projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
stat: fixing bw_agg reporting
[fio.git] / stat.c
... / ...
CommitLineData
1#include <stdio.h>
2#include <string.h>
3#include <sys/time.h>
4#include <sys/types.h>
5#include <sys/stat.h>
6#include <dirent.h>
7#include <libgen.h>
8#include <math.h>
9
10#include "fio.h"
11#include "diskutil.h"
12#include "lib/ieee754.h"
13#include "json.h"
14#include "lib/getrusage.h"
15#include "idletime.h"
16
17static struct fio_mutex *stat_mutex;
18
19void update_rusage_stat(struct thread_data *td)
20{
21 struct thread_stat *ts = &td->ts;
22
23 fio_getrusage(&td->ru_end);
24 ts->usr_time += mtime_since(&td->ru_start.ru_utime,
25 &td->ru_end.ru_utime);
26 ts->sys_time += mtime_since(&td->ru_start.ru_stime,
27 &td->ru_end.ru_stime);
28 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
29 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
30 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
31 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
32
33 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
34}
35
36/*
37 * Given a latency, return the index of the corresponding bucket in
38 * the structure tracking percentiles.
39 *
40 * (1) find the group (and error bits) that the value (latency)
41 * belongs to by looking at its MSB. (2) find the bucket number in the
42 * group by looking at the index bits.
43 *
44 */
45static unsigned int plat_val_to_idx(unsigned int val)
46{
47 unsigned int msb, error_bits, base, offset, idx;
48
49 /* Find MSB starting from bit 0 */
50 if (val == 0)
51 msb = 0;
52 else
53 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
54
55 /*
56 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
57 * all bits of the sample as index
58 */
59 if (msb <= FIO_IO_U_PLAT_BITS)
60 return val;
61
62 /* Compute the number of error bits to discard*/
63 error_bits = msb - FIO_IO_U_PLAT_BITS;
64
65 /* Compute the number of buckets before the group */
66 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
67
68 /*
69 * Discard the error bits and apply the mask to find the
70 * index for the buckets in the group
71 */
72 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
73
74 /* Make sure the index does not exceed (array size - 1) */
75 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
76 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
77
78 return idx;
79}
80
81/*
82 * Convert the given index of the bucket array to the value
83 * represented by the bucket
84 */
85static unsigned int plat_idx_to_val(unsigned int idx)
86{
87 unsigned int error_bits, k, base;
88
89 assert(idx < FIO_IO_U_PLAT_NR);
90
91 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
92 * all bits of the sample as index */
93 if (idx < (FIO_IO_U_PLAT_VAL << 1))
94 return idx;
95
96 /* Find the group and compute the minimum value of that group */
97 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
98 base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
99
100 /* Find its bucket number of the group */
101 k = idx % FIO_IO_U_PLAT_VAL;
102
103 /* Return the mean of the range of the bucket */
104 return base + ((k + 0.5) * (1 << error_bits));
105}
106
107static int double_cmp(const void *a, const void *b)
108{
109 const fio_fp64_t fa = *(const fio_fp64_t *) a;
110 const fio_fp64_t fb = *(const fio_fp64_t *) b;
111 int cmp = 0;
112
113 if (fa.u.f > fb.u.f)
114 cmp = 1;
115 else if (fa.u.f < fb.u.f)
116 cmp = -1;
117
118 return cmp;
119}
120
121unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
122 fio_fp64_t *plist, unsigned int **output,
123 unsigned int *maxv, unsigned int *minv)
124{
125 unsigned long sum = 0;
126 unsigned int len, i, j = 0;
127 unsigned int oval_len = 0;
128 unsigned int *ovals = NULL;
129 int is_last;
130
131 *minv = -1U;
132 *maxv = 0;
133
134 len = 0;
135 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
136 len++;
137
138 if (!len)
139 return 0;
140
141 /*
142 * Sort the percentile list. Note that it may already be sorted if
143 * we are using the default values, but since it's a short list this
144 * isn't a worry. Also note that this does not work for NaN values.
145 */
146 if (len > 1)
147 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
148
149 /*
150 * Calculate bucket values, note down max and min values
151 */
152 is_last = 0;
153 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
154 sum += io_u_plat[i];
155 while (sum >= (plist[j].u.f / 100.0 * nr)) {
156 assert(plist[j].u.f <= 100.0);
157
158 if (j == oval_len) {
159 oval_len += 100;
160 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
161 }
162
163 ovals[j] = plat_idx_to_val(i);
164 if (ovals[j] < *minv)
165 *minv = ovals[j];
166 if (ovals[j] > *maxv)
167 *maxv = ovals[j];
168
169 is_last = (j == len - 1);
170 if (is_last)
171 break;
172
173 j++;
174 }
175 }
176
177 *output = ovals;
178 return len;
179}
180
181/*
182 * Find and display the p-th percentile of clat
183 */
184static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
185 fio_fp64_t *plist, unsigned int precision)
186{
187 unsigned int len, j = 0, minv, maxv;
188 unsigned int *ovals;
189 int is_last, per_line, scale_down;
190 char fmt[32];
191
192 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
193 if (!len)
194 goto out;
195
196 /*
197 * We default to usecs, but if the value range is such that we
198 * should scale down to msecs, do that.
199 */
200 if (minv > 2000 && maxv > 99999) {
201 scale_down = 1;
202 log_info(" clat percentiles (msec):\n |");
203 } else {
204 scale_down = 0;
205 log_info(" clat percentiles (usec):\n |");
206 }
207
208 snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
209 per_line = (80 - 7) / (precision + 14);
210
211 for (j = 0; j < len; j++) {
212 char fbuf[16], *ptr = fbuf;
213
214 /* for formatting */
215 if (j != 0 && (j % per_line) == 0)
216 log_info(" |");
217
218 /* end of the list */
219 is_last = (j == len - 1);
220
221 if (plist[j].u.f < 10.0)
222 ptr += sprintf(fbuf, " ");
223
224 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
225
226 if (scale_down)
227 ovals[j] = (ovals[j] + 999) / 1000;
228
229 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
230
231 if (is_last)
232 break;
233
234 if ((j % per_line) == per_line - 1) /* for formatting */
235 log_info("\n");
236 }
237
238out:
239 if (ovals)
240 free(ovals);
241}
242
243int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
244 double *mean, double *dev)
245{
246 double n = (double) is->samples;
247
248 if (n == 0)
249 return 0;
250
251 *min = is->min_val;
252 *max = is->max_val;
253 *mean = is->mean.u.f;
254
255 if (n > 1.0)
256 *dev = sqrt(is->S.u.f / (n - 1.0));
257 else
258 *dev = 0;
259
260 return 1;
261}
262
263void show_group_stats(struct group_run_stats *rs)
264{
265 char *p1, *p2, *p3, *p4;
266 const char *ddir_str[] = { " READ", " WRITE" , " TRIM"};
267 int i;
268
269 log_info("\nRun status group %d (all jobs):\n", rs->groupid);
270
271 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
272 const int i2p = is_power_of_2(rs->kb_base);
273
274 if (!rs->max_run[i])
275 continue;
276
277 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
278 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
279 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
280 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
281
282 log_info("%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
283 " mint=%llumsec, maxt=%llumsec\n",
284 rs->unified_rw_rep ? " MIXED" : ddir_str[i],
285 p1, p2, p3, p4,
286 (unsigned long long) rs->min_run[i],
287 (unsigned long long) rs->max_run[i]);
288
289 free(p1);
290 free(p2);
291 free(p3);
292 free(p4);
293 }
294}
295
296void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
297{
298 int i;
299
300 /*
301 * Do depth distribution calculations
302 */
303 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
304 if (total) {
305 io_u_dist[i] = (double) map[i] / (double) total;
306 io_u_dist[i] *= 100.0;
307 if (io_u_dist[i] < 0.1 && map[i])
308 io_u_dist[i] = 0.1;
309 } else
310 io_u_dist[i] = 0.0;
311 }
312}
313
314static void stat_calc_lat(struct thread_stat *ts, double *dst,
315 unsigned int *src, int nr)
316{
317 unsigned long total = ddir_rw_sum(ts->total_io_u);
318 int i;
319
320 /*
321 * Do latency distribution calculations
322 */
323 for (i = 0; i < nr; i++) {
324 if (total) {
325 dst[i] = (double) src[i] / (double) total;
326 dst[i] *= 100.0;
327 if (dst[i] < 0.01 && src[i])
328 dst[i] = 0.01;
329 } else
330 dst[i] = 0.0;
331 }
332}
333
334void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
335{
336 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
337}
338
339void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
340{
341 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
342}
343
344static void display_lat(const char *name, unsigned long min, unsigned long max,
345 double mean, double dev)
346{
347 const char *base = "(usec)";
348 char *minp, *maxp;
349
350 if (!usec_to_msec(&min, &max, &mean, &dev))
351 base = "(msec)";
352
353 minp = num2str(min, 6, 1, 0, 0);
354 maxp = num2str(max, 6, 1, 0, 0);
355
356 log_info(" %s %s: min=%s, max=%s, avg=%5.02f,"
357 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
358
359 free(minp);
360 free(maxp);
361}
362
363static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
364 int ddir)
365{
366 const char *ddir_str[] = { "read ", "write", "trim" };
367 unsigned long min, max, runt;
368 unsigned long long bw, iops;
369 double mean, dev;
370 char *io_p, *bw_p, *iops_p;
371 int i2p;
372
373 assert(ddir_rw(ddir));
374
375 if (!ts->runtime[ddir])
376 return;
377
378 i2p = is_power_of_2(rs->kb_base);
379 runt = ts->runtime[ddir];
380
381 bw = (1000 * ts->io_bytes[ddir]) / runt;
382 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
383 bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
384
385 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
386 iops_p = num2str(iops, 6, 1, 0, 0);
387
388 log_info(" %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
389 rs->unified_rw_rep ? "mixed" : ddir_str[ddir],
390 io_p, bw_p, iops_p,
391 (unsigned long long) ts->runtime[ddir]);
392
393 free(io_p);
394 free(bw_p);
395 free(iops_p);
396
397 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
398 display_lat("slat", min, max, mean, dev);
399 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
400 display_lat("clat", min, max, mean, dev);
401 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
402 display_lat(" lat", min, max, mean, dev);
403
404 if (ts->clat_percentiles) {
405 show_clat_percentiles(ts->io_u_plat[ddir],
406 ts->clat_stat[ddir].samples,
407 ts->percentile_list,
408 ts->percentile_precision);
409 }
410 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
411 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
412 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
413
414 if (rs->unit_base == 1) {
415 min *= 8.0;
416 max *= 8.0;
417 mean *= 8.0;
418 dev *= 8.0;
419 }
420
421 if (rs->agg[ddir]) {
422 p_of_agg = mean * 100 / (double) rs->agg[ddir];
423 if (p_of_agg > 100.0)
424 p_of_agg = 100.0;
425 }
426
427 if (mean > fkb_base * fkb_base) {
428 min /= fkb_base;
429 max /= fkb_base;
430 mean /= fkb_base;
431 dev /= fkb_base;
432 bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
433 }
434
435 log_info(" bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
436 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
437 p_of_agg, mean, dev);
438 }
439}
440
441static int show_lat(double *io_u_lat, int nr, const char **ranges,
442 const char *msg)
443{
444 int new_line = 1, i, line = 0, shown = 0;
445
446 for (i = 0; i < nr; i++) {
447 if (io_u_lat[i] <= 0.0)
448 continue;
449 shown = 1;
450 if (new_line) {
451 if (line)
452 log_info("\n");
453 log_info(" lat (%s) : ", msg);
454 new_line = 0;
455 line = 0;
456 }
457 if (line)
458 log_info(", ");
459 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
460 line++;
461 if (line == 5)
462 new_line = 1;
463 }
464
465 if (shown)
466 log_info("\n");
467
468 return shown;
469}
470
471static void show_lat_u(double *io_u_lat_u)
472{
473 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
474 "250=", "500=", "750=", "1000=", };
475
476 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
477}
478
479static void show_lat_m(double *io_u_lat_m)
480{
481 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
482 "250=", "500=", "750=", "1000=", "2000=",
483 ">=2000=", };
484
485 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
486}
487
488static void show_latencies(struct thread_stat *ts)
489{
490 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
491 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
492
493 stat_calc_lat_u(ts, io_u_lat_u);
494 stat_calc_lat_m(ts, io_u_lat_m);
495
496 show_lat_u(io_u_lat_u);
497 show_lat_m(io_u_lat_m);
498}
499
500void show_thread_status_normal(struct thread_stat *ts, struct group_run_stats *rs)
501{
502 double usr_cpu, sys_cpu;
503 unsigned long runtime;
504 double io_u_dist[FIO_IO_U_MAP_NR];
505 time_t time_p;
506 char time_buf[64];
507
508 if (!(ts->io_bytes[DDIR_READ] + ts->io_bytes[DDIR_WRITE] +
509 ts->io_bytes[DDIR_TRIM]) && !(ts->total_io_u[DDIR_READ] +
510 ts->total_io_u[DDIR_WRITE] + ts->total_io_u[DDIR_TRIM]))
511 return;
512
513 time(&time_p);
514 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
515
516 if (!ts->error) {
517 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
518 ts->name, ts->groupid, ts->members,
519 ts->error, (int) ts->pid, time_buf);
520 } else {
521 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
522 ts->name, ts->groupid, ts->members,
523 ts->error, ts->verror, (int) ts->pid,
524 time_buf);
525 }
526
527 if (strlen(ts->description))
528 log_info(" Description : [%s]\n", ts->description);
529
530 if (ts->io_bytes[DDIR_READ])
531 show_ddir_status(rs, ts, DDIR_READ);
532 if (ts->io_bytes[DDIR_WRITE])
533 show_ddir_status(rs, ts, DDIR_WRITE);
534 if (ts->io_bytes[DDIR_TRIM])
535 show_ddir_status(rs, ts, DDIR_TRIM);
536
537 show_latencies(ts);
538
539 runtime = ts->total_run_time;
540 if (runtime) {
541 double runt = (double) runtime;
542
543 usr_cpu = (double) ts->usr_time * 100 / runt;
544 sys_cpu = (double) ts->sys_time * 100 / runt;
545 } else {
546 usr_cpu = 0;
547 sys_cpu = 0;
548 }
549
550 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
551 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
552 (unsigned long long) ts->ctx,
553 (unsigned long long) ts->majf,
554 (unsigned long long) ts->minf);
555
556 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
557 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
558 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
559 io_u_dist[1], io_u_dist[2],
560 io_u_dist[3], io_u_dist[4],
561 io_u_dist[5], io_u_dist[6]);
562
563 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
564 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
565 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
566 io_u_dist[1], io_u_dist[2],
567 io_u_dist[3], io_u_dist[4],
568 io_u_dist[5], io_u_dist[6]);
569 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
570 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
571 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
572 io_u_dist[1], io_u_dist[2],
573 io_u_dist[3], io_u_dist[4],
574 io_u_dist[5], io_u_dist[6]);
575 log_info(" issued : total=r=%llu/w=%llu/d=%llu,"
576 " short=r=%llu/w=%llu/d=%llu\n",
577 (unsigned long long) ts->total_io_u[0],
578 (unsigned long long) ts->total_io_u[1],
579 (unsigned long long) ts->total_io_u[2],
580 (unsigned long long) ts->short_io_u[0],
581 (unsigned long long) ts->short_io_u[1],
582 (unsigned long long) ts->short_io_u[2]);
583 if (ts->continue_on_error) {
584 log_info(" errors : total=%llu, first_error=%d/<%s>\n",
585 (unsigned long long)ts->total_err_count,
586 ts->first_error,
587 strerror(ts->first_error));
588 }
589 if (ts->latency_depth) {
590 log_info(" latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
591 (unsigned long long)ts->latency_target,
592 (unsigned long long)ts->latency_window,
593 ts->latency_percentile.u.f,
594 ts->latency_depth);
595 }
596}
597
598static void show_ddir_status_terse(struct thread_stat *ts,
599 struct group_run_stats *rs, int ddir)
600{
601 unsigned long min, max;
602 unsigned long long bw, iops;
603 unsigned int *ovals = NULL;
604 double mean, dev;
605 unsigned int len, minv, maxv;
606 int i;
607
608 assert(ddir_rw(ddir));
609
610 iops = bw = 0;
611 if (ts->runtime[ddir]) {
612 uint64_t runt = ts->runtime[ddir];
613
614 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
615 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
616 }
617
618 log_info(";%llu;%llu;%llu;%llu",
619 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
620 (unsigned long long) ts->runtime[ddir]);
621
622 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
623 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
624 else
625 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
626
627 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
628 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
629 else
630 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
631
632 if (ts->clat_percentiles) {
633 len = calc_clat_percentiles(ts->io_u_plat[ddir],
634 ts->clat_stat[ddir].samples,
635 ts->percentile_list, &ovals, &maxv,
636 &minv);
637 } else
638 len = 0;
639
640 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
641 if (i >= len) {
642 log_info(";0%%=0");
643 continue;
644 }
645 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
646 }
647
648 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
649 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
650 else
651 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
652
653 if (ovals)
654 free(ovals);
655
656 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
657 double p_of_agg = 100.0;
658
659 if (rs->agg[ddir]) {
660 p_of_agg = mean * 100 / (double) rs->agg[ddir];
661 if (p_of_agg > 100.0)
662 p_of_agg = 100.0;
663 }
664
665 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
666 } else
667 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
668}
669
670static void add_ddir_status_json(struct thread_stat *ts,
671 struct group_run_stats *rs, int ddir, struct json_object *parent)
672{
673 unsigned long min, max;
674 unsigned long long bw, iops;
675 unsigned int *ovals = NULL;
676 double mean, dev;
677 unsigned int len, minv, maxv;
678 int i;
679 const char *ddirname[] = {"read", "write", "trim"};
680 struct json_object *dir_object, *tmp_object, *percentile_object;
681 char buf[120];
682 double p_of_agg = 100.0;
683
684 assert(ddir_rw(ddir));
685
686 if (ts->unified_rw_rep && ddir != DDIR_READ)
687 return;
688
689 dir_object = json_create_object();
690 json_object_add_value_object(parent,
691 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
692
693 iops = bw = 0;
694 if (ts->runtime[ddir]) {
695 uint64_t runt = ts->runtime[ddir];
696
697 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
698 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
699 }
700
701 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
702 json_object_add_value_int(dir_object, "bw", bw);
703 json_object_add_value_int(dir_object, "iops", iops);
704 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
705
706 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
707 min = max = 0;
708 mean = dev = 0.0;
709 }
710 tmp_object = json_create_object();
711 json_object_add_value_object(dir_object, "slat", tmp_object);
712 json_object_add_value_int(tmp_object, "min", min);
713 json_object_add_value_int(tmp_object, "max", max);
714 json_object_add_value_float(tmp_object, "mean", mean);
715 json_object_add_value_float(tmp_object, "stddev", dev);
716
717 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
718 min = max = 0;
719 mean = dev = 0.0;
720 }
721 tmp_object = json_create_object();
722 json_object_add_value_object(dir_object, "clat", tmp_object);
723 json_object_add_value_int(tmp_object, "min", min);
724 json_object_add_value_int(tmp_object, "max", max);
725 json_object_add_value_float(tmp_object, "mean", mean);
726 json_object_add_value_float(tmp_object, "stddev", dev);
727
728 if (ts->clat_percentiles) {
729 len = calc_clat_percentiles(ts->io_u_plat[ddir],
730 ts->clat_stat[ddir].samples,
731 ts->percentile_list, &ovals, &maxv,
732 &minv);
733 } else
734 len = 0;
735
736 percentile_object = json_create_object();
737 json_object_add_value_object(tmp_object, "percentile", percentile_object);
738 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
739 if (i >= len) {
740 json_object_add_value_int(percentile_object, "0.00", 0);
741 continue;
742 }
743 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
744 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
745 }
746
747 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
748 min = max = 0;
749 mean = dev = 0.0;
750 }
751 tmp_object = json_create_object();
752 json_object_add_value_object(dir_object, "lat", tmp_object);
753 json_object_add_value_int(tmp_object, "min", min);
754 json_object_add_value_int(tmp_object, "max", max);
755 json_object_add_value_float(tmp_object, "mean", mean);
756 json_object_add_value_float(tmp_object, "stddev", dev);
757 if (ovals)
758 free(ovals);
759
760 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
761 if (rs->agg[ddir]) {
762 p_of_agg = mean * 100 / (double) rs->agg[ddir];
763 if (p_of_agg > 100.0)
764 p_of_agg = 100.0;
765 }
766 } else {
767 min = max = 0;
768 p_of_agg = mean = dev = 0.0;
769 }
770 json_object_add_value_int(dir_object, "bw_min", min);
771 json_object_add_value_int(dir_object, "bw_max", max);
772 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
773 json_object_add_value_float(dir_object, "bw_mean", mean);
774 json_object_add_value_float(dir_object, "bw_dev", dev);
775}
776
777static void show_thread_status_terse_v2(struct thread_stat *ts,
778 struct group_run_stats *rs)
779{
780 double io_u_dist[FIO_IO_U_MAP_NR];
781 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
782 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
783 double usr_cpu, sys_cpu;
784 int i;
785
786 /* General Info */
787 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
788 /* Log Read Status */
789 show_ddir_status_terse(ts, rs, DDIR_READ);
790 /* Log Write Status */
791 show_ddir_status_terse(ts, rs, DDIR_WRITE);
792 /* Log Trim Status */
793 show_ddir_status_terse(ts, rs, DDIR_TRIM);
794
795 /* CPU Usage */
796 if (ts->total_run_time) {
797 double runt = (double) ts->total_run_time;
798
799 usr_cpu = (double) ts->usr_time * 100 / runt;
800 sys_cpu = (double) ts->sys_time * 100 / runt;
801 } else {
802 usr_cpu = 0;
803 sys_cpu = 0;
804 }
805
806 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
807 (unsigned long long) ts->ctx,
808 (unsigned long long) ts->majf,
809 (unsigned long long) ts->minf);
810
811 /* Calc % distribution of IO depths, usecond, msecond latency */
812 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
813 stat_calc_lat_u(ts, io_u_lat_u);
814 stat_calc_lat_m(ts, io_u_lat_m);
815
816 /* Only show fixed 7 I/O depth levels*/
817 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
818 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
819 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
820
821 /* Microsecond latency */
822 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
823 log_info(";%3.2f%%", io_u_lat_u[i]);
824 /* Millisecond latency */
825 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
826 log_info(";%3.2f%%", io_u_lat_m[i]);
827 /* Additional output if continue_on_error set - default off*/
828 if (ts->continue_on_error)
829 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
830 log_info("\n");
831
832 /* Additional output if description is set */
833 if (ts->description)
834 log_info(";%s", ts->description);
835
836 log_info("\n");
837}
838
839static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
840 struct group_run_stats *rs, int ver)
841{
842 double io_u_dist[FIO_IO_U_MAP_NR];
843 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
844 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
845 double usr_cpu, sys_cpu;
846 int i;
847
848 /* General Info */
849 log_info("%d;%s;%s;%d;%d", ver, fio_version_string,
850 ts->name, ts->groupid, ts->error);
851 /* Log Read Status */
852 show_ddir_status_terse(ts, rs, DDIR_READ);
853 /* Log Write Status */
854 show_ddir_status_terse(ts, rs, DDIR_WRITE);
855 /* Log Trim Status */
856 if (ver == 4)
857 show_ddir_status_terse(ts, rs, DDIR_TRIM);
858
859 /* CPU Usage */
860 if (ts->total_run_time) {
861 double runt = (double) ts->total_run_time;
862
863 usr_cpu = (double) ts->usr_time * 100 / runt;
864 sys_cpu = (double) ts->sys_time * 100 / runt;
865 } else {
866 usr_cpu = 0;
867 sys_cpu = 0;
868 }
869
870 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
871 (unsigned long long) ts->ctx,
872 (unsigned long long) ts->majf,
873 (unsigned long long) ts->minf);
874
875 /* Calc % distribution of IO depths, usecond, msecond latency */
876 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
877 stat_calc_lat_u(ts, io_u_lat_u);
878 stat_calc_lat_m(ts, io_u_lat_m);
879
880 /* Only show fixed 7 I/O depth levels*/
881 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
882 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
883 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
884
885 /* Microsecond latency */
886 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
887 log_info(";%3.2f%%", io_u_lat_u[i]);
888 /* Millisecond latency */
889 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
890 log_info(";%3.2f%%", io_u_lat_m[i]);
891
892 /* disk util stats, if any */
893 if (is_backend)
894 show_disk_util(1, NULL);
895
896 /* Additional output if continue_on_error set - default off*/
897 if (ts->continue_on_error)
898 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
899
900 /* Additional output if description is set */
901 if (strlen(ts->description))
902 log_info(";%s", ts->description);
903
904 log_info("\n");
905}
906
907static struct json_object *show_thread_status_json(struct thread_stat *ts,
908 struct group_run_stats *rs)
909{
910 struct json_object *root, *tmp;
911 double io_u_dist[FIO_IO_U_MAP_NR];
912 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
913 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
914 double usr_cpu, sys_cpu;
915 int i;
916
917 root = json_create_object();
918 json_object_add_value_string(root, "jobname", ts->name);
919 json_object_add_value_int(root, "groupid", ts->groupid);
920 json_object_add_value_int(root, "error", ts->error);
921
922 add_ddir_status_json(ts, rs, DDIR_READ, root);
923 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
924 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
925
926 /* CPU Usage */
927 if (ts->total_run_time) {
928 double runt = (double) ts->total_run_time;
929
930 usr_cpu = (double) ts->usr_time * 100 / runt;
931 sys_cpu = (double) ts->sys_time * 100 / runt;
932 } else {
933 usr_cpu = 0;
934 sys_cpu = 0;
935 }
936 json_object_add_value_float(root, "usr_cpu", usr_cpu);
937 json_object_add_value_float(root, "sys_cpu", sys_cpu);
938 json_object_add_value_int(root, "ctx", ts->ctx);
939 json_object_add_value_int(root, "majf", ts->majf);
940 json_object_add_value_int(root, "minf", ts->minf);
941
942
943 /* Calc % distribution of IO depths, usecond, msecond latency */
944 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
945 stat_calc_lat_u(ts, io_u_lat_u);
946 stat_calc_lat_m(ts, io_u_lat_m);
947
948 tmp = json_create_object();
949 json_object_add_value_object(root, "iodepth_level", tmp);
950 /* Only show fixed 7 I/O depth levels*/
951 for (i = 0; i < 7; i++) {
952 char name[20];
953 if (i < 6)
954 snprintf(name, 20, "%d", 1 << i);
955 else
956 snprintf(name, 20, ">=%d", 1 << i);
957 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
958 }
959
960 tmp = json_create_object();
961 json_object_add_value_object(root, "latency_us", tmp);
962 /* Microsecond latency */
963 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
964 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
965 "250", "500", "750", "1000", };
966 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
967 }
968 /* Millisecond latency */
969 tmp = json_create_object();
970 json_object_add_value_object(root, "latency_ms", tmp);
971 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
972 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
973 "250", "500", "750", "1000", "2000",
974 ">=2000", };
975 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
976 }
977
978 /* Additional output if continue_on_error set - default off*/
979 if (ts->continue_on_error) {
980 json_object_add_value_int(root, "total_err", ts->total_err_count);
981 json_object_add_value_int(root, "first_error", ts->first_error);
982 }
983
984 if (ts->latency_depth) {
985 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
986 json_object_add_value_int(root, "latency_target", ts->latency_target);
987 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
988 json_object_add_value_int(root, "latency_window", ts->latency_window);
989 }
990
991 /* Additional output if description is set */
992 if (strlen(ts->description))
993 json_object_add_value_string(root, "desc", ts->description);
994
995 return root;
996}
997
998static void show_thread_status_terse(struct thread_stat *ts,
999 struct group_run_stats *rs)
1000{
1001 if (terse_version == 2)
1002 show_thread_status_terse_v2(ts, rs);
1003 else if (terse_version == 3 || terse_version == 4)
1004 show_thread_status_terse_v3_v4(ts, rs, terse_version);
1005 else
1006 log_err("fio: bad terse version!? %d\n", terse_version);
1007}
1008
1009struct json_object *show_thread_status(struct thread_stat *ts,
1010 struct group_run_stats *rs)
1011{
1012 if (output_format == FIO_OUTPUT_TERSE)
1013 show_thread_status_terse(ts, rs);
1014 else if (output_format == FIO_OUTPUT_JSON)
1015 return(show_thread_status_json(ts, rs));
1016 else
1017 show_thread_status_normal(ts, rs);
1018 return NULL;
1019}
1020
1021static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
1022{
1023 double mean, S;
1024
1025 if (src->samples == 0)
1026 return;
1027
1028 dst->min_val = min(dst->min_val, src->min_val);
1029 dst->max_val = max(dst->max_val, src->max_val);
1030
1031 /*
1032 * Compute new mean and S after the merge
1033 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1034 * #Parallel_algorithm>
1035 */
1036 if (nr == 1) {
1037 mean = src->mean.u.f;
1038 S = src->S.u.f;
1039 } else {
1040 double delta = src->mean.u.f - dst->mean.u.f;
1041
1042 mean = ((src->mean.u.f * src->samples) +
1043 (dst->mean.u.f * dst->samples)) /
1044 (dst->samples + src->samples);
1045
1046 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1047 (dst->samples * src->samples) /
1048 (dst->samples + src->samples);
1049 }
1050
1051 dst->samples += src->samples;
1052 dst->mean.u.f = mean;
1053 dst->S.u.f = S;
1054}
1055
1056void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1057{
1058 int i;
1059
1060 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1061 if (dst->max_run[i] < src->max_run[i])
1062 dst->max_run[i] = src->max_run[i];
1063 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1064 dst->min_run[i] = src->min_run[i];
1065 if (dst->max_bw[i] < src->max_bw[i])
1066 dst->max_bw[i] = src->max_bw[i];
1067 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1068 dst->min_bw[i] = src->min_bw[i];
1069
1070 dst->io_kb[i] += src->io_kb[i];
1071 dst->agg[i] += src->agg[i];
1072 }
1073
1074}
1075
1076void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1077{
1078 int l, k;
1079
1080 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1081 if (!dst->unified_rw_rep) {
1082 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1083 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1084 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1085 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1086
1087 dst->io_bytes[l] += src->io_bytes[l];
1088
1089 if (dst->runtime[l] < src->runtime[l])
1090 dst->runtime[l] = src->runtime[l];
1091 } else {
1092 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1093 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1094 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1095 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1096
1097 dst->io_bytes[0] += src->io_bytes[l];
1098
1099 if (dst->runtime[0] < src->runtime[l])
1100 dst->runtime[0] = src->runtime[l];
1101 }
1102 }
1103
1104 dst->usr_time += src->usr_time;
1105 dst->sys_time += src->sys_time;
1106 dst->ctx += src->ctx;
1107 dst->majf += src->majf;
1108 dst->minf += src->minf;
1109
1110 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1111 dst->io_u_map[k] += src->io_u_map[k];
1112 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1113 dst->io_u_submit[k] += src->io_u_submit[k];
1114 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1115 dst->io_u_complete[k] += src->io_u_complete[k];
1116 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1117 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1118 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1119 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1120
1121 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1122 if (!dst->unified_rw_rep) {
1123 dst->total_io_u[k] += src->total_io_u[k];
1124 dst->short_io_u[k] += src->short_io_u[k];
1125 } else {
1126 dst->total_io_u[0] += src->total_io_u[k];
1127 dst->short_io_u[0] += src->short_io_u[k];
1128 }
1129 }
1130
1131 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1132 int m;
1133
1134 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1135 if (!dst->unified_rw_rep)
1136 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1137 else
1138 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1139 }
1140 }
1141
1142 dst->total_run_time += src->total_run_time;
1143 dst->total_submit += src->total_submit;
1144 dst->total_complete += src->total_complete;
1145}
1146
1147void init_group_run_stat(struct group_run_stats *gs)
1148{
1149 int i;
1150 memset(gs, 0, sizeof(*gs));
1151
1152 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1153 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1154}
1155
1156void init_thread_stat(struct thread_stat *ts)
1157{
1158 int j;
1159
1160 memset(ts, 0, sizeof(*ts));
1161
1162 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1163 ts->lat_stat[j].min_val = -1UL;
1164 ts->clat_stat[j].min_val = -1UL;
1165 ts->slat_stat[j].min_val = -1UL;
1166 ts->bw_stat[j].min_val = -1UL;
1167 }
1168 ts->groupid = -1;
1169}
1170
1171static void __show_run_stats(void)
1172{
1173 struct group_run_stats *runstats, *rs;
1174 struct thread_data *td;
1175 struct thread_stat *threadstats, *ts;
1176 int i, j, nr_ts, last_ts, idx;
1177 int kb_base_warned = 0;
1178 int unit_base_warned = 0;
1179 struct json_object *root = NULL;
1180 struct json_array *array = NULL;
1181
1182 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1183
1184 for (i = 0; i < groupid + 1; i++)
1185 init_group_run_stat(&runstats[i]);
1186
1187 /*
1188 * find out how many threads stats we need. if group reporting isn't
1189 * enabled, it's one-per-td.
1190 */
1191 nr_ts = 0;
1192 last_ts = -1;
1193 for_each_td(td, i) {
1194 if (!td->o.group_reporting) {
1195 nr_ts++;
1196 continue;
1197 }
1198 if (last_ts == td->groupid)
1199 continue;
1200
1201 last_ts = td->groupid;
1202 nr_ts++;
1203 }
1204
1205 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1206
1207 for (i = 0; i < nr_ts; i++)
1208 init_thread_stat(&threadstats[i]);
1209
1210 j = 0;
1211 last_ts = -1;
1212 idx = 0;
1213 for_each_td(td, i) {
1214 if (idx && (!td->o.group_reporting ||
1215 (td->o.group_reporting && last_ts != td->groupid))) {
1216 idx = 0;
1217 j++;
1218 }
1219
1220 last_ts = td->groupid;
1221
1222 ts = &threadstats[j];
1223
1224 ts->clat_percentiles = td->o.clat_percentiles;
1225 ts->percentile_precision = td->o.percentile_precision;
1226 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1227
1228 idx++;
1229 ts->members++;
1230
1231 if (ts->groupid == -1) {
1232 /*
1233 * These are per-group shared already
1234 */
1235 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1236 if (td->o.description)
1237 strncpy(ts->description, td->o.description,
1238 FIO_JOBDESC_SIZE - 1);
1239 else
1240 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1241
1242 /*
1243 * If multiple entries in this group, this is
1244 * the first member.
1245 */
1246 ts->thread_number = td->thread_number;
1247 ts->groupid = td->groupid;
1248
1249 /*
1250 * first pid in group, not very useful...
1251 */
1252 ts->pid = td->pid;
1253
1254 ts->kb_base = td->o.kb_base;
1255 ts->unit_base = td->o.unit_base;
1256 ts->unified_rw_rep = td->o.unified_rw_rep;
1257 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1258 log_info("fio: kb_base differs for jobs in group, using"
1259 " %u as the base\n", ts->kb_base);
1260 kb_base_warned = 1;
1261 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1262 log_info("fio: unit_base differs for jobs in group, using"
1263 " %u as the base\n", ts->unit_base);
1264 unit_base_warned = 1;
1265 }
1266
1267 ts->continue_on_error = td->o.continue_on_error;
1268 ts->total_err_count += td->total_err_count;
1269 ts->first_error = td->first_error;
1270 if (!ts->error) {
1271 if (!td->error && td->o.continue_on_error &&
1272 td->first_error) {
1273 ts->error = td->first_error;
1274 strcpy(ts->verror, td->verror);
1275 } else if (td->error) {
1276 ts->error = td->error;
1277 strcpy(ts->verror, td->verror);
1278 }
1279 }
1280
1281 ts->latency_depth = td->latency_qd;
1282 ts->latency_target = td->o.latency_target;
1283 ts->latency_percentile = td->o.latency_percentile;
1284 ts->latency_window = td->o.latency_window;
1285
1286 sum_thread_stats(ts, &td->ts, idx);
1287 }
1288
1289 for (i = 0; i < nr_ts; i++) {
1290 unsigned long long bw;
1291
1292 ts = &threadstats[i];
1293 rs = &runstats[ts->groupid];
1294 rs->kb_base = ts->kb_base;
1295 rs->unit_base = ts->unit_base;
1296 rs->unified_rw_rep += ts->unified_rw_rep;
1297
1298 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1299 if (!ts->runtime[j])
1300 continue;
1301 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1302 rs->min_run[j] = ts->runtime[j];
1303 if (ts->runtime[j] > rs->max_run[j])
1304 rs->max_run[j] = ts->runtime[j];
1305
1306 bw = 0;
1307 if (ts->runtime[j]) {
1308 unsigned long runt = ts->runtime[j];
1309 unsigned long long kb;
1310
1311 kb = ts->io_bytes[j] / rs->kb_base;
1312 bw = kb * 1000 / runt;
1313 }
1314 if (bw < rs->min_bw[j])
1315 rs->min_bw[j] = bw;
1316 if (bw > rs->max_bw[j])
1317 rs->max_bw[j] = bw;
1318
1319 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1320 }
1321 }
1322
1323 for (i = 0; i < groupid + 1; i++) {
1324 int ddir;
1325
1326 rs = &runstats[i];
1327
1328 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1329 if (rs->max_run[ddir])
1330 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1331 rs->max_run[ddir];
1332 }
1333 }
1334
1335 /*
1336 * don't overwrite last signal output
1337 */
1338 if (output_format == FIO_OUTPUT_NORMAL)
1339 log_info("\n");
1340 else if (output_format == FIO_OUTPUT_JSON) {
1341 root = json_create_object();
1342 json_object_add_value_string(root, "fio version", fio_version_string);
1343 array = json_create_array();
1344 json_object_add_value_array(root, "jobs", array);
1345 }
1346
1347 for (i = 0; i < nr_ts; i++) {
1348 ts = &threadstats[i];
1349 rs = &runstats[ts->groupid];
1350
1351 if (is_backend)
1352 fio_server_send_ts(ts, rs);
1353 else if (output_format == FIO_OUTPUT_TERSE)
1354 show_thread_status_terse(ts, rs);
1355 else if (output_format == FIO_OUTPUT_JSON) {
1356 struct json_object *tmp = show_thread_status_json(ts, rs);
1357 json_array_add_value_object(array, tmp);
1358 } else
1359 show_thread_status_normal(ts, rs);
1360 }
1361 if (output_format == FIO_OUTPUT_JSON) {
1362 /* disk util stats, if any */
1363 show_disk_util(1, root);
1364
1365 show_idle_prof_stats(FIO_OUTPUT_JSON, root);
1366
1367 json_print_object(root);
1368 log_info("\n");
1369 json_free_object(root);
1370 }
1371
1372 for (i = 0; i < groupid + 1; i++) {
1373 rs = &runstats[i];
1374
1375 rs->groupid = i;
1376 if (is_backend)
1377 fio_server_send_gs(rs);
1378 else if (output_format == FIO_OUTPUT_NORMAL)
1379 show_group_stats(rs);
1380 }
1381
1382 if (is_backend)
1383 fio_server_send_du();
1384 else if (output_format == FIO_OUTPUT_NORMAL) {
1385 show_disk_util(0, NULL);
1386 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
1387 }
1388
1389 if ( !(output_format == FIO_OUTPUT_TERSE) && append_terse_output) {
1390 log_info("\nAdditional Terse Output:\n");
1391
1392 for (i = 0; i < nr_ts; i++) {
1393 ts = &threadstats[i];
1394 rs = &runstats[ts->groupid];
1395 show_thread_status_terse(ts, rs);
1396 }
1397 }
1398
1399 log_info_flush();
1400 free(runstats);
1401 free(threadstats);
1402}
1403
1404void show_run_stats(void)
1405{
1406 fio_mutex_down(stat_mutex);
1407 __show_run_stats();
1408 fio_mutex_up(stat_mutex);
1409}
1410
1411static void *__show_running_run_stats(void fio_unused *arg)
1412{
1413 struct thread_data *td;
1414 unsigned long long *rt;
1415 struct timeval tv;
1416 int i;
1417
1418 rt = malloc(thread_number * sizeof(unsigned long long));
1419 fio_gettime(&tv, NULL);
1420
1421 for_each_td(td, i) {
1422 rt[i] = mtime_since(&td->start, &tv);
1423 if (td_read(td) && td->io_bytes[DDIR_READ])
1424 td->ts.runtime[DDIR_READ] += rt[i];
1425 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1426 td->ts.runtime[DDIR_WRITE] += rt[i];
1427 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1428 td->ts.runtime[DDIR_TRIM] += rt[i];
1429
1430 td->update_rusage = 1;
1431 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1432 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1433 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1434 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1435 }
1436
1437 for_each_td(td, i) {
1438 if (td->rusage_sem) {
1439 td->update_rusage = 1;
1440 fio_mutex_down(td->rusage_sem);
1441 }
1442 td->update_rusage = 0;
1443 }
1444
1445 __show_run_stats();
1446
1447 for_each_td(td, i) {
1448 if (td_read(td) && td->io_bytes[DDIR_READ])
1449 td->ts.runtime[DDIR_READ] -= rt[i];
1450 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1451 td->ts.runtime[DDIR_WRITE] -= rt[i];
1452 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1453 td->ts.runtime[DDIR_TRIM] -= rt[i];
1454 }
1455
1456 free(rt);
1457 fio_mutex_up(stat_mutex);
1458 return NULL;
1459}
1460
1461/*
1462 * Called from signal handler. It _should_ be safe to just run this inline
1463 * in the sig handler, but we should be disturbing the system less by just
1464 * creating a thread to do it.
1465 */
1466void show_running_run_stats(void)
1467{
1468 pthread_t thread;
1469
1470 fio_mutex_down(stat_mutex);
1471
1472 if (!pthread_create(&thread, NULL, __show_running_run_stats, NULL)) {
1473 pthread_detach(thread);
1474 return;
1475 }
1476
1477 fio_mutex_up(stat_mutex);
1478}
1479
1480static int status_interval_init;
1481static struct timeval status_time;
1482static int status_file_disabled;
1483
1484#define FIO_STATUS_FILE "fio-dump-status"
1485
1486static int check_status_file(void)
1487{
1488 struct stat sb;
1489 const char *temp_dir;
1490 char fio_status_file_path[PATH_MAX];
1491
1492 if (status_file_disabled)
1493 return 0;
1494
1495 temp_dir = getenv("TMPDIR");
1496 if (temp_dir == NULL)
1497 temp_dir = getenv("TEMP");
1498 if (temp_dir == NULL)
1499 temp_dir = "/tmp";
1500
1501 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1502
1503 if (stat(fio_status_file_path, &sb))
1504 return 0;
1505
1506 if (unlink(fio_status_file_path) < 0) {
1507 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1508 strerror(errno));
1509 log_err("fio: disabling status file updates\n");
1510 status_file_disabled = 1;
1511 }
1512
1513 return 1;
1514}
1515
1516void check_for_running_stats(void)
1517{
1518 if (status_interval) {
1519 if (!status_interval_init) {
1520 fio_gettime(&status_time, NULL);
1521 status_interval_init = 1;
1522 } else if (mtime_since_now(&status_time) >= status_interval) {
1523 show_running_run_stats();
1524 fio_gettime(&status_time, NULL);
1525 return;
1526 }
1527 }
1528 if (check_status_file()) {
1529 show_running_run_stats();
1530 return;
1531 }
1532}
1533
1534static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1535{
1536 double val = data;
1537 double delta;
1538
1539 if (data > is->max_val)
1540 is->max_val = data;
1541 if (data < is->min_val)
1542 is->min_val = data;
1543
1544 delta = val - is->mean.u.f;
1545 if (delta) {
1546 is->mean.u.f += delta / (is->samples + 1.0);
1547 is->S.u.f += delta * (val - is->mean.u.f);
1548 }
1549
1550 is->samples++;
1551}
1552
1553static void __add_log_sample(struct io_log *iolog, unsigned long val,
1554 enum fio_ddir ddir, unsigned int bs,
1555 unsigned long t)
1556{
1557 const int nr_samples = iolog->nr_samples;
1558
1559 if (iolog->disabled)
1560 return;
1561
1562 if (!iolog->nr_samples)
1563 iolog->avg_last = t;
1564
1565 if (iolog->nr_samples == iolog->max_samples) {
1566 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
1567 void *new_log;
1568
1569 new_log = realloc(iolog->log, new_size);
1570 if (!new_log) {
1571 log_err("fio: failed extending iolog! Will stop logging.\n");
1572 iolog->disabled = 1;
1573 return;
1574 }
1575 iolog->log = new_log;
1576 iolog->max_samples <<= 1;
1577 }
1578
1579 iolog->log[nr_samples].val = val;
1580 iolog->log[nr_samples].time = t;
1581 iolog->log[nr_samples].ddir = ddir;
1582 iolog->log[nr_samples].bs = bs;
1583 iolog->nr_samples++;
1584}
1585
1586static inline void reset_io_stat(struct io_stat *ios)
1587{
1588 ios->max_val = ios->min_val = ios->samples = 0;
1589 ios->mean.u.f = ios->S.u.f = 0;
1590}
1591
1592void reset_io_stats(struct thread_data *td)
1593{
1594 struct thread_stat *ts = &td->ts;
1595 int i, j;
1596
1597 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1598 reset_io_stat(&ts->clat_stat[i]);
1599 reset_io_stat(&ts->slat_stat[i]);
1600 reset_io_stat(&ts->lat_stat[i]);
1601 reset_io_stat(&ts->bw_stat[i]);
1602 reset_io_stat(&ts->iops_stat[i]);
1603
1604 ts->io_bytes[i] = 0;
1605 ts->runtime[i] = 0;
1606
1607 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1608 ts->io_u_plat[i][j] = 0;
1609 }
1610
1611 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1612 ts->io_u_map[i] = 0;
1613 ts->io_u_submit[i] = 0;
1614 ts->io_u_complete[i] = 0;
1615 ts->io_u_lat_u[i] = 0;
1616 ts->io_u_lat_m[i] = 0;
1617 ts->total_submit = 0;
1618 ts->total_complete = 0;
1619 }
1620
1621 for (i = 0; i < 3; i++) {
1622 ts->total_io_u[i] = 0;
1623 ts->short_io_u[i] = 0;
1624 }
1625}
1626
1627static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
1628{
1629 /*
1630 * Note an entry in the log. Use the mean from the logged samples,
1631 * making sure to properly round up. Only write a log entry if we
1632 * had actual samples done.
1633 */
1634 if (iolog->avg_window[DDIR_READ].samples) {
1635 unsigned long mr;
1636
1637 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1638 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed);
1639 }
1640 if (iolog->avg_window[DDIR_WRITE].samples) {
1641 unsigned long mw;
1642
1643 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1644 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed);
1645 }
1646 if (iolog->avg_window[DDIR_TRIM].samples) {
1647 unsigned long mw;
1648
1649 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1650 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed);
1651 }
1652
1653 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1654 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1655 reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1656}
1657
1658static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1659 unsigned long val, enum fio_ddir ddir,
1660 unsigned int bs)
1661{
1662 unsigned long elapsed, this_window;
1663
1664 if (!ddir_rw(ddir))
1665 return;
1666
1667 elapsed = mtime_since_now(&td->epoch);
1668
1669 /*
1670 * If no time averaging, just add the log sample.
1671 */
1672 if (!iolog->avg_msec) {
1673 __add_log_sample(iolog, val, ddir, bs, elapsed);
1674 return;
1675 }
1676
1677 /*
1678 * Add the sample. If the time period has passed, then
1679 * add that entry to the log and clear.
1680 */
1681 add_stat_sample(&iolog->avg_window[ddir], val);
1682
1683 /*
1684 * If period hasn't passed, adding the above sample is all we
1685 * need to do.
1686 */
1687 this_window = elapsed - iolog->avg_last;
1688 if (this_window < iolog->avg_msec)
1689 return;
1690
1691 _add_stat_to_log(iolog, elapsed);
1692
1693 iolog->avg_last = elapsed;
1694}
1695
1696void finalize_logs(struct thread_data *td)
1697{
1698 unsigned long elapsed;
1699
1700 elapsed = mtime_since_now(&td->epoch);
1701
1702 if (td->clat_log)
1703 _add_stat_to_log(td->clat_log, elapsed);
1704 if (td->slat_log)
1705 _add_stat_to_log(td->slat_log, elapsed);
1706 if (td->lat_log)
1707 _add_stat_to_log(td->lat_log, elapsed);
1708 if (td->bw_log)
1709 _add_stat_to_log(td->bw_log, elapsed);
1710 if (td->iops_log)
1711 _add_stat_to_log(td->iops_log, elapsed);
1712}
1713
1714void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1715{
1716 struct io_log *iolog;
1717
1718 if (!ddir_rw(ddir))
1719 return;
1720
1721 iolog = agg_io_log[ddir];
1722 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
1723}
1724
1725static void add_clat_percentile_sample(struct thread_stat *ts,
1726 unsigned long usec, enum fio_ddir ddir)
1727{
1728 unsigned int idx = plat_val_to_idx(usec);
1729 assert(idx < FIO_IO_U_PLAT_NR);
1730
1731 ts->io_u_plat[ddir][idx]++;
1732}
1733
1734void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1735 unsigned long usec, unsigned int bs)
1736{
1737 struct thread_stat *ts = &td->ts;
1738
1739 if (!ddir_rw(ddir))
1740 return;
1741
1742 add_stat_sample(&ts->clat_stat[ddir], usec);
1743
1744 if (td->clat_log)
1745 add_log_sample(td, td->clat_log, usec, ddir, bs);
1746
1747 if (ts->clat_percentiles)
1748 add_clat_percentile_sample(ts, usec, ddir);
1749}
1750
1751void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1752 unsigned long usec, unsigned int bs)
1753{
1754 struct thread_stat *ts = &td->ts;
1755
1756 if (!ddir_rw(ddir))
1757 return;
1758
1759 add_stat_sample(&ts->slat_stat[ddir], usec);
1760
1761 if (td->slat_log)
1762 add_log_sample(td, td->slat_log, usec, ddir, bs);
1763}
1764
1765void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1766 unsigned long usec, unsigned int bs)
1767{
1768 struct thread_stat *ts = &td->ts;
1769
1770 if (!ddir_rw(ddir))
1771 return;
1772
1773 add_stat_sample(&ts->lat_stat[ddir], usec);
1774
1775 if (td->lat_log)
1776 add_log_sample(td, td->lat_log, usec, ddir, bs);
1777}
1778
1779void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1780 struct timeval *t)
1781{
1782 struct thread_stat *ts = &td->ts;
1783 unsigned long spent, rate;
1784
1785 if (!ddir_rw(ddir))
1786 return;
1787
1788 spent = mtime_since(&td->bw_sample_time, t);
1789 if (spent < td->o.bw_avg_time)
1790 return;
1791
1792 /*
1793 * Compute both read and write rates for the interval.
1794 */
1795 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1796 uint64_t delta;
1797
1798 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1799 if (!delta)
1800 continue; /* No entries for interval */
1801
1802 rate = delta * 1000 / spent / 1024;
1803 add_stat_sample(&ts->bw_stat[ddir], rate);
1804
1805 if (td->bw_log)
1806 add_log_sample(td, td->bw_log, rate, ddir, bs);
1807
1808 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1809 }
1810
1811 fio_gettime(&td->bw_sample_time, NULL);
1812}
1813
1814void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1815 struct timeval *t)
1816{
1817 struct thread_stat *ts = &td->ts;
1818 unsigned long spent, iops;
1819
1820 if (!ddir_rw(ddir))
1821 return;
1822
1823 spent = mtime_since(&td->iops_sample_time, t);
1824 if (spent < td->o.iops_avg_time)
1825 return;
1826
1827 /*
1828 * Compute both read and write rates for the interval.
1829 */
1830 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1831 uint64_t delta;
1832
1833 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1834 if (!delta)
1835 continue; /* No entries for interval */
1836
1837 iops = (delta * 1000) / spent;
1838 add_stat_sample(&ts->iops_stat[ddir], iops);
1839
1840 if (td->iops_log)
1841 add_log_sample(td, td->iops_log, iops, ddir, bs);
1842
1843 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1844 }
1845
1846 fio_gettime(&td->iops_sample_time, NULL);
1847}
1848
1849void stat_init(void)
1850{
1851 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1852}
1853
1854void stat_exit(void)
1855{
1856 /*
1857 * When we have the mutex, we know out-of-band access to it
1858 * have ended.
1859 */
1860 fio_mutex_down(stat_mutex);
1861 fio_mutex_remove(stat_mutex);
1862}
fio - Flexible IO Tester