projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Add json output for client/server mode
[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
500
501void show_thread_status_normal(struct thread_stat *ts, struct group_run_stats *rs)
502{
503 double usr_cpu, sys_cpu;
504 unsigned long runtime;
505 double io_u_dist[FIO_IO_U_MAP_NR];
506 time_t time_p;
507 char time_buf[64];
508
509 if (!(ts->io_bytes[DDIR_READ] + ts->io_bytes[DDIR_WRITE] +
510 ts->io_bytes[DDIR_TRIM]) && !(ts->total_io_u[DDIR_READ] +
511 ts->total_io_u[DDIR_WRITE] + ts->total_io_u[DDIR_TRIM]))
512 return;
513
514 time(&time_p);
515 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
516
517 if (!ts->error) {
518 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
519 ts->name, ts->groupid, ts->members,
520 ts->error, (int) ts->pid, time_buf);
521 } else {
522 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
523 ts->name, ts->groupid, ts->members,
524 ts->error, ts->verror, (int) ts->pid,
525 time_buf);
526 }
527
528 if (strlen(ts->description))
529 log_info(" Description : [%s]\n", ts->description);
530
531 if (ts->io_bytes[DDIR_READ])
532 show_ddir_status(rs, ts, DDIR_READ);
533 if (ts->io_bytes[DDIR_WRITE])
534 show_ddir_status(rs, ts, DDIR_WRITE);
535 if (ts->io_bytes[DDIR_TRIM])
536 show_ddir_status(rs, ts, DDIR_TRIM);
537
538 show_latencies(ts);
539
540 runtime = ts->total_run_time;
541 if (runtime) {
542 double runt = (double) runtime;
543
544 usr_cpu = (double) ts->usr_time * 100 / runt;
545 sys_cpu = (double) ts->sys_time * 100 / runt;
546 } else {
547 usr_cpu = 0;
548 sys_cpu = 0;
549 }
550
551 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
552 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
553 (unsigned long long) ts->ctx,
554 (unsigned long long) ts->majf,
555 (unsigned long long) ts->minf);
556
557 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
558 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
559 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
560 io_u_dist[1], io_u_dist[2],
561 io_u_dist[3], io_u_dist[4],
562 io_u_dist[5], io_u_dist[6]);
563
564 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
565 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
566 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
567 io_u_dist[1], io_u_dist[2],
568 io_u_dist[3], io_u_dist[4],
569 io_u_dist[5], io_u_dist[6]);
570 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
571 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
572 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
573 io_u_dist[1], io_u_dist[2],
574 io_u_dist[3], io_u_dist[4],
575 io_u_dist[5], io_u_dist[6]);
576 log_info(" issued : total=r=%llu/w=%llu/d=%llu,"
577 " short=r=%llu/w=%llu/d=%llu\n",
578 (unsigned long long) ts->total_io_u[0],
579 (unsigned long long) ts->total_io_u[1],
580 (unsigned long long) ts->total_io_u[2],
581 (unsigned long long) ts->short_io_u[0],
582 (unsigned long long) ts->short_io_u[1],
583 (unsigned long long) ts->short_io_u[2]);
584 if (ts->continue_on_error) {
585 log_info(" errors : total=%llu, first_error=%d/<%s>\n",
586 (unsigned long long)ts->total_err_count,
587 ts->first_error,
588 strerror(ts->first_error));
589 }
590}
591
592static void show_ddir_status_terse(struct thread_stat *ts,
593 struct group_run_stats *rs, int ddir)
594{
595 unsigned long min, max;
596 unsigned long long bw, iops;
597 unsigned int *ovals = NULL;
598 double mean, dev;
599 unsigned int len, minv, maxv;
600 int i;
601
602 assert(ddir_rw(ddir));
603
604 iops = bw = 0;
605 if (ts->runtime[ddir]) {
606 uint64_t runt = ts->runtime[ddir];
607
608 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
609 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
610 }
611
612 log_info(";%llu;%llu;%llu;%llu",
613 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
614 (unsigned long long) ts->runtime[ddir]);
615
616 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
617 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
618 else
619 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
620
621 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
622 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
623 else
624 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
625
626 if (ts->clat_percentiles) {
627 len = calc_clat_percentiles(ts->io_u_plat[ddir],
628 ts->clat_stat[ddir].samples,
629 ts->percentile_list, &ovals, &maxv,
630 &minv);
631 } else
632 len = 0;
633
634 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
635 if (i >= len) {
636 log_info(";0%%=0");
637 continue;
638 }
639 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
640 }
641
642 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
643 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
644 else
645 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
646
647 if (ovals)
648 free(ovals);
649
650 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
651 double p_of_agg = 100.0;
652
653 if (rs->agg[ddir]) {
654 p_of_agg = mean * 100 / (double) rs->agg[ddir];
655 if (p_of_agg > 100.0)
656 p_of_agg = 100.0;
657 }
658
659 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
660 } else
661 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
662}
663
664static void add_ddir_status_json(struct thread_stat *ts,
665 struct group_run_stats *rs, int ddir, struct json_object *parent)
666{
667 unsigned long min, max;
668 unsigned long long bw, iops;
669 unsigned int *ovals = NULL;
670 double mean, dev;
671 unsigned int len, minv, maxv;
672 int i;
673 const char *ddirname[] = {"read", "write", "trim"};
674 struct json_object *dir_object, *tmp_object, *percentile_object;
675 char buf[120];
676 double p_of_agg = 100.0;
677
678 assert(ddir_rw(ddir));
679
680 if (ts->unified_rw_rep && ddir != DDIR_READ)
681 return;
682
683 dir_object = json_create_object();
684 json_object_add_value_object(parent,
685 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
686
687 iops = bw = 0;
688 if (ts->runtime[ddir]) {
689 uint64_t runt = ts->runtime[ddir];
690
691 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
692 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
693 }
694
695 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
696 json_object_add_value_int(dir_object, "bw", bw);
697 json_object_add_value_int(dir_object, "iops", iops);
698 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
699
700 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
701 min = max = 0;
702 mean = dev = 0.0;
703 }
704 tmp_object = json_create_object();
705 json_object_add_value_object(dir_object, "slat", tmp_object);
706 json_object_add_value_int(tmp_object, "min", min);
707 json_object_add_value_int(tmp_object, "max", max);
708 json_object_add_value_float(tmp_object, "mean", mean);
709 json_object_add_value_float(tmp_object, "stddev", dev);
710
711 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
712 min = max = 0;
713 mean = dev = 0.0;
714 }
715 tmp_object = json_create_object();
716 json_object_add_value_object(dir_object, "clat", tmp_object);
717 json_object_add_value_int(tmp_object, "min", min);
718 json_object_add_value_int(tmp_object, "max", max);
719 json_object_add_value_float(tmp_object, "mean", mean);
720 json_object_add_value_float(tmp_object, "stddev", dev);
721
722 if (ts->clat_percentiles) {
723 len = calc_clat_percentiles(ts->io_u_plat[ddir],
724 ts->clat_stat[ddir].samples,
725 ts->percentile_list, &ovals, &maxv,
726 &minv);
727 } else
728 len = 0;
729
730 percentile_object = json_create_object();
731 json_object_add_value_object(tmp_object, "percentile", percentile_object);
732 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
733 if (i >= len) {
734 json_object_add_value_int(percentile_object, "0.00", 0);
735 continue;
736 }
737 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
738 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
739 }
740
741 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
742 min = max = 0;
743 mean = dev = 0.0;
744 }
745 tmp_object = json_create_object();
746 json_object_add_value_object(dir_object, "lat", tmp_object);
747 json_object_add_value_int(tmp_object, "min", min);
748 json_object_add_value_int(tmp_object, "max", max);
749 json_object_add_value_float(tmp_object, "mean", mean);
750 json_object_add_value_float(tmp_object, "stddev", dev);
751 if (ovals)
752 free(ovals);
753
754 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
755 if (rs->agg[ddir]) {
756 p_of_agg = mean * 100 / (double) rs->agg[ddir];
757 if (p_of_agg > 100.0)
758 p_of_agg = 100.0;
759 }
760 } else {
761 min = max = 0;
762 p_of_agg = mean = dev = 0.0;
763 }
764 json_object_add_value_int(dir_object, "bw_min", min);
765 json_object_add_value_int(dir_object, "bw_max", max);
766 json_object_add_value_float(dir_object, "bw_agg", mean);
767 json_object_add_value_float(dir_object, "bw_mean", mean);
768 json_object_add_value_float(dir_object, "bw_dev", dev);
769}
770
771static void show_thread_status_terse_v2(struct thread_stat *ts,
772 struct group_run_stats *rs)
773{
774 double io_u_dist[FIO_IO_U_MAP_NR];
775 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
776 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
777 double usr_cpu, sys_cpu;
778 int i;
779
780 /* General Info */
781 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
782 /* Log Read Status */
783 show_ddir_status_terse(ts, rs, DDIR_READ);
784 /* Log Write Status */
785 show_ddir_status_terse(ts, rs, DDIR_WRITE);
786 /* Log Trim Status */
787 show_ddir_status_terse(ts, rs, DDIR_TRIM);
788
789 /* CPU Usage */
790 if (ts->total_run_time) {
791 double runt = (double) ts->total_run_time;
792
793 usr_cpu = (double) ts->usr_time * 100 / runt;
794 sys_cpu = (double) ts->sys_time * 100 / runt;
795 } else {
796 usr_cpu = 0;
797 sys_cpu = 0;
798 }
799
800 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
801 (unsigned long long) ts->ctx,
802 (unsigned long long) ts->majf,
803 (unsigned long long) ts->minf);
804
805 /* Calc % distribution of IO depths, usecond, msecond latency */
806 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
807 stat_calc_lat_u(ts, io_u_lat_u);
808 stat_calc_lat_m(ts, io_u_lat_m);
809
810 /* Only show fixed 7 I/O depth levels*/
811 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
812 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
813 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
814
815 /* Microsecond latency */
816 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
817 log_info(";%3.2f%%", io_u_lat_u[i]);
818 /* Millisecond latency */
819 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
820 log_info(";%3.2f%%", io_u_lat_m[i]);
821 /* Additional output if continue_on_error set - default off*/
822 if (ts->continue_on_error)
823 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
824 log_info("\n");
825
826 /* Additional output if description is set */
827 if (ts->description)
828 log_info(";%s", ts->description);
829
830 log_info("\n");
831}
832
833static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
834 struct group_run_stats *rs, int ver)
835{
836 double io_u_dist[FIO_IO_U_MAP_NR];
837 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
838 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
839 double usr_cpu, sys_cpu;
840 int i;
841
842 /* General Info */
843 log_info("%d;%s;%s;%d;%d", ver, fio_version_string,
844 ts->name, ts->groupid, ts->error);
845 /* Log Read Status */
846 show_ddir_status_terse(ts, rs, DDIR_READ);
847 /* Log Write Status */
848 show_ddir_status_terse(ts, rs, DDIR_WRITE);
849 /* Log Trim Status */
850 if (ver == 4)
851 show_ddir_status_terse(ts, rs, DDIR_TRIM);
852
853 /* CPU Usage */
854 if (ts->total_run_time) {
855 double runt = (double) ts->total_run_time;
856
857 usr_cpu = (double) ts->usr_time * 100 / runt;
858 sys_cpu = (double) ts->sys_time * 100 / runt;
859 } else {
860 usr_cpu = 0;
861 sys_cpu = 0;
862 }
863
864 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
865 (unsigned long long) ts->ctx,
866 (unsigned long long) ts->majf,
867 (unsigned long long) ts->minf);
868
869 /* Calc % distribution of IO depths, usecond, msecond latency */
870 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
871 stat_calc_lat_u(ts, io_u_lat_u);
872 stat_calc_lat_m(ts, io_u_lat_m);
873
874 /* Only show fixed 7 I/O depth levels*/
875 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
876 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
877 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
878
879 /* Microsecond latency */
880 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
881 log_info(";%3.2f%%", io_u_lat_u[i]);
882 /* Millisecond latency */
883 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
884 log_info(";%3.2f%%", io_u_lat_m[i]);
885
886 /* disk util stats, if any */
887 if (is_backend)
888 show_disk_util(1, NULL);
889
890 /* Additional output if continue_on_error set - default off*/
891 if (ts->continue_on_error)
892 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
893
894 /* Additional output if description is set */
895 if (strlen(ts->description))
896 log_info(";%s", ts->description);
897
898 log_info("\n");
899}
900
901static struct json_object *show_thread_status_json(struct thread_stat *ts,
902 struct group_run_stats *rs)
903{
904 struct json_object *root, *tmp;
905 double io_u_dist[FIO_IO_U_MAP_NR];
906 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
907 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
908 double usr_cpu, sys_cpu;
909 int i;
910
911 root = json_create_object();
912 json_object_add_value_string(root, "jobname", ts->name);
913 json_object_add_value_int(root, "groupid", ts->groupid);
914 json_object_add_value_int(root, "error", ts->error);
915
916 add_ddir_status_json(ts, rs, DDIR_READ, root);
917 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
918 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
919
920 /* CPU Usage */
921 if (ts->total_run_time) {
922 double runt = (double) ts->total_run_time;
923
924 usr_cpu = (double) ts->usr_time * 100 / runt;
925 sys_cpu = (double) ts->sys_time * 100 / runt;
926 } else {
927 usr_cpu = 0;
928 sys_cpu = 0;
929 }
930 json_object_add_value_float(root, "usr_cpu", usr_cpu);
931 json_object_add_value_float(root, "sys_cpu", sys_cpu);
932 json_object_add_value_int(root, "ctx", ts->ctx);
933 json_object_add_value_int(root, "majf", ts->majf);
934 json_object_add_value_int(root, "minf", ts->minf);
935
936
937 /* Calc % distribution of IO depths, usecond, msecond latency */
938 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
939 stat_calc_lat_u(ts, io_u_lat_u);
940 stat_calc_lat_m(ts, io_u_lat_m);
941
942 tmp = json_create_object();
943 json_object_add_value_object(root, "iodepth_level", tmp);
944 /* Only show fixed 7 I/O depth levels*/
945 for (i = 0; i < 7; i++) {
946 char name[20];
947 if (i < 6)
948 snprintf(name, 20, "%d", 1 << i);
949 else
950 snprintf(name, 20, ">=%d", 1 << i);
951 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
952 }
953
954 tmp = json_create_object();
955 json_object_add_value_object(root, "latency_us", tmp);
956 /* Microsecond latency */
957 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
958 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
959 "250", "500", "750", "1000", };
960 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
961 }
962 /* Millisecond latency */
963 tmp = json_create_object();
964 json_object_add_value_object(root, "latency_ms", tmp);
965 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
966 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
967 "250", "500", "750", "1000", "2000",
968 ">=2000", };
969 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
970 }
971
972 /* Additional output if continue_on_error set - default off*/
973 if (ts->continue_on_error) {
974 json_object_add_value_int(root, "total_err", ts->total_err_count);
975 json_object_add_value_int(root, "first_error", ts->first_error);
976 }
977
978 /* Additional output if description is set */
979 if (strlen(ts->description))
980 json_object_add_value_string(root, "desc", ts->description);
981
982 return root;
983}
984
985static void show_thread_status_terse(struct thread_stat *ts,
986 struct group_run_stats *rs)
987{
988 if (terse_version == 2)
989 show_thread_status_terse_v2(ts, rs);
990 else if (terse_version == 3 || terse_version == 4)
991 show_thread_status_terse_v3_v4(ts, rs, terse_version);
992 else
993 log_err("fio: bad terse version!? %d\n", terse_version);
994}
995
996struct json_object *show_thread_status(struct thread_stat *ts,
997 struct group_run_stats *rs)
998{
999 if (output_format == FIO_OUTPUT_TERSE)
1000 show_thread_status_terse(ts, rs);
1001 else if (output_format == FIO_OUTPUT_JSON)
1002 return(show_thread_status_json(ts, rs));
1003 else
1004 show_thread_status_normal(ts, rs);
1005 return NULL;
1006}
1007
1008static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
1009{
1010 double mean, S;
1011
1012 if (src->samples == 0)
1013 return;
1014
1015 dst->min_val = min(dst->min_val, src->min_val);
1016 dst->max_val = max(dst->max_val, src->max_val);
1017
1018 /*
1019 * Compute new mean and S after the merge
1020 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1021 * #Parallel_algorithm>
1022 */
1023 if (nr == 1) {
1024 mean = src->mean.u.f;
1025 S = src->S.u.f;
1026 } else {
1027 double delta = src->mean.u.f - dst->mean.u.f;
1028
1029 mean = ((src->mean.u.f * src->samples) +
1030 (dst->mean.u.f * dst->samples)) /
1031 (dst->samples + src->samples);
1032
1033 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1034 (dst->samples * src->samples) /
1035 (dst->samples + src->samples);
1036 }
1037
1038 dst->samples += src->samples;
1039 dst->mean.u.f = mean;
1040 dst->S.u.f = S;
1041}
1042
1043void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1044{
1045 int i;
1046
1047 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1048 if (dst->max_run[i] < src->max_run[i])
1049 dst->max_run[i] = src->max_run[i];
1050 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1051 dst->min_run[i] = src->min_run[i];
1052 if (dst->max_bw[i] < src->max_bw[i])
1053 dst->max_bw[i] = src->max_bw[i];
1054 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1055 dst->min_bw[i] = src->min_bw[i];
1056
1057 dst->io_kb[i] += src->io_kb[i];
1058 dst->agg[i] += src->agg[i];
1059 }
1060
1061}
1062
1063void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1064{
1065 int l, k;
1066
1067 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1068 if (!dst->unified_rw_rep) {
1069 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1070 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1071 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1072 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1073
1074 dst->io_bytes[l] += src->io_bytes[l];
1075
1076 if (dst->runtime[l] < src->runtime[l])
1077 dst->runtime[l] = src->runtime[l];
1078 } else {
1079 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1080 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1081 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1082 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1083
1084 dst->io_bytes[0] += src->io_bytes[l];
1085
1086 if (dst->runtime[0] < src->runtime[l])
1087 dst->runtime[0] = src->runtime[l];
1088 }
1089 }
1090
1091 dst->usr_time += src->usr_time;
1092 dst->sys_time += src->sys_time;
1093 dst->ctx += src->ctx;
1094 dst->majf += src->majf;
1095 dst->minf += src->minf;
1096
1097 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1098 dst->io_u_map[k] += src->io_u_map[k];
1099 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1100 dst->io_u_submit[k] += src->io_u_submit[k];
1101 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1102 dst->io_u_complete[k] += src->io_u_complete[k];
1103 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1104 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1105 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1106 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1107
1108 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1109 if (!dst->unified_rw_rep) {
1110 dst->total_io_u[k] += src->total_io_u[k];
1111 dst->short_io_u[k] += src->short_io_u[k];
1112 } else {
1113 dst->total_io_u[0] += src->total_io_u[k];
1114 dst->short_io_u[0] += src->short_io_u[k];
1115 }
1116 }
1117
1118 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1119 int m;
1120
1121 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1122 if (!dst->unified_rw_rep)
1123 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1124 else
1125 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1126 }
1127 }
1128
1129 dst->total_run_time += src->total_run_time;
1130 dst->total_submit += src->total_submit;
1131 dst->total_complete += src->total_complete;
1132}
1133
1134void init_group_run_stat(struct group_run_stats *gs)
1135{
1136 int i;
1137 memset(gs, 0, sizeof(*gs));
1138
1139 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1140 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1141}
1142
1143void init_thread_stat(struct thread_stat *ts)
1144{
1145 int j;
1146
1147 memset(ts, 0, sizeof(*ts));
1148
1149 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1150 ts->lat_stat[j].min_val = -1UL;
1151 ts->clat_stat[j].min_val = -1UL;
1152 ts->slat_stat[j].min_val = -1UL;
1153 ts->bw_stat[j].min_val = -1UL;
1154 }
1155 ts->groupid = -1;
1156}
1157
1158static void __show_run_stats(void)
1159{
1160 struct group_run_stats *runstats, *rs;
1161 struct thread_data *td;
1162 struct thread_stat *threadstats, *ts;
1163 int i, j, nr_ts, last_ts, idx;
1164 int kb_base_warned = 0;
1165 int unit_base_warned = 0;
1166 struct json_object *root = NULL;
1167 struct json_array *array = NULL;
1168
1169 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1170
1171 for (i = 0; i < groupid + 1; i++)
1172 init_group_run_stat(&runstats[i]);
1173
1174 /*
1175 * find out how many threads stats we need. if group reporting isn't
1176 * enabled, it's one-per-td.
1177 */
1178 nr_ts = 0;
1179 last_ts = -1;
1180 for_each_td(td, i) {
1181 if (!td->o.group_reporting) {
1182 nr_ts++;
1183 continue;
1184 }
1185 if (last_ts == td->groupid)
1186 continue;
1187
1188 last_ts = td->groupid;
1189 nr_ts++;
1190 }
1191
1192 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1193
1194 for (i = 0; i < nr_ts; i++)
1195 init_thread_stat(&threadstats[i]);
1196
1197 j = 0;
1198 last_ts = -1;
1199 idx = 0;
1200 for_each_td(td, i) {
1201 if (idx && (!td->o.group_reporting ||
1202 (td->o.group_reporting && last_ts != td->groupid))) {
1203 idx = 0;
1204 j++;
1205 }
1206
1207 last_ts = td->groupid;
1208
1209 ts = &threadstats[j];
1210
1211 ts->clat_percentiles = td->o.clat_percentiles;
1212 ts->percentile_precision = td->o.percentile_precision;
1213 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1214
1215 idx++;
1216 ts->members++;
1217
1218 if (ts->groupid == -1) {
1219 /*
1220 * These are per-group shared already
1221 */
1222 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
1223 if (td->o.description)
1224 strncpy(ts->description, td->o.description,
1225 FIO_JOBNAME_SIZE);
1226 else
1227 memset(ts->description, 0, FIO_JOBNAME_SIZE);
1228
1229 /*
1230 * If multiple entries in this group, this is
1231 * the first member.
1232 */
1233 ts->thread_number = td->thread_number;
1234 ts->groupid = td->groupid;
1235
1236 /*
1237 * first pid in group, not very useful...
1238 */
1239 ts->pid = td->pid;
1240
1241 ts->kb_base = td->o.kb_base;
1242 ts->unit_base = td->o.unit_base;
1243 ts->unified_rw_rep = td->o.unified_rw_rep;
1244 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1245 log_info("fio: kb_base differs for jobs in group, using"
1246 " %u as the base\n", ts->kb_base);
1247 kb_base_warned = 1;
1248 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1249 log_info("fio: unit_base differs for jobs in group, using"
1250 " %u as the base\n", ts->unit_base);
1251 unit_base_warned = 1;
1252 }
1253
1254 ts->continue_on_error = td->o.continue_on_error;
1255 ts->total_err_count += td->total_err_count;
1256 ts->first_error = td->first_error;
1257 if (!ts->error) {
1258 if (!td->error && td->o.continue_on_error &&
1259 td->first_error) {
1260 ts->error = td->first_error;
1261 strcpy(ts->verror, td->verror);
1262 } else if (td->error) {
1263 ts->error = td->error;
1264 strcpy(ts->verror, td->verror);
1265 }
1266 }
1267
1268 sum_thread_stats(ts, &td->ts, idx);
1269 }
1270
1271 for (i = 0; i < nr_ts; i++) {
1272 unsigned long long bw;
1273
1274 ts = &threadstats[i];
1275 rs = &runstats[ts->groupid];
1276 rs->kb_base = ts->kb_base;
1277 rs->unit_base = ts->unit_base;
1278 rs->unified_rw_rep += ts->unified_rw_rep;
1279
1280 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1281 if (!ts->runtime[j])
1282 continue;
1283 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1284 rs->min_run[j] = ts->runtime[j];
1285 if (ts->runtime[j] > rs->max_run[j])
1286 rs->max_run[j] = ts->runtime[j];
1287
1288 bw = 0;
1289 if (ts->runtime[j]) {
1290 unsigned long runt = ts->runtime[j];
1291 unsigned long long kb;
1292
1293 kb = ts->io_bytes[j] / rs->kb_base;
1294 bw = kb * 1000 / runt;
1295 }
1296 if (bw < rs->min_bw[j])
1297 rs->min_bw[j] = bw;
1298 if (bw > rs->max_bw[j])
1299 rs->max_bw[j] = bw;
1300
1301 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1302 }
1303 }
1304
1305 for (i = 0; i < groupid + 1; i++) {
1306 int ddir;
1307
1308 rs = &runstats[i];
1309
1310 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1311 if (rs->max_run[ddir])
1312 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1313 rs->max_run[ddir];
1314 }
1315 }
1316
1317 /*
1318 * don't overwrite last signal output
1319 */
1320 if (output_format == FIO_OUTPUT_NORMAL)
1321 log_info("\n");
1322 else if (output_format == FIO_OUTPUT_JSON) {
1323 root = json_create_object();
1324 json_object_add_value_string(root, "fio version", fio_version_string);
1325 array = json_create_array();
1326 json_object_add_value_array(root, "jobs", array);
1327 }
1328
1329 for (i = 0; i < nr_ts; i++) {
1330 ts = &threadstats[i];
1331 rs = &runstats[ts->groupid];
1332
1333 if (is_backend)
1334 fio_server_send_ts(ts, rs);
1335 else if (output_format == FIO_OUTPUT_TERSE)
1336 show_thread_status_terse(ts, rs);
1337 else if (output_format == FIO_OUTPUT_JSON) {
1338 struct json_object *tmp = show_thread_status_json(ts, rs);
1339 json_array_add_value_object(array, tmp);
1340 } else
1341 show_thread_status_normal(ts, rs);
1342 }
1343 if (output_format == FIO_OUTPUT_JSON) {
1344 /* disk util stats, if any */
1345 show_disk_util(1, root);
1346
1347 show_idle_prof_stats(FIO_OUTPUT_JSON, root);
1348
1349 json_print_object(root);
1350 log_info("\n");
1351 json_free_object(root);
1352 }
1353
1354 for (i = 0; i < groupid + 1; i++) {
1355 rs = &runstats[i];
1356
1357 rs->groupid = i;
1358 if (is_backend)
1359 fio_server_send_gs(rs);
1360 else if (output_format == FIO_OUTPUT_NORMAL)
1361 show_group_stats(rs);
1362 }
1363
1364 if (is_backend)
1365 fio_server_send_du();
1366 else if (output_format == FIO_OUTPUT_NORMAL) {
1367 show_disk_util(0, NULL);
1368 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
1369 }
1370
1371 log_info_flush();
1372 free(runstats);
1373 free(threadstats);
1374}
1375
1376void show_run_stats(void)
1377{
1378 fio_mutex_down(stat_mutex);
1379 __show_run_stats();
1380 fio_mutex_up(stat_mutex);
1381}
1382
1383static void *__show_running_run_stats(void fio_unused *arg)
1384{
1385 struct thread_data *td;
1386 unsigned long long *rt;
1387 struct timeval tv;
1388 int i;
1389
1390 rt = malloc(thread_number * sizeof(unsigned long long));
1391 fio_gettime(&tv, NULL);
1392
1393 for_each_td(td, i) {
1394 rt[i] = mtime_since(&td->start, &tv);
1395 if (td_read(td) && td->io_bytes[DDIR_READ])
1396 td->ts.runtime[DDIR_READ] += rt[i];
1397 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1398 td->ts.runtime[DDIR_WRITE] += rt[i];
1399 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1400 td->ts.runtime[DDIR_TRIM] += rt[i];
1401
1402 td->update_rusage = 1;
1403 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1404 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1405 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1406 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1407 }
1408
1409 for_each_td(td, i) {
1410 if (td->rusage_sem) {
1411 td->update_rusage = 1;
1412 fio_mutex_down(td->rusage_sem);
1413 }
1414 td->update_rusage = 0;
1415 }
1416
1417 __show_run_stats();
1418
1419 for_each_td(td, i) {
1420 if (td_read(td) && td->io_bytes[DDIR_READ])
1421 td->ts.runtime[DDIR_READ] -= rt[i];
1422 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1423 td->ts.runtime[DDIR_WRITE] -= rt[i];
1424 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1425 td->ts.runtime[DDIR_TRIM] -= rt[i];
1426 }
1427
1428 free(rt);
1429 fio_mutex_up(stat_mutex);
1430 return NULL;
1431}
1432
1433/*
1434 * Called from signal handler. It _should_ be safe to just run this inline
1435 * in the sig handler, but we should be disturbing the system less by just
1436 * creating a thread to do it.
1437 */
1438void show_running_run_stats(void)
1439{
1440 pthread_t thread;
1441
1442 fio_mutex_down(stat_mutex);
1443
1444 if (!pthread_create(&thread, NULL, __show_running_run_stats, NULL)) {
1445 pthread_detach(thread);
1446 return;
1447 }
1448
1449 fio_mutex_up(stat_mutex);
1450}
1451
1452static int status_interval_init;
1453static struct timeval status_time;
1454
1455#define FIO_STATUS_FILE "/tmp/fio-dump-status"
1456
1457static int check_status_file(void)
1458{
1459 struct stat sb;
1460 const char *temp_dir;
1461 char fio_status_file_path[PATH_MAX];
1462
1463 temp_dir = getenv("TMPDIR");
1464 if (temp_dir == NULL)
1465 temp_dir = getenv("TEMP");
1466 if (temp_dir == NULL)
1467 temp_dir = "/tmp";
1468
1469 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1470
1471 if (stat(fio_status_file_path, &sb))
1472 return 0;
1473
1474 unlink(fio_status_file_path);
1475 return 1;
1476}
1477
1478void check_for_running_stats(void)
1479{
1480 if (status_interval) {
1481 if (!status_interval_init) {
1482 fio_gettime(&status_time, NULL);
1483 status_interval_init = 1;
1484 } else if (mtime_since_now(&status_time) >= status_interval) {
1485 show_running_run_stats();
1486 fio_gettime(&status_time, NULL);
1487 return;
1488 }
1489 }
1490 if (check_status_file()) {
1491 show_running_run_stats();
1492 return;
1493 }
1494}
1495
1496static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1497{
1498 double val = data;
1499 double delta;
1500
1501 if (data > is->max_val)
1502 is->max_val = data;
1503 if (data < is->min_val)
1504 is->min_val = data;
1505
1506 delta = val - is->mean.u.f;
1507 if (delta) {
1508 is->mean.u.f += delta / (is->samples + 1.0);
1509 is->S.u.f += delta * (val - is->mean.u.f);
1510 }
1511
1512 is->samples++;
1513}
1514
1515static void __add_log_sample(struct io_log *iolog, unsigned long val,
1516 enum fio_ddir ddir, unsigned int bs,
1517 unsigned long t)
1518{
1519 const int nr_samples = iolog->nr_samples;
1520
1521 if (iolog->disabled)
1522 return;
1523
1524 if (!iolog->nr_samples)
1525 iolog->avg_last = t;
1526
1527 if (iolog->nr_samples == iolog->max_samples) {
1528 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
1529 void *new_log;
1530
1531 new_log = realloc(iolog->log, new_size);
1532 if (!new_log) {
1533 log_err("fio: failed extending iolog! Will stop logging.\n");
1534 iolog->disabled = 1;
1535 return;
1536 }
1537 iolog->log = new_log;
1538 iolog->max_samples <<= 1;
1539 }
1540
1541 iolog->log[nr_samples].val = val;
1542 iolog->log[nr_samples].time = t;
1543 iolog->log[nr_samples].ddir = ddir;
1544 iolog->log[nr_samples].bs = bs;
1545 iolog->nr_samples++;
1546}
1547
1548static inline void reset_io_stat(struct io_stat *ios)
1549{
1550 ios->max_val = ios->min_val = ios->samples = 0;
1551 ios->mean.u.f = ios->S.u.f = 0;
1552}
1553
1554static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1555 unsigned long val, enum fio_ddir ddir,
1556 unsigned int bs)
1557{
1558 unsigned long elapsed, this_window;
1559
1560 if (!ddir_rw(ddir))
1561 return;
1562
1563 elapsed = mtime_since_now(&td->epoch);
1564
1565 /*
1566 * If no time averaging, just add the log sample.
1567 */
1568 if (!iolog->avg_msec) {
1569 __add_log_sample(iolog, val, ddir, bs, elapsed);
1570 return;
1571 }
1572
1573 /*
1574 * Add the sample. If the time period has passed, then
1575 * add that entry to the log and clear.
1576 */
1577 add_stat_sample(&iolog->avg_window[ddir], val);
1578
1579 /*
1580 * If period hasn't passed, adding the above sample is all we
1581 * need to do.
1582 */
1583 this_window = elapsed - iolog->avg_last;
1584 if (this_window < iolog->avg_msec)
1585 return;
1586
1587 /*
1588 * Note an entry in the log. Use the mean from the logged samples,
1589 * making sure to properly round up. Only write a log entry if we
1590 * had actual samples done.
1591 */
1592 if (iolog->avg_window[DDIR_READ].samples) {
1593 unsigned long mr;
1594
1595 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1596 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed);
1597 }
1598 if (iolog->avg_window[DDIR_WRITE].samples) {
1599 unsigned long mw;
1600
1601 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1602 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed);
1603 }
1604 if (iolog->avg_window[DDIR_TRIM].samples) {
1605 unsigned long mw;
1606
1607 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1608 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed);
1609 }
1610
1611
1612 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1613 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1614 reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1615 iolog->avg_last = elapsed;
1616}
1617
1618void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1619{
1620 struct io_log *iolog;
1621
1622 if (!ddir_rw(ddir))
1623 return;
1624
1625 iolog = agg_io_log[ddir];
1626 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
1627}
1628
1629static void add_clat_percentile_sample(struct thread_stat *ts,
1630 unsigned long usec, enum fio_ddir ddir)
1631{
1632 unsigned int idx = plat_val_to_idx(usec);
1633 assert(idx < FIO_IO_U_PLAT_NR);
1634
1635 ts->io_u_plat[ddir][idx]++;
1636}
1637
1638void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1639 unsigned long usec, unsigned int bs)
1640{
1641 struct thread_stat *ts = &td->ts;
1642
1643 if (!ddir_rw(ddir))
1644 return;
1645
1646 add_stat_sample(&ts->clat_stat[ddir], usec);
1647
1648 if (td->clat_log)
1649 add_log_sample(td, td->clat_log, usec, ddir, bs);
1650
1651 if (ts->clat_percentiles)
1652 add_clat_percentile_sample(ts, usec, ddir);
1653}
1654
1655void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1656 unsigned long usec, unsigned int bs)
1657{
1658 struct thread_stat *ts = &td->ts;
1659
1660 if (!ddir_rw(ddir))
1661 return;
1662
1663 add_stat_sample(&ts->slat_stat[ddir], usec);
1664
1665 if (td->slat_log)
1666 add_log_sample(td, td->slat_log, usec, ddir, bs);
1667}
1668
1669void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1670 unsigned long usec, unsigned int bs)
1671{
1672 struct thread_stat *ts = &td->ts;
1673
1674 if (!ddir_rw(ddir))
1675 return;
1676
1677 add_stat_sample(&ts->lat_stat[ddir], usec);
1678
1679 if (td->lat_log)
1680 add_log_sample(td, td->lat_log, usec, ddir, bs);
1681}
1682
1683void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1684 struct timeval *t)
1685{
1686 struct thread_stat *ts = &td->ts;
1687 unsigned long spent, rate;
1688
1689 if (!ddir_rw(ddir))
1690 return;
1691
1692 spent = mtime_since(&td->bw_sample_time, t);
1693 if (spent < td->o.bw_avg_time)
1694 return;
1695
1696 /*
1697 * Compute both read and write rates for the interval.
1698 */
1699 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1700 uint64_t delta;
1701
1702 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1703 if (!delta)
1704 continue; /* No entries for interval */
1705
1706 rate = delta * 1000 / spent / 1024;
1707 add_stat_sample(&ts->bw_stat[ddir], rate);
1708
1709 if (td->bw_log)
1710 add_log_sample(td, td->bw_log, rate, ddir, bs);
1711
1712 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1713 }
1714
1715 fio_gettime(&td->bw_sample_time, NULL);
1716}
1717
1718void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1719 struct timeval *t)
1720{
1721 struct thread_stat *ts = &td->ts;
1722 unsigned long spent, iops;
1723
1724 if (!ddir_rw(ddir))
1725 return;
1726
1727 spent = mtime_since(&td->iops_sample_time, t);
1728 if (spent < td->o.iops_avg_time)
1729 return;
1730
1731 /*
1732 * Compute both read and write rates for the interval.
1733 */
1734 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1735 uint64_t delta;
1736
1737 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1738 if (!delta)
1739 continue; /* No entries for interval */
1740
1741 iops = (delta * 1000) / spent;
1742 add_stat_sample(&ts->iops_stat[ddir], iops);
1743
1744 if (td->iops_log)
1745 add_log_sample(td, td->iops_log, iops, ddir, bs);
1746
1747 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1748 }
1749
1750 fio_gettime(&td->iops_sample_time, NULL);
1751}
1752
1753void stat_init(void)
1754{
1755 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1756}
1757
1758void stat_exit(void)
1759{
1760 /*
1761 * When we have the mutex, we know out-of-band access to it
1762 * have ended.
1763 */
1764 fio_mutex_down(stat_mutex);
1765 fio_mutex_remove(stat_mutex);
1766}
fio - Flexible IO Tester