stat: make add lat percentile functions inline
[fio.git] / stat.c
1 #include <stdio.h>
2 #include <string.h>
3 #include <sys/time.h>
4 #include <sys/stat.h>
5 #include <math.h>
6
7 #include "fio.h"
8 #include "diskutil.h"
9 #include "lib/ieee754.h"
10 #include "json.h"
11 #include "lib/getrusage.h"
12 #include "idletime.h"
13 #include "lib/pow2.h"
14 #include "lib/output_buffer.h"
15 #include "helper_thread.h"
16 #include "smalloc.h"
17 #include "zbd.h"
18 #include "oslib/asprintf.h"
19
20 #ifdef WIN32
21 #define LOG_MSEC_SLACK  2
22 #else
23 #define LOG_MSEC_SLACK  1
24 #endif
25
26 struct fio_sem *stat_sem;
27
28 void clear_rusage_stat(struct thread_data *td)
29 {
30         struct thread_stat *ts = &td->ts;
31
32         fio_getrusage(&td->ru_start);
33         ts->usr_time = ts->sys_time = 0;
34         ts->ctx = 0;
35         ts->minf = ts->majf = 0;
36 }
37
38 void update_rusage_stat(struct thread_data *td)
39 {
40         struct thread_stat *ts = &td->ts;
41
42         fio_getrusage(&td->ru_end);
43         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
44                                         &td->ru_end.ru_utime);
45         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
46                                         &td->ru_end.ru_stime);
47         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
48                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
49         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
50         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
51
52         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
53 }
54
55 /*
56  * Given a latency, return the index of the corresponding bucket in
57  * the structure tracking percentiles.
58  *
59  * (1) find the group (and error bits) that the value (latency)
60  * belongs to by looking at its MSB. (2) find the bucket number in the
61  * group by looking at the index bits.
62  *
63  */
64 static unsigned int plat_val_to_idx(unsigned long long val)
65 {
66         unsigned int msb, error_bits, base, offset, idx;
67
68         /* Find MSB starting from bit 0 */
69         if (val == 0)
70                 msb = 0;
71         else
72                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
73
74         /*
75          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
76          * all bits of the sample as index
77          */
78         if (msb <= FIO_IO_U_PLAT_BITS)
79                 return val;
80
81         /* Compute the number of error bits to discard*/
82         error_bits = msb - FIO_IO_U_PLAT_BITS;
83
84         /* Compute the number of buckets before the group */
85         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
86
87         /*
88          * Discard the error bits and apply the mask to find the
89          * index for the buckets in the group
90          */
91         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
92
93         /* Make sure the index does not exceed (array size - 1) */
94         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
95                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
96
97         return idx;
98 }
99
100 /*
101  * Convert the given index of the bucket array to the value
102  * represented by the bucket
103  */
104 static unsigned long long plat_idx_to_val(unsigned int idx)
105 {
106         unsigned int error_bits;
107         unsigned long long k, base;
108
109         assert(idx < FIO_IO_U_PLAT_NR);
110
111         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
112          * all bits of the sample as index */
113         if (idx < (FIO_IO_U_PLAT_VAL << 1))
114                 return idx;
115
116         /* Find the group and compute the minimum value of that group */
117         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
118         base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
119
120         /* Find its bucket number of the group */
121         k = idx % FIO_IO_U_PLAT_VAL;
122
123         /* Return the mean of the range of the bucket */
124         return base + ((k + 0.5) * (1 << error_bits));
125 }
126
127 static int double_cmp(const void *a, const void *b)
128 {
129         const fio_fp64_t fa = *(const fio_fp64_t *) a;
130         const fio_fp64_t fb = *(const fio_fp64_t *) b;
131         int cmp = 0;
132
133         if (fa.u.f > fb.u.f)
134                 cmp = 1;
135         else if (fa.u.f < fb.u.f)
136                 cmp = -1;
137
138         return cmp;
139 }
140
141 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
142                                    fio_fp64_t *plist, unsigned long long **output,
143                                    unsigned long long *maxv, unsigned long long *minv)
144 {
145         unsigned long long sum = 0;
146         unsigned int len, i, j = 0;
147         unsigned long long *ovals = NULL;
148         bool is_last;
149
150         *minv = -1ULL;
151         *maxv = 0;
152
153         len = 0;
154         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
155                 len++;
156
157         if (!len)
158                 return 0;
159
160         /*
161          * Sort the percentile list. Note that it may already be sorted if
162          * we are using the default values, but since it's a short list this
163          * isn't a worry. Also note that this does not work for NaN values.
164          */
165         if (len > 1)
166                 qsort(plist, len, sizeof(plist[0]), double_cmp);
167
168         ovals = malloc(len * sizeof(*ovals));
169         if (!ovals)
170                 return 0;
171
172         /*
173          * Calculate bucket values, note down max and min values
174          */
175         is_last = false;
176         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
177                 sum += io_u_plat[i];
178                 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
179                         assert(plist[j].u.f <= 100.0);
180
181                         ovals[j] = plat_idx_to_val(i);
182                         if (ovals[j] < *minv)
183                                 *minv = ovals[j];
184                         if (ovals[j] > *maxv)
185                                 *maxv = ovals[j];
186
187                         is_last = (j == len - 1) != 0;
188                         if (is_last)
189                                 break;
190
191                         j++;
192                 }
193         }
194
195         if (!is_last)
196                 log_err("fio: error calculating latency percentiles\n");
197
198         *output = ovals;
199         return len;
200 }
201
202 /*
203  * Find and display the p-th percentile of clat
204  */
205 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
206                                   fio_fp64_t *plist, unsigned int precision,
207                                   const char *pre, struct buf_output *out)
208 {
209         unsigned int divisor, len, i, j = 0;
210         unsigned long long minv, maxv;
211         unsigned long long *ovals;
212         int per_line, scale_down, time_width;
213         bool is_last;
214         char fmt[32];
215
216         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
217         if (!len || !ovals)
218                 return;
219
220         /*
221          * We default to nsecs, but if the value range is such that we
222          * should scale down to usecs or msecs, do that.
223          */
224         if (minv > 2000000 && maxv > 99999999ULL) {
225                 scale_down = 2;
226                 divisor = 1000000;
227                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
228         } else if (minv > 2000 && maxv > 99999) {
229                 scale_down = 1;
230                 divisor = 1000;
231                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
232         } else {
233                 scale_down = 0;
234                 divisor = 1;
235                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
236         }
237
238
239         time_width = max(5, (int) (log10(maxv / divisor) + 1));
240         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
241                         precision, time_width);
242         /* fmt will be something like " %5.2fth=[%4llu]%c" */
243         per_line = (80 - 7) / (precision + 10 + time_width);
244
245         for (j = 0; j < len; j++) {
246                 /* for formatting */
247                 if (j != 0 && (j % per_line) == 0)
248                         log_buf(out, "     |");
249
250                 /* end of the list */
251                 is_last = (j == len - 1) != 0;
252
253                 for (i = 0; i < scale_down; i++)
254                         ovals[j] = (ovals[j] + 999) / 1000;
255
256                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
257
258                 if (is_last)
259                         break;
260
261                 if ((j % per_line) == per_line - 1)     /* for formatting */
262                         log_buf(out, "\n");
263         }
264
265         free(ovals);
266 }
267
268 bool calc_lat(struct io_stat *is, unsigned long long *min,
269               unsigned long long *max, double *mean, double *dev)
270 {
271         double n = (double) is->samples;
272
273         if (n == 0)
274                 return false;
275
276         *min = is->min_val;
277         *max = is->max_val;
278         *mean = is->mean.u.f;
279
280         if (n > 1.0)
281                 *dev = sqrt(is->S.u.f / (n - 1.0));
282         else
283                 *dev = 0;
284
285         return true;
286 }
287
288 void show_mixed_group_stats(struct group_run_stats *rs, struct buf_output *out) 
289 {
290         char *io, *agg, *min, *max;
291         char *ioalt, *aggalt, *minalt, *maxalt;
292         uint64_t io_mix = 0, agg_mix = 0, min_mix = -1, max_mix = 0, min_run = -1, max_run = 0;
293         int i;
294         const int i2p = is_power_of_2(rs->kb_base);
295
296         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
297                 if (!rs->max_run[i])
298                         continue;
299                 io_mix += rs->iobytes[i];
300                 agg_mix += rs->agg[i];
301                 min_mix = min_mix < rs->min_bw[i] ? min_mix : rs->min_bw[i];
302                 max_mix = max_mix > rs->max_bw[i] ? max_mix : rs->max_bw[i];
303                 min_run = min_run < rs->min_run[i] ? min_run : rs->min_run[i];
304                 max_run = max_run > rs->max_run[i] ? max_run : rs->max_run[i];
305         }
306         io = num2str(io_mix, rs->sig_figs, 1, i2p, N2S_BYTE);
307         ioalt = num2str(io_mix, rs->sig_figs, 1, !i2p, N2S_BYTE);
308         agg = num2str(agg_mix, rs->sig_figs, 1, i2p, rs->unit_base);
309         aggalt = num2str(agg_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
310         min = num2str(min_mix, rs->sig_figs, 1, i2p, rs->unit_base);
311         minalt = num2str(min_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
312         max = num2str(max_mix, rs->sig_figs, 1, i2p, rs->unit_base);
313         maxalt = num2str(max_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
314         log_buf(out, "  MIXED: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
315                         agg, aggalt, min, max, minalt, maxalt, io, ioalt,
316                         (unsigned long long) min_run,
317                         (unsigned long long) max_run);
318         free(io);
319         free(agg);
320         free(min);
321         free(max);
322         free(ioalt);
323         free(aggalt);
324         free(minalt);
325         free(maxalt);
326 }
327
328 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
329 {
330         char *io, *agg, *min, *max;
331         char *ioalt, *aggalt, *minalt, *maxalt;
332         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
333         int i;
334
335         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
336
337         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
338                 const int i2p = is_power_of_2(rs->kb_base);
339
340                 if (!rs->max_run[i])
341                         continue;
342
343                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
344                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
345                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
346                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
347                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
348                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
349                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
350                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
351                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
352                                 (rs->unified_rw_rep == UNIFIED_MIXED) ? "  MIXED" : str[i],
353                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
354                                 (unsigned long long) rs->min_run[i],
355                                 (unsigned long long) rs->max_run[i]);
356
357                 free(io);
358                 free(agg);
359                 free(min);
360                 free(max);
361                 free(ioalt);
362                 free(aggalt);
363                 free(minalt);
364                 free(maxalt);
365         }
366         
367         /* Need to aggregate statisitics to show mixed values */
368         if (rs->unified_rw_rep == UNIFIED_BOTH) 
369                 show_mixed_group_stats(rs, out);
370 }
371
372 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
373 {
374         int i;
375
376         /*
377          * Do depth distribution calculations
378          */
379         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
380                 if (total) {
381                         io_u_dist[i] = (double) map[i] / (double) total;
382                         io_u_dist[i] *= 100.0;
383                         if (io_u_dist[i] < 0.1 && map[i])
384                                 io_u_dist[i] = 0.1;
385                 } else
386                         io_u_dist[i] = 0.0;
387         }
388 }
389
390 static void stat_calc_lat(struct thread_stat *ts, double *dst,
391                           uint64_t *src, int nr)
392 {
393         unsigned long total = ddir_rw_sum(ts->total_io_u);
394         int i;
395
396         /*
397          * Do latency distribution calculations
398          */
399         for (i = 0; i < nr; i++) {
400                 if (total) {
401                         dst[i] = (double) src[i] / (double) total;
402                         dst[i] *= 100.0;
403                         if (dst[i] < 0.01 && src[i])
404                                 dst[i] = 0.01;
405                 } else
406                         dst[i] = 0.0;
407         }
408 }
409
410 /*
411  * To keep the terse format unaltered, add all of the ns latency
412  * buckets to the first us latency bucket
413  */
414 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
415 {
416         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
417         int i;
418
419         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
420
421         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
422                 ntotal += ts->io_u_lat_n[i];
423
424         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
425 }
426
427 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
428 {
429         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
430 }
431
432 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
433 {
434         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
435 }
436
437 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
438 {
439         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
440 }
441
442 static void display_lat(const char *name, unsigned long long min,
443                         unsigned long long max, double mean, double dev,
444                         struct buf_output *out)
445 {
446         const char *base = "(nsec)";
447         char *minp, *maxp;
448
449         if (nsec_to_msec(&min, &max, &mean, &dev))
450                 base = "(msec)";
451         else if (nsec_to_usec(&min, &max, &mean, &dev))
452                 base = "(usec)";
453
454         minp = num2str(min, 6, 1, 0, N2S_NONE);
455         maxp = num2str(max, 6, 1, 0, N2S_NONE);
456
457         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
458                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
459
460         free(minp);
461         free(maxp);
462 }
463
464 static double convert_agg_kbytes_percent(struct group_run_stats *rs, int ddir, int mean)
465 {
466         double p_of_agg = 100.0;
467         if (rs && rs->agg[ddir] > 1024) {
468                 p_of_agg = mean * 100.0 / (double) (rs->agg[ddir] / 1024.0);
469
470                 if (p_of_agg > 100.0)
471                         p_of_agg = 100.0;
472         }
473         return p_of_agg;
474 }
475
476 static void show_mixed_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
477                              struct buf_output *out)
478 {
479         unsigned long runt;
480         unsigned long long min, max, bw, iops;
481         double mean, dev;
482         char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
483         struct thread_stat *ts_lcl;
484
485         int i2p;
486         int ddir = 0;
487
488         /* Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and Trims (ddir = 2) */
489         ts_lcl = malloc(sizeof(struct thread_stat));
490         memset((void *)ts_lcl, 0, sizeof(struct thread_stat));
491         ts_lcl->unified_rw_rep = UNIFIED_MIXED;               /* calculate mixed stats  */
492         init_thread_stat_min_vals(ts_lcl);
493
494         sum_thread_stats(ts_lcl, ts, 1);
495
496         assert(ddir_rw(ddir));
497
498         if (!ts_lcl->runtime[ddir])
499                 return;
500
501         i2p = is_power_of_2(rs->kb_base);
502         runt = ts_lcl->runtime[ddir];
503
504         bw = (1000 * ts_lcl->io_bytes[ddir]) / runt;
505         io_p = num2str(ts_lcl->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
506         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
507         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
508
509         iops = (1000 * ts_lcl->total_io_u[ddir]) / runt;
510         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
511
512         log_buf(out, "  mixed: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
513                         iops_p, bw_p, bw_p_alt, io_p,
514                         (unsigned long long) ts_lcl->runtime[ddir],
515                         post_st ? : "");
516
517         free(post_st);
518         free(io_p);
519         free(bw_p);
520         free(bw_p_alt);
521         free(iops_p);
522
523         if (calc_lat(&ts_lcl->slat_stat[ddir], &min, &max, &mean, &dev))
524                 display_lat("slat", min, max, mean, dev, out);
525         if (calc_lat(&ts_lcl->clat_stat[ddir], &min, &max, &mean, &dev))
526                 display_lat("clat", min, max, mean, dev, out);
527         if (calc_lat(&ts_lcl->lat_stat[ddir], &min, &max, &mean, &dev))
528                 display_lat(" lat", min, max, mean, dev, out);
529         if (calc_lat(&ts_lcl->clat_high_prio_stat[ddir], &min, &max, &mean, &dev)) {
530                 display_lat(ts_lcl->lat_percentiles ? "high prio_lat" : "high prio_clat",
531                                 min, max, mean, dev, out);
532                 if (calc_lat(&ts_lcl->clat_low_prio_stat[ddir], &min, &max, &mean, &dev))
533                         display_lat(ts_lcl->lat_percentiles ? "low prio_lat" : "low prio_clat",
534                                         min, max, mean, dev, out);
535         }
536
537         if (ts->slat_percentiles && ts_lcl->slat_stat[ddir].samples > 0)
538                 show_clat_percentiles(ts_lcl->io_u_plat[FIO_SLAT][ddir],
539                                 ts_lcl->slat_stat[ddir].samples,
540                                 ts->percentile_list,
541                                 ts->percentile_precision, "slat", out);
542         if (ts->clat_percentiles && ts_lcl->clat_stat[ddir].samples > 0)
543                 show_clat_percentiles(ts_lcl->io_u_plat[FIO_CLAT][ddir],
544                                 ts_lcl->clat_stat[ddir].samples,
545                                 ts->percentile_list,
546                                 ts->percentile_precision, "clat", out);
547         if (ts->lat_percentiles && ts_lcl->lat_stat[ddir].samples > 0)
548                 show_clat_percentiles(ts_lcl->io_u_plat[FIO_LAT][ddir],
549                                 ts_lcl->lat_stat[ddir].samples,
550                                 ts->percentile_list,
551                                 ts->percentile_precision, "lat", out);
552
553         if (ts->clat_percentiles || ts->lat_percentiles) {
554                 const char *name = ts->lat_percentiles ? "lat" : "clat";
555                 char prio_name[32];
556                 uint64_t samples;
557
558                 if (ts->lat_percentiles)
559                         samples = ts_lcl->lat_stat[ddir].samples;
560                 else
561                         samples = ts_lcl->clat_stat[ddir].samples;
562
563                 /* Only print this if some high and low priority stats were collected */
564                 if (ts_lcl->clat_high_prio_stat[ddir].samples > 0 &&
565                                 ts_lcl->clat_low_prio_stat[ddir].samples > 0)
566                 {
567                         sprintf(prio_name, "high prio (%.2f%%) %s",
568                                         100. * (double) ts_lcl->clat_high_prio_stat[ddir].samples / (double) samples,
569                                         name);
570                         show_clat_percentiles(ts_lcl->io_u_plat_high_prio[ddir],
571                                         ts_lcl->clat_high_prio_stat[ddir].samples,
572                                         ts->percentile_list,
573                                         ts->percentile_precision, prio_name, out);
574
575                         sprintf(prio_name, "low prio (%.2f%%) %s",
576                                         100. * (double) ts_lcl->clat_low_prio_stat[ddir].samples / (double) samples,
577                                         name);
578                         show_clat_percentiles(ts_lcl->io_u_plat_low_prio[ddir],
579                                         ts_lcl->clat_low_prio_stat[ddir].samples,
580                                         ts->percentile_list,
581                                         ts->percentile_precision, prio_name, out);
582                 }
583         }
584
585         if (calc_lat(&ts_lcl->bw_stat[ddir], &min, &max, &mean, &dev)) {
586                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
587                 const char *bw_str;
588
589                 if ((rs->unit_base == 1) && i2p)
590                         bw_str = "Kibit";
591                 else if (rs->unit_base == 1)
592                         bw_str = "kbit";
593                 else if (i2p)
594                         bw_str = "KiB";
595                 else
596                         bw_str = "kB";
597
598                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
599
600                 if (rs->unit_base == 1) {
601                         min *= 8.0;
602                         max *= 8.0;
603                         mean *= 8.0;
604                         dev *= 8.0;
605                 }
606
607                 if (mean > fkb_base * fkb_base) {
608                         min /= fkb_base;
609                         max /= fkb_base;
610                         mean /= fkb_base;
611                         dev /= fkb_base;
612                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
613                 }
614
615                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
616                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
617                         bw_str, min, max, p_of_agg, mean, dev,
618                         (&ts_lcl->bw_stat[ddir])->samples);
619         }
620         if (calc_lat(&ts_lcl->iops_stat[ddir], &min, &max, &mean, &dev)) {
621                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
622                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
623                         min, max, mean, dev, (&ts_lcl->iops_stat[ddir])->samples);
624         }
625
626         free(ts_lcl);
627 }
628
629 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
630                              int ddir, struct buf_output *out)
631 {
632         unsigned long runt;
633         unsigned long long min, max, bw, iops;
634         double mean, dev;
635         char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
636         int i2p;
637
638         if (ddir_sync(ddir)) {
639                 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
640                         log_buf(out, "  %s:\n", "fsync/fdatasync/sync_file_range");
641                         display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
642                         show_clat_percentiles(ts->io_u_sync_plat,
643                                                 ts->sync_stat.samples,
644                                                 ts->percentile_list,
645                                                 ts->percentile_precision,
646                                                 io_ddir_name(ddir), out);
647                 }
648                 return;
649         }
650
651         assert(ddir_rw(ddir));
652
653         if (!ts->runtime[ddir])
654                 return;
655
656         i2p = is_power_of_2(rs->kb_base);
657         runt = ts->runtime[ddir];
658
659         bw = (1000 * ts->io_bytes[ddir]) / runt;
660         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
661         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
662         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
663
664         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
665         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
666         if (ddir == DDIR_WRITE)
667                 post_st = zbd_write_status(ts);
668         else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
669                 uint64_t total;
670                 double hit;
671
672                 total = ts->cachehit + ts->cachemiss;
673                 hit = (double) ts->cachehit / (double) total;
674                 hit *= 100.0;
675                 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
676                         post_st = NULL;
677         }
678
679         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
680                         (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir),
681                         iops_p, bw_p, bw_p_alt, io_p,
682                         (unsigned long long) ts->runtime[ddir],
683                         post_st ? : "");
684
685         free(post_st);
686         free(io_p);
687         free(bw_p);
688         free(bw_p_alt);
689         free(iops_p);
690
691         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
692                 display_lat("slat", min, max, mean, dev, out);
693         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
694                 display_lat("clat", min, max, mean, dev, out);
695         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
696                 display_lat(" lat", min, max, mean, dev, out);
697         if (calc_lat(&ts->clat_high_prio_stat[ddir], &min, &max, &mean, &dev)) {
698                 display_lat(ts->lat_percentiles ? "high prio_lat" : "high prio_clat",
699                                 min, max, mean, dev, out);
700                 if (calc_lat(&ts->clat_low_prio_stat[ddir], &min, &max, &mean, &dev))
701                         display_lat(ts->lat_percentiles ? "low prio_lat" : "low prio_clat",
702                                         min, max, mean, dev, out);
703         }
704
705         if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
706                 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
707                                         ts->slat_stat[ddir].samples,
708                                         ts->percentile_list,
709                                         ts->percentile_precision, "slat", out);
710         if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
711                 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
712                                         ts->clat_stat[ddir].samples,
713                                         ts->percentile_list,
714                                         ts->percentile_precision, "clat", out);
715         if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
716                 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
717                                         ts->lat_stat[ddir].samples,
718                                         ts->percentile_list,
719                                         ts->percentile_precision, "lat", out);
720
721         if (ts->clat_percentiles || ts->lat_percentiles) {
722                 const char *name = ts->lat_percentiles ? "lat" : "clat";
723                 char prio_name[32];
724                 uint64_t samples;
725
726                 if (ts->lat_percentiles)
727                         samples = ts->lat_stat[ddir].samples;
728                 else
729                         samples = ts->clat_stat[ddir].samples;
730
731                 /* Only print this if some high and low priority stats were collected */
732                 if (ts->clat_high_prio_stat[ddir].samples > 0 &&
733                         ts->clat_low_prio_stat[ddir].samples > 0)
734                 {
735                         sprintf(prio_name, "high prio (%.2f%%) %s",
736                                         100. * (double) ts->clat_high_prio_stat[ddir].samples / (double) samples,
737                                         name);
738                         show_clat_percentiles(ts->io_u_plat_high_prio[ddir],
739                                                 ts->clat_high_prio_stat[ddir].samples,
740                                                 ts->percentile_list,
741                                                 ts->percentile_precision, prio_name, out);
742
743                         sprintf(prio_name, "low prio (%.2f%%) %s",
744                                         100. * (double) ts->clat_low_prio_stat[ddir].samples / (double) samples,
745                                         name);
746                         show_clat_percentiles(ts->io_u_plat_low_prio[ddir],
747                                                 ts->clat_low_prio_stat[ddir].samples,
748                                                 ts->percentile_list,
749                                                 ts->percentile_precision, prio_name, out);
750                 }
751         }
752
753         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
754                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
755                 const char *bw_str;
756
757                 if ((rs->unit_base == 1) && i2p)
758                         bw_str = "Kibit";
759                 else if (rs->unit_base == 1)
760                         bw_str = "kbit";
761                 else if (i2p)
762                         bw_str = "KiB";
763                 else
764                         bw_str = "kB";
765
766                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
767
768                 if (rs->unit_base == 1) {
769                         min *= 8.0;
770                         max *= 8.0;
771                         mean *= 8.0;
772                         dev *= 8.0;
773                 }
774
775                 if (mean > fkb_base * fkb_base) {
776                         min /= fkb_base;
777                         max /= fkb_base;
778                         mean /= fkb_base;
779                         dev /= fkb_base;
780                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
781                 }
782
783                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
784                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
785                         bw_str, min, max, p_of_agg, mean, dev,
786                         (&ts->bw_stat[ddir])->samples);
787         }
788         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
789                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
790                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
791                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
792         }
793 }
794
795 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
796                      const char *msg, struct buf_output *out)
797 {
798         bool new_line = true, shown = false;
799         int i, line = 0;
800
801         for (i = 0; i < nr; i++) {
802                 if (io_u_lat[i] <= 0.0)
803                         continue;
804                 shown = true;
805                 if (new_line) {
806                         if (line)
807                                 log_buf(out, "\n");
808                         log_buf(out, "  lat (%s)   : ", msg);
809                         new_line = false;
810                         line = 0;
811                 }
812                 if (line)
813                         log_buf(out, ", ");
814                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
815                 line++;
816                 if (line == 5)
817                         new_line = true;
818         }
819
820         if (shown)
821                 log_buf(out, "\n");
822
823         return true;
824 }
825
826 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
827 {
828         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
829                                  "250=", "500=", "750=", "1000=", };
830
831         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
832 }
833
834 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
835 {
836         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
837                                  "250=", "500=", "750=", "1000=", };
838
839         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
840 }
841
842 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
843 {
844         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
845                                  "250=", "500=", "750=", "1000=", "2000=",
846                                  ">=2000=", };
847
848         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
849 }
850
851 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
852 {
853         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
854         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
855         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
856
857         stat_calc_lat_n(ts, io_u_lat_n);
858         stat_calc_lat_u(ts, io_u_lat_u);
859         stat_calc_lat_m(ts, io_u_lat_m);
860
861         show_lat_n(io_u_lat_n, out);
862         show_lat_u(io_u_lat_u, out);
863         show_lat_m(io_u_lat_m, out);
864 }
865
866 static int block_state_category(int block_state)
867 {
868         switch (block_state) {
869         case BLOCK_STATE_UNINIT:
870                 return 0;
871         case BLOCK_STATE_TRIMMED:
872         case BLOCK_STATE_WRITTEN:
873                 return 1;
874         case BLOCK_STATE_WRITE_FAILURE:
875         case BLOCK_STATE_TRIM_FAILURE:
876                 return 2;
877         default:
878                 /* Silence compile warning on some BSDs and have a return */
879                 assert(0);
880                 return -1;
881         }
882 }
883
884 static int compare_block_infos(const void *bs1, const void *bs2)
885 {
886         uint64_t block1 = *(uint64_t *)bs1;
887         uint64_t block2 = *(uint64_t *)bs2;
888         int state1 = BLOCK_INFO_STATE(block1);
889         int state2 = BLOCK_INFO_STATE(block2);
890         int bscat1 = block_state_category(state1);
891         int bscat2 = block_state_category(state2);
892         int cycles1 = BLOCK_INFO_TRIMS(block1);
893         int cycles2 = BLOCK_INFO_TRIMS(block2);
894
895         if (bscat1 < bscat2)
896                 return -1;
897         if (bscat1 > bscat2)
898                 return 1;
899
900         if (cycles1 < cycles2)
901                 return -1;
902         if (cycles1 > cycles2)
903                 return 1;
904
905         if (state1 < state2)
906                 return -1;
907         if (state1 > state2)
908                 return 1;
909
910         assert(block1 == block2);
911         return 0;
912 }
913
914 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
915                                   fio_fp64_t *plist, unsigned int **percentiles,
916                                   unsigned int *types)
917 {
918         int len = 0;
919         int i, nr_uninit;
920
921         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
922
923         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
924                 len++;
925
926         if (!len)
927                 return 0;
928
929         /*
930          * Sort the percentile list. Note that it may already be sorted if
931          * we are using the default values, but since it's a short list this
932          * isn't a worry. Also note that this does not work for NaN values.
933          */
934         if (len > 1)
935                 qsort(plist, len, sizeof(plist[0]), double_cmp);
936
937         /* Start only after the uninit entries end */
938         for (nr_uninit = 0;
939              nr_uninit < nr_block_infos
940                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
941              nr_uninit ++)
942                 ;
943
944         if (nr_uninit == nr_block_infos)
945                 return 0;
946
947         *percentiles = calloc(len, sizeof(**percentiles));
948
949         for (i = 0; i < len; i++) {
950                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
951                                 + nr_uninit;
952                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
953         }
954
955         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
956         for (i = 0; i < nr_block_infos; i++)
957                 types[BLOCK_INFO_STATE(block_infos[i])]++;
958
959         return len;
960 }
961
962 static const char *block_state_names[] = {
963         [BLOCK_STATE_UNINIT] = "unwritten",
964         [BLOCK_STATE_TRIMMED] = "trimmed",
965         [BLOCK_STATE_WRITTEN] = "written",
966         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
967         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
968 };
969
970 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
971                              fio_fp64_t *plist, struct buf_output *out)
972 {
973         int len, pos, i;
974         unsigned int *percentiles = NULL;
975         unsigned int block_state_counts[BLOCK_STATE_COUNT];
976
977         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
978                                      &percentiles, block_state_counts);
979
980         log_buf(out, "  block lifetime percentiles :\n   |");
981         pos = 0;
982         for (i = 0; i < len; i++) {
983                 uint32_t block_info = percentiles[i];
984 #define LINE_LENGTH     75
985                 char str[LINE_LENGTH];
986                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
987                                      plist[i].u.f, block_info,
988                                      i == len - 1 ? '\n' : ',');
989                 assert(strln < LINE_LENGTH);
990                 if (pos + strln > LINE_LENGTH) {
991                         pos = 0;
992                         log_buf(out, "\n   |");
993                 }
994                 log_buf(out, "%s", str);
995                 pos += strln;
996 #undef LINE_LENGTH
997         }
998         if (percentiles)
999                 free(percentiles);
1000
1001         log_buf(out, "        states               :");
1002         for (i = 0; i < BLOCK_STATE_COUNT; i++)
1003                 log_buf(out, " %s=%u%c",
1004                          block_state_names[i], block_state_counts[i],
1005                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
1006 }
1007
1008 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
1009 {
1010         char *p1, *p1alt, *p2;
1011         unsigned long long bw_mean, iops_mean;
1012         const int i2p = is_power_of_2(ts->kb_base);
1013
1014         if (!ts->ss_dur)
1015                 return;
1016
1017         bw_mean = steadystate_bw_mean(ts);
1018         iops_mean = steadystate_iops_mean(ts);
1019
1020         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
1021         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
1022         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
1023
1024         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
1025                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
1026                 p1, p1alt, p2,
1027                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1028                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
1029                 ts->ss_criterion.u.f,
1030                 ts->ss_state & FIO_SS_PCT ? "%" : "");
1031
1032         free(p1);
1033         free(p1alt);
1034         free(p2);
1035 }
1036
1037 static void show_agg_stats(struct disk_util_agg *agg, int terse,
1038                            struct buf_output *out)
1039 {
1040         if (!agg->slavecount)
1041                 return;
1042
1043         if (!terse) {
1044                 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
1045                          "aggrticks=%llu/%llu, aggrin_queue=%llu, "
1046                          "aggrutil=%3.2f%%",
1047                         (unsigned long long) agg->ios[0] / agg->slavecount,
1048                         (unsigned long long) agg->ios[1] / agg->slavecount,
1049                         (unsigned long long) agg->merges[0] / agg->slavecount,
1050                         (unsigned long long) agg->merges[1] / agg->slavecount,
1051                         (unsigned long long) agg->ticks[0] / agg->slavecount,
1052                         (unsigned long long) agg->ticks[1] / agg->slavecount,
1053                         (unsigned long long) agg->time_in_queue / agg->slavecount,
1054                         agg->max_util.u.f);
1055         } else {
1056                 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1057                         (unsigned long long) agg->ios[0] / agg->slavecount,
1058                         (unsigned long long) agg->ios[1] / agg->slavecount,
1059                         (unsigned long long) agg->merges[0] / agg->slavecount,
1060                         (unsigned long long) agg->merges[1] / agg->slavecount,
1061                         (unsigned long long) agg->ticks[0] / agg->slavecount,
1062                         (unsigned long long) agg->ticks[1] / agg->slavecount,
1063                         (unsigned long long) agg->time_in_queue / agg->slavecount,
1064                         agg->max_util.u.f);
1065         }
1066 }
1067
1068 static void aggregate_slaves_stats(struct disk_util *masterdu)
1069 {
1070         struct disk_util_agg *agg = &masterdu->agg;
1071         struct disk_util_stat *dus;
1072         struct flist_head *entry;
1073         struct disk_util *slavedu;
1074         double util;
1075
1076         flist_for_each(entry, &masterdu->slaves) {
1077                 slavedu = flist_entry(entry, struct disk_util, slavelist);
1078                 dus = &slavedu->dus;
1079                 agg->ios[0] += dus->s.ios[0];
1080                 agg->ios[1] += dus->s.ios[1];
1081                 agg->merges[0] += dus->s.merges[0];
1082                 agg->merges[1] += dus->s.merges[1];
1083                 agg->sectors[0] += dus->s.sectors[0];
1084                 agg->sectors[1] += dus->s.sectors[1];
1085                 agg->ticks[0] += dus->s.ticks[0];
1086                 agg->ticks[1] += dus->s.ticks[1];
1087                 agg->time_in_queue += dus->s.time_in_queue;
1088                 agg->slavecount++;
1089
1090                 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
1091                 /* System utilization is the utilization of the
1092                  * component with the highest utilization.
1093                  */
1094                 if (util > agg->max_util.u.f)
1095                         agg->max_util.u.f = util;
1096
1097         }
1098
1099         if (agg->max_util.u.f > 100.0)
1100                 agg->max_util.u.f = 100.0;
1101 }
1102
1103 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
1104                      int terse, struct buf_output *out)
1105 {
1106         double util = 0;
1107
1108         if (dus->s.msec)
1109                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1110         if (util > 100.0)
1111                 util = 100.0;
1112
1113         if (!terse) {
1114                 if (agg->slavecount)
1115                         log_buf(out, "  ");
1116
1117                 log_buf(out, "  %s: ios=%llu/%llu, merge=%llu/%llu, "
1118                          "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
1119                                 dus->name,
1120                                 (unsigned long long) dus->s.ios[0],
1121                                 (unsigned long long) dus->s.ios[1],
1122                                 (unsigned long long) dus->s.merges[0],
1123                                 (unsigned long long) dus->s.merges[1],
1124                                 (unsigned long long) dus->s.ticks[0],
1125                                 (unsigned long long) dus->s.ticks[1],
1126                                 (unsigned long long) dus->s.time_in_queue,
1127                                 util);
1128         } else {
1129                 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1130                                 dus->name,
1131                                 (unsigned long long) dus->s.ios[0],
1132                                 (unsigned long long) dus->s.ios[1],
1133                                 (unsigned long long) dus->s.merges[0],
1134                                 (unsigned long long) dus->s.merges[1],
1135                                 (unsigned long long) dus->s.ticks[0],
1136                                 (unsigned long long) dus->s.ticks[1],
1137                                 (unsigned long long) dus->s.time_in_queue,
1138                                 util);
1139         }
1140
1141         /*
1142          * If the device has slaves, aggregate the stats for
1143          * those slave devices also.
1144          */
1145         show_agg_stats(agg, terse, out);
1146
1147         if (!terse)
1148                 log_buf(out, "\n");
1149 }
1150
1151 void json_array_add_disk_util(struct disk_util_stat *dus,
1152                 struct disk_util_agg *agg, struct json_array *array)
1153 {
1154         struct json_object *obj;
1155         double util = 0;
1156
1157         if (dus->s.msec)
1158                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1159         if (util > 100.0)
1160                 util = 100.0;
1161
1162         obj = json_create_object();
1163         json_array_add_value_object(array, obj);
1164
1165         json_object_add_value_string(obj, "name", (const char *)dus->name);
1166         json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1167         json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1168         json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1169         json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1170         json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1171         json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1172         json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1173         json_object_add_value_float(obj, "util", util);
1174
1175         /*
1176          * If the device has slaves, aggregate the stats for
1177          * those slave devices also.
1178          */
1179         if (!agg->slavecount)
1180                 return;
1181         json_object_add_value_int(obj, "aggr_read_ios",
1182                                 agg->ios[0] / agg->slavecount);
1183         json_object_add_value_int(obj, "aggr_write_ios",
1184                                 agg->ios[1] / agg->slavecount);
1185         json_object_add_value_int(obj, "aggr_read_merges",
1186                                 agg->merges[0] / agg->slavecount);
1187         json_object_add_value_int(obj, "aggr_write_merge",
1188                                 agg->merges[1] / agg->slavecount);
1189         json_object_add_value_int(obj, "aggr_read_ticks",
1190                                 agg->ticks[0] / agg->slavecount);
1191         json_object_add_value_int(obj, "aggr_write_ticks",
1192                                 agg->ticks[1] / agg->slavecount);
1193         json_object_add_value_int(obj, "aggr_in_queue",
1194                                 agg->time_in_queue / agg->slavecount);
1195         json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1196 }
1197
1198 static void json_object_add_disk_utils(struct json_object *obj,
1199                                        struct flist_head *head)
1200 {
1201         struct json_array *array = json_create_array();
1202         struct flist_head *entry;
1203         struct disk_util *du;
1204
1205         json_object_add_value_array(obj, "disk_util", array);
1206
1207         flist_for_each(entry, head) {
1208                 du = flist_entry(entry, struct disk_util, list);
1209
1210                 aggregate_slaves_stats(du);
1211                 json_array_add_disk_util(&du->dus, &du->agg, array);
1212         }
1213 }
1214
1215 void show_disk_util(int terse, struct json_object *parent,
1216                     struct buf_output *out)
1217 {
1218         struct flist_head *entry;
1219         struct disk_util *du;
1220         bool do_json;
1221
1222         if (!is_running_backend())
1223                 return;
1224
1225         if (flist_empty(&disk_list)) {
1226                 return;
1227         }
1228
1229         if ((output_format & FIO_OUTPUT_JSON) && parent)
1230                 do_json = true;
1231         else
1232                 do_json = false;
1233
1234         if (!terse && !do_json)
1235                 log_buf(out, "\nDisk stats (read/write):\n");
1236
1237         if (do_json)
1238                 json_object_add_disk_utils(parent, &disk_list);
1239         else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1240                 flist_for_each(entry, &disk_list) {
1241                         du = flist_entry(entry, struct disk_util, list);
1242
1243                         aggregate_slaves_stats(du);
1244                         print_disk_util(&du->dus, &du->agg, terse, out);
1245                 }
1246         }
1247 }
1248
1249 static void show_thread_status_normal(struct thread_stat *ts,
1250                                       struct group_run_stats *rs,
1251                                       struct buf_output *out)
1252 {
1253         double usr_cpu, sys_cpu;
1254         unsigned long runtime;
1255         double io_u_dist[FIO_IO_U_MAP_NR];
1256         time_t time_p;
1257         char time_buf[32];
1258
1259         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1260                 return;
1261
1262         memset(time_buf, 0, sizeof(time_buf));
1263
1264         time(&time_p);
1265         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1266
1267         if (!ts->error) {
1268                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1269                                         ts->name, ts->groupid, ts->members,
1270                                         ts->error, (int) ts->pid, time_buf);
1271         } else {
1272                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1273                                         ts->name, ts->groupid, ts->members,
1274                                         ts->error, ts->verror, (int) ts->pid,
1275                                         time_buf);
1276         }
1277
1278         if (strlen(ts->description))
1279                 log_buf(out, "  Description  : [%s]\n", ts->description);
1280
1281         for_each_rw_ddir(ddir) {
1282                 if (ts->io_bytes[ddir])
1283                         show_ddir_status(rs, ts, ddir, out);
1284         }
1285
1286         if (ts->unified_rw_rep == UNIFIED_BOTH)
1287                 show_mixed_ddir_status(rs, ts, out);
1288
1289         show_latencies(ts, out);
1290
1291         if (ts->sync_stat.samples)
1292                 show_ddir_status(rs, ts, DDIR_SYNC, out);
1293
1294         runtime = ts->total_run_time;
1295         if (runtime) {
1296                 double runt = (double) runtime;
1297
1298                 usr_cpu = (double) ts->usr_time * 100 / runt;
1299                 sys_cpu = (double) ts->sys_time * 100 / runt;
1300         } else {
1301                 usr_cpu = 0;
1302                 sys_cpu = 0;
1303         }
1304
1305         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1306                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1307                         (unsigned long long) ts->ctx,
1308                         (unsigned long long) ts->majf,
1309                         (unsigned long long) ts->minf);
1310
1311         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1312         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1313                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1314                                         io_u_dist[1], io_u_dist[2],
1315                                         io_u_dist[3], io_u_dist[4],
1316                                         io_u_dist[5], io_u_dist[6]);
1317
1318         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1319         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1320                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1321                                         io_u_dist[1], io_u_dist[2],
1322                                         io_u_dist[3], io_u_dist[4],
1323                                         io_u_dist[5], io_u_dist[6]);
1324         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1325         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1326                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1327                                         io_u_dist[1], io_u_dist[2],
1328                                         io_u_dist[3], io_u_dist[4],
1329                                         io_u_dist[5], io_u_dist[6]);
1330         log_buf(out, "     issued rwts: total=%llu,%llu,%llu,%llu"
1331                                  " short=%llu,%llu,%llu,0"
1332                                  " dropped=%llu,%llu,%llu,0\n",
1333                                         (unsigned long long) ts->total_io_u[0],
1334                                         (unsigned long long) ts->total_io_u[1],
1335                                         (unsigned long long) ts->total_io_u[2],
1336                                         (unsigned long long) ts->total_io_u[3],
1337                                         (unsigned long long) ts->short_io_u[0],
1338                                         (unsigned long long) ts->short_io_u[1],
1339                                         (unsigned long long) ts->short_io_u[2],
1340                                         (unsigned long long) ts->drop_io_u[0],
1341                                         (unsigned long long) ts->drop_io_u[1],
1342                                         (unsigned long long) ts->drop_io_u[2]);
1343         if (ts->continue_on_error) {
1344                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
1345                                         (unsigned long long)ts->total_err_count,
1346                                         ts->first_error,
1347                                         strerror(ts->first_error));
1348         }
1349         if (ts->latency_depth) {
1350                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1351                                         (unsigned long long)ts->latency_target,
1352                                         (unsigned long long)ts->latency_window,
1353                                         ts->latency_percentile.u.f,
1354                                         ts->latency_depth);
1355         }
1356
1357         if (ts->nr_block_infos)
1358                 show_block_infos(ts->nr_block_infos, ts->block_infos,
1359                                   ts->percentile_list, out);
1360
1361         if (ts->ss_dur)
1362                 show_ss_normal(ts, out);
1363 }
1364
1365 static void show_ddir_status_terse(struct thread_stat *ts,
1366                                    struct group_run_stats *rs, int ddir,
1367                                    int ver, struct buf_output *out)
1368 {
1369         unsigned long long min, max, minv, maxv, bw, iops;
1370         unsigned long long *ovals = NULL;
1371         double mean, dev;
1372         unsigned int len;
1373         int i, bw_stat;
1374
1375         assert(ddir_rw(ddir));
1376
1377         iops = bw = 0;
1378         if (ts->runtime[ddir]) {
1379                 uint64_t runt = ts->runtime[ddir];
1380
1381                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1382                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1383         }
1384
1385         log_buf(out, ";%llu;%llu;%llu;%llu",
1386                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1387                                         (unsigned long long) ts->runtime[ddir]);
1388
1389         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1390                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1391         else
1392                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1393
1394         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1395                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1396         else
1397                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1398
1399         if (ts->lat_percentiles)
1400                 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1401                                         ts->lat_stat[ddir].samples,
1402                                         ts->percentile_list, &ovals, &maxv,
1403                                         &minv);
1404         else if (ts->clat_percentiles)
1405                 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1406                                         ts->clat_stat[ddir].samples,
1407                                         ts->percentile_list, &ovals, &maxv,
1408                                         &minv);
1409         else
1410                 len = 0;
1411         
1412         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1413                 if (i >= len) {
1414                         log_buf(out, ";0%%=0");
1415                         continue;
1416                 }
1417                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1418         }
1419
1420         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1421                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1422         else
1423                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1424
1425         free(ovals);
1426
1427         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1428         if (bw_stat) {
1429                 double p_of_agg = 100.0;
1430
1431                 if (rs->agg[ddir]) {
1432                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1433                         if (p_of_agg > 100.0)
1434                                 p_of_agg = 100.0;
1435                 }
1436
1437                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1438         } else
1439                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1440
1441         if (ver == 5) {
1442                 if (bw_stat)
1443                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1444                 else
1445                         log_buf(out, ";%lu", 0UL);
1446
1447                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1448                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1449                                 mean, dev, (&ts->iops_stat[ddir])->samples);
1450                 else
1451                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1452         }
1453 }
1454
1455 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1456                                    struct group_run_stats *rs,
1457                                    int ver, struct buf_output *out)
1458 {
1459         struct thread_stat *ts_lcl;
1460
1461         /* Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and Trims (ddir = 2) */
1462         ts_lcl = malloc(sizeof(struct thread_stat));
1463         memset((void *)ts_lcl, 0, sizeof(struct thread_stat));
1464         ts_lcl->unified_rw_rep = UNIFIED_MIXED;               /* calculate mixed stats  */
1465         init_thread_stat_min_vals(ts_lcl);
1466         ts_lcl->lat_percentiles = ts->lat_percentiles;
1467         ts_lcl->clat_percentiles = ts->clat_percentiles;
1468         ts_lcl->slat_percentiles = ts->slat_percentiles;
1469         ts_lcl->percentile_precision = ts->percentile_precision;                
1470         memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
1471         
1472         sum_thread_stats(ts_lcl, ts, 1);
1473
1474         /* add the aggregated stats to json parent */
1475         show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1476         free(ts_lcl);
1477 }
1478
1479 static struct json_object *add_ddir_lat_json(struct thread_stat *ts, uint32_t percentiles,
1480                 struct io_stat *lat_stat, uint64_t *io_u_plat)
1481 {
1482         char buf[120];
1483         double mean, dev;
1484         unsigned int i, len;
1485         struct json_object *lat_object, *percentile_object, *clat_bins_object;
1486         unsigned long long min, max, maxv, minv, *ovals = NULL;
1487
1488         if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1489                 min = max = 0;
1490                 mean = dev = 0.0;
1491         }
1492         lat_object = json_create_object();
1493         json_object_add_value_int(lat_object, "min", min);
1494         json_object_add_value_int(lat_object, "max", max);
1495         json_object_add_value_float(lat_object, "mean", mean);
1496         json_object_add_value_float(lat_object, "stddev", dev);
1497         json_object_add_value_int(lat_object, "N", lat_stat->samples);
1498
1499         if (percentiles && lat_stat->samples) {
1500                 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1501                                 ts->percentile_list, &ovals, &maxv, &minv);
1502
1503                 if (len > FIO_IO_U_LIST_MAX_LEN)
1504                         len = FIO_IO_U_LIST_MAX_LEN;
1505
1506                 percentile_object = json_create_object();
1507                 json_object_add_value_object(lat_object, "percentile", percentile_object);
1508                 for (i = 0; i < len; i++) {
1509                         snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1510                         json_object_add_value_int(percentile_object, buf, ovals[i]);
1511                 }
1512                 free(ovals);
1513
1514                 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1515                         clat_bins_object = json_create_object();
1516                         json_object_add_value_object(lat_object, "bins", clat_bins_object);
1517
1518                         for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1519                                 if (io_u_plat[i]) {
1520                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1521                                         json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1522                                 }
1523                 }
1524         }
1525
1526         return lat_object;
1527 }
1528
1529 static void add_ddir_status_json(struct thread_stat *ts,
1530                 struct group_run_stats *rs, int ddir, struct json_object *parent)
1531 {
1532         unsigned long long min, max;
1533         unsigned long long bw_bytes, bw;
1534         double mean, dev, iops;
1535         struct json_object *dir_object, *tmp_object;
1536         double p_of_agg = 100.0;
1537
1538         assert(ddir_rw(ddir) || ddir_sync(ddir));
1539
1540         if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1541                 return;
1542
1543         dir_object = json_create_object();
1544         json_object_add_value_object(parent,
1545                 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1546
1547         if (ddir_rw(ddir)) {
1548                 bw_bytes = 0;
1549                 bw = 0;
1550                 iops = 0.0;
1551                 if (ts->runtime[ddir]) {
1552                         uint64_t runt = ts->runtime[ddir];
1553
1554                         bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1555                         bw = bw_bytes / 1024; /* KiB/s */
1556                         iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1557                 }
1558
1559                 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1560                 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1561                 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1562                 json_object_add_value_int(dir_object, "bw", bw);
1563                 json_object_add_value_float(dir_object, "iops", iops);
1564                 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1565                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1566                 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1567                 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1568
1569                 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1570                                 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1571                 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1572
1573                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1574                                 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1575                 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1576
1577                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1578                                 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1579                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1580         } else {
1581                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1582                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1583                                 &ts->sync_stat, ts->io_u_sync_plat);
1584                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1585         }
1586
1587         if (!ddir_rw(ddir))
1588                 return;
1589
1590         /* Only print PRIO latencies if some high priority samples were gathered */
1591         if (ts->clat_high_prio_stat[ddir].samples > 0) {
1592                 const char *high, *low;
1593
1594                 if (ts->lat_percentiles) {
1595                         high = "lat_high_prio";
1596                         low = "lat_low_prio";
1597                 } else {
1598                         high = "clat_high_prio";
1599                         low = "clat_low_prio";
1600                 }
1601
1602                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1603                                 &ts->clat_high_prio_stat[ddir], ts->io_u_plat_high_prio[ddir]);
1604                 json_object_add_value_object(dir_object, high, tmp_object);
1605
1606                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1607                                 &ts->clat_low_prio_stat[ddir], ts->io_u_plat_low_prio[ddir]);
1608                 json_object_add_value_object(dir_object, low, tmp_object);
1609         }
1610
1611         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1612                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1613         } else {
1614                 min = max = 0;
1615                 p_of_agg = mean = dev = 0.0;
1616         }
1617
1618         json_object_add_value_int(dir_object, "bw_min", min);
1619         json_object_add_value_int(dir_object, "bw_max", max);
1620         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1621         json_object_add_value_float(dir_object, "bw_mean", mean);
1622         json_object_add_value_float(dir_object, "bw_dev", dev);
1623         json_object_add_value_int(dir_object, "bw_samples",
1624                                 (&ts->bw_stat[ddir])->samples);
1625
1626         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1627                 min = max = 0;
1628                 mean = dev = 0.0;
1629         }
1630         json_object_add_value_int(dir_object, "iops_min", min);
1631         json_object_add_value_int(dir_object, "iops_max", max);
1632         json_object_add_value_float(dir_object, "iops_mean", mean);
1633         json_object_add_value_float(dir_object, "iops_stddev", dev);
1634         json_object_add_value_int(dir_object, "iops_samples",
1635                                 (&ts->iops_stat[ddir])->samples);
1636
1637         if (ts->cachehit + ts->cachemiss) {
1638                 uint64_t total;
1639                 double hit;
1640
1641                 total = ts->cachehit + ts->cachemiss;
1642                 hit = (double) ts->cachehit / (double) total;
1643                 hit *= 100.0;
1644                 json_object_add_value_float(dir_object, "cachehit", hit);
1645         }
1646 }
1647
1648 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1649                 struct group_run_stats *rs, struct json_object *parent)
1650 {
1651         struct thread_stat *ts_lcl;
1652
1653         /* Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and Trims (ddir = 2) */
1654         ts_lcl = malloc(sizeof(struct thread_stat));
1655         memset((void *)ts_lcl, 0, sizeof(struct thread_stat));
1656         ts_lcl->unified_rw_rep = UNIFIED_MIXED;               /* calculate mixed stats  */
1657         init_thread_stat_min_vals(ts_lcl);
1658         ts_lcl->lat_percentiles = ts->lat_percentiles;
1659         ts_lcl->clat_percentiles = ts->clat_percentiles;
1660         ts_lcl->slat_percentiles = ts->slat_percentiles;
1661         ts_lcl->percentile_precision = ts->percentile_precision;                
1662         memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
1663
1664         sum_thread_stats(ts_lcl, ts, 1);
1665
1666         /* add the aggregated stats to json parent */
1667         add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1668         free(ts_lcl);
1669 }
1670
1671 static void show_thread_status_terse_all(struct thread_stat *ts,
1672                                          struct group_run_stats *rs, int ver,
1673                                          struct buf_output *out)
1674 {
1675         double io_u_dist[FIO_IO_U_MAP_NR];
1676         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1677         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1678         double usr_cpu, sys_cpu;
1679         int i;
1680
1681         /* General Info */
1682         if (ver == 2)
1683                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1684         else
1685                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1686                         ts->name, ts->groupid, ts->error);
1687
1688         /* Log Read Status, or mixed if unified_rw_rep = 1 */
1689         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1690         if (ts->unified_rw_rep != UNIFIED_MIXED) {
1691                 /* Log Write Status */
1692                 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1693                 /* Log Trim Status */
1694                 if (ver == 2 || ver == 4 || ver == 5)
1695                         show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1696         }
1697         if (ts->unified_rw_rep == UNIFIED_BOTH)
1698                 show_mixed_ddir_status_terse(ts, rs, ver, out);
1699         /* CPU Usage */
1700         if (ts->total_run_time) {
1701                 double runt = (double) ts->total_run_time;
1702
1703                 usr_cpu = (double) ts->usr_time * 100 / runt;
1704                 sys_cpu = (double) ts->sys_time * 100 / runt;
1705         } else {
1706                 usr_cpu = 0;
1707                 sys_cpu = 0;
1708         }
1709
1710         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1711                                                 (unsigned long long) ts->ctx,
1712                                                 (unsigned long long) ts->majf,
1713                                                 (unsigned long long) ts->minf);
1714
1715         /* Calc % distribution of IO depths, usecond, msecond latency */
1716         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1717         stat_calc_lat_nu(ts, io_u_lat_u);
1718         stat_calc_lat_m(ts, io_u_lat_m);
1719
1720         /* Only show fixed 7 I/O depth levels*/
1721         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1722                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1723                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1724
1725         /* Microsecond latency */
1726         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1727                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1728         /* Millisecond latency */
1729         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1730                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1731
1732         /* disk util stats, if any */
1733         if (ver >= 3 && is_running_backend())
1734                 show_disk_util(1, NULL, out);
1735
1736         /* Additional output if continue_on_error set - default off*/
1737         if (ts->continue_on_error)
1738                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1739
1740         /* Additional output if description is set */
1741         if (strlen(ts->description)) {
1742                 if (ver == 2)
1743                         log_buf(out, "\n");
1744                 log_buf(out, ";%s", ts->description);
1745         }
1746
1747         log_buf(out, "\n");
1748 }
1749
1750 static void json_add_job_opts(struct json_object *root, const char *name,
1751                               struct flist_head *opt_list)
1752 {
1753         struct json_object *dir_object;
1754         struct flist_head *entry;
1755         struct print_option *p;
1756
1757         if (flist_empty(opt_list))
1758                 return;
1759
1760         dir_object = json_create_object();
1761         json_object_add_value_object(root, name, dir_object);
1762
1763         flist_for_each(entry, opt_list) {
1764                 p = flist_entry(entry, struct print_option, list);
1765                 json_object_add_value_string(dir_object, p->name, p->value);
1766         }
1767 }
1768
1769 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1770                                                    struct group_run_stats *rs,
1771                                                    struct flist_head *opt_list)
1772 {
1773         struct json_object *root, *tmp;
1774         struct jobs_eta *je;
1775         double io_u_dist[FIO_IO_U_MAP_NR];
1776         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1777         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1778         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1779         double usr_cpu, sys_cpu;
1780         int i;
1781         size_t size;
1782
1783         root = json_create_object();
1784         json_object_add_value_string(root, "jobname", ts->name);
1785         json_object_add_value_int(root, "groupid", ts->groupid);
1786         json_object_add_value_int(root, "error", ts->error);
1787
1788         /* ETA Info */
1789         je = get_jobs_eta(true, &size);
1790         if (je) {
1791                 json_object_add_value_int(root, "eta", je->eta_sec);
1792                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1793         }
1794
1795         if (opt_list)
1796                 json_add_job_opts(root, "job options", opt_list);
1797
1798         add_ddir_status_json(ts, rs, DDIR_READ, root);
1799         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1800         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1801         add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1802
1803         if (ts->unified_rw_rep == UNIFIED_BOTH)
1804                 add_mixed_ddir_status_json(ts, rs, root);
1805
1806         /* CPU Usage */
1807         if (ts->total_run_time) {
1808                 double runt = (double) ts->total_run_time;
1809
1810                 usr_cpu = (double) ts->usr_time * 100 / runt;
1811                 sys_cpu = (double) ts->sys_time * 100 / runt;
1812         } else {
1813                 usr_cpu = 0;
1814                 sys_cpu = 0;
1815         }
1816         json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1817         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1818         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1819         json_object_add_value_int(root, "ctx", ts->ctx);
1820         json_object_add_value_int(root, "majf", ts->majf);
1821         json_object_add_value_int(root, "minf", ts->minf);
1822
1823         /* Calc % distribution of IO depths */
1824         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1825         tmp = json_create_object();
1826         json_object_add_value_object(root, "iodepth_level", tmp);
1827         /* Only show fixed 7 I/O depth levels*/
1828         for (i = 0; i < 7; i++) {
1829                 char name[20];
1830                 if (i < 6)
1831                         snprintf(name, 20, "%d", 1 << i);
1832                 else
1833                         snprintf(name, 20, ">=%d", 1 << i);
1834                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1835         }
1836
1837         /* Calc % distribution of submit IO depths */
1838         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1839         tmp = json_create_object();
1840         json_object_add_value_object(root, "iodepth_submit", tmp);
1841         /* Only show fixed 7 I/O depth levels*/
1842         for (i = 0; i < 7; i++) {
1843                 char name[20];
1844                 if (i == 0)
1845                         snprintf(name, 20, "0");
1846                 else if (i < 6)
1847                         snprintf(name, 20, "%d", 1 << (i+1));
1848                 else
1849                         snprintf(name, 20, ">=%d", 1 << i);
1850                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1851         }
1852
1853         /* Calc % distribution of completion IO depths */
1854         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1855         tmp = json_create_object();
1856         json_object_add_value_object(root, "iodepth_complete", tmp);
1857         /* Only show fixed 7 I/O depth levels*/
1858         for (i = 0; i < 7; i++) {
1859                 char name[20];
1860                 if (i == 0)
1861                         snprintf(name, 20, "0");
1862                 else if (i < 6)
1863                         snprintf(name, 20, "%d", 1 << (i+1));
1864                 else
1865                         snprintf(name, 20, ">=%d", 1 << i);
1866                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1867         }
1868
1869         /* Calc % distribution of nsecond, usecond, msecond latency */
1870         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1871         stat_calc_lat_n(ts, io_u_lat_n);
1872         stat_calc_lat_u(ts, io_u_lat_u);
1873         stat_calc_lat_m(ts, io_u_lat_m);
1874
1875         /* Nanosecond latency */
1876         tmp = json_create_object();
1877         json_object_add_value_object(root, "latency_ns", tmp);
1878         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1879                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1880                                  "250", "500", "750", "1000", };
1881                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1882         }
1883         /* Microsecond latency */
1884         tmp = json_create_object();
1885         json_object_add_value_object(root, "latency_us", tmp);
1886         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1887                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1888                                  "250", "500", "750", "1000", };
1889                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1890         }
1891         /* Millisecond latency */
1892         tmp = json_create_object();
1893         json_object_add_value_object(root, "latency_ms", tmp);
1894         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1895                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1896                                  "250", "500", "750", "1000", "2000",
1897                                  ">=2000", };
1898                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1899         }
1900
1901         /* Additional output if continue_on_error set - default off*/
1902         if (ts->continue_on_error) {
1903                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1904                 json_object_add_value_int(root, "first_error", ts->first_error);
1905         }
1906
1907         if (ts->latency_depth) {
1908                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1909                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1910                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1911                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1912         }
1913
1914         /* Additional output if description is set */
1915         if (strlen(ts->description))
1916                 json_object_add_value_string(root, "desc", ts->description);
1917
1918         if (ts->nr_block_infos) {
1919                 /* Block error histogram and types */
1920                 int len;
1921                 unsigned int *percentiles = NULL;
1922                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1923
1924                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1925                                              ts->percentile_list,
1926                                              &percentiles, block_state_counts);
1927
1928                 if (len) {
1929                         struct json_object *block, *percentile_object, *states;
1930                         int state;
1931                         block = json_create_object();
1932                         json_object_add_value_object(root, "block", block);
1933
1934                         percentile_object = json_create_object();
1935                         json_object_add_value_object(block, "percentiles",
1936                                                      percentile_object);
1937                         for (i = 0; i < len; i++) {
1938                                 char buf[20];
1939                                 snprintf(buf, sizeof(buf), "%f",
1940                                          ts->percentile_list[i].u.f);
1941                                 json_object_add_value_int(percentile_object,
1942                                                           buf,
1943                                                           percentiles[i]);
1944                         }
1945
1946                         states = json_create_object();
1947                         json_object_add_value_object(block, "states", states);
1948                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1949                                 json_object_add_value_int(states,
1950                                         block_state_names[state],
1951                                         block_state_counts[state]);
1952                         }
1953                         free(percentiles);
1954                 }
1955         }
1956
1957         if (ts->ss_dur) {
1958                 struct json_object *data;
1959                 struct json_array *iops, *bw;
1960                 int j, k, l;
1961                 char ss_buf[64];
1962
1963                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1964                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1965                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1966                         (float) ts->ss_limit.u.f,
1967                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1968
1969                 tmp = json_create_object();
1970                 json_object_add_value_object(root, "steadystate", tmp);
1971                 json_object_add_value_string(tmp, "ss", ss_buf);
1972                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1973                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1974
1975                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1976                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1977                 json_object_add_value_string(tmp, "criterion", ss_buf);
1978                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1979                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1980
1981                 data = json_create_object();
1982                 json_object_add_value_object(tmp, "data", data);
1983                 bw = json_create_array();
1984                 iops = json_create_array();
1985
1986                 /*
1987                 ** if ss was attained or the buffer is not full,
1988                 ** ss->head points to the first element in the list.
1989                 ** otherwise it actually points to the second element
1990                 ** in the list
1991                 */
1992                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1993                         j = ts->ss_head;
1994                 else
1995                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1996                 for (l = 0; l < ts->ss_dur; l++) {
1997                         k = (j + l) % ts->ss_dur;
1998                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1999                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
2000                 }
2001                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
2002                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
2003                 json_object_add_value_array(data, "iops", iops);
2004                 json_object_add_value_array(data, "bw", bw);
2005         }
2006
2007         return root;
2008 }
2009
2010 static void show_thread_status_terse(struct thread_stat *ts,
2011                                      struct group_run_stats *rs,
2012                                      struct buf_output *out)
2013 {
2014         if (terse_version >= 2 && terse_version <= 5)
2015                 show_thread_status_terse_all(ts, rs, terse_version, out);
2016         else
2017                 log_err("fio: bad terse version!? %d\n", terse_version);
2018 }
2019
2020 struct json_object *show_thread_status(struct thread_stat *ts,
2021                                        struct group_run_stats *rs,
2022                                        struct flist_head *opt_list,
2023                                        struct buf_output *out)
2024 {
2025         struct json_object *ret = NULL;
2026
2027         if (output_format & FIO_OUTPUT_TERSE)
2028                 show_thread_status_terse(ts, rs,  out);
2029         if (output_format & FIO_OUTPUT_JSON)
2030                 ret = show_thread_status_json(ts, rs, opt_list);
2031         if (output_format & FIO_OUTPUT_NORMAL)
2032                 show_thread_status_normal(ts, rs,  out);
2033
2034         return ret;
2035 }
2036
2037 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
2038 {
2039         double mean, S;
2040
2041         dst->min_val = min(dst->min_val, src->min_val);
2042         dst->max_val = max(dst->max_val, src->max_val);
2043
2044         /*
2045          * Compute new mean and S after the merge
2046          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
2047          *  #Parallel_algorithm>
2048          */
2049         if (first) {
2050                 mean = src->mean.u.f;
2051                 S = src->S.u.f;
2052         } else {
2053                 double delta = src->mean.u.f - dst->mean.u.f;
2054
2055                 mean = ((src->mean.u.f * src->samples) +
2056                         (dst->mean.u.f * dst->samples)) /
2057                         (dst->samples + src->samples);
2058
2059                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
2060                         (dst->samples * src->samples) /
2061                         (dst->samples + src->samples);
2062         }
2063
2064         dst->samples += src->samples;
2065         dst->mean.u.f = mean;
2066         dst->S.u.f = S;
2067
2068 }
2069
2070 /*
2071  * We sum two kinds of stats - one that is time based, in which case we
2072  * apply the proper summing technique, and then one that is iops/bw
2073  * numbers. For group_reporting, we should just add those up, not make
2074  * them the mean of everything.
2075  */
2076 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first,
2077                      bool pure_sum)
2078 {
2079         if (src->samples == 0)
2080                 return;
2081
2082         if (!pure_sum) {
2083                 __sum_stat(dst, src, first);
2084                 return;
2085         }
2086
2087         if (first) {
2088                 dst->min_val = src->min_val;
2089                 dst->max_val = src->max_val;
2090                 dst->samples = src->samples;
2091                 dst->mean.u.f = src->mean.u.f;
2092                 dst->S.u.f = src->S.u.f;
2093         } else {
2094                 dst->min_val += src->min_val;
2095                 dst->max_val += src->max_val;
2096                 dst->samples += src->samples;
2097                 dst->mean.u.f += src->mean.u.f;
2098                 dst->S.u.f += src->S.u.f;
2099         }
2100 }
2101
2102 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
2103 {
2104         int i;
2105
2106         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2107                 if (dst->max_run[i] < src->max_run[i])
2108                         dst->max_run[i] = src->max_run[i];
2109                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
2110                         dst->min_run[i] = src->min_run[i];
2111                 if (dst->max_bw[i] < src->max_bw[i])
2112                         dst->max_bw[i] = src->max_bw[i];
2113                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
2114                         dst->min_bw[i] = src->min_bw[i];
2115
2116                 dst->iobytes[i] += src->iobytes[i];
2117                 dst->agg[i] += src->agg[i];
2118         }
2119
2120         if (!dst->kb_base)
2121                 dst->kb_base = src->kb_base;
2122         if (!dst->unit_base)
2123                 dst->unit_base = src->unit_base;
2124         if (!dst->sig_figs)
2125                 dst->sig_figs = src->sig_figs;
2126 }
2127
2128 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
2129                       bool first)
2130 {
2131         int k, l, m;
2132
2133         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2134                 if (!(dst->unified_rw_rep == UNIFIED_MIXED)) {
2135                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first, false);
2136                         sum_stat(&dst->clat_high_prio_stat[l], &src->clat_high_prio_stat[l], first, false);
2137                         sum_stat(&dst->clat_low_prio_stat[l], &src->clat_low_prio_stat[l], first, false);
2138                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first, false);
2139                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first, false);
2140                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first, true);
2141                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first, true);
2142
2143                         dst->io_bytes[l] += src->io_bytes[l];
2144
2145                         if (dst->runtime[l] < src->runtime[l])
2146                                 dst->runtime[l] = src->runtime[l];
2147                 } else {
2148                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first, false);
2149                         sum_stat(&dst->clat_high_prio_stat[0], &src->clat_high_prio_stat[l], first, false);
2150                         sum_stat(&dst->clat_low_prio_stat[0], &src->clat_low_prio_stat[l], first, false);
2151                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first, false);
2152                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first, false);
2153                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first, true);
2154                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first, true);
2155
2156                         dst->io_bytes[0] += src->io_bytes[l];
2157
2158                         if (dst->runtime[0] < src->runtime[l])
2159                                 dst->runtime[0] = src->runtime[l];
2160
2161                         /*
2162                          * We're summing to the same destination, so override
2163                          * 'first' after the first iteration of the loop
2164                          */
2165                         first = false;
2166                 }
2167         }
2168
2169         sum_stat(&dst->sync_stat, &src->sync_stat, first, false);
2170         dst->usr_time += src->usr_time;
2171         dst->sys_time += src->sys_time;
2172         dst->ctx += src->ctx;
2173         dst->majf += src->majf;
2174         dst->minf += src->minf;
2175
2176         for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2177                 dst->io_u_map[k] += src->io_u_map[k];
2178                 dst->io_u_submit[k] += src->io_u_submit[k];
2179                 dst->io_u_complete[k] += src->io_u_complete[k];
2180         }
2181
2182         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2183                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2184         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2185                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2186         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2187                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2188
2189         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2190                 if (!(dst->unified_rw_rep == UNIFIED_MIXED)) {
2191                         dst->total_io_u[k] += src->total_io_u[k];
2192                         dst->short_io_u[k] += src->short_io_u[k];
2193                         dst->drop_io_u[k] += src->drop_io_u[k];
2194                 } else {
2195                         dst->total_io_u[0] += src->total_io_u[k];
2196                         dst->short_io_u[0] += src->short_io_u[k];
2197                         dst->drop_io_u[0] += src->drop_io_u[k];
2198                 }
2199         }
2200
2201         dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2202
2203         for (k = 0; k < FIO_LAT_CNT; k++)
2204                 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2205                         for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2206                                 if (!(dst->unified_rw_rep == UNIFIED_MIXED))
2207                                         dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2208                                 else
2209                                         dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2210
2211         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2212                 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2213
2214         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2215                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
2216                         if (!(dst->unified_rw_rep == UNIFIED_MIXED)) {
2217                                 dst->io_u_plat_high_prio[k][m] += src->io_u_plat_high_prio[k][m];
2218                                 dst->io_u_plat_low_prio[k][m] += src->io_u_plat_low_prio[k][m];
2219                         } else {
2220                                 dst->io_u_plat_high_prio[0][m] += src->io_u_plat_high_prio[k][m];
2221                                 dst->io_u_plat_low_prio[0][m] += src->io_u_plat_low_prio[k][m];
2222                         }
2223
2224                 }
2225         }
2226
2227         dst->total_run_time += src->total_run_time;
2228         dst->total_submit += src->total_submit;
2229         dst->total_complete += src->total_complete;
2230         dst->nr_zone_resets += src->nr_zone_resets;
2231         dst->cachehit += src->cachehit;
2232         dst->cachemiss += src->cachemiss;
2233 }
2234
2235 void init_group_run_stat(struct group_run_stats *gs)
2236 {
2237         int i;
2238         memset(gs, 0, sizeof(*gs));
2239
2240         for (i = 0; i < DDIR_RWDIR_CNT; i++)
2241                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2242 }
2243
2244 void init_thread_stat_min_vals(struct thread_stat *ts)
2245 {
2246         int i;
2247
2248         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2249                 ts->clat_stat[i].min_val = ULONG_MAX;
2250                 ts->slat_stat[i].min_val = ULONG_MAX;
2251                 ts->lat_stat[i].min_val = ULONG_MAX;
2252                 ts->bw_stat[i].min_val = ULONG_MAX;
2253                 ts->iops_stat[i].min_val = ULONG_MAX;
2254                 ts->clat_high_prio_stat[i].min_val = ULONG_MAX;
2255                 ts->clat_low_prio_stat[i].min_val = ULONG_MAX;
2256         }
2257         ts->sync_stat.min_val = ULONG_MAX;
2258 }
2259
2260 void init_thread_stat(struct thread_stat *ts)
2261 {
2262         memset(ts, 0, sizeof(*ts));
2263
2264         init_thread_stat_min_vals(ts);
2265         ts->groupid = -1;
2266 }
2267
2268 void __show_run_stats(void)
2269 {
2270         struct group_run_stats *runstats, *rs;
2271         struct thread_data *td;
2272         struct thread_stat *threadstats, *ts;
2273         int i, j, k, nr_ts, last_ts, idx;
2274         bool kb_base_warned = false;
2275         bool unit_base_warned = false;
2276         struct json_object *root = NULL;
2277         struct json_array *array = NULL;
2278         struct buf_output output[FIO_OUTPUT_NR];
2279         struct flist_head **opt_lists;
2280
2281         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2282
2283         for (i = 0; i < groupid + 1; i++)
2284                 init_group_run_stat(&runstats[i]);
2285
2286         /*
2287          * find out how many threads stats we need. if group reporting isn't
2288          * enabled, it's one-per-td.
2289          */
2290         nr_ts = 0;
2291         last_ts = -1;
2292         for_each_td(td, i) {
2293                 if (!td->o.group_reporting) {
2294                         nr_ts++;
2295                         continue;
2296                 }
2297                 if (last_ts == td->groupid)
2298                         continue;
2299                 if (!td->o.stats)
2300                         continue;
2301
2302                 last_ts = td->groupid;
2303                 nr_ts++;
2304         }
2305
2306         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2307         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2308
2309         for (i = 0; i < nr_ts; i++) {
2310                 init_thread_stat(&threadstats[i]);
2311                 opt_lists[i] = NULL;
2312         }
2313
2314         j = 0;
2315         last_ts = -1;
2316         idx = 0;
2317         for_each_td(td, i) {
2318                 if (!td->o.stats)
2319                         continue;
2320                 if (idx && (!td->o.group_reporting ||
2321                     (td->o.group_reporting && last_ts != td->groupid))) {
2322                         idx = 0;
2323                         j++;
2324                 }
2325
2326                 last_ts = td->groupid;
2327
2328                 ts = &threadstats[j];
2329
2330                 ts->clat_percentiles = td->o.clat_percentiles;
2331                 ts->lat_percentiles = td->o.lat_percentiles;
2332                 ts->slat_percentiles = td->o.slat_percentiles;
2333                 ts->percentile_precision = td->o.percentile_precision;
2334                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2335                 opt_lists[j] = &td->opt_list;
2336
2337                 idx++;
2338                 ts->members++;
2339
2340                 if (ts->groupid == -1) {
2341                         /*
2342                          * These are per-group shared already
2343                          */
2344                         snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2345                         if (td->o.description)
2346                                 snprintf(ts->description,
2347                                          sizeof(ts->description), "%s",
2348                                          td->o.description);
2349                         else
2350                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2351
2352                         /*
2353                          * If multiple entries in this group, this is
2354                          * the first member.
2355                          */
2356                         ts->thread_number = td->thread_number;
2357                         ts->groupid = td->groupid;
2358
2359                         /*
2360                          * first pid in group, not very useful...
2361                          */
2362                         ts->pid = td->pid;
2363
2364                         ts->kb_base = td->o.kb_base;
2365                         ts->unit_base = td->o.unit_base;
2366                         ts->sig_figs = td->o.sig_figs;
2367                         ts->unified_rw_rep = td->o.unified_rw_rep;
2368                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2369                         log_info("fio: kb_base differs for jobs in group, using"
2370                                  " %u as the base\n", ts->kb_base);
2371                         kb_base_warned = true;
2372                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2373                         log_info("fio: unit_base differs for jobs in group, using"
2374                                  " %u as the base\n", ts->unit_base);
2375                         unit_base_warned = true;
2376                 }
2377
2378                 ts->continue_on_error = td->o.continue_on_error;
2379                 ts->total_err_count += td->total_err_count;
2380                 ts->first_error = td->first_error;
2381                 if (!ts->error) {
2382                         if (!td->error && td->o.continue_on_error &&
2383                             td->first_error) {
2384                                 ts->error = td->first_error;
2385                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2386                                          td->verror);
2387                         } else  if (td->error) {
2388                                 ts->error = td->error;
2389                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2390                                          td->verror);
2391                         }
2392                 }
2393
2394                 ts->latency_depth = td->latency_qd;
2395                 ts->latency_target = td->o.latency_target;
2396                 ts->latency_percentile = td->o.latency_percentile;
2397                 ts->latency_window = td->o.latency_window;
2398
2399                 ts->nr_block_infos = td->ts.nr_block_infos;
2400                 for (k = 0; k < ts->nr_block_infos; k++)
2401                         ts->block_infos[k] = td->ts.block_infos[k];
2402
2403                 sum_thread_stats(ts, &td->ts, idx == 1);
2404
2405                 if (td->o.ss_dur) {
2406                         ts->ss_state = td->ss.state;
2407                         ts->ss_dur = td->ss.dur;
2408                         ts->ss_head = td->ss.head;
2409                         ts->ss_bw_data = td->ss.bw_data;
2410                         ts->ss_iops_data = td->ss.iops_data;
2411                         ts->ss_limit.u.f = td->ss.limit;
2412                         ts->ss_slope.u.f = td->ss.slope;
2413                         ts->ss_deviation.u.f = td->ss.deviation;
2414                         ts->ss_criterion.u.f = td->ss.criterion;
2415                 }
2416                 else
2417                         ts->ss_dur = ts->ss_state = 0;
2418         }
2419
2420         for (i = 0; i < nr_ts; i++) {
2421                 unsigned long long bw;
2422
2423                 ts = &threadstats[i];
2424                 if (ts->groupid == -1)
2425                         continue;
2426                 rs = &runstats[ts->groupid];
2427                 rs->kb_base = ts->kb_base;
2428                 rs->unit_base = ts->unit_base;
2429                 rs->sig_figs = ts->sig_figs;
2430                 rs->unified_rw_rep |= ts->unified_rw_rep;
2431
2432                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2433                         if (!ts->runtime[j])
2434                                 continue;
2435                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2436                                 rs->min_run[j] = ts->runtime[j];
2437                         if (ts->runtime[j] > rs->max_run[j])
2438                                 rs->max_run[j] = ts->runtime[j];
2439
2440                         bw = 0;
2441                         if (ts->runtime[j])
2442                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2443                         if (bw < rs->min_bw[j])
2444                                 rs->min_bw[j] = bw;
2445                         if (bw > rs->max_bw[j])
2446                                 rs->max_bw[j] = bw;
2447
2448                         rs->iobytes[j] += ts->io_bytes[j];
2449                 }
2450         }
2451
2452         for (i = 0; i < groupid + 1; i++) {
2453                 int ddir;
2454
2455                 rs = &runstats[i];
2456
2457                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2458                         if (rs->max_run[ddir])
2459                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2460                                                 rs->max_run[ddir];
2461                 }
2462         }
2463
2464         for (i = 0; i < FIO_OUTPUT_NR; i++)
2465                 buf_output_init(&output[i]);
2466
2467         /*
2468          * don't overwrite last signal output
2469          */
2470         if (output_format & FIO_OUTPUT_NORMAL)
2471                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2472         if (output_format & FIO_OUTPUT_JSON) {
2473                 struct thread_data *global;
2474                 char time_buf[32];
2475                 struct timeval now;
2476                 unsigned long long ms_since_epoch;
2477                 time_t tv_sec;
2478
2479                 gettimeofday(&now, NULL);
2480                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2481                                  (unsigned long long)(now.tv_usec) / 1000;
2482
2483                 tv_sec = now.tv_sec;
2484                 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2485                 if (time_buf[strlen(time_buf) - 1] == '\n')
2486                         time_buf[strlen(time_buf) - 1] = '\0';
2487
2488                 root = json_create_object();
2489                 json_object_add_value_string(root, "fio version", fio_version_string);
2490                 json_object_add_value_int(root, "timestamp", now.tv_sec);
2491                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2492                 json_object_add_value_string(root, "time", time_buf);
2493                 global = get_global_options();
2494                 json_add_job_opts(root, "global options", &global->opt_list);
2495                 array = json_create_array();
2496                 json_object_add_value_array(root, "jobs", array);
2497         }
2498
2499         if (is_backend)
2500                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2501
2502         for (i = 0; i < nr_ts; i++) {
2503                 ts = &threadstats[i];
2504                 rs = &runstats[ts->groupid];
2505
2506                 if (is_backend) {
2507                         fio_server_send_job_options(opt_lists[i], i);
2508                         fio_server_send_ts(ts, rs);
2509                 } else {
2510                         if (output_format & FIO_OUTPUT_TERSE)
2511                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2512                         if (output_format & FIO_OUTPUT_JSON) {
2513                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2514                                 json_array_add_value_object(array, tmp);
2515                         }
2516                         if (output_format & FIO_OUTPUT_NORMAL)
2517                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2518                 }
2519         }
2520         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2521                 /* disk util stats, if any */
2522                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2523
2524                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2525
2526                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2527                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2528                 json_free_object(root);
2529         }
2530
2531         for (i = 0; i < groupid + 1; i++) {
2532                 rs = &runstats[i];
2533
2534                 rs->groupid = i;
2535                 if (is_backend)
2536                         fio_server_send_gs(rs);
2537                 else if (output_format & FIO_OUTPUT_NORMAL)
2538                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2539         }
2540
2541         if (is_backend)
2542                 fio_server_send_du();
2543         else if (output_format & FIO_OUTPUT_NORMAL) {
2544                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2545                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2546         }
2547
2548         for (i = 0; i < FIO_OUTPUT_NR; i++) {
2549                 struct buf_output *out = &output[i];
2550
2551                 log_info_buf(out->buf, out->buflen);
2552                 buf_output_free(out);
2553         }
2554
2555         fio_idle_prof_cleanup();
2556
2557         log_info_flush();
2558         free(runstats);
2559         free(threadstats);
2560         free(opt_lists);
2561 }
2562
2563 int __show_running_run_stats(void)
2564 {
2565         struct thread_data *td;
2566         unsigned long long *rt;
2567         struct timespec ts;
2568         int i;
2569
2570         fio_sem_down(stat_sem);
2571
2572         rt = malloc(thread_number * sizeof(unsigned long long));
2573         fio_gettime(&ts, NULL);
2574
2575         for_each_td(td, i) {
2576                 td->update_rusage = 1;
2577                 for_each_rw_ddir(ddir) {
2578                         td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2579                 }
2580                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2581
2582                 rt[i] = mtime_since(&td->start, &ts);
2583                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2584                         td->ts.runtime[DDIR_READ] += rt[i];
2585                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2586                         td->ts.runtime[DDIR_WRITE] += rt[i];
2587                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2588                         td->ts.runtime[DDIR_TRIM] += rt[i];
2589         }
2590
2591         for_each_td(td, i) {
2592                 if (td->runstate >= TD_EXITED)
2593                         continue;
2594                 if (td->rusage_sem) {
2595                         td->update_rusage = 1;
2596                         fio_sem_down(td->rusage_sem);
2597                 }
2598                 td->update_rusage = 0;
2599         }
2600
2601         __show_run_stats();
2602
2603         for_each_td(td, i) {
2604                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2605                         td->ts.runtime[DDIR_READ] -= rt[i];
2606                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2607                         td->ts.runtime[DDIR_WRITE] -= rt[i];
2608                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2609                         td->ts.runtime[DDIR_TRIM] -= rt[i];
2610         }
2611
2612         free(rt);
2613         fio_sem_up(stat_sem);
2614
2615         return 0;
2616 }
2617
2618 static bool status_file_disabled;
2619
2620 #define FIO_STATUS_FILE         "fio-dump-status"
2621
2622 static int check_status_file(void)
2623 {
2624         struct stat sb;
2625         const char *temp_dir;
2626         char fio_status_file_path[PATH_MAX];
2627
2628         if (status_file_disabled)
2629                 return 0;
2630
2631         temp_dir = getenv("TMPDIR");
2632         if (temp_dir == NULL) {
2633                 temp_dir = getenv("TEMP");
2634                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2635                         temp_dir = NULL;
2636         }
2637         if (temp_dir == NULL)
2638                 temp_dir = "/tmp";
2639 #ifdef __COVERITY__
2640         __coverity_tainted_data_sanitize__(temp_dir);
2641 #endif
2642
2643         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2644
2645         if (stat(fio_status_file_path, &sb))
2646                 return 0;
2647
2648         if (unlink(fio_status_file_path) < 0) {
2649                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2650                                                         strerror(errno));
2651                 log_err("fio: disabling status file updates\n");
2652                 status_file_disabled = true;
2653         }
2654
2655         return 1;
2656 }
2657
2658 void check_for_running_stats(void)
2659 {
2660         if (check_status_file()) {
2661                 show_running_run_stats();
2662                 return;
2663         }
2664 }
2665
2666 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2667 {
2668         double val = data;
2669         double delta;
2670
2671         if (data > is->max_val)
2672                 is->max_val = data;
2673         if (data < is->min_val)
2674                 is->min_val = data;
2675
2676         delta = val - is->mean.u.f;
2677         if (delta) {
2678                 is->mean.u.f += delta / (is->samples + 1.0);
2679                 is->S.u.f += delta * (val - is->mean.u.f);
2680         }
2681
2682         is->samples++;
2683 }
2684
2685 /*
2686  * Return a struct io_logs, which is added to the tail of the log
2687  * list for 'iolog'.
2688  */
2689 static struct io_logs *get_new_log(struct io_log *iolog)
2690 {
2691         size_t new_samples;
2692         struct io_logs *cur_log;
2693
2694         /*
2695          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2696          * forever
2697          */
2698         if (!iolog->cur_log_max) {
2699                 new_samples = iolog->td->o.log_entries;
2700         } else {
2701                 new_samples = iolog->cur_log_max * 2;
2702                 if (new_samples > MAX_LOG_ENTRIES)
2703                         new_samples = MAX_LOG_ENTRIES;
2704         }
2705
2706         cur_log = smalloc(sizeof(*cur_log));
2707         if (cur_log) {
2708                 INIT_FLIST_HEAD(&cur_log->list);
2709                 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2710                 if (cur_log->log) {
2711                         cur_log->nr_samples = 0;
2712                         cur_log->max_samples = new_samples;
2713                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2714                         iolog->cur_log_max = new_samples;
2715                         return cur_log;
2716                 }
2717                 sfree(cur_log);
2718         }
2719
2720         return NULL;
2721 }
2722
2723 /*
2724  * Add and return a new log chunk, or return current log if big enough
2725  */
2726 static struct io_logs *regrow_log(struct io_log *iolog)
2727 {
2728         struct io_logs *cur_log;
2729         int i;
2730
2731         if (!iolog || iolog->disabled)
2732                 goto disable;
2733
2734         cur_log = iolog_cur_log(iolog);
2735         if (!cur_log) {
2736                 cur_log = get_new_log(iolog);
2737                 if (!cur_log)
2738                         return NULL;
2739         }
2740
2741         if (cur_log->nr_samples < cur_log->max_samples)
2742                 return cur_log;
2743
2744         /*
2745          * No room for a new sample. If we're compressing on the fly, flush
2746          * out the current chunk
2747          */
2748         if (iolog->log_gz) {
2749                 if (iolog_cur_flush(iolog, cur_log)) {
2750                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2751                         return NULL;
2752                 }
2753         }
2754
2755         /*
2756          * Get a new log array, and add to our list
2757          */
2758         cur_log = get_new_log(iolog);
2759         if (!cur_log) {
2760                 log_err("fio: failed extending iolog! Will stop logging.\n");
2761                 return NULL;
2762         }
2763
2764         if (!iolog->pending || !iolog->pending->nr_samples)
2765                 return cur_log;
2766
2767         /*
2768          * Flush pending items to new log
2769          */
2770         for (i = 0; i < iolog->pending->nr_samples; i++) {
2771                 struct io_sample *src, *dst;
2772
2773                 src = get_sample(iolog, iolog->pending, i);
2774                 dst = get_sample(iolog, cur_log, i);
2775                 memcpy(dst, src, log_entry_sz(iolog));
2776         }
2777         cur_log->nr_samples = iolog->pending->nr_samples;
2778
2779         iolog->pending->nr_samples = 0;
2780         return cur_log;
2781 disable:
2782         if (iolog)
2783                 iolog->disabled = true;
2784         return NULL;
2785 }
2786
2787 void regrow_logs(struct thread_data *td)
2788 {
2789         regrow_log(td->slat_log);
2790         regrow_log(td->clat_log);
2791         regrow_log(td->clat_hist_log);
2792         regrow_log(td->lat_log);
2793         regrow_log(td->bw_log);
2794         regrow_log(td->iops_log);
2795         td->flags &= ~TD_F_REGROW_LOGS;
2796 }
2797
2798 void regrow_agg_logs(void)
2799 {
2800         enum fio_ddir ddir;
2801
2802         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2803                 regrow_log(agg_io_log[ddir]);
2804 }
2805
2806 static struct io_logs *get_cur_log(struct io_log *iolog)
2807 {
2808         struct io_logs *cur_log;
2809
2810         cur_log = iolog_cur_log(iolog);
2811         if (!cur_log) {
2812                 cur_log = get_new_log(iolog);
2813                 if (!cur_log)
2814                         return NULL;
2815         }
2816
2817         if (cur_log->nr_samples < cur_log->max_samples)
2818                 return cur_log;
2819
2820         /*
2821          * Out of space. If we're in IO offload mode, or we're not doing
2822          * per unit logging (hence logging happens outside of the IO thread
2823          * as well), add a new log chunk inline. If we're doing inline
2824          * submissions, flag 'td' as needing a log regrow and we'll take
2825          * care of it on the submission side.
2826          */
2827         if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2828             !per_unit_log(iolog))
2829                 return regrow_log(iolog);
2830
2831         if (iolog->td)
2832                 iolog->td->flags |= TD_F_REGROW_LOGS;
2833         if (iolog->pending)
2834                 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2835         return iolog->pending;
2836 }
2837
2838 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2839                              enum fio_ddir ddir, unsigned long long bs,
2840                              unsigned long t, uint64_t offset,
2841                              unsigned int priority)
2842 {
2843         struct io_logs *cur_log;
2844
2845         if (iolog->disabled)
2846                 return;
2847         if (flist_empty(&iolog->io_logs))
2848                 iolog->avg_last[ddir] = t;
2849
2850         cur_log = get_cur_log(iolog);
2851         if (cur_log) {
2852                 struct io_sample *s;
2853
2854                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2855
2856                 s->data = data;
2857                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2858                 io_sample_set_ddir(iolog, s, ddir);
2859                 s->bs = bs;
2860                 s->priority = priority;
2861
2862                 if (iolog->log_offset) {
2863                         struct io_sample_offset *so = (void *) s;
2864
2865                         so->offset = offset;
2866                 }
2867
2868                 cur_log->nr_samples++;
2869                 return;
2870         }
2871
2872         iolog->disabled = true;
2873 }
2874
2875 static inline void reset_io_stat(struct io_stat *ios)
2876 {
2877         ios->min_val = -1ULL;
2878         ios->max_val = ios->samples = 0;
2879         ios->mean.u.f = ios->S.u.f = 0;
2880 }
2881
2882 void reset_io_stats(struct thread_data *td)
2883 {
2884         struct thread_stat *ts = &td->ts;
2885         int i, j, k;
2886
2887         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2888                 reset_io_stat(&ts->clat_high_prio_stat[i]);
2889                 reset_io_stat(&ts->clat_low_prio_stat[i]);
2890                 reset_io_stat(&ts->clat_stat[i]);
2891                 reset_io_stat(&ts->slat_stat[i]);
2892                 reset_io_stat(&ts->lat_stat[i]);
2893                 reset_io_stat(&ts->bw_stat[i]);
2894                 reset_io_stat(&ts->iops_stat[i]);
2895
2896                 ts->io_bytes[i] = 0;
2897                 ts->runtime[i] = 0;
2898                 ts->total_io_u[i] = 0;
2899                 ts->short_io_u[i] = 0;
2900                 ts->drop_io_u[i] = 0;
2901
2902                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2903                         ts->io_u_plat_high_prio[i][j] = 0;
2904                         ts->io_u_plat_low_prio[i][j] = 0;
2905                         if (!i)
2906                                 ts->io_u_sync_plat[j] = 0;
2907                 }
2908         }
2909
2910         for (i = 0; i < FIO_LAT_CNT; i++)
2911                 for (j = 0; j < DDIR_RWDIR_CNT; j++)
2912                         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2913                                 ts->io_u_plat[i][j][k] = 0;
2914
2915         ts->total_io_u[DDIR_SYNC] = 0;
2916
2917         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2918                 ts->io_u_map[i] = 0;
2919                 ts->io_u_submit[i] = 0;
2920                 ts->io_u_complete[i] = 0;
2921         }
2922
2923         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2924                 ts->io_u_lat_n[i] = 0;
2925         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2926                 ts->io_u_lat_u[i] = 0;
2927         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2928                 ts->io_u_lat_m[i] = 0;
2929
2930         ts->total_submit = 0;
2931         ts->total_complete = 0;
2932         ts->nr_zone_resets = 0;
2933         ts->cachehit = ts->cachemiss = 0;
2934 }
2935
2936 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2937                               unsigned long elapsed, bool log_max)
2938 {
2939         /*
2940          * Note an entry in the log. Use the mean from the logged samples,
2941          * making sure to properly round up. Only write a log entry if we
2942          * had actual samples done.
2943          */
2944         if (iolog->avg_window[ddir].samples) {
2945                 union io_sample_data data;
2946
2947                 if (log_max)
2948                         data.val = iolog->avg_window[ddir].max_val;
2949                 else
2950                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2951
2952                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
2953         }
2954
2955         reset_io_stat(&iolog->avg_window[ddir]);
2956 }
2957
2958 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2959                              bool log_max)
2960 {
2961         int ddir;
2962
2963         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2964                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2965 }
2966
2967 static unsigned long add_log_sample(struct thread_data *td,
2968                                     struct io_log *iolog,
2969                                     union io_sample_data data,
2970                                     enum fio_ddir ddir, unsigned long long bs,
2971                                     uint64_t offset, unsigned int ioprio)
2972 {
2973         unsigned long elapsed, this_window;
2974
2975         if (!ddir_rw(ddir))
2976                 return 0;
2977
2978         elapsed = mtime_since_now(&td->epoch);
2979
2980         /*
2981          * If no time averaging, just add the log sample.
2982          */
2983         if (!iolog->avg_msec) {
2984                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
2985                                  ioprio);
2986                 return 0;
2987         }
2988
2989         /*
2990          * Add the sample. If the time period has passed, then
2991          * add that entry to the log and clear.
2992          */
2993         add_stat_sample(&iolog->avg_window[ddir], data.val);
2994
2995         /*
2996          * If period hasn't passed, adding the above sample is all we
2997          * need to do.
2998          */
2999         this_window = elapsed - iolog->avg_last[ddir];
3000         if (elapsed < iolog->avg_last[ddir])
3001                 return iolog->avg_last[ddir] - elapsed;
3002         else if (this_window < iolog->avg_msec) {
3003                 unsigned long diff = iolog->avg_msec - this_window;
3004
3005                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
3006                         return diff;
3007         }
3008
3009         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
3010
3011         iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
3012
3013         return iolog->avg_msec;
3014 }
3015
3016 void finalize_logs(struct thread_data *td, bool unit_logs)
3017 {
3018         unsigned long elapsed;
3019
3020         elapsed = mtime_since_now(&td->epoch);
3021
3022         if (td->clat_log && unit_logs)
3023                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
3024         if (td->slat_log && unit_logs)
3025                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
3026         if (td->lat_log && unit_logs)
3027                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
3028         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
3029                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
3030         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
3031                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
3032 }
3033
3034 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
3035                     unsigned long long bs)
3036 {
3037         struct io_log *iolog;
3038
3039         if (!ddir_rw(ddir))
3040                 return;
3041
3042         iolog = agg_io_log[ddir];
3043         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
3044 }
3045
3046 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
3047 {
3048         unsigned int idx = plat_val_to_idx(nsec);
3049         assert(idx < FIO_IO_U_PLAT_NR);
3050
3051         ts->io_u_sync_plat[idx]++;
3052         add_stat_sample(&ts->sync_stat, nsec);
3053 }
3054
3055 static inline void add_lat_percentile_sample(struct thread_stat *ts,
3056                                              unsigned long long nsec,
3057                                              enum fio_ddir ddir,
3058                                              enum fio_lat lat)
3059 {
3060         unsigned int idx = plat_val_to_idx(nsec);
3061         assert(idx < FIO_IO_U_PLAT_NR);
3062
3063         ts->io_u_plat[lat][ddir][idx]++;
3064 }
3065
3066 static inline void add_lat_percentile_prio_sample(struct thread_stat *ts,
3067                                                   unsigned long long nsec,
3068                                                   enum fio_ddir ddir,
3069                                                   bool high_prio)
3070 {
3071         unsigned int idx = plat_val_to_idx(nsec);
3072
3073         if (!high_prio)
3074                 ts->io_u_plat_low_prio[ddir][idx]++;
3075         else
3076                 ts->io_u_plat_high_prio[ddir][idx]++;
3077 }
3078
3079 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
3080                      unsigned long long nsec, unsigned long long bs,
3081                      uint64_t offset, unsigned int ioprio, bool high_prio)
3082 {
3083         const bool needs_lock = td_async_processing(td);
3084         unsigned long elapsed, this_window;
3085         struct thread_stat *ts = &td->ts;
3086         struct io_log *iolog = td->clat_hist_log;
3087
3088         if (needs_lock)
3089                 __td_io_u_lock(td);
3090
3091         add_stat_sample(&ts->clat_stat[ddir], nsec);
3092
3093         /*
3094          * When lat_percentiles=1 (default 0), the reported high/low priority
3095          * percentiles and stats are used for describing total latency values,
3096          * even though the variable names themselves start with clat_.
3097          *
3098          * Because of the above definition, add a prio stat sample only when
3099          * lat_percentiles=0. add_lat_sample() will add the prio stat sample
3100          * when lat_percentiles=1.
3101          */
3102         if (!ts->lat_percentiles) {
3103                 if (high_prio)
3104                         add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
3105                 else
3106                         add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
3107         }
3108
3109         if (td->clat_log)
3110                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
3111                                offset, ioprio);
3112
3113         if (ts->clat_percentiles) {
3114                 /*
3115                  * Because of the above definition, add a prio lat percentile
3116                  * sample only when lat_percentiles=0. add_lat_sample() will add
3117                  * the prio lat percentile sample when lat_percentiles=1.
3118                  */
3119                 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3120                 if (!ts->lat_percentiles)
3121                         add_lat_percentile_prio_sample(ts, nsec, ddir,
3122                                                        high_prio);
3123         }
3124
3125         if (iolog && iolog->hist_msec) {
3126                 struct io_hist *hw = &iolog->hist_window[ddir];
3127
3128                 hw->samples++;
3129                 elapsed = mtime_since_now(&td->epoch);
3130                 if (!hw->hist_last)
3131                         hw->hist_last = elapsed;
3132                 this_window = elapsed - hw->hist_last;
3133
3134                 if (this_window >= iolog->hist_msec) {
3135                         uint64_t *io_u_plat;
3136                         struct io_u_plat_entry *dst;
3137
3138                         /*
3139                          * Make a byte-for-byte copy of the latency histogram
3140                          * stored in td->ts.io_u_plat[ddir], recording it in a
3141                          * log sample. Note that the matching call to free() is
3142                          * located in iolog.c after printing this sample to the
3143                          * log file.
3144                          */
3145                         io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3146                         dst = malloc(sizeof(struct io_u_plat_entry));
3147                         memcpy(&(dst->io_u_plat), io_u_plat,
3148                                 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3149                         flist_add(&dst->list, &hw->list);
3150                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3151                                          elapsed, offset, ioprio);
3152
3153                         /*
3154                          * Update the last time we recorded as being now, minus
3155                          * any drift in time we encountered before actually
3156                          * making the record.
3157                          */
3158                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3159                         hw->samples = 0;
3160                 }
3161         }
3162
3163         if (needs_lock)
3164                 __td_io_u_unlock(td);
3165 }
3166
3167 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3168                      unsigned long long nsec, unsigned long long bs,
3169                      uint64_t offset, unsigned int ioprio)
3170 {
3171         const bool needs_lock = td_async_processing(td);
3172         struct thread_stat *ts = &td->ts;
3173
3174         if (!ddir_rw(ddir))
3175                 return;
3176
3177         if (needs_lock)
3178                 __td_io_u_lock(td);
3179
3180         add_stat_sample(&ts->slat_stat[ddir], nsec);
3181
3182         if (td->slat_log)
3183                 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3184                                offset, ioprio);
3185
3186         if (ts->slat_percentiles)
3187                 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3188
3189         if (needs_lock)
3190                 __td_io_u_unlock(td);
3191 }
3192
3193 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3194                     unsigned long long nsec, unsigned long long bs,
3195                     uint64_t offset, unsigned int ioprio, bool high_prio)
3196 {
3197         const bool needs_lock = td_async_processing(td);
3198         struct thread_stat *ts = &td->ts;
3199
3200         if (!ddir_rw(ddir))
3201                 return;
3202
3203         if (needs_lock)
3204                 __td_io_u_lock(td);
3205
3206         add_stat_sample(&ts->lat_stat[ddir], nsec);
3207
3208         if (td->lat_log)
3209                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3210                                offset, ioprio);
3211
3212         /*
3213          * When lat_percentiles=1 (default 0), the reported high/low priority
3214          * percentiles and stats are used for describing total latency values,
3215          * even though the variable names themselves start with clat_.
3216          *
3217          * Because of the above definition, add a prio stat and prio lat
3218          * percentile sample only when lat_percentiles=1. add_clat_sample() will
3219          * add the prio stat and prio lat percentile sample when
3220          * lat_percentiles=0.
3221          */
3222         if (ts->lat_percentiles) {
3223                 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3224                 add_lat_percentile_prio_sample(ts, nsec, ddir, high_prio);
3225                 if (high_prio)
3226                         add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
3227                 else
3228                         add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
3229
3230         }
3231         if (needs_lock)
3232                 __td_io_u_unlock(td);
3233 }
3234
3235 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3236                    unsigned int bytes, unsigned long long spent)
3237 {
3238         const bool needs_lock = td_async_processing(td);
3239         struct thread_stat *ts = &td->ts;
3240         unsigned long rate;
3241
3242         if (spent)
3243                 rate = (unsigned long) (bytes * 1000000ULL / spent);
3244         else
3245                 rate = 0;
3246
3247         if (needs_lock)
3248                 __td_io_u_lock(td);
3249
3250         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3251
3252         if (td->bw_log)
3253                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3254                                bytes, io_u->offset, io_u->ioprio);
3255
3256         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3257
3258         if (needs_lock)
3259                 __td_io_u_unlock(td);
3260 }
3261
3262 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3263                          struct timespec *t, unsigned int avg_time,
3264                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3265                          struct io_stat *stat, struct io_log *log,
3266                          bool is_kb)
3267 {
3268         const bool needs_lock = td_async_processing(td);
3269         unsigned long spent, rate;
3270         enum fio_ddir ddir;
3271         unsigned long next, next_log;
3272
3273         next_log = avg_time;
3274
3275         spent = mtime_since(parent_tv, t);
3276         if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3277                 return avg_time - spent;
3278
3279         if (needs_lock)
3280                 __td_io_u_lock(td);
3281
3282         /*
3283          * Compute both read and write rates for the interval.
3284          */
3285         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3286                 uint64_t delta;
3287
3288                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3289                 if (!delta)
3290                         continue; /* No entries for interval */
3291
3292                 if (spent) {
3293                         if (is_kb)
3294                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
3295                         else
3296                                 rate = (delta * 1000) / spent;
3297                 } else
3298                         rate = 0;
3299
3300                 add_stat_sample(&stat[ddir], rate);
3301
3302                 if (log) {
3303                         unsigned long long bs = 0;
3304
3305                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3306                                 bs = td->o.min_bs[ddir];
3307
3308                         next = add_log_sample(td, log, sample_val(rate), ddir,
3309                                               bs, 0, 0);
3310                         next_log = min(next_log, next);
3311                 }
3312
3313                 stat_io_bytes[ddir] = this_io_bytes[ddir];
3314         }
3315
3316         *parent_tv = *t;
3317
3318         if (needs_lock)
3319                 __td_io_u_unlock(td);
3320
3321         if (spent <= avg_time)
3322                 next = avg_time;
3323         else
3324                 next = avg_time - (1 + spent - avg_time);
3325
3326         return min(next, next_log);
3327 }
3328
3329 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3330 {
3331         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3332                                 td->this_io_bytes, td->stat_io_bytes,
3333                                 td->ts.bw_stat, td->bw_log, true);
3334 }
3335
3336 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3337                      unsigned int bytes)
3338 {
3339         const bool needs_lock = td_async_processing(td);
3340         struct thread_stat *ts = &td->ts;
3341
3342         if (needs_lock)
3343                 __td_io_u_lock(td);
3344
3345         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3346
3347         if (td->iops_log)
3348                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3349                                bytes, io_u->offset, io_u->ioprio);
3350
3351         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3352
3353         if (needs_lock)
3354                 __td_io_u_unlock(td);
3355 }
3356
3357 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3358 {
3359         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3360                                 td->this_io_blocks, td->stat_io_blocks,
3361                                 td->ts.iops_stat, td->iops_log, false);
3362 }
3363
3364 /*
3365  * Returns msecs to next event
3366  */
3367 int calc_log_samples(void)
3368 {
3369         struct thread_data *td;
3370         unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3371         struct timespec now;
3372         int i;
3373         long elapsed_time = 0;
3374
3375         fio_gettime(&now, NULL);
3376
3377         for_each_td(td, i) {
3378                 elapsed_time = mtime_since_now(&td->epoch);
3379
3380                 if (!td->o.stats)
3381                         continue;
3382                 if (in_ramp_time(td) ||
3383                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3384                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3385                         continue;
3386                 }
3387                 if (!td->bw_log ||
3388                         (td->bw_log && !per_unit_log(td->bw_log))) {
3389                         tmp = add_bw_samples(td, &now);
3390
3391                         if (td->bw_log)
3392                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3393                 }
3394                 if (!td->iops_log ||
3395                         (td->iops_log && !per_unit_log(td->iops_log))) {
3396                         tmp = add_iops_samples(td, &now);
3397
3398                         if (td->iops_log)
3399                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3400                 }
3401
3402                 if (tmp < next)
3403                         next = tmp;
3404         }
3405
3406         /* if log_avg_msec_min has not been changed, set it to 0 */
3407         if (log_avg_msec_min == -1U)
3408                 log_avg_msec_min = 0;
3409
3410         if (log_avg_msec_min == 0)
3411                 next_mod = elapsed_time;
3412         else
3413                 next_mod = elapsed_time % log_avg_msec_min;
3414
3415         /* correction to keep the time on the log avg msec boundary */
3416         next = min(next, (log_avg_msec_min - next_mod));
3417
3418         return next == ~0U ? 0 : next;
3419 }
3420
3421 void stat_init(void)
3422 {
3423         stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3424 }
3425
3426 void stat_exit(void)
3427 {
3428         /*
3429          * When we have the mutex, we know out-of-band access to it
3430          * have ended.
3431          */
3432         fio_sem_down(stat_sem);
3433         fio_sem_remove(stat_sem);
3434 }
3435
3436 /*
3437  * Called from signal handler. Wake up status thread.
3438  */
3439 void show_running_run_stats(void)
3440 {
3441         helper_do_stat();
3442 }
3443
3444 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3445 {
3446         /* Ignore io_u's which span multiple blocks--they will just get
3447          * inaccurate counts. */
3448         int idx = (io_u->offset - io_u->file->file_offset)
3449                         / td->o.bs[DDIR_TRIM];
3450         uint32_t *info = &td->ts.block_infos[idx];
3451         assert(idx < td->ts.nr_block_infos);
3452         return info;
3453 }
3454