2438 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2439 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2440 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2441}
2442
2443void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned long long bs)
2444{
2445 struct io_log *iolog;
2446
2447 if (!ddir_rw(ddir))
2448 return;
2449
2450 iolog = agg_io_log[ddir];
2451 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2452}
2453
2454void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2455{
2456 unsigned int idx = plat_val_to_idx(nsec);
2457 assert(idx < FIO_IO_U_PLAT_NR);
2458
2459 ts->io_u_sync_plat[idx]++;
2460 add_stat_sample(&ts->sync_stat, nsec);
2461}
2462
2463static void add_clat_percentile_sample(struct thread_stat *ts,
2464 unsigned long long nsec, enum fio_ddir ddir)
2465{
2466 unsigned int idx = plat_val_to_idx(nsec);
2467 assert(idx < FIO_IO_U_PLAT_NR);
2468
2469 ts->io_u_plat[ddir][idx]++;
2470}
2471
2472void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2473 unsigned long long nsec, unsigned long long bs,
2474 uint64_t offset)
2475{
2476 const bool needs_lock = td_async_processing(td);
2477 unsigned long elapsed, this_window;
2478 struct thread_stat *ts = &td->ts;
2479 struct io_log *iolog = td->clat_hist_log;
2480
2481 if (needs_lock)
2482 __td_io_u_lock(td);
2483
2484 add_stat_sample(&ts->clat_stat[ddir], nsec);
2485
2486 if (td->clat_log)
2487 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2488 offset);
2489
2490 if (ts->clat_percentiles)
2491 add_clat_percentile_sample(ts, nsec, ddir);
2492
2493 if (iolog && iolog->hist_msec) {
2494 struct io_hist *hw = &iolog->hist_window[ddir];
2495
2496 hw->samples++;
2497 elapsed = mtime_since_now(&td->epoch);
2498 if (!hw->hist_last)
2499 hw->hist_last = elapsed;
2500 this_window = elapsed - hw->hist_last;
2501
2502 if (this_window >= iolog->hist_msec) {
2503 uint64_t *io_u_plat;
2504 struct io_u_plat_entry *dst;
2505
2506 /*
2507 * Make a byte-for-byte copy of the latency histogram
2508 * stored in td->ts.io_u_plat[ddir], recording it in a
2509 * log sample. Note that the matching call to free() is
2510 * located in iolog.c after printing this sample to the
2511 * log file.
2512 */
2513 io_u_plat = (uint64_t *) td->ts.io_u_plat[ddir];
2514 dst = malloc(sizeof(struct io_u_plat_entry));
2515 memcpy(&(dst->io_u_plat), io_u_plat,
2516 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2517 flist_add(&dst->list, &hw->list);
2518 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2519 elapsed, offset);
2520
2521 /*
2522 * Update the last time we recorded as being now, minus
2523 * any drift in time we encountered before actually
2524 * making the record.
2525 */
2526 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2527 hw->samples = 0;
2528 }
2529 }
2530
2531 if (needs_lock)
2532 __td_io_u_unlock(td);
2533}
2534
2535void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2536 unsigned long usec, unsigned long long bs, uint64_t offset)
2537{
2538 const bool needs_lock = td_async_processing(td);
2539 struct thread_stat *ts = &td->ts;
2540
2541 if (!ddir_rw(ddir))
2542 return;
2543
2544 if (needs_lock)
2545 __td_io_u_lock(td);
2546
2547 add_stat_sample(&ts->slat_stat[ddir], usec);
2548
2549 if (td->slat_log)
2550 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2551
2552 if (needs_lock)
2553 __td_io_u_unlock(td);
2554}
2555
2556void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2557 unsigned long long nsec, unsigned long long bs,
2558 uint64_t offset)
2559{
2560 const bool needs_lock = td_async_processing(td);
2561 struct thread_stat *ts = &td->ts;
2562
2563 if (!ddir_rw(ddir))
2564 return;
2565
2566 if (needs_lock)
2567 __td_io_u_lock(td);
2568
2569 add_stat_sample(&ts->lat_stat[ddir], nsec);
2570
2571 if (td->lat_log)
2572 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2573 offset);
2574
2575 if (ts->lat_percentiles)
2576 add_clat_percentile_sample(ts, nsec, ddir);
2577
2578 if (needs_lock)
2579 __td_io_u_unlock(td);
2580}
2581
2582void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2583 unsigned int bytes, unsigned long long spent)
2584{
2585 const bool needs_lock = td_async_processing(td);
2586 struct thread_stat *ts = &td->ts;
2587 unsigned long rate;
2588
2589 if (spent)
2590 rate = (unsigned long) (bytes * 1000000ULL / spent);
2591 else
2592 rate = 0;
2593
2594 if (needs_lock)
2595 __td_io_u_lock(td);
2596
2597 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2598
2599 if (td->bw_log)
2600 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2601 bytes, io_u->offset);
2602
2603 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2604
2605 if (needs_lock)
2606 __td_io_u_unlock(td);
2607}
2608
2609static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2610 struct timespec *t, unsigned int avg_time,
2611 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2612 struct io_stat *stat, struct io_log *log,
2613 bool is_kb)
2614{
2615 const bool needs_lock = td_async_processing(td);
2616 unsigned long spent, rate;
2617 enum fio_ddir ddir;
2618 unsigned long next, next_log;
2619
2620 next_log = avg_time;
2621
2622 spent = mtime_since(parent_tv, t);
2623 if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2624 return avg_time - spent;
2625
2626 if (needs_lock)
2627 __td_io_u_lock(td);
2628
2629 /*
2630 * Compute both read and write rates for the interval.
2631 */
2632 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2633 uint64_t delta;
2634
2635 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2636 if (!delta)
2637 continue; /* No entries for interval */
2638
2639 if (spent) {
2640 if (is_kb)
2641 rate = delta * 1000 / spent / 1024; /* KiB/s */
2642 else
2643 rate = (delta * 1000) / spent;
2644 } else
2645 rate = 0;
2646
2647 add_stat_sample(&stat[ddir], rate);
2648
2649 if (log) {
2650 unsigned long long bs = 0;
2651
2652 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2653 bs = td->o.min_bs[ddir];
2654
2655 next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2656 next_log = min(next_log, next);
2657 }
2658
2659 stat_io_bytes[ddir] = this_io_bytes[ddir];
2660 }
2661
2662 timespec_add_msec(parent_tv, avg_time);
2663
2664 if (needs_lock)
2665 __td_io_u_unlock(td);
2666
2667 if (spent <= avg_time)
2668 next = avg_time;
2669 else
2670 next = avg_time - (1 + spent - avg_time);
2671
2672 return min(next, next_log);
2673}
2674
2675static int add_bw_samples(struct thread_data *td, struct timespec *t)
2676{
2677 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2678 td->this_io_bytes, td->stat_io_bytes,
2679 td->ts.bw_stat, td->bw_log, true);
2680}
2681
2682void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2683 unsigned int bytes)
2684{
2685 const bool needs_lock = td_async_processing(td);
2686 struct thread_stat *ts = &td->ts;
2687
2688 if (needs_lock)
2689 __td_io_u_lock(td);
2690
2691 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2692
2693 if (td->iops_log)
2694 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2695 bytes, io_u->offset);
2696
2697 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2698
2699 if (needs_lock)
2700 __td_io_u_unlock(td);
2701}
2702
2703static int add_iops_samples(struct thread_data *td, struct timespec *t)
2704{
2705 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2706 td->this_io_blocks, td->stat_io_blocks,
2707 td->ts.iops_stat, td->iops_log, false);
2708}
2709
2710/*
2711 * Returns msecs to next event
2712 */
2713int calc_log_samples(void)
2714{
2715 struct thread_data *td;
2716 unsigned int next = ~0U, tmp;
2717 struct timespec now;
2718 int i;
2719
2720 fio_gettime(&now, NULL);
2721
2722 for_each_td(td, i) {
2723 if (!td->o.stats)
2724 continue;
2725 if (in_ramp_time(td) ||
2726 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2727 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2728 continue;
2729 }
2730 if (!td->bw_log ||
2731 (td->bw_log && !per_unit_log(td->bw_log))) {
2732 tmp = add_bw_samples(td, &now);
2733 if (tmp < next)
2734 next = tmp;
2735 }
2736 if (!td->iops_log ||
2737 (td->iops_log && !per_unit_log(td->iops_log))) {
2738 tmp = add_iops_samples(td, &now);
2739 if (tmp < next)
2740 next = tmp;
2741 }
2742 }
2743
2744 return next == ~0U ? 0 : next;
2745}
2746
2747void stat_init(void)
2748{
2749 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
2750}
2751
2752void stat_exit(void)
2753{
2754 /*
2755 * When we have the mutex, we know out-of-band access to it
2756 * have ended.
2757 */
2758 fio_sem_down(stat_sem);
2759 fio_sem_remove(stat_sem);
2760}
2761
2762/*
2763 * Called from signal handler. Wake up status thread.
2764 */
2765void show_running_run_stats(void)
2766{
2767 helper_do_stat();
2768}
2769
2770uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2771{
2772 /* Ignore io_u's which span multiple blocks--they will just get
2773 * inaccurate counts. */
2774 int idx = (io_u->offset - io_u->file->file_offset)
2775 / td->o.bs[DDIR_TRIM];
2776 uint32_t *info = &td->ts.block_infos[idx];
2777 assert(idx < td->ts.nr_block_infos);
2778 return info;
2779}
fio - Flexible IO Tester