projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Patch to make per-thread IOPS more accurate
[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
17struct 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 *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" : 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 *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" : 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
500static void show_thread_status_normal(struct thread_stat *ts,
501 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[32];
508
509 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
510 return;
511
512 time(&time_p);
513 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
514
515 if (!ts->error) {
516 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
517 ts->name, ts->groupid, ts->members,
518 ts->error, (int) ts->pid, time_buf);
519 } else {
520 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
521 ts->name, ts->groupid, ts->members,
522 ts->error, ts->verror, (int) ts->pid,
523 time_buf);
524 }
525
526 if (strlen(ts->description))
527 log_info(" Description : [%s]\n", ts->description);
528
529 if (ts->io_bytes[DDIR_READ])
530 show_ddir_status(rs, ts, DDIR_READ);
531 if (ts->io_bytes[DDIR_WRITE])
532 show_ddir_status(rs, ts, DDIR_WRITE);
533 if (ts->io_bytes[DDIR_TRIM])
534 show_ddir_status(rs, ts, DDIR_TRIM);
535
536 show_latencies(ts);
537
538 runtime = ts->total_run_time;
539 if (runtime) {
540 double runt = (double) runtime;
541
542 usr_cpu = (double) ts->usr_time * 100 / runt;
543 sys_cpu = (double) ts->sys_time * 100 / runt;
544 } else {
545 usr_cpu = 0;
546 sys_cpu = 0;
547 }
548
549 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
550 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
551 (unsigned long long) ts->ctx,
552 (unsigned long long) ts->majf,
553 (unsigned long long) ts->minf);
554
555 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
556 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
557 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
558 io_u_dist[1], io_u_dist[2],
559 io_u_dist[3], io_u_dist[4],
560 io_u_dist[5], io_u_dist[6]);
561
562 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
563 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
564 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
565 io_u_dist[1], io_u_dist[2],
566 io_u_dist[3], io_u_dist[4],
567 io_u_dist[5], io_u_dist[6]);
568 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
569 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
570 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
571 io_u_dist[1], io_u_dist[2],
572 io_u_dist[3], io_u_dist[4],
573 io_u_dist[5], io_u_dist[6]);
574 log_info(" issued : total=r=%llu/w=%llu/d=%llu,"
575 " short=r=%llu/w=%llu/d=%llu,"
576 " drop=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 (unsigned long long) ts->drop_io_u[0],
584 (unsigned long long) ts->drop_io_u[1],
585 (unsigned long long) ts->drop_io_u[2]);
586 if (ts->continue_on_error) {
587 log_info(" errors : total=%llu, first_error=%d/<%s>\n",
588 (unsigned long long)ts->total_err_count,
589 ts->first_error,
590 strerror(ts->first_error));
591 }
592 if (ts->latency_depth) {
593 log_info(" latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
594 (unsigned long long)ts->latency_target,
595 (unsigned long long)ts->latency_window,
596 ts->latency_percentile.u.f,
597 ts->latency_depth);
598 }
599}
600
601static void show_ddir_status_terse(struct thread_stat *ts,
602 struct group_run_stats *rs, int ddir)
603{
604 unsigned long min, max;
605 unsigned long long bw, iops;
606 unsigned int *ovals = NULL;
607 double mean, dev;
608 unsigned int len, minv, maxv;
609 int i;
610
611 assert(ddir_rw(ddir));
612
613 iops = bw = 0;
614 if (ts->runtime[ddir]) {
615 uint64_t runt = ts->runtime[ddir];
616
617 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
618 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
619 }
620
621 log_info(";%llu;%llu;%llu;%llu",
622 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
623 (unsigned long long) ts->runtime[ddir]);
624
625 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
626 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
627 else
628 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
629
630 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
631 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
632 else
633 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
634
635 if (ts->clat_percentiles) {
636 len = calc_clat_percentiles(ts->io_u_plat[ddir],
637 ts->clat_stat[ddir].samples,
638 ts->percentile_list, &ovals, &maxv,
639 &minv);
640 } else
641 len = 0;
642
643 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
644 if (i >= len) {
645 log_info(";0%%=0");
646 continue;
647 }
648 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
649 }
650
651 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
652 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
653 else
654 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
655
656 if (ovals)
657 free(ovals);
658
659 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
660 double p_of_agg = 100.0;
661
662 if (rs->agg[ddir]) {
663 p_of_agg = mean * 100 / (double) rs->agg[ddir];
664 if (p_of_agg > 100.0)
665 p_of_agg = 100.0;
666 }
667
668 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
669 } else
670 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
671}
672
673static void add_ddir_status_json(struct thread_stat *ts,
674 struct group_run_stats *rs, int ddir, struct json_object *parent)
675{
676 unsigned long min, max;
677 unsigned long long bw;
678 unsigned int *ovals = NULL;
679 double mean, dev, iops;
680 unsigned int len, minv, maxv;
681 int i;
682 const char *ddirname[] = {"read", "write", "trim"};
683 struct json_object *dir_object, *tmp_object, *percentile_object;
684 char buf[120];
685 double p_of_agg = 100.0;
686
687 assert(ddir_rw(ddir));
688
689 if (ts->unified_rw_rep && ddir != DDIR_READ)
690 return;
691
692 dir_object = json_create_object();
693 json_object_add_value_object(parent,
694 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
695
696 bw = 0;
697 iops = 0.0;
698 if (ts->runtime[ddir]) {
699 uint64_t runt = ts->runtime[ddir];
700
701 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
702 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
703 }
704
705 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
706 json_object_add_value_int(dir_object, "bw", bw);
707 json_object_add_value_float(dir_object, "iops", iops);
708 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
709 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
710 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
711 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
712
713 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
714 min = max = 0;
715 mean = dev = 0.0;
716 }
717 tmp_object = json_create_object();
718 json_object_add_value_object(dir_object, "slat", tmp_object);
719 json_object_add_value_int(tmp_object, "min", min);
720 json_object_add_value_int(tmp_object, "max", max);
721 json_object_add_value_float(tmp_object, "mean", mean);
722 json_object_add_value_float(tmp_object, "stddev", dev);
723
724 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
725 min = max = 0;
726 mean = dev = 0.0;
727 }
728 tmp_object = json_create_object();
729 json_object_add_value_object(dir_object, "clat", tmp_object);
730 json_object_add_value_int(tmp_object, "min", min);
731 json_object_add_value_int(tmp_object, "max", max);
732 json_object_add_value_float(tmp_object, "mean", mean);
733 json_object_add_value_float(tmp_object, "stddev", dev);
734
735 if (ts->clat_percentiles) {
736 len = calc_clat_percentiles(ts->io_u_plat[ddir],
737 ts->clat_stat[ddir].samples,
738 ts->percentile_list, &ovals, &maxv,
739 &minv);
740 } else
741 len = 0;
742
743 percentile_object = json_create_object();
744 json_object_add_value_object(tmp_object, "percentile", percentile_object);
745 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
746 if (i >= len) {
747 json_object_add_value_int(percentile_object, "0.00", 0);
748 continue;
749 }
750 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
751 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
752 }
753
754 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
755 min = max = 0;
756 mean = dev = 0.0;
757 }
758 tmp_object = json_create_object();
759 json_object_add_value_object(dir_object, "lat", tmp_object);
760 json_object_add_value_int(tmp_object, "min", min);
761 json_object_add_value_int(tmp_object, "max", max);
762 json_object_add_value_float(tmp_object, "mean", mean);
763 json_object_add_value_float(tmp_object, "stddev", dev);
764 if (ovals)
765 free(ovals);
766
767 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
768 if (rs->agg[ddir]) {
769 p_of_agg = mean * 100 / (double) rs->agg[ddir];
770 if (p_of_agg > 100.0)
771 p_of_agg = 100.0;
772 }
773 } else {
774 min = max = 0;
775 p_of_agg = mean = dev = 0.0;
776 }
777 json_object_add_value_int(dir_object, "bw_min", min);
778 json_object_add_value_int(dir_object, "bw_max", max);
779 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
780 json_object_add_value_float(dir_object, "bw_mean", mean);
781 json_object_add_value_float(dir_object, "bw_dev", dev);
782}
783
784static void show_thread_status_terse_v2(struct thread_stat *ts,
785 struct group_run_stats *rs)
786{
787 double io_u_dist[FIO_IO_U_MAP_NR];
788 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
789 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
790 double usr_cpu, sys_cpu;
791 int i;
792
793 /* General Info */
794 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
795 /* Log Read Status */
796 show_ddir_status_terse(ts, rs, DDIR_READ);
797 /* Log Write Status */
798 show_ddir_status_terse(ts, rs, DDIR_WRITE);
799 /* Log Trim Status */
800 show_ddir_status_terse(ts, rs, DDIR_TRIM);
801
802 /* CPU Usage */
803 if (ts->total_run_time) {
804 double runt = (double) ts->total_run_time;
805
806 usr_cpu = (double) ts->usr_time * 100 / runt;
807 sys_cpu = (double) ts->sys_time * 100 / runt;
808 } else {
809 usr_cpu = 0;
810 sys_cpu = 0;
811 }
812
813 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
814 (unsigned long long) ts->ctx,
815 (unsigned long long) ts->majf,
816 (unsigned long long) ts->minf);
817
818 /* Calc % distribution of IO depths, usecond, msecond latency */
819 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
820 stat_calc_lat_u(ts, io_u_lat_u);
821 stat_calc_lat_m(ts, io_u_lat_m);
822
823 /* Only show fixed 7 I/O depth levels*/
824 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
825 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
826 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
827
828 /* Microsecond latency */
829 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
830 log_info(";%3.2f%%", io_u_lat_u[i]);
831 /* Millisecond latency */
832 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
833 log_info(";%3.2f%%", io_u_lat_m[i]);
834 /* Additional output if continue_on_error set - default off*/
835 if (ts->continue_on_error)
836 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
837 log_info("\n");
838
839 /* Additional output if description is set */
840 if (strlen(ts->description))
841 log_info(";%s", ts->description);
842
843 log_info("\n");
844}
845
846static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
847 struct group_run_stats *rs, int ver)
848{
849 double io_u_dist[FIO_IO_U_MAP_NR];
850 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
851 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
852 double usr_cpu, sys_cpu;
853 int i;
854
855 /* General Info */
856 log_info("%d;%s;%s;%d;%d", ver, fio_version_string,
857 ts->name, ts->groupid, ts->error);
858 /* Log Read Status */
859 show_ddir_status_terse(ts, rs, DDIR_READ);
860 /* Log Write Status */
861 show_ddir_status_terse(ts, rs, DDIR_WRITE);
862 /* Log Trim Status */
863 if (ver == 4)
864 show_ddir_status_terse(ts, rs, DDIR_TRIM);
865
866 /* CPU Usage */
867 if (ts->total_run_time) {
868 double runt = (double) ts->total_run_time;
869
870 usr_cpu = (double) ts->usr_time * 100 / runt;
871 sys_cpu = (double) ts->sys_time * 100 / runt;
872 } else {
873 usr_cpu = 0;
874 sys_cpu = 0;
875 }
876
877 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
878 (unsigned long long) ts->ctx,
879 (unsigned long long) ts->majf,
880 (unsigned long long) ts->minf);
881
882 /* Calc % distribution of IO depths, usecond, msecond latency */
883 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
884 stat_calc_lat_u(ts, io_u_lat_u);
885 stat_calc_lat_m(ts, io_u_lat_m);
886
887 /* Only show fixed 7 I/O depth levels*/
888 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
889 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
890 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
891
892 /* Microsecond latency */
893 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
894 log_info(";%3.2f%%", io_u_lat_u[i]);
895 /* Millisecond latency */
896 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
897 log_info(";%3.2f%%", io_u_lat_m[i]);
898
899 /* disk util stats, if any */
900 show_disk_util(1, NULL);
901
902 /* Additional output if continue_on_error set - default off*/
903 if (ts->continue_on_error)
904 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
905
906 /* Additional output if description is set */
907 if (strlen(ts->description))
908 log_info(";%s", ts->description);
909
910 log_info("\n");
911}
912
913static struct json_object *show_thread_status_json(struct thread_stat *ts,
914 struct group_run_stats *rs)
915{
916 struct json_object *root, *tmp;
917 double io_u_dist[FIO_IO_U_MAP_NR];
918 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
919 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
920 double usr_cpu, sys_cpu;
921 int i;
922
923 root = json_create_object();
924 json_object_add_value_string(root, "jobname", ts->name);
925 json_object_add_value_int(root, "groupid", ts->groupid);
926 json_object_add_value_int(root, "error", ts->error);
927
928 add_ddir_status_json(ts, rs, DDIR_READ, root);
929 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
930 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
931
932 /* CPU Usage */
933 if (ts->total_run_time) {
934 double runt = (double) ts->total_run_time;
935
936 usr_cpu = (double) ts->usr_time * 100 / runt;
937 sys_cpu = (double) ts->sys_time * 100 / runt;
938 } else {
939 usr_cpu = 0;
940 sys_cpu = 0;
941 }
942 json_object_add_value_float(root, "usr_cpu", usr_cpu);
943 json_object_add_value_float(root, "sys_cpu", sys_cpu);
944 json_object_add_value_int(root, "ctx", ts->ctx);
945 json_object_add_value_int(root, "majf", ts->majf);
946 json_object_add_value_int(root, "minf", ts->minf);
947
948
949 /* Calc % distribution of IO depths, usecond, msecond latency */
950 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
951 stat_calc_lat_u(ts, io_u_lat_u);
952 stat_calc_lat_m(ts, io_u_lat_m);
953
954 tmp = json_create_object();
955 json_object_add_value_object(root, "iodepth_level", tmp);
956 /* Only show fixed 7 I/O depth levels*/
957 for (i = 0; i < 7; i++) {
958 char name[20];
959 if (i < 6)
960 snprintf(name, 20, "%d", 1 << i);
961 else
962 snprintf(name, 20, ">=%d", 1 << i);
963 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
964 }
965
966 tmp = json_create_object();
967 json_object_add_value_object(root, "latency_us", tmp);
968 /* Microsecond latency */
969 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
970 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
971 "250", "500", "750", "1000", };
972 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
973 }
974 /* Millisecond latency */
975 tmp = json_create_object();
976 json_object_add_value_object(root, "latency_ms", tmp);
977 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
978 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
979 "250", "500", "750", "1000", "2000",
980 ">=2000", };
981 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
982 }
983
984 /* Additional output if continue_on_error set - default off*/
985 if (ts->continue_on_error) {
986 json_object_add_value_int(root, "total_err", ts->total_err_count);
987 json_object_add_value_int(root, "first_error", ts->first_error);
988 }
989
990 if (ts->latency_depth) {
991 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
992 json_object_add_value_int(root, "latency_target", ts->latency_target);
993 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
994 json_object_add_value_int(root, "latency_window", ts->latency_window);
995 }
996
997 /* Additional output if description is set */
998 if (strlen(ts->description))
999 json_object_add_value_string(root, "desc", ts->description);
1000
1001 return root;
1002}
1003
1004static void show_thread_status_terse(struct thread_stat *ts,
1005 struct group_run_stats *rs)
1006{
1007 if (terse_version == 2)
1008 show_thread_status_terse_v2(ts, rs);
1009 else if (terse_version == 3 || terse_version == 4)
1010 show_thread_status_terse_v3_v4(ts, rs, terse_version);
1011 else
1012 log_err("fio: bad terse version!? %d\n", terse_version);
1013}
1014
1015struct json_object *show_thread_status(struct thread_stat *ts,
1016 struct group_run_stats *rs)
1017{
1018 if (output_format == FIO_OUTPUT_TERSE)
1019 show_thread_status_terse(ts, rs);
1020 else if (output_format == FIO_OUTPUT_JSON)
1021 return show_thread_status_json(ts, rs);
1022 else
1023 show_thread_status_normal(ts, rs);
1024 return NULL;
1025}
1026
1027static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
1028{
1029 double mean, S;
1030
1031 if (src->samples == 0)
1032 return;
1033
1034 dst->min_val = min(dst->min_val, src->min_val);
1035 dst->max_val = max(dst->max_val, src->max_val);
1036
1037 /*
1038 * Compute new mean and S after the merge
1039 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1040 * #Parallel_algorithm>
1041 */
1042 if (nr == 1) {
1043 mean = src->mean.u.f;
1044 S = src->S.u.f;
1045 } else {
1046 double delta = src->mean.u.f - dst->mean.u.f;
1047
1048 mean = ((src->mean.u.f * src->samples) +
1049 (dst->mean.u.f * dst->samples)) /
1050 (dst->samples + src->samples);
1051
1052 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1053 (dst->samples * src->samples) /
1054 (dst->samples + src->samples);
1055 }
1056
1057 dst->samples += src->samples;
1058 dst->mean.u.f = mean;
1059 dst->S.u.f = S;
1060}
1061
1062void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1063{
1064 int i;
1065
1066 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1067 if (dst->max_run[i] < src->max_run[i])
1068 dst->max_run[i] = src->max_run[i];
1069 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1070 dst->min_run[i] = src->min_run[i];
1071 if (dst->max_bw[i] < src->max_bw[i])
1072 dst->max_bw[i] = src->max_bw[i];
1073 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1074 dst->min_bw[i] = src->min_bw[i];
1075
1076 dst->io_kb[i] += src->io_kb[i];
1077 dst->agg[i] += src->agg[i];
1078 }
1079
1080 if (!dst->kb_base)
1081 dst->kb_base = src->kb_base;
1082 if (!dst->unit_base)
1083 dst->unit_base = src->unit_base;
1084}
1085
1086void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1087{
1088 int l, k;
1089
1090 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1091 if (!dst->unified_rw_rep) {
1092 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1093 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1094 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1095 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1096
1097 dst->io_bytes[l] += src->io_bytes[l];
1098
1099 if (dst->runtime[l] < src->runtime[l])
1100 dst->runtime[l] = src->runtime[l];
1101 } else {
1102 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1103 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1104 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1105 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1106
1107 dst->io_bytes[0] += src->io_bytes[l];
1108
1109 if (dst->runtime[0] < src->runtime[l])
1110 dst->runtime[0] = src->runtime[l];
1111 }
1112 }
1113
1114 dst->usr_time += src->usr_time;
1115 dst->sys_time += src->sys_time;
1116 dst->ctx += src->ctx;
1117 dst->majf += src->majf;
1118 dst->minf += src->minf;
1119
1120 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1121 dst->io_u_map[k] += src->io_u_map[k];
1122 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1123 dst->io_u_submit[k] += src->io_u_submit[k];
1124 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1125 dst->io_u_complete[k] += src->io_u_complete[k];
1126 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1127 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1128 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1129 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1130
1131 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1132 if (!dst->unified_rw_rep) {
1133 dst->total_io_u[k] += src->total_io_u[k];
1134 dst->short_io_u[k] += src->short_io_u[k];
1135 dst->drop_io_u[k] += src->drop_io_u[k];
1136 } else {
1137 dst->total_io_u[0] += src->total_io_u[k];
1138 dst->short_io_u[0] += src->short_io_u[k];
1139 dst->drop_io_u[0] += src->drop_io_u[k];
1140 }
1141 }
1142
1143 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1144 int m;
1145
1146 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1147 if (!dst->unified_rw_rep)
1148 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1149 else
1150 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1151 }
1152 }
1153
1154 dst->total_run_time += src->total_run_time;
1155 dst->total_submit += src->total_submit;
1156 dst->total_complete += src->total_complete;
1157}
1158
1159void init_group_run_stat(struct group_run_stats *gs)
1160{
1161 int i;
1162 memset(gs, 0, sizeof(*gs));
1163
1164 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1165 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1166}
1167
1168void init_thread_stat(struct thread_stat *ts)
1169{
1170 int j;
1171
1172 memset(ts, 0, sizeof(*ts));
1173
1174 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1175 ts->lat_stat[j].min_val = -1UL;
1176 ts->clat_stat[j].min_val = -1UL;
1177 ts->slat_stat[j].min_val = -1UL;
1178 ts->bw_stat[j].min_val = -1UL;
1179 }
1180 ts->groupid = -1;
1181}
1182
1183void __show_run_stats(void)
1184{
1185 struct group_run_stats *runstats, *rs;
1186 struct thread_data *td;
1187 struct thread_stat *threadstats, *ts;
1188 int i, j, nr_ts, last_ts, idx;
1189 int kb_base_warned = 0;
1190 int unit_base_warned = 0;
1191 struct json_object *root = NULL;
1192 struct json_array *array = NULL;
1193 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1194
1195 for (i = 0; i < groupid + 1; i++)
1196 init_group_run_stat(&runstats[i]);
1197
1198 /*
1199 * find out how many threads stats we need. if group reporting isn't
1200 * enabled, it's one-per-td.
1201 */
1202 nr_ts = 0;
1203 last_ts = -1;
1204 for_each_td(td, i) {
1205 if (!td->o.group_reporting) {
1206 nr_ts++;
1207 continue;
1208 }
1209 if (last_ts == td->groupid)
1210 continue;
1211
1212 last_ts = td->groupid;
1213 nr_ts++;
1214 }
1215
1216 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1217
1218 for (i = 0; i < nr_ts; i++)
1219 init_thread_stat(&threadstats[i]);
1220
1221 j = 0;
1222 last_ts = -1;
1223 idx = 0;
1224 for_each_td(td, i) {
1225 if (idx && (!td->o.group_reporting ||
1226 (td->o.group_reporting && last_ts != td->groupid))) {
1227 idx = 0;
1228 j++;
1229 }
1230
1231 last_ts = td->groupid;
1232
1233 ts = &threadstats[j];
1234
1235 ts->clat_percentiles = td->o.clat_percentiles;
1236 ts->percentile_precision = td->o.percentile_precision;
1237 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1238
1239 idx++;
1240 ts->members++;
1241
1242 if (ts->groupid == -1) {
1243 /*
1244 * These are per-group shared already
1245 */
1246 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1247 if (td->o.description)
1248 strncpy(ts->description, td->o.description,
1249 FIO_JOBDESC_SIZE - 1);
1250 else
1251 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1252
1253 /*
1254 * If multiple entries in this group, this is
1255 * the first member.
1256 */
1257 ts->thread_number = td->thread_number;
1258 ts->groupid = td->groupid;
1259
1260 /*
1261 * first pid in group, not very useful...
1262 */
1263 ts->pid = td->pid;
1264
1265 ts->kb_base = td->o.kb_base;
1266 ts->unit_base = td->o.unit_base;
1267 ts->unified_rw_rep = td->o.unified_rw_rep;
1268 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1269 log_info("fio: kb_base differs for jobs in group, using"
1270 " %u as the base\n", ts->kb_base);
1271 kb_base_warned = 1;
1272 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1273 log_info("fio: unit_base differs for jobs in group, using"
1274 " %u as the base\n", ts->unit_base);
1275 unit_base_warned = 1;
1276 }
1277
1278 ts->continue_on_error = td->o.continue_on_error;
1279 ts->total_err_count += td->total_err_count;
1280 ts->first_error = td->first_error;
1281 if (!ts->error) {
1282 if (!td->error && td->o.continue_on_error &&
1283 td->first_error) {
1284 ts->error = td->first_error;
1285 ts->verror[sizeof(ts->verror) - 1] = '\0';
1286 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1287 } else if (td->error) {
1288 ts->error = td->error;
1289 ts->verror[sizeof(ts->verror) - 1] = '\0';
1290 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1291 }
1292 }
1293
1294 ts->latency_depth = td->latency_qd;
1295 ts->latency_target = td->o.latency_target;
1296 ts->latency_percentile = td->o.latency_percentile;
1297 ts->latency_window = td->o.latency_window;
1298
1299 sum_thread_stats(ts, &td->ts, idx);
1300 }
1301
1302 for (i = 0; i < nr_ts; i++) {
1303 unsigned long long bw;
1304
1305 ts = &threadstats[i];
1306 rs = &runstats[ts->groupid];
1307 rs->kb_base = ts->kb_base;
1308 rs->unit_base = ts->unit_base;
1309 rs->unified_rw_rep += ts->unified_rw_rep;
1310
1311 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1312 if (!ts->runtime[j])
1313 continue;
1314 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1315 rs->min_run[j] = ts->runtime[j];
1316 if (ts->runtime[j] > rs->max_run[j])
1317 rs->max_run[j] = ts->runtime[j];
1318
1319 bw = 0;
1320 if (ts->runtime[j]) {
1321 unsigned long runt = ts->runtime[j];
1322 unsigned long long kb;
1323
1324 kb = ts->io_bytes[j] / rs->kb_base;
1325 bw = kb * 1000 / runt;
1326 }
1327 if (bw < rs->min_bw[j])
1328 rs->min_bw[j] = bw;
1329 if (bw > rs->max_bw[j])
1330 rs->max_bw[j] = bw;
1331
1332 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1333 }
1334 }
1335
1336 for (i = 0; i < groupid + 1; i++) {
1337 int ddir;
1338
1339 rs = &runstats[i];
1340
1341 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1342 if (rs->max_run[ddir])
1343 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1344 rs->max_run[ddir];
1345 }
1346 }
1347
1348 /*
1349 * don't overwrite last signal output
1350 */
1351 if (output_format == FIO_OUTPUT_NORMAL)
1352 log_info("\n");
1353 else if (output_format == FIO_OUTPUT_JSON) {
1354 char time_buf[32];
1355 time_t time_p;
1356
1357 time(&time_p);
1358 os_ctime_r((const time_t *) &time_p, time_buf,
1359 sizeof(time_buf));
1360 time_buf[strlen(time_buf) - 1] = '\0';
1361
1362 root = json_create_object();
1363 json_object_add_value_string(root, "fio version", fio_version_string);
1364 json_object_add_value_int(root, "timestamp", time_p);
1365 json_object_add_value_string(root, "time", time_buf);
1366 array = json_create_array();
1367 json_object_add_value_array(root, "jobs", array);
1368 }
1369
1370 for (i = 0; i < nr_ts; i++) {
1371 ts = &threadstats[i];
1372 rs = &runstats[ts->groupid];
1373
1374 if (is_backend)
1375 fio_server_send_ts(ts, rs);
1376 else if (output_format == FIO_OUTPUT_TERSE)
1377 show_thread_status_terse(ts, rs);
1378 else if (output_format == FIO_OUTPUT_JSON) {
1379 struct json_object *tmp = show_thread_status_json(ts, rs);
1380 json_array_add_value_object(array, tmp);
1381 } else
1382 show_thread_status_normal(ts, rs);
1383 }
1384 if (output_format == FIO_OUTPUT_JSON) {
1385 /* disk util stats, if any */
1386 show_disk_util(1, root);
1387
1388 show_idle_prof_stats(FIO_OUTPUT_JSON, root);
1389
1390 json_print_object(root);
1391 log_info("\n");
1392 json_free_object(root);
1393 }
1394
1395 for (i = 0; i < groupid + 1; i++) {
1396 rs = &runstats[i];
1397
1398 rs->groupid = i;
1399 if (is_backend)
1400 fio_server_send_gs(rs);
1401 else if (output_format == FIO_OUTPUT_NORMAL)
1402 show_group_stats(rs);
1403 }
1404
1405 if (is_backend)
1406 fio_server_send_du();
1407 else if (output_format == FIO_OUTPUT_NORMAL) {
1408 show_disk_util(0, NULL);
1409 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
1410 }
1411
1412 if ( !(output_format == FIO_OUTPUT_TERSE) && append_terse_output) {
1413 log_info("\nAdditional Terse Output:\n");
1414
1415 for (i = 0; i < nr_ts; i++) {
1416 ts = &threadstats[i];
1417 rs = &runstats[ts->groupid];
1418 show_thread_status_terse(ts, rs);
1419 }
1420 }
1421
1422 log_info_flush();
1423 free(runstats);
1424 free(threadstats);
1425}
1426
1427void show_run_stats(void)
1428{
1429 fio_mutex_down(stat_mutex);
1430 __show_run_stats();
1431 fio_mutex_up(stat_mutex);
1432}
1433
1434void __show_running_run_stats(void)
1435{
1436 struct thread_data *td;
1437 unsigned long long *rt;
1438 struct timeval tv;
1439 int i;
1440
1441 fio_mutex_down(stat_mutex);
1442
1443 rt = malloc(thread_number * sizeof(unsigned long long));
1444 fio_gettime(&tv, NULL);
1445
1446 for_each_td(td, i) {
1447 rt[i] = mtime_since(&td->start, &tv);
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 td->update_rusage = 1;
1456 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1457 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1458 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1459 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1460 }
1461
1462 for_each_td(td, i) {
1463 if (td->runstate >= TD_EXITED)
1464 continue;
1465 if (td->rusage_sem) {
1466 td->update_rusage = 1;
1467 fio_mutex_down(td->rusage_sem);
1468 }
1469 td->update_rusage = 0;
1470 }
1471
1472 __show_run_stats();
1473
1474 for_each_td(td, i) {
1475 if (td_read(td) && td->io_bytes[DDIR_READ])
1476 td->ts.runtime[DDIR_READ] -= rt[i];
1477 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1478 td->ts.runtime[DDIR_WRITE] -= rt[i];
1479 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1480 td->ts.runtime[DDIR_TRIM] -= rt[i];
1481 }
1482
1483 free(rt);
1484 fio_mutex_up(stat_mutex);
1485}
1486
1487static int status_interval_init;
1488static struct timeval status_time;
1489static int status_file_disabled;
1490
1491#define FIO_STATUS_FILE "fio-dump-status"
1492
1493static int check_status_file(void)
1494{
1495 struct stat sb;
1496 const char *temp_dir;
1497 char fio_status_file_path[PATH_MAX];
1498
1499 if (status_file_disabled)
1500 return 0;
1501
1502 temp_dir = getenv("TMPDIR");
1503 if (temp_dir == NULL) {
1504 temp_dir = getenv("TEMP");
1505 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1506 temp_dir = NULL;
1507 }
1508 if (temp_dir == NULL)
1509 temp_dir = "/tmp";
1510
1511 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1512
1513 if (stat(fio_status_file_path, &sb))
1514 return 0;
1515
1516 if (unlink(fio_status_file_path) < 0) {
1517 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1518 strerror(errno));
1519 log_err("fio: disabling status file updates\n");
1520 status_file_disabled = 1;
1521 }
1522
1523 return 1;
1524}
1525
1526void check_for_running_stats(void)
1527{
1528 if (status_interval) {
1529 if (!status_interval_init) {
1530 fio_gettime(&status_time, NULL);
1531 status_interval_init = 1;
1532 } else if (mtime_since_now(&status_time) >= status_interval) {
1533 show_running_run_stats();
1534 fio_gettime(&status_time, NULL);
1535 return;
1536 }
1537 }
1538 if (check_status_file()) {
1539 show_running_run_stats();
1540 return;
1541 }
1542}
1543
1544static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1545{
1546 double val = data;
1547 double delta;
1548
1549 if (data > is->max_val)
1550 is->max_val = data;
1551 if (data < is->min_val)
1552 is->min_val = data;
1553
1554 delta = val - is->mean.u.f;
1555 if (delta) {
1556 is->mean.u.f += delta / (is->samples + 1.0);
1557 is->S.u.f += delta * (val - is->mean.u.f);
1558 }
1559
1560 is->samples++;
1561}
1562
1563static void __add_log_sample(struct io_log *iolog, unsigned long val,
1564 enum fio_ddir ddir, unsigned int bs,
1565 unsigned long t, uint64_t offset)
1566{
1567 uint64_t nr_samples = iolog->nr_samples;
1568 struct io_sample *s;
1569
1570 if (iolog->disabled)
1571 return;
1572
1573 if (!iolog->nr_samples)
1574 iolog->avg_last = t;
1575
1576 if (iolog->nr_samples == iolog->max_samples) {
1577 size_t new_size;
1578 void *new_log;
1579
1580 new_size = 2 * iolog->max_samples * log_entry_sz(iolog);
1581
1582 if (iolog->log_gz && (new_size > iolog->log_gz)) {
1583 if (iolog_flush(iolog, 0)) {
1584 log_err("fio: failed flushing iolog! Will stop logging.\n");
1585 iolog->disabled = 1;
1586 return;
1587 }
1588 nr_samples = iolog->nr_samples;
1589 } else {
1590 new_log = realloc(iolog->log, new_size);
1591 if (!new_log) {
1592 log_err("fio: failed extending iolog! Will stop logging.\n");
1593 iolog->disabled = 1;
1594 return;
1595 }
1596 iolog->log = new_log;
1597 iolog->max_samples <<= 1;
1598 }
1599 }
1600
1601 s = get_sample(iolog, nr_samples);
1602
1603 s->val = val;
1604 s->time = t;
1605 io_sample_set_ddir(iolog, s, ddir);
1606 s->bs = bs;
1607
1608 if (iolog->log_offset) {
1609 struct io_sample_offset *so = (void *) s;
1610
1611 so->offset = offset;
1612 }
1613
1614 iolog->nr_samples++;
1615}
1616
1617static inline void reset_io_stat(struct io_stat *ios)
1618{
1619 ios->max_val = ios->min_val = ios->samples = 0;
1620 ios->mean.u.f = ios->S.u.f = 0;
1621}
1622
1623void reset_io_stats(struct thread_data *td)
1624{
1625 struct thread_stat *ts = &td->ts;
1626 int i, j;
1627
1628 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1629 reset_io_stat(&ts->clat_stat[i]);
1630 reset_io_stat(&ts->slat_stat[i]);
1631 reset_io_stat(&ts->lat_stat[i]);
1632 reset_io_stat(&ts->bw_stat[i]);
1633 reset_io_stat(&ts->iops_stat[i]);
1634
1635 ts->io_bytes[i] = 0;
1636 ts->runtime[i] = 0;
1637
1638 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1639 ts->io_u_plat[i][j] = 0;
1640 }
1641
1642 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1643 ts->io_u_map[i] = 0;
1644 ts->io_u_submit[i] = 0;
1645 ts->io_u_complete[i] = 0;
1646 ts->io_u_lat_u[i] = 0;
1647 ts->io_u_lat_m[i] = 0;
1648 ts->total_submit = 0;
1649 ts->total_complete = 0;
1650 }
1651
1652 for (i = 0; i < 3; i++) {
1653 ts->total_io_u[i] = 0;
1654 ts->short_io_u[i] = 0;
1655 ts->drop_io_u[i] = 0;
1656 }
1657}
1658
1659static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
1660{
1661 /*
1662 * Note an entry in the log. Use the mean from the logged samples,
1663 * making sure to properly round up. Only write a log entry if we
1664 * had actual samples done.
1665 */
1666 if (iolog->avg_window[DDIR_READ].samples) {
1667 unsigned long mr;
1668
1669 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1670 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed, 0);
1671 }
1672 if (iolog->avg_window[DDIR_WRITE].samples) {
1673 unsigned long mw;
1674
1675 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1676 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed, 0);
1677 }
1678 if (iolog->avg_window[DDIR_TRIM].samples) {
1679 unsigned long mw;
1680
1681 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1682 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed, 0);
1683 }
1684
1685 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1686 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1687 reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1688}
1689
1690static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1691 unsigned long val, enum fio_ddir ddir,
1692 unsigned int bs, uint64_t offset)
1693{
1694 unsigned long elapsed, this_window;
1695
1696 if (!ddir_rw(ddir))
1697 return;
1698
1699 elapsed = mtime_since_now(&td->epoch);
1700
1701 /*
1702 * If no time averaging, just add the log sample.
1703 */
1704 if (!iolog->avg_msec) {
1705 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
1706 return;
1707 }
1708
1709 /*
1710 * Add the sample. If the time period has passed, then
1711 * add that entry to the log and clear.
1712 */
1713 add_stat_sample(&iolog->avg_window[ddir], val);
1714
1715 /*
1716 * If period hasn't passed, adding the above sample is all we
1717 * need to do.
1718 */
1719 this_window = elapsed - iolog->avg_last;
1720 if (this_window < iolog->avg_msec)
1721 return;
1722
1723 _add_stat_to_log(iolog, elapsed);
1724
1725 iolog->avg_last = elapsed;
1726}
1727
1728void finalize_logs(struct thread_data *td)
1729{
1730 unsigned long elapsed;
1731
1732 elapsed = mtime_since_now(&td->epoch);
1733
1734 if (td->clat_log)
1735 _add_stat_to_log(td->clat_log, elapsed);
1736 if (td->slat_log)
1737 _add_stat_to_log(td->slat_log, elapsed);
1738 if (td->lat_log)
1739 _add_stat_to_log(td->lat_log, elapsed);
1740 if (td->bw_log)
1741 _add_stat_to_log(td->bw_log, elapsed);
1742 if (td->iops_log)
1743 _add_stat_to_log(td->iops_log, elapsed);
1744}
1745
1746void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1747{
1748 struct io_log *iolog;
1749
1750 if (!ddir_rw(ddir))
1751 return;
1752
1753 iolog = agg_io_log[ddir];
1754 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
1755}
1756
1757static void add_clat_percentile_sample(struct thread_stat *ts,
1758 unsigned long usec, enum fio_ddir ddir)
1759{
1760 unsigned int idx = plat_val_to_idx(usec);
1761 assert(idx < FIO_IO_U_PLAT_NR);
1762
1763 ts->io_u_plat[ddir][idx]++;
1764}
1765
1766void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1767 unsigned long usec, unsigned int bs, uint64_t offset)
1768{
1769 struct thread_stat *ts = &td->ts;
1770
1771 if (!ddir_rw(ddir))
1772 return;
1773
1774 add_stat_sample(&ts->clat_stat[ddir], usec);
1775
1776 if (td->clat_log)
1777 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
1778
1779 if (ts->clat_percentiles)
1780 add_clat_percentile_sample(ts, usec, ddir);
1781}
1782
1783void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1784 unsigned long usec, unsigned int bs, uint64_t offset)
1785{
1786 struct thread_stat *ts = &td->ts;
1787
1788 if (!ddir_rw(ddir))
1789 return;
1790
1791 add_stat_sample(&ts->slat_stat[ddir], usec);
1792
1793 if (td->slat_log)
1794 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
1795}
1796
1797void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1798 unsigned long usec, unsigned int bs, uint64_t offset)
1799{
1800 struct thread_stat *ts = &td->ts;
1801
1802 if (!ddir_rw(ddir))
1803 return;
1804
1805 add_stat_sample(&ts->lat_stat[ddir], usec);
1806
1807 if (td->lat_log)
1808 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
1809}
1810
1811void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1812 struct timeval *t)
1813{
1814 struct thread_stat *ts = &td->ts;
1815 unsigned long spent, rate;
1816
1817 if (!ddir_rw(ddir))
1818 return;
1819
1820 spent = mtime_since(&td->bw_sample_time, t);
1821 if (spent < td->o.bw_avg_time)
1822 return;
1823
1824 /*
1825 * Compute both read and write rates for the interval.
1826 */
1827 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1828 uint64_t delta;
1829
1830 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1831 if (!delta)
1832 continue; /* No entries for interval */
1833
1834 if (spent)
1835 rate = delta * 1000 / spent / 1024;
1836 else
1837 rate = 0;
1838
1839 add_stat_sample(&ts->bw_stat[ddir], rate);
1840
1841 if (td->bw_log)
1842 add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
1843
1844 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1845 }
1846
1847 fio_gettime(&td->bw_sample_time, NULL);
1848}
1849
1850void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1851 struct timeval *t)
1852{
1853 struct thread_stat *ts = &td->ts;
1854 unsigned long spent, iops;
1855
1856 if (!ddir_rw(ddir))
1857 return;
1858
1859 spent = mtime_since(&td->iops_sample_time, t);
1860 if (spent < td->o.iops_avg_time)
1861 return;
1862
1863 /*
1864 * Compute both read and write rates for the interval.
1865 */
1866 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1867 uint64_t delta;
1868
1869 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1870 if (!delta)
1871 continue; /* No entries for interval */
1872
1873 if (spent)
1874 iops = (delta * 1000) / spent;
1875 else
1876 iops = 0;
1877
1878 add_stat_sample(&ts->iops_stat[ddir], iops);
1879
1880 if (td->iops_log)
1881 add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
1882
1883 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1884 }
1885
1886 fio_gettime(&td->iops_sample_time, NULL);
1887}
1888
1889void stat_init(void)
1890{
1891 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1892}
1893
1894void stat_exit(void)
1895{
1896 /*
1897 * When we have the mutex, we know out-of-band access to it
1898 * have ended.
1899 */
1900 fio_mutex_down(stat_mutex);
1901 fio_mutex_remove(stat_mutex);
1902}
1903
1904/*
1905 * Called from signal handler. Wake up status thread.
1906 */
1907void show_running_run_stats(void)
1908{
1909 helper_do_stat = 1;
1910 pthread_cond_signal(&helper_cond);
1911}
fio - Flexible IO Tester