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