Makefile: include linux-blkzoned.c for Android, if set
[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                 goto out;
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 out:
262         free(ovals);
263 }
264
265 bool calc_lat(struct io_stat *is, unsigned long long *min,
266               unsigned long long *max, double *mean, double *dev)
267 {
268         double n = (double) is->samples;
269
270         if (n == 0)
271                 return false;
272
273         *min = is->min_val;
274         *max = is->max_val;
275         *mean = is->mean.u.f;
276
277         if (n > 1.0)
278                 *dev = sqrt(is->S.u.f / (n - 1.0));
279         else
280                 *dev = 0;
281
282         return true;
283 }
284
285 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
286 {
287         char *io, *agg, *min, *max;
288         char *ioalt, *aggalt, *minalt, *maxalt;
289         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
290         int i;
291
292         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
293
294         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
295                 const int i2p = is_power_of_2(rs->kb_base);
296
297                 if (!rs->max_run[i])
298                         continue;
299
300                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
301                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
302                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
303                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
304                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
305                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
306                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
307                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
308                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
309                                 rs->unified_rw_rep ? "  MIXED" : str[i],
310                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
311                                 (unsigned long long) rs->min_run[i],
312                                 (unsigned long long) rs->max_run[i]);
313
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
325 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
326 {
327         int i;
328
329         /*
330          * Do depth distribution calculations
331          */
332         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
333                 if (total) {
334                         io_u_dist[i] = (double) map[i] / (double) total;
335                         io_u_dist[i] *= 100.0;
336                         if (io_u_dist[i] < 0.1 && map[i])
337                                 io_u_dist[i] = 0.1;
338                 } else
339                         io_u_dist[i] = 0.0;
340         }
341 }
342
343 static void stat_calc_lat(struct thread_stat *ts, double *dst,
344                           uint64_t *src, int nr)
345 {
346         unsigned long total = ddir_rw_sum(ts->total_io_u);
347         int i;
348
349         /*
350          * Do latency distribution calculations
351          */
352         for (i = 0; i < nr; i++) {
353                 if (total) {
354                         dst[i] = (double) src[i] / (double) total;
355                         dst[i] *= 100.0;
356                         if (dst[i] < 0.01 && src[i])
357                                 dst[i] = 0.01;
358                 } else
359                         dst[i] = 0.0;
360         }
361 }
362
363 /*
364  * To keep the terse format unaltered, add all of the ns latency
365  * buckets to the first us latency bucket
366  */
367 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
368 {
369         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
370         int i;
371
372         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
373
374         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
375                 ntotal += ts->io_u_lat_n[i];
376
377         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
378 }
379
380 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
381 {
382         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
383 }
384
385 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
386 {
387         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
388 }
389
390 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
391 {
392         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
393 }
394
395 static void display_lat(const char *name, unsigned long long min,
396                         unsigned long long max, double mean, double dev,
397                         struct buf_output *out)
398 {
399         const char *base = "(nsec)";
400         char *minp, *maxp;
401
402         if (nsec_to_msec(&min, &max, &mean, &dev))
403                 base = "(msec)";
404         else if (nsec_to_usec(&min, &max, &mean, &dev))
405                 base = "(usec)";
406
407         minp = num2str(min, 6, 1, 0, N2S_NONE);
408         maxp = num2str(max, 6, 1, 0, N2S_NONE);
409
410         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
411                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
412
413         free(minp);
414         free(maxp);
415 }
416
417 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
418                              int ddir, struct buf_output *out)
419 {
420         unsigned long runt;
421         unsigned long long min, max, bw, iops;
422         double mean, dev;
423         char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
424         int i2p;
425
426         if (ddir_sync(ddir)) {
427                 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
428                         log_buf(out, "  %s:\n", "fsync/fdatasync/sync_file_range");
429                         display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
430                         show_clat_percentiles(ts->io_u_sync_plat,
431                                                 ts->sync_stat.samples,
432                                                 ts->percentile_list,
433                                                 ts->percentile_precision,
434                                                 io_ddir_name(ddir), out);
435                 }
436                 return;
437         }
438
439         assert(ddir_rw(ddir));
440
441         if (!ts->runtime[ddir])
442                 return;
443
444         i2p = is_power_of_2(rs->kb_base);
445         runt = ts->runtime[ddir];
446
447         bw = (1000 * ts->io_bytes[ddir]) / runt;
448         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
449         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
450         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
451
452         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
453         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
454         if (ddir == DDIR_WRITE)
455                 post_st = zbd_write_status(ts);
456         else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
457                 uint64_t total;
458                 double hit;
459
460                 total = ts->cachehit + ts->cachemiss;
461                 hit = (double) ts->cachehit / (double) total;
462                 hit *= 100.0;
463                 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
464                         post_st = NULL;
465         }
466
467         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
468                         rs->unified_rw_rep ? "mixed" : io_ddir_name(ddir),
469                         iops_p, bw_p, bw_p_alt, io_p,
470                         (unsigned long long) ts->runtime[ddir],
471                         post_st ? : "");
472
473         free(post_st);
474         free(io_p);
475         free(bw_p);
476         free(bw_p_alt);
477         free(iops_p);
478
479         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
480                 display_lat("slat", min, max, mean, dev, out);
481         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
482                 display_lat("clat", min, max, mean, dev, out);
483         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
484                 display_lat(" lat", min, max, mean, dev, out);
485         if (calc_lat(&ts->clat_high_prio_stat[ddir], &min, &max, &mean, &dev)) {
486                 display_lat(ts->lat_percentiles ? "high prio_lat" : "high prio_clat",
487                                 min, max, mean, dev, out);
488                 if (calc_lat(&ts->clat_low_prio_stat[ddir], &min, &max, &mean, &dev))
489                         display_lat(ts->lat_percentiles ? "low prio_lat" : "low prio_clat",
490                                         min, max, mean, dev, out);
491         }
492
493         if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
494                 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
495                                         ts->slat_stat[ddir].samples,
496                                         ts->percentile_list,
497                                         ts->percentile_precision, "slat", out);
498         if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
499                 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
500                                         ts->clat_stat[ddir].samples,
501                                         ts->percentile_list,
502                                         ts->percentile_precision, "clat", out);
503         if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
504                 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
505                                         ts->lat_stat[ddir].samples,
506                                         ts->percentile_list,
507                                         ts->percentile_precision, "lat", out);
508
509         if (ts->clat_percentiles || ts->lat_percentiles) {
510                 const char *name = ts->lat_percentiles ? "lat" : "clat";
511                 char prio_name[32];
512                 uint64_t samples;
513
514                 if (ts->lat_percentiles)
515                         samples = ts->lat_stat[ddir].samples;
516                 else
517                         samples = ts->clat_stat[ddir].samples;
518
519                 /* Only print this if some high and low priority stats were collected */
520                 if (ts->clat_high_prio_stat[ddir].samples > 0 &&
521                         ts->clat_low_prio_stat[ddir].samples > 0)
522                 {
523                         sprintf(prio_name, "high prio (%.2f%%) %s",
524                                         100. * (double) ts->clat_high_prio_stat[ddir].samples / (double) samples,
525                                         name);
526                         show_clat_percentiles(ts->io_u_plat_high_prio[ddir],
527                                                 ts->clat_high_prio_stat[ddir].samples,
528                                                 ts->percentile_list,
529                                                 ts->percentile_precision, prio_name, out);
530
531                         sprintf(prio_name, "low prio (%.2f%%) %s",
532                                         100. * (double) ts->clat_low_prio_stat[ddir].samples / (double) samples,
533                                         name);
534                         show_clat_percentiles(ts->io_u_plat_low_prio[ddir],
535                                                 ts->clat_low_prio_stat[ddir].samples,
536                                                 ts->percentile_list,
537                                                 ts->percentile_precision, prio_name, out);
538                 }
539         }
540
541         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
542                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
543                 const char *bw_str;
544
545                 if ((rs->unit_base == 1) && i2p)
546                         bw_str = "Kibit";
547                 else if (rs->unit_base == 1)
548                         bw_str = "kbit";
549                 else if (i2p)
550                         bw_str = "KiB";
551                 else
552                         bw_str = "kB";
553
554                 if (rs->agg[ddir]) {
555                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
556                         if (p_of_agg > 100.0)
557                                 p_of_agg = 100.0;
558                 }
559
560                 if (rs->unit_base == 1) {
561                         min *= 8.0;
562                         max *= 8.0;
563                         mean *= 8.0;
564                         dev *= 8.0;
565                 }
566
567                 if (mean > fkb_base * fkb_base) {
568                         min /= fkb_base;
569                         max /= fkb_base;
570                         mean /= fkb_base;
571                         dev /= fkb_base;
572                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
573                 }
574
575                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
576                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
577                         bw_str, min, max, p_of_agg, mean, dev,
578                         (&ts->bw_stat[ddir])->samples);
579         }
580         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
581                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
582                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
583                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
584         }
585 }
586
587 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
588                      const char *msg, struct buf_output *out)
589 {
590         bool new_line = true, shown = false;
591         int i, line = 0;
592
593         for (i = 0; i < nr; i++) {
594                 if (io_u_lat[i] <= 0.0)
595                         continue;
596                 shown = true;
597                 if (new_line) {
598                         if (line)
599                                 log_buf(out, "\n");
600                         log_buf(out, "  lat (%s)   : ", msg);
601                         new_line = false;
602                         line = 0;
603                 }
604                 if (line)
605                         log_buf(out, ", ");
606                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
607                 line++;
608                 if (line == 5)
609                         new_line = true;
610         }
611
612         if (shown)
613                 log_buf(out, "\n");
614
615         return true;
616 }
617
618 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
619 {
620         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
621                                  "250=", "500=", "750=", "1000=", };
622
623         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
624 }
625
626 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
627 {
628         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
629                                  "250=", "500=", "750=", "1000=", };
630
631         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
632 }
633
634 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
635 {
636         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
637                                  "250=", "500=", "750=", "1000=", "2000=",
638                                  ">=2000=", };
639
640         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
641 }
642
643 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
644 {
645         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
646         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
647         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
648
649         stat_calc_lat_n(ts, io_u_lat_n);
650         stat_calc_lat_u(ts, io_u_lat_u);
651         stat_calc_lat_m(ts, io_u_lat_m);
652
653         show_lat_n(io_u_lat_n, out);
654         show_lat_u(io_u_lat_u, out);
655         show_lat_m(io_u_lat_m, out);
656 }
657
658 static int block_state_category(int block_state)
659 {
660         switch (block_state) {
661         case BLOCK_STATE_UNINIT:
662                 return 0;
663         case BLOCK_STATE_TRIMMED:
664         case BLOCK_STATE_WRITTEN:
665                 return 1;
666         case BLOCK_STATE_WRITE_FAILURE:
667         case BLOCK_STATE_TRIM_FAILURE:
668                 return 2;
669         default:
670                 /* Silence compile warning on some BSDs and have a return */
671                 assert(0);
672                 return -1;
673         }
674 }
675
676 static int compare_block_infos(const void *bs1, const void *bs2)
677 {
678         uint64_t block1 = *(uint64_t *)bs1;
679         uint64_t block2 = *(uint64_t *)bs2;
680         int state1 = BLOCK_INFO_STATE(block1);
681         int state2 = BLOCK_INFO_STATE(block2);
682         int bscat1 = block_state_category(state1);
683         int bscat2 = block_state_category(state2);
684         int cycles1 = BLOCK_INFO_TRIMS(block1);
685         int cycles2 = BLOCK_INFO_TRIMS(block2);
686
687         if (bscat1 < bscat2)
688                 return -1;
689         if (bscat1 > bscat2)
690                 return 1;
691
692         if (cycles1 < cycles2)
693                 return -1;
694         if (cycles1 > cycles2)
695                 return 1;
696
697         if (state1 < state2)
698                 return -1;
699         if (state1 > state2)
700                 return 1;
701
702         assert(block1 == block2);
703         return 0;
704 }
705
706 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
707                                   fio_fp64_t *plist, unsigned int **percentiles,
708                                   unsigned int *types)
709 {
710         int len = 0;
711         int i, nr_uninit;
712
713         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
714
715         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
716                 len++;
717
718         if (!len)
719                 return 0;
720
721         /*
722          * Sort the percentile list. Note that it may already be sorted if
723          * we are using the default values, but since it's a short list this
724          * isn't a worry. Also note that this does not work for NaN values.
725          */
726         if (len > 1)
727                 qsort(plist, len, sizeof(plist[0]), double_cmp);
728
729         /* Start only after the uninit entries end */
730         for (nr_uninit = 0;
731              nr_uninit < nr_block_infos
732                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
733              nr_uninit ++)
734                 ;
735
736         if (nr_uninit == nr_block_infos)
737                 return 0;
738
739         *percentiles = calloc(len, sizeof(**percentiles));
740
741         for (i = 0; i < len; i++) {
742                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
743                                 + nr_uninit;
744                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
745         }
746
747         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
748         for (i = 0; i < nr_block_infos; i++)
749                 types[BLOCK_INFO_STATE(block_infos[i])]++;
750
751         return len;
752 }
753
754 static const char *block_state_names[] = {
755         [BLOCK_STATE_UNINIT] = "unwritten",
756         [BLOCK_STATE_TRIMMED] = "trimmed",
757         [BLOCK_STATE_WRITTEN] = "written",
758         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
759         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
760 };
761
762 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
763                              fio_fp64_t *plist, struct buf_output *out)
764 {
765         int len, pos, i;
766         unsigned int *percentiles = NULL;
767         unsigned int block_state_counts[BLOCK_STATE_COUNT];
768
769         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
770                                      &percentiles, block_state_counts);
771
772         log_buf(out, "  block lifetime percentiles :\n   |");
773         pos = 0;
774         for (i = 0; i < len; i++) {
775                 uint32_t block_info = percentiles[i];
776 #define LINE_LENGTH     75
777                 char str[LINE_LENGTH];
778                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
779                                      plist[i].u.f, block_info,
780                                      i == len - 1 ? '\n' : ',');
781                 assert(strln < LINE_LENGTH);
782                 if (pos + strln > LINE_LENGTH) {
783                         pos = 0;
784                         log_buf(out, "\n   |");
785                 }
786                 log_buf(out, "%s", str);
787                 pos += strln;
788 #undef LINE_LENGTH
789         }
790         if (percentiles)
791                 free(percentiles);
792
793         log_buf(out, "        states               :");
794         for (i = 0; i < BLOCK_STATE_COUNT; i++)
795                 log_buf(out, " %s=%u%c",
796                          block_state_names[i], block_state_counts[i],
797                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
798 }
799
800 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
801 {
802         char *p1, *p1alt, *p2;
803         unsigned long long bw_mean, iops_mean;
804         const int i2p = is_power_of_2(ts->kb_base);
805
806         if (!ts->ss_dur)
807                 return;
808
809         bw_mean = steadystate_bw_mean(ts);
810         iops_mean = steadystate_iops_mean(ts);
811
812         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
813         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
814         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
815
816         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
817                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
818                 p1, p1alt, p2,
819                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
820                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
821                 ts->ss_criterion.u.f,
822                 ts->ss_state & FIO_SS_PCT ? "%" : "");
823
824         free(p1);
825         free(p1alt);
826         free(p2);
827 }
828
829 static void show_agg_stats(struct disk_util_agg *agg, int terse,
830                            struct buf_output *out)
831 {
832         if (!agg->slavecount)
833                 return;
834
835         if (!terse) {
836                 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
837                          "aggrticks=%llu/%llu, aggrin_queue=%llu, "
838                          "aggrutil=%3.2f%%",
839                         (unsigned long long) agg->ios[0] / agg->slavecount,
840                         (unsigned long long) agg->ios[1] / agg->slavecount,
841                         (unsigned long long) agg->merges[0] / agg->slavecount,
842                         (unsigned long long) agg->merges[1] / agg->slavecount,
843                         (unsigned long long) agg->ticks[0] / agg->slavecount,
844                         (unsigned long long) agg->ticks[1] / agg->slavecount,
845                         (unsigned long long) agg->time_in_queue / agg->slavecount,
846                         agg->max_util.u.f);
847         } else {
848                 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
849                         (unsigned long long) agg->ios[0] / agg->slavecount,
850                         (unsigned long long) agg->ios[1] / agg->slavecount,
851                         (unsigned long long) agg->merges[0] / agg->slavecount,
852                         (unsigned long long) agg->merges[1] / agg->slavecount,
853                         (unsigned long long) agg->ticks[0] / agg->slavecount,
854                         (unsigned long long) agg->ticks[1] / agg->slavecount,
855                         (unsigned long long) agg->time_in_queue / agg->slavecount,
856                         agg->max_util.u.f);
857         }
858 }
859
860 static void aggregate_slaves_stats(struct disk_util *masterdu)
861 {
862         struct disk_util_agg *agg = &masterdu->agg;
863         struct disk_util_stat *dus;
864         struct flist_head *entry;
865         struct disk_util *slavedu;
866         double util;
867
868         flist_for_each(entry, &masterdu->slaves) {
869                 slavedu = flist_entry(entry, struct disk_util, slavelist);
870                 dus = &slavedu->dus;
871                 agg->ios[0] += dus->s.ios[0];
872                 agg->ios[1] += dus->s.ios[1];
873                 agg->merges[0] += dus->s.merges[0];
874                 agg->merges[1] += dus->s.merges[1];
875                 agg->sectors[0] += dus->s.sectors[0];
876                 agg->sectors[1] += dus->s.sectors[1];
877                 agg->ticks[0] += dus->s.ticks[0];
878                 agg->ticks[1] += dus->s.ticks[1];
879                 agg->time_in_queue += dus->s.time_in_queue;
880                 agg->slavecount++;
881
882                 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
883                 /* System utilization is the utilization of the
884                  * component with the highest utilization.
885                  */
886                 if (util > agg->max_util.u.f)
887                         agg->max_util.u.f = util;
888
889         }
890
891         if (agg->max_util.u.f > 100.0)
892                 agg->max_util.u.f = 100.0;
893 }
894
895 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
896                      int terse, struct buf_output *out)
897 {
898         double util = 0;
899
900         if (dus->s.msec)
901                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
902         if (util > 100.0)
903                 util = 100.0;
904
905         if (!terse) {
906                 if (agg->slavecount)
907                         log_buf(out, "  ");
908
909                 log_buf(out, "  %s: ios=%llu/%llu, merge=%llu/%llu, "
910                          "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
911                                 dus->name,
912                                 (unsigned long long) dus->s.ios[0],
913                                 (unsigned long long) dus->s.ios[1],
914                                 (unsigned long long) dus->s.merges[0],
915                                 (unsigned long long) dus->s.merges[1],
916                                 (unsigned long long) dus->s.ticks[0],
917                                 (unsigned long long) dus->s.ticks[1],
918                                 (unsigned long long) dus->s.time_in_queue,
919                                 util);
920         } else {
921                 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
922                                 dus->name,
923                                 (unsigned long long) dus->s.ios[0],
924                                 (unsigned long long) dus->s.ios[1],
925                                 (unsigned long long) dus->s.merges[0],
926                                 (unsigned long long) dus->s.merges[1],
927                                 (unsigned long long) dus->s.ticks[0],
928                                 (unsigned long long) dus->s.ticks[1],
929                                 (unsigned long long) dus->s.time_in_queue,
930                                 util);
931         }
932
933         /*
934          * If the device has slaves, aggregate the stats for
935          * those slave devices also.
936          */
937         show_agg_stats(agg, terse, out);
938
939         if (!terse)
940                 log_buf(out, "\n");
941 }
942
943 void json_array_add_disk_util(struct disk_util_stat *dus,
944                 struct disk_util_agg *agg, struct json_array *array)
945 {
946         struct json_object *obj;
947         double util = 0;
948
949         if (dus->s.msec)
950                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
951         if (util > 100.0)
952                 util = 100.0;
953
954         obj = json_create_object();
955         json_array_add_value_object(array, obj);
956
957         json_object_add_value_string(obj, "name", (const char *)dus->name);
958         json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
959         json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
960         json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
961         json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
962         json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
963         json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
964         json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
965         json_object_add_value_float(obj, "util", util);
966
967         /*
968          * If the device has slaves, aggregate the stats for
969          * those slave devices also.
970          */
971         if (!agg->slavecount)
972                 return;
973         json_object_add_value_int(obj, "aggr_read_ios",
974                                 agg->ios[0] / agg->slavecount);
975         json_object_add_value_int(obj, "aggr_write_ios",
976                                 agg->ios[1] / agg->slavecount);
977         json_object_add_value_int(obj, "aggr_read_merges",
978                                 agg->merges[0] / agg->slavecount);
979         json_object_add_value_int(obj, "aggr_write_merge",
980                                 agg->merges[1] / agg->slavecount);
981         json_object_add_value_int(obj, "aggr_read_ticks",
982                                 agg->ticks[0] / agg->slavecount);
983         json_object_add_value_int(obj, "aggr_write_ticks",
984                                 agg->ticks[1] / agg->slavecount);
985         json_object_add_value_int(obj, "aggr_in_queue",
986                                 agg->time_in_queue / agg->slavecount);
987         json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
988 }
989
990 static void json_object_add_disk_utils(struct json_object *obj,
991                                        struct flist_head *head)
992 {
993         struct json_array *array = json_create_array();
994         struct flist_head *entry;
995         struct disk_util *du;
996
997         json_object_add_value_array(obj, "disk_util", array);
998
999         flist_for_each(entry, head) {
1000                 du = flist_entry(entry, struct disk_util, list);
1001
1002                 aggregate_slaves_stats(du);
1003                 json_array_add_disk_util(&du->dus, &du->agg, array);
1004         }
1005 }
1006
1007 void show_disk_util(int terse, struct json_object *parent,
1008                     struct buf_output *out)
1009 {
1010         struct flist_head *entry;
1011         struct disk_util *du;
1012         bool do_json;
1013
1014         if (!is_running_backend())
1015                 return;
1016
1017         if (flist_empty(&disk_list)) {
1018                 return;
1019         }
1020
1021         if ((output_format & FIO_OUTPUT_JSON) && parent)
1022                 do_json = true;
1023         else
1024                 do_json = false;
1025
1026         if (!terse && !do_json)
1027                 log_buf(out, "\nDisk stats (read/write):\n");
1028
1029         if (do_json)
1030                 json_object_add_disk_utils(parent, &disk_list);
1031         else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1032                 flist_for_each(entry, &disk_list) {
1033                         du = flist_entry(entry, struct disk_util, list);
1034
1035                         aggregate_slaves_stats(du);
1036                         print_disk_util(&du->dus, &du->agg, terse, out);
1037                 }
1038         }
1039 }
1040
1041 static void show_thread_status_normal(struct thread_stat *ts,
1042                                       struct group_run_stats *rs,
1043                                       struct buf_output *out)
1044 {
1045         double usr_cpu, sys_cpu;
1046         unsigned long runtime;
1047         double io_u_dist[FIO_IO_U_MAP_NR];
1048         time_t time_p;
1049         char time_buf[32];
1050
1051         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1052                 return;
1053
1054         memset(time_buf, 0, sizeof(time_buf));
1055
1056         time(&time_p);
1057         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1058
1059         if (!ts->error) {
1060                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1061                                         ts->name, ts->groupid, ts->members,
1062                                         ts->error, (int) ts->pid, time_buf);
1063         } else {
1064                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1065                                         ts->name, ts->groupid, ts->members,
1066                                         ts->error, ts->verror, (int) ts->pid,
1067                                         time_buf);
1068         }
1069
1070         if (strlen(ts->description))
1071                 log_buf(out, "  Description  : [%s]\n", ts->description);
1072
1073         if (ts->io_bytes[DDIR_READ])
1074                 show_ddir_status(rs, ts, DDIR_READ, out);
1075         if (ts->io_bytes[DDIR_WRITE])
1076                 show_ddir_status(rs, ts, DDIR_WRITE, out);
1077         if (ts->io_bytes[DDIR_TRIM])
1078                 show_ddir_status(rs, ts, DDIR_TRIM, out);
1079
1080         show_latencies(ts, out);
1081
1082         if (ts->sync_stat.samples)
1083                 show_ddir_status(rs, ts, DDIR_SYNC, out);
1084
1085         runtime = ts->total_run_time;
1086         if (runtime) {
1087                 double runt = (double) runtime;
1088
1089                 usr_cpu = (double) ts->usr_time * 100 / runt;
1090                 sys_cpu = (double) ts->sys_time * 100 / runt;
1091         } else {
1092                 usr_cpu = 0;
1093                 sys_cpu = 0;
1094         }
1095
1096         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1097                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1098                         (unsigned long long) ts->ctx,
1099                         (unsigned long long) ts->majf,
1100                         (unsigned long long) ts->minf);
1101
1102         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1103         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1104                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1105                                         io_u_dist[1], io_u_dist[2],
1106                                         io_u_dist[3], io_u_dist[4],
1107                                         io_u_dist[5], io_u_dist[6]);
1108
1109         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1110         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1111                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1112                                         io_u_dist[1], io_u_dist[2],
1113                                         io_u_dist[3], io_u_dist[4],
1114                                         io_u_dist[5], io_u_dist[6]);
1115         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1116         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1117                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1118                                         io_u_dist[1], io_u_dist[2],
1119                                         io_u_dist[3], io_u_dist[4],
1120                                         io_u_dist[5], io_u_dist[6]);
1121         log_buf(out, "     issued rwts: total=%llu,%llu,%llu,%llu"
1122                                  " short=%llu,%llu,%llu,0"
1123                                  " dropped=%llu,%llu,%llu,0\n",
1124                                         (unsigned long long) ts->total_io_u[0],
1125                                         (unsigned long long) ts->total_io_u[1],
1126                                         (unsigned long long) ts->total_io_u[2],
1127                                         (unsigned long long) ts->total_io_u[3],
1128                                         (unsigned long long) ts->short_io_u[0],
1129                                         (unsigned long long) ts->short_io_u[1],
1130                                         (unsigned long long) ts->short_io_u[2],
1131                                         (unsigned long long) ts->drop_io_u[0],
1132                                         (unsigned long long) ts->drop_io_u[1],
1133                                         (unsigned long long) ts->drop_io_u[2]);
1134         if (ts->continue_on_error) {
1135                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
1136                                         (unsigned long long)ts->total_err_count,
1137                                         ts->first_error,
1138                                         strerror(ts->first_error));
1139         }
1140         if (ts->latency_depth) {
1141                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1142                                         (unsigned long long)ts->latency_target,
1143                                         (unsigned long long)ts->latency_window,
1144                                         ts->latency_percentile.u.f,
1145                                         ts->latency_depth);
1146         }
1147
1148         if (ts->nr_block_infos)
1149                 show_block_infos(ts->nr_block_infos, ts->block_infos,
1150                                   ts->percentile_list, out);
1151
1152         if (ts->ss_dur)
1153                 show_ss_normal(ts, out);
1154 }
1155
1156 static void show_ddir_status_terse(struct thread_stat *ts,
1157                                    struct group_run_stats *rs, int ddir,
1158                                    int ver, struct buf_output *out)
1159 {
1160         unsigned long long min, max, minv, maxv, bw, iops;
1161         unsigned long long *ovals = NULL;
1162         double mean, dev;
1163         unsigned int len;
1164         int i, bw_stat;
1165
1166         assert(ddir_rw(ddir));
1167
1168         iops = bw = 0;
1169         if (ts->runtime[ddir]) {
1170                 uint64_t runt = ts->runtime[ddir];
1171
1172                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1173                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1174         }
1175
1176         log_buf(out, ";%llu;%llu;%llu;%llu",
1177                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1178                                         (unsigned long long) ts->runtime[ddir]);
1179
1180         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1181                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1182         else
1183                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1184
1185         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1186                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1187         else
1188                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1189
1190         if (ts->lat_percentiles)
1191                 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1192                                         ts->lat_stat[ddir].samples,
1193                                         ts->percentile_list, &ovals, &maxv,
1194                                         &minv);
1195         else if (ts->clat_percentiles)
1196                 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1197                                         ts->clat_stat[ddir].samples,
1198                                         ts->percentile_list, &ovals, &maxv,
1199                                         &minv);
1200         else
1201                 len = 0;
1202
1203         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1204                 if (i >= len) {
1205                         log_buf(out, ";0%%=0");
1206                         continue;
1207                 }
1208                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1209         }
1210
1211         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1212                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1213         else
1214                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1215
1216         free(ovals);
1217
1218         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1219         if (bw_stat) {
1220                 double p_of_agg = 100.0;
1221
1222                 if (rs->agg[ddir]) {
1223                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1224                         if (p_of_agg > 100.0)
1225                                 p_of_agg = 100.0;
1226                 }
1227
1228                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1229         } else
1230                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1231
1232         if (ver == 5) {
1233                 if (bw_stat)
1234                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1235                 else
1236                         log_buf(out, ";%lu", 0UL);
1237
1238                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1239                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1240                                 mean, dev, (&ts->iops_stat[ddir])->samples);
1241                 else
1242                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1243         }
1244 }
1245
1246 static struct json_object *add_ddir_lat_json(struct thread_stat *ts, uint32_t percentiles,
1247                 struct io_stat *lat_stat, uint64_t *io_u_plat)
1248 {
1249         char buf[120];
1250         double mean, dev;
1251         unsigned int i, len;
1252         struct json_object *lat_object, *percentile_object, *clat_bins_object;
1253         unsigned long long min, max, maxv, minv, *ovals = NULL;
1254
1255         if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1256                 min = max = 0;
1257                 mean = dev = 0.0;
1258         }
1259         lat_object = json_create_object();
1260         json_object_add_value_int(lat_object, "min", min);
1261         json_object_add_value_int(lat_object, "max", max);
1262         json_object_add_value_float(lat_object, "mean", mean);
1263         json_object_add_value_float(lat_object, "stddev", dev);
1264         json_object_add_value_int(lat_object, "N", lat_stat->samples);
1265
1266         if (percentiles && lat_stat->samples) {
1267                 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1268                                 ts->percentile_list, &ovals, &maxv, &minv);
1269
1270                 if (len > FIO_IO_U_LIST_MAX_LEN)
1271                         len = FIO_IO_U_LIST_MAX_LEN;
1272
1273                 percentile_object = json_create_object();
1274                 json_object_add_value_object(lat_object, "percentile", percentile_object);
1275                 for (i = 0; i < len; i++) {
1276                         snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1277                         json_object_add_value_int(percentile_object, buf, ovals[i]);
1278                 }
1279                 free(ovals);
1280
1281                 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1282                         clat_bins_object = json_create_object();
1283                         json_object_add_value_object(lat_object, "bins", clat_bins_object);
1284
1285                         for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1286                                 if (io_u_plat[i]) {
1287                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1288                                         json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1289                                 }
1290                 }
1291         }
1292
1293         return lat_object;
1294 }
1295
1296 static void add_ddir_status_json(struct thread_stat *ts,
1297                 struct group_run_stats *rs, int ddir, struct json_object *parent)
1298 {
1299         unsigned long long min, max;
1300         unsigned long long bw_bytes, bw;
1301         double mean, dev, iops;
1302         struct json_object *dir_object, *tmp_object;
1303         double p_of_agg = 100.0;
1304
1305         assert(ddir_rw(ddir) || ddir_sync(ddir));
1306
1307         if (ts->unified_rw_rep && ddir != DDIR_READ)
1308                 return;
1309
1310         dir_object = json_create_object();
1311         json_object_add_value_object(parent,
1312                 ts->unified_rw_rep ? "mixed" : io_ddir_name(ddir), dir_object);
1313
1314         if (ddir_rw(ddir)) {
1315                 bw_bytes = 0;
1316                 bw = 0;
1317                 iops = 0.0;
1318                 if (ts->runtime[ddir]) {
1319                         uint64_t runt = ts->runtime[ddir];
1320
1321                         bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1322                         bw = bw_bytes / 1024; /* KiB/s */
1323                         iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1324                 }
1325
1326                 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1327                 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1328                 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1329                 json_object_add_value_int(dir_object, "bw", bw);
1330                 json_object_add_value_float(dir_object, "iops", iops);
1331                 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1332                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1333                 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1334                 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1335
1336                 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1337                                 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1338                 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1339
1340                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1341                                 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1342                 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1343
1344                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1345                                 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1346                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1347         } else {
1348                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1349                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1350                                 &ts->sync_stat, ts->io_u_sync_plat);
1351                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1352         }
1353
1354         if (!ddir_rw(ddir))
1355                 return;
1356
1357         /* Only print PRIO latencies if some high priority samples were gathered */
1358         if (ts->clat_high_prio_stat[ddir].samples > 0) {
1359                 const char *high, *low;
1360
1361                 if (ts->lat_percentiles) {
1362                         high = "lat_high_prio";
1363                         low = "lat_low_prio";
1364                 } else {
1365                         high = "clat_high_prio";
1366                         low = "clat_low_prio";
1367                 }
1368
1369                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1370                                 &ts->clat_high_prio_stat[ddir], ts->io_u_plat_high_prio[ddir]);
1371                 json_object_add_value_object(dir_object, high, tmp_object);
1372
1373                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1374                                 &ts->clat_low_prio_stat[ddir], ts->io_u_plat_low_prio[ddir]);
1375                 json_object_add_value_object(dir_object, low, tmp_object);
1376         }
1377
1378         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1379                 if (rs->agg[ddir]) {
1380                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1381                         if (p_of_agg > 100.0)
1382                                 p_of_agg = 100.0;
1383                 }
1384         } else {
1385                 min = max = 0;
1386                 p_of_agg = mean = dev = 0.0;
1387         }
1388
1389         json_object_add_value_int(dir_object, "bw_min", min);
1390         json_object_add_value_int(dir_object, "bw_max", max);
1391         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1392         json_object_add_value_float(dir_object, "bw_mean", mean);
1393         json_object_add_value_float(dir_object, "bw_dev", dev);
1394         json_object_add_value_int(dir_object, "bw_samples",
1395                                 (&ts->bw_stat[ddir])->samples);
1396
1397         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1398                 min = max = 0;
1399                 mean = dev = 0.0;
1400         }
1401         json_object_add_value_int(dir_object, "iops_min", min);
1402         json_object_add_value_int(dir_object, "iops_max", max);
1403         json_object_add_value_float(dir_object, "iops_mean", mean);
1404         json_object_add_value_float(dir_object, "iops_stddev", dev);
1405         json_object_add_value_int(dir_object, "iops_samples",
1406                                 (&ts->iops_stat[ddir])->samples);
1407
1408         if (ts->cachehit + ts->cachemiss) {
1409                 uint64_t total;
1410                 double hit;
1411
1412                 total = ts->cachehit + ts->cachemiss;
1413                 hit = (double) ts->cachehit / (double) total;
1414                 hit *= 100.0;
1415                 json_object_add_value_float(dir_object, "cachehit", hit);
1416         }
1417 }
1418
1419 static void show_thread_status_terse_all(struct thread_stat *ts,
1420                                          struct group_run_stats *rs, int ver,
1421                                          struct buf_output *out)
1422 {
1423         double io_u_dist[FIO_IO_U_MAP_NR];
1424         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1425         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1426         double usr_cpu, sys_cpu;
1427         int i;
1428
1429         /* General Info */
1430         if (ver == 2)
1431                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1432         else
1433                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1434                         ts->name, ts->groupid, ts->error);
1435
1436         /* Log Read Status */
1437         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1438         /* Log Write Status */
1439         show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1440         /* Log Trim Status */
1441         if (ver == 2 || ver == 4 || ver == 5)
1442                 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1443
1444         /* CPU Usage */
1445         if (ts->total_run_time) {
1446                 double runt = (double) ts->total_run_time;
1447
1448                 usr_cpu = (double) ts->usr_time * 100 / runt;
1449                 sys_cpu = (double) ts->sys_time * 100 / runt;
1450         } else {
1451                 usr_cpu = 0;
1452                 sys_cpu = 0;
1453         }
1454
1455         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1456                                                 (unsigned long long) ts->ctx,
1457                                                 (unsigned long long) ts->majf,
1458                                                 (unsigned long long) ts->minf);
1459
1460         /* Calc % distribution of IO depths, usecond, msecond latency */
1461         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1462         stat_calc_lat_nu(ts, io_u_lat_u);
1463         stat_calc_lat_m(ts, io_u_lat_m);
1464
1465         /* Only show fixed 7 I/O depth levels*/
1466         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1467                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1468                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1469
1470         /* Microsecond latency */
1471         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1472                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1473         /* Millisecond latency */
1474         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1475                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1476
1477         /* disk util stats, if any */
1478         if (ver >= 3 && is_running_backend())
1479                 show_disk_util(1, NULL, out);
1480
1481         /* Additional output if continue_on_error set - default off*/
1482         if (ts->continue_on_error)
1483                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1484
1485         /* Additional output if description is set */
1486         if (strlen(ts->description)) {
1487                 if (ver == 2)
1488                         log_buf(out, "\n");
1489                 log_buf(out, ";%s", ts->description);
1490         }
1491
1492         log_buf(out, "\n");
1493 }
1494
1495 static void json_add_job_opts(struct json_object *root, const char *name,
1496                               struct flist_head *opt_list)
1497 {
1498         struct json_object *dir_object;
1499         struct flist_head *entry;
1500         struct print_option *p;
1501
1502         if (flist_empty(opt_list))
1503                 return;
1504
1505         dir_object = json_create_object();
1506         json_object_add_value_object(root, name, dir_object);
1507
1508         flist_for_each(entry, opt_list) {
1509                 const char *pos = "";
1510
1511                 p = flist_entry(entry, struct print_option, list);
1512                 if (p->value)
1513                         pos = p->value;
1514                 json_object_add_value_string(dir_object, p->name, pos);
1515         }
1516 }
1517
1518 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1519                                                    struct group_run_stats *rs,
1520                                                    struct flist_head *opt_list)
1521 {
1522         struct json_object *root, *tmp;
1523         struct jobs_eta *je;
1524         double io_u_dist[FIO_IO_U_MAP_NR];
1525         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1526         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1527         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1528         double usr_cpu, sys_cpu;
1529         int i;
1530         size_t size;
1531
1532         root = json_create_object();
1533         json_object_add_value_string(root, "jobname", ts->name);
1534         json_object_add_value_int(root, "groupid", ts->groupid);
1535         json_object_add_value_int(root, "error", ts->error);
1536
1537         /* ETA Info */
1538         je = get_jobs_eta(true, &size);
1539         if (je) {
1540                 json_object_add_value_int(root, "eta", je->eta_sec);
1541                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1542         }
1543
1544         if (opt_list)
1545                 json_add_job_opts(root, "job options", opt_list);
1546
1547         add_ddir_status_json(ts, rs, DDIR_READ, root);
1548         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1549         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1550         add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1551
1552         /* CPU Usage */
1553         if (ts->total_run_time) {
1554                 double runt = (double) ts->total_run_time;
1555
1556                 usr_cpu = (double) ts->usr_time * 100 / runt;
1557                 sys_cpu = (double) ts->sys_time * 100 / runt;
1558         } else {
1559                 usr_cpu = 0;
1560                 sys_cpu = 0;
1561         }
1562         json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1563         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1564         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1565         json_object_add_value_int(root, "ctx", ts->ctx);
1566         json_object_add_value_int(root, "majf", ts->majf);
1567         json_object_add_value_int(root, "minf", ts->minf);
1568
1569         /* Calc % distribution of IO depths */
1570         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1571         tmp = json_create_object();
1572         json_object_add_value_object(root, "iodepth_level", tmp);
1573         /* Only show fixed 7 I/O depth levels*/
1574         for (i = 0; i < 7; i++) {
1575                 char name[20];
1576                 if (i < 6)
1577                         snprintf(name, 20, "%d", 1 << i);
1578                 else
1579                         snprintf(name, 20, ">=%d", 1 << i);
1580                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1581         }
1582
1583         /* Calc % distribution of submit IO depths */
1584         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1585         tmp = json_create_object();
1586         json_object_add_value_object(root, "iodepth_submit", tmp);
1587         /* Only show fixed 7 I/O depth levels*/
1588         for (i = 0; i < 7; i++) {
1589                 char name[20];
1590                 if (i == 0)
1591                         snprintf(name, 20, "0");
1592                 else if (i < 6)
1593                         snprintf(name, 20, "%d", 1 << (i+1));
1594                 else
1595                         snprintf(name, 20, ">=%d", 1 << i);
1596                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1597         }
1598
1599         /* Calc % distribution of completion IO depths */
1600         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1601         tmp = json_create_object();
1602         json_object_add_value_object(root, "iodepth_complete", tmp);
1603         /* Only show fixed 7 I/O depth levels*/
1604         for (i = 0; i < 7; i++) {
1605                 char name[20];
1606                 if (i == 0)
1607                         snprintf(name, 20, "0");
1608                 else if (i < 6)
1609                         snprintf(name, 20, "%d", 1 << (i+1));
1610                 else
1611                         snprintf(name, 20, ">=%d", 1 << i);
1612                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1613         }
1614
1615         /* Calc % distribution of nsecond, usecond, msecond latency */
1616         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1617         stat_calc_lat_n(ts, io_u_lat_n);
1618         stat_calc_lat_u(ts, io_u_lat_u);
1619         stat_calc_lat_m(ts, io_u_lat_m);
1620
1621         /* Nanosecond latency */
1622         tmp = json_create_object();
1623         json_object_add_value_object(root, "latency_ns", tmp);
1624         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1625                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1626                                  "250", "500", "750", "1000", };
1627                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1628         }
1629         /* Microsecond latency */
1630         tmp = json_create_object();
1631         json_object_add_value_object(root, "latency_us", tmp);
1632         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1633                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1634                                  "250", "500", "750", "1000", };
1635                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1636         }
1637         /* Millisecond latency */
1638         tmp = json_create_object();
1639         json_object_add_value_object(root, "latency_ms", tmp);
1640         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1641                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1642                                  "250", "500", "750", "1000", "2000",
1643                                  ">=2000", };
1644                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1645         }
1646
1647         /* Additional output if continue_on_error set - default off*/
1648         if (ts->continue_on_error) {
1649                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1650                 json_object_add_value_int(root, "first_error", ts->first_error);
1651         }
1652
1653         if (ts->latency_depth) {
1654                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1655                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1656                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1657                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1658         }
1659
1660         /* Additional output if description is set */
1661         if (strlen(ts->description))
1662                 json_object_add_value_string(root, "desc", ts->description);
1663
1664         if (ts->nr_block_infos) {
1665                 /* Block error histogram and types */
1666                 int len;
1667                 unsigned int *percentiles = NULL;
1668                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1669
1670                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1671                                              ts->percentile_list,
1672                                              &percentiles, block_state_counts);
1673
1674                 if (len) {
1675                         struct json_object *block, *percentile_object, *states;
1676                         int state;
1677                         block = json_create_object();
1678                         json_object_add_value_object(root, "block", block);
1679
1680                         percentile_object = json_create_object();
1681                         json_object_add_value_object(block, "percentiles",
1682                                                      percentile_object);
1683                         for (i = 0; i < len; i++) {
1684                                 char buf[20];
1685                                 snprintf(buf, sizeof(buf), "%f",
1686                                          ts->percentile_list[i].u.f);
1687                                 json_object_add_value_int(percentile_object,
1688                                                           buf,
1689                                                           percentiles[i]);
1690                         }
1691
1692                         states = json_create_object();
1693                         json_object_add_value_object(block, "states", states);
1694                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1695                                 json_object_add_value_int(states,
1696                                         block_state_names[state],
1697                                         block_state_counts[state]);
1698                         }
1699                         free(percentiles);
1700                 }
1701         }
1702
1703         if (ts->ss_dur) {
1704                 struct json_object *data;
1705                 struct json_array *iops, *bw;
1706                 int j, k, l;
1707                 char ss_buf[64];
1708
1709                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1710                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1711                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1712                         (float) ts->ss_limit.u.f,
1713                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1714
1715                 tmp = json_create_object();
1716                 json_object_add_value_object(root, "steadystate", tmp);
1717                 json_object_add_value_string(tmp, "ss", ss_buf);
1718                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1719                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1720
1721                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1722                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1723                 json_object_add_value_string(tmp, "criterion", ss_buf);
1724                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1725                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1726
1727                 data = json_create_object();
1728                 json_object_add_value_object(tmp, "data", data);
1729                 bw = json_create_array();
1730                 iops = json_create_array();
1731
1732                 /*
1733                 ** if ss was attained or the buffer is not full,
1734                 ** ss->head points to the first element in the list.
1735                 ** otherwise it actually points to the second element
1736                 ** in the list
1737                 */
1738                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1739                         j = ts->ss_head;
1740                 else
1741                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1742                 for (l = 0; l < ts->ss_dur; l++) {
1743                         k = (j + l) % ts->ss_dur;
1744                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1745                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1746                 }
1747                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1748                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1749                 json_object_add_value_array(data, "iops", iops);
1750                 json_object_add_value_array(data, "bw", bw);
1751         }
1752
1753         return root;
1754 }
1755
1756 static void show_thread_status_terse(struct thread_stat *ts,
1757                                      struct group_run_stats *rs,
1758                                      struct buf_output *out)
1759 {
1760         if (terse_version >= 2 && terse_version <= 5)
1761                 show_thread_status_terse_all(ts, rs, terse_version, out);
1762         else
1763                 log_err("fio: bad terse version!? %d\n", terse_version);
1764 }
1765
1766 struct json_object *show_thread_status(struct thread_stat *ts,
1767                                        struct group_run_stats *rs,
1768                                        struct flist_head *opt_list,
1769                                        struct buf_output *out)
1770 {
1771         struct json_object *ret = NULL;
1772
1773         if (output_format & FIO_OUTPUT_TERSE)
1774                 show_thread_status_terse(ts, rs,  out);
1775         if (output_format & FIO_OUTPUT_JSON)
1776                 ret = show_thread_status_json(ts, rs, opt_list);
1777         if (output_format & FIO_OUTPUT_NORMAL)
1778                 show_thread_status_normal(ts, rs,  out);
1779
1780         return ret;
1781 }
1782
1783 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1784 {
1785         double mean, S;
1786
1787         dst->min_val = min(dst->min_val, src->min_val);
1788         dst->max_val = max(dst->max_val, src->max_val);
1789
1790         /*
1791          * Compute new mean and S after the merge
1792          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1793          *  #Parallel_algorithm>
1794          */
1795         if (first) {
1796                 mean = src->mean.u.f;
1797                 S = src->S.u.f;
1798         } else {
1799                 double delta = src->mean.u.f - dst->mean.u.f;
1800
1801                 mean = ((src->mean.u.f * src->samples) +
1802                         (dst->mean.u.f * dst->samples)) /
1803                         (dst->samples + src->samples);
1804
1805                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1806                         (dst->samples * src->samples) /
1807                         (dst->samples + src->samples);
1808         }
1809
1810         dst->samples += src->samples;
1811         dst->mean.u.f = mean;
1812         dst->S.u.f = S;
1813
1814 }
1815
1816 /*
1817  * We sum two kinds of stats - one that is time based, in which case we
1818  * apply the proper summing technique, and then one that is iops/bw
1819  * numbers. For group_reporting, we should just add those up, not make
1820  * them the mean of everything.
1821  */
1822 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first,
1823                      bool pure_sum)
1824 {
1825         if (src->samples == 0)
1826                 return;
1827
1828         if (!pure_sum) {
1829                 __sum_stat(dst, src, first);
1830                 return;
1831         }
1832
1833         if (first) {
1834                 dst->min_val = src->min_val;
1835                 dst->max_val = src->max_val;
1836                 dst->samples = src->samples;
1837                 dst->mean.u.f = src->mean.u.f;
1838                 dst->S.u.f = src->S.u.f;
1839         } else {
1840                 dst->min_val += src->min_val;
1841                 dst->max_val += src->max_val;
1842                 dst->samples += src->samples;
1843                 dst->mean.u.f += src->mean.u.f;
1844                 dst->S.u.f += src->S.u.f;
1845         }
1846 }
1847
1848 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1849 {
1850         int i;
1851
1852         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1853                 if (dst->max_run[i] < src->max_run[i])
1854                         dst->max_run[i] = src->max_run[i];
1855                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1856                         dst->min_run[i] = src->min_run[i];
1857                 if (dst->max_bw[i] < src->max_bw[i])
1858                         dst->max_bw[i] = src->max_bw[i];
1859                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1860                         dst->min_bw[i] = src->min_bw[i];
1861
1862                 dst->iobytes[i] += src->iobytes[i];
1863                 dst->agg[i] += src->agg[i];
1864         }
1865
1866         if (!dst->kb_base)
1867                 dst->kb_base = src->kb_base;
1868         if (!dst->unit_base)
1869                 dst->unit_base = src->unit_base;
1870         if (!dst->sig_figs)
1871                 dst->sig_figs = src->sig_figs;
1872 }
1873
1874 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1875                       bool first)
1876 {
1877         int k, l, m;
1878
1879         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1880                 if (!dst->unified_rw_rep) {
1881                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first, false);
1882                         sum_stat(&dst->clat_high_prio_stat[l], &src->clat_high_prio_stat[l], first, false);
1883                         sum_stat(&dst->clat_low_prio_stat[l], &src->clat_low_prio_stat[l], first, false);
1884                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first, false);
1885                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first, false);
1886                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first, true);
1887                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first, true);
1888
1889                         dst->io_bytes[l] += src->io_bytes[l];
1890
1891                         if (dst->runtime[l] < src->runtime[l])
1892                                 dst->runtime[l] = src->runtime[l];
1893                 } else {
1894                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first, false);
1895                         sum_stat(&dst->clat_high_prio_stat[0], &src->clat_high_prio_stat[l], first, false);
1896                         sum_stat(&dst->clat_low_prio_stat[0], &src->clat_low_prio_stat[l], first, false);
1897                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first, false);
1898                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first, false);
1899                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first, true);
1900                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first, true);
1901
1902                         dst->io_bytes[0] += src->io_bytes[l];
1903
1904                         if (dst->runtime[0] < src->runtime[l])
1905                                 dst->runtime[0] = src->runtime[l];
1906
1907                         /*
1908                          * We're summing to the same destination, so override
1909                          * 'first' after the first iteration of the loop
1910                          */
1911                         first = false;
1912                 }
1913         }
1914
1915         sum_stat(&dst->sync_stat, &src->sync_stat, first, false);
1916         dst->usr_time += src->usr_time;
1917         dst->sys_time += src->sys_time;
1918         dst->ctx += src->ctx;
1919         dst->majf += src->majf;
1920         dst->minf += src->minf;
1921
1922         for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
1923                 dst->io_u_map[k] += src->io_u_map[k];
1924                 dst->io_u_submit[k] += src->io_u_submit[k];
1925                 dst->io_u_complete[k] += src->io_u_complete[k];
1926         }
1927
1928         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1929                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1930         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1931                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1932         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1933                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1934
1935         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1936                 if (!dst->unified_rw_rep) {
1937                         dst->total_io_u[k] += src->total_io_u[k];
1938                         dst->short_io_u[k] += src->short_io_u[k];
1939                         dst->drop_io_u[k] += src->drop_io_u[k];
1940                 } else {
1941                         dst->total_io_u[0] += src->total_io_u[k];
1942                         dst->short_io_u[0] += src->short_io_u[k];
1943                         dst->drop_io_u[0] += src->drop_io_u[k];
1944                 }
1945         }
1946
1947         dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
1948
1949         for (k = 0; k < FIO_LAT_CNT; k++)
1950                 for (l = 0; l < DDIR_RWDIR_CNT; l++)
1951                         for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
1952                                 if (!dst->unified_rw_rep)
1953                                         dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
1954                                 else
1955                                         dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
1956
1957         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1958                 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
1959
1960         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1961                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1962                         if (!dst->unified_rw_rep) {
1963                                 dst->io_u_plat_high_prio[k][m] += src->io_u_plat_high_prio[k][m];
1964                                 dst->io_u_plat_low_prio[k][m] += src->io_u_plat_low_prio[k][m];
1965                         } else {
1966                                 dst->io_u_plat_high_prio[0][m] += src->io_u_plat_high_prio[k][m];
1967                                 dst->io_u_plat_low_prio[0][m] += src->io_u_plat_low_prio[k][m];
1968                         }
1969
1970                 }
1971         }
1972
1973         dst->total_run_time += src->total_run_time;
1974         dst->total_submit += src->total_submit;
1975         dst->total_complete += src->total_complete;
1976         dst->nr_zone_resets += src->nr_zone_resets;
1977         dst->cachehit += src->cachehit;
1978         dst->cachemiss += src->cachemiss;
1979 }
1980
1981 void init_group_run_stat(struct group_run_stats *gs)
1982 {
1983         int i;
1984         memset(gs, 0, sizeof(*gs));
1985
1986         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1987                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1988 }
1989
1990 void init_thread_stat(struct thread_stat *ts)
1991 {
1992         int j;
1993
1994         memset(ts, 0, sizeof(*ts));
1995
1996         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1997                 ts->lat_stat[j].min_val = -1UL;
1998                 ts->clat_stat[j].min_val = -1UL;
1999                 ts->slat_stat[j].min_val = -1UL;
2000                 ts->bw_stat[j].min_val = -1UL;
2001                 ts->iops_stat[j].min_val = -1UL;
2002                 ts->clat_high_prio_stat[j].min_val = -1UL;
2003                 ts->clat_low_prio_stat[j].min_val = -1UL;
2004         }
2005         ts->sync_stat.min_val = -1UL;
2006         ts->groupid = -1;
2007 }
2008
2009 void __show_run_stats(void)
2010 {
2011         struct group_run_stats *runstats, *rs;
2012         struct thread_data *td;
2013         struct thread_stat *threadstats, *ts;
2014         int i, j, k, nr_ts, last_ts, idx;
2015         bool kb_base_warned = false;
2016         bool unit_base_warned = false;
2017         struct json_object *root = NULL;
2018         struct json_array *array = NULL;
2019         struct buf_output output[FIO_OUTPUT_NR];
2020         struct flist_head **opt_lists;
2021
2022         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2023
2024         for (i = 0; i < groupid + 1; i++)
2025                 init_group_run_stat(&runstats[i]);
2026
2027         /*
2028          * find out how many threads stats we need. if group reporting isn't
2029          * enabled, it's one-per-td.
2030          */
2031         nr_ts = 0;
2032         last_ts = -1;
2033         for_each_td(td, i) {
2034                 if (!td->o.group_reporting) {
2035                         nr_ts++;
2036                         continue;
2037                 }
2038                 if (last_ts == td->groupid)
2039                         continue;
2040                 if (!td->o.stats)
2041                         continue;
2042
2043                 last_ts = td->groupid;
2044                 nr_ts++;
2045         }
2046
2047         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2048         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2049
2050         for (i = 0; i < nr_ts; i++) {
2051                 init_thread_stat(&threadstats[i]);
2052                 opt_lists[i] = NULL;
2053         }
2054
2055         j = 0;
2056         last_ts = -1;
2057         idx = 0;
2058         for_each_td(td, i) {
2059                 if (!td->o.stats)
2060                         continue;
2061                 if (idx && (!td->o.group_reporting ||
2062                     (td->o.group_reporting && last_ts != td->groupid))) {
2063                         idx = 0;
2064                         j++;
2065                 }
2066
2067                 last_ts = td->groupid;
2068
2069                 ts = &threadstats[j];
2070
2071                 ts->clat_percentiles = td->o.clat_percentiles;
2072                 ts->lat_percentiles = td->o.lat_percentiles;
2073                 ts->slat_percentiles = td->o.slat_percentiles;
2074                 ts->percentile_precision = td->o.percentile_precision;
2075                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2076                 opt_lists[j] = &td->opt_list;
2077
2078                 idx++;
2079                 ts->members++;
2080
2081                 if (ts->groupid == -1) {
2082                         /*
2083                          * These are per-group shared already
2084                          */
2085                         snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2086                         if (td->o.description)
2087                                 snprintf(ts->description,
2088                                          sizeof(ts->description), "%s",
2089                                          td->o.description);
2090                         else
2091                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2092
2093                         /*
2094                          * If multiple entries in this group, this is
2095                          * the first member.
2096                          */
2097                         ts->thread_number = td->thread_number;
2098                         ts->groupid = td->groupid;
2099
2100                         /*
2101                          * first pid in group, not very useful...
2102                          */
2103                         ts->pid = td->pid;
2104
2105                         ts->kb_base = td->o.kb_base;
2106                         ts->unit_base = td->o.unit_base;
2107                         ts->sig_figs = td->o.sig_figs;
2108                         ts->unified_rw_rep = td->o.unified_rw_rep;
2109                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2110                         log_info("fio: kb_base differs for jobs in group, using"
2111                                  " %u as the base\n", ts->kb_base);
2112                         kb_base_warned = true;
2113                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2114                         log_info("fio: unit_base differs for jobs in group, using"
2115                                  " %u as the base\n", ts->unit_base);
2116                         unit_base_warned = true;
2117                 }
2118
2119                 ts->continue_on_error = td->o.continue_on_error;
2120                 ts->total_err_count += td->total_err_count;
2121                 ts->first_error = td->first_error;
2122                 if (!ts->error) {
2123                         if (!td->error && td->o.continue_on_error &&
2124                             td->first_error) {
2125                                 ts->error = td->first_error;
2126                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2127                                          td->verror);
2128                         } else  if (td->error) {
2129                                 ts->error = td->error;
2130                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2131                                          td->verror);
2132                         }
2133                 }
2134
2135                 ts->latency_depth = td->latency_qd;
2136                 ts->latency_target = td->o.latency_target;
2137                 ts->latency_percentile = td->o.latency_percentile;
2138                 ts->latency_window = td->o.latency_window;
2139
2140                 ts->nr_block_infos = td->ts.nr_block_infos;
2141                 for (k = 0; k < ts->nr_block_infos; k++)
2142                         ts->block_infos[k] = td->ts.block_infos[k];
2143
2144                 sum_thread_stats(ts, &td->ts, idx == 1);
2145
2146                 if (td->o.ss_dur) {
2147                         ts->ss_state = td->ss.state;
2148                         ts->ss_dur = td->ss.dur;
2149                         ts->ss_head = td->ss.head;
2150                         ts->ss_bw_data = td->ss.bw_data;
2151                         ts->ss_iops_data = td->ss.iops_data;
2152                         ts->ss_limit.u.f = td->ss.limit;
2153                         ts->ss_slope.u.f = td->ss.slope;
2154                         ts->ss_deviation.u.f = td->ss.deviation;
2155                         ts->ss_criterion.u.f = td->ss.criterion;
2156                 }
2157                 else
2158                         ts->ss_dur = ts->ss_state = 0;
2159         }
2160
2161         for (i = 0; i < nr_ts; i++) {
2162                 unsigned long long bw;
2163
2164                 ts = &threadstats[i];
2165                 if (ts->groupid == -1)
2166                         continue;
2167                 rs = &runstats[ts->groupid];
2168                 rs->kb_base = ts->kb_base;
2169                 rs->unit_base = ts->unit_base;
2170                 rs->sig_figs = ts->sig_figs;
2171                 rs->unified_rw_rep += ts->unified_rw_rep;
2172
2173                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2174                         if (!ts->runtime[j])
2175                                 continue;
2176                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2177                                 rs->min_run[j] = ts->runtime[j];
2178                         if (ts->runtime[j] > rs->max_run[j])
2179                                 rs->max_run[j] = ts->runtime[j];
2180
2181                         bw = 0;
2182                         if (ts->runtime[j])
2183                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2184                         if (bw < rs->min_bw[j])
2185                                 rs->min_bw[j] = bw;
2186                         if (bw > rs->max_bw[j])
2187                                 rs->max_bw[j] = bw;
2188
2189                         rs->iobytes[j] += ts->io_bytes[j];
2190                 }
2191         }
2192
2193         for (i = 0; i < groupid + 1; i++) {
2194                 int ddir;
2195
2196                 rs = &runstats[i];
2197
2198                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2199                         if (rs->max_run[ddir])
2200                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2201                                                 rs->max_run[ddir];
2202                 }
2203         }
2204
2205         for (i = 0; i < FIO_OUTPUT_NR; i++)
2206                 buf_output_init(&output[i]);
2207
2208         /*
2209          * don't overwrite last signal output
2210          */
2211         if (output_format & FIO_OUTPUT_NORMAL)
2212                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2213         if (output_format & FIO_OUTPUT_JSON) {
2214                 struct thread_data *global;
2215                 char time_buf[32];
2216                 struct timeval now;
2217                 unsigned long long ms_since_epoch;
2218                 time_t tv_sec;
2219
2220                 gettimeofday(&now, NULL);
2221                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2222                                  (unsigned long long)(now.tv_usec) / 1000;
2223
2224                 tv_sec = now.tv_sec;
2225                 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2226                 if (time_buf[strlen(time_buf) - 1] == '\n')
2227                         time_buf[strlen(time_buf) - 1] = '\0';
2228
2229                 root = json_create_object();
2230                 json_object_add_value_string(root, "fio version", fio_version_string);
2231                 json_object_add_value_int(root, "timestamp", now.tv_sec);
2232                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2233                 json_object_add_value_string(root, "time", time_buf);
2234                 global = get_global_options();
2235                 json_add_job_opts(root, "global options", &global->opt_list);
2236                 array = json_create_array();
2237                 json_object_add_value_array(root, "jobs", array);
2238         }
2239
2240         if (is_backend)
2241                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2242
2243         for (i = 0; i < nr_ts; i++) {
2244                 ts = &threadstats[i];
2245                 rs = &runstats[ts->groupid];
2246
2247                 if (is_backend) {
2248                         fio_server_send_job_options(opt_lists[i], i);
2249                         fio_server_send_ts(ts, rs);
2250                 } else {
2251                         if (output_format & FIO_OUTPUT_TERSE)
2252                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2253                         if (output_format & FIO_OUTPUT_JSON) {
2254                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2255                                 json_array_add_value_object(array, tmp);
2256                         }
2257                         if (output_format & FIO_OUTPUT_NORMAL)
2258                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2259                 }
2260         }
2261         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2262                 /* disk util stats, if any */
2263                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2264
2265                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2266
2267                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2268                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2269                 json_free_object(root);
2270         }
2271
2272         for (i = 0; i < groupid + 1; i++) {
2273                 rs = &runstats[i];
2274
2275                 rs->groupid = i;
2276                 if (is_backend)
2277                         fio_server_send_gs(rs);
2278                 else if (output_format & FIO_OUTPUT_NORMAL)
2279                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2280         }
2281
2282         if (is_backend)
2283                 fio_server_send_du();
2284         else if (output_format & FIO_OUTPUT_NORMAL) {
2285                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2286                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2287         }
2288
2289         for (i = 0; i < FIO_OUTPUT_NR; i++) {
2290                 struct buf_output *out = &output[i];
2291
2292                 log_info_buf(out->buf, out->buflen);
2293                 buf_output_free(out);
2294         }
2295
2296         fio_idle_prof_cleanup();
2297
2298         log_info_flush();
2299         free(runstats);
2300         free(threadstats);
2301         free(opt_lists);
2302 }
2303
2304 void __show_running_run_stats(void)
2305 {
2306         struct thread_data *td;
2307         unsigned long long *rt;
2308         struct timespec ts;
2309         int i;
2310
2311         fio_sem_down(stat_sem);
2312
2313         rt = malloc(thread_number * sizeof(unsigned long long));
2314         fio_gettime(&ts, NULL);
2315
2316         for_each_td(td, i) {
2317                 td->update_rusage = 1;
2318                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
2319                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
2320                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2321                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2322
2323                 rt[i] = mtime_since(&td->start, &ts);
2324                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2325                         td->ts.runtime[DDIR_READ] += rt[i];
2326                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2327                         td->ts.runtime[DDIR_WRITE] += rt[i];
2328                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2329                         td->ts.runtime[DDIR_TRIM] += rt[i];
2330         }
2331
2332         for_each_td(td, i) {
2333                 if (td->runstate >= TD_EXITED)
2334                         continue;
2335                 if (td->rusage_sem) {
2336                         td->update_rusage = 1;
2337                         fio_sem_down(td->rusage_sem);
2338                 }
2339                 td->update_rusage = 0;
2340         }
2341
2342         __show_run_stats();
2343
2344         for_each_td(td, i) {
2345                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2346                         td->ts.runtime[DDIR_READ] -= rt[i];
2347                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2348                         td->ts.runtime[DDIR_WRITE] -= rt[i];
2349                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2350                         td->ts.runtime[DDIR_TRIM] -= rt[i];
2351         }
2352
2353         free(rt);
2354         fio_sem_up(stat_sem);
2355 }
2356
2357 static bool status_interval_init;
2358 static struct timespec status_time;
2359 static bool status_file_disabled;
2360
2361 #define FIO_STATUS_FILE         "fio-dump-status"
2362
2363 static int check_status_file(void)
2364 {
2365         struct stat sb;
2366         const char *temp_dir;
2367         char fio_status_file_path[PATH_MAX];
2368
2369         if (status_file_disabled)
2370                 return 0;
2371
2372         temp_dir = getenv("TMPDIR");
2373         if (temp_dir == NULL) {
2374                 temp_dir = getenv("TEMP");
2375                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2376                         temp_dir = NULL;
2377         }
2378         if (temp_dir == NULL)
2379                 temp_dir = "/tmp";
2380 #ifdef __COVERITY__
2381         __coverity_tainted_data_sanitize__(temp_dir);
2382 #endif
2383
2384         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2385
2386         if (stat(fio_status_file_path, &sb))
2387                 return 0;
2388
2389         if (unlink(fio_status_file_path) < 0) {
2390                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2391                                                         strerror(errno));
2392                 log_err("fio: disabling status file updates\n");
2393                 status_file_disabled = true;
2394         }
2395
2396         return 1;
2397 }
2398
2399 void check_for_running_stats(void)
2400 {
2401         if (status_interval) {
2402                 if (!status_interval_init) {
2403                         fio_gettime(&status_time, NULL);
2404                         status_interval_init = true;
2405                 } else if (mtime_since_now(&status_time) >= status_interval) {
2406                         show_running_run_stats();
2407                         fio_gettime(&status_time, NULL);
2408                         return;
2409                 }
2410         }
2411         if (check_status_file()) {
2412                 show_running_run_stats();
2413                 return;
2414         }
2415 }
2416
2417 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2418 {
2419         double val = data;
2420         double delta;
2421
2422         if (data > is->max_val)
2423                 is->max_val = data;
2424         if (data < is->min_val)
2425                 is->min_val = data;
2426
2427         delta = val - is->mean.u.f;
2428         if (delta) {
2429                 is->mean.u.f += delta / (is->samples + 1.0);
2430                 is->S.u.f += delta * (val - is->mean.u.f);
2431         }
2432
2433         is->samples++;
2434 }
2435
2436 /*
2437  * Return a struct io_logs, which is added to the tail of the log
2438  * list for 'iolog'.
2439  */
2440 static struct io_logs *get_new_log(struct io_log *iolog)
2441 {
2442         size_t new_size, new_samples;
2443         struct io_logs *cur_log;
2444
2445         /*
2446          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2447          * forever
2448          */
2449         if (!iolog->cur_log_max)
2450                 new_samples = DEF_LOG_ENTRIES;
2451         else {
2452                 new_samples = iolog->cur_log_max * 2;
2453                 if (new_samples > MAX_LOG_ENTRIES)
2454                         new_samples = MAX_LOG_ENTRIES;
2455         }
2456
2457         new_size = new_samples * log_entry_sz(iolog);
2458
2459         cur_log = smalloc(sizeof(*cur_log));
2460         if (cur_log) {
2461                 INIT_FLIST_HEAD(&cur_log->list);
2462                 cur_log->log = malloc(new_size);
2463                 if (cur_log->log) {
2464                         cur_log->nr_samples = 0;
2465                         cur_log->max_samples = new_samples;
2466                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2467                         iolog->cur_log_max = new_samples;
2468                         return cur_log;
2469                 }
2470                 sfree(cur_log);
2471         }
2472
2473         return NULL;
2474 }
2475
2476 /*
2477  * Add and return a new log chunk, or return current log if big enough
2478  */
2479 static struct io_logs *regrow_log(struct io_log *iolog)
2480 {
2481         struct io_logs *cur_log;
2482         int i;
2483
2484         if (!iolog || iolog->disabled)
2485                 goto disable;
2486
2487         cur_log = iolog_cur_log(iolog);
2488         if (!cur_log) {
2489                 cur_log = get_new_log(iolog);
2490                 if (!cur_log)
2491                         return NULL;
2492         }
2493
2494         if (cur_log->nr_samples < cur_log->max_samples)
2495                 return cur_log;
2496
2497         /*
2498          * No room for a new sample. If we're compressing on the fly, flush
2499          * out the current chunk
2500          */
2501         if (iolog->log_gz) {
2502                 if (iolog_cur_flush(iolog, cur_log)) {
2503                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2504                         return NULL;
2505                 }
2506         }
2507
2508         /*
2509          * Get a new log array, and add to our list
2510          */
2511         cur_log = get_new_log(iolog);
2512         if (!cur_log) {
2513                 log_err("fio: failed extending iolog! Will stop logging.\n");
2514                 return NULL;
2515         }
2516
2517         if (!iolog->pending || !iolog->pending->nr_samples)
2518                 return cur_log;
2519
2520         /*
2521          * Flush pending items to new log
2522          */
2523         for (i = 0; i < iolog->pending->nr_samples; i++) {
2524                 struct io_sample *src, *dst;
2525
2526                 src = get_sample(iolog, iolog->pending, i);
2527                 dst = get_sample(iolog, cur_log, i);
2528                 memcpy(dst, src, log_entry_sz(iolog));
2529         }
2530         cur_log->nr_samples = iolog->pending->nr_samples;
2531
2532         iolog->pending->nr_samples = 0;
2533         return cur_log;
2534 disable:
2535         if (iolog)
2536                 iolog->disabled = true;
2537         return NULL;
2538 }
2539
2540 void regrow_logs(struct thread_data *td)
2541 {
2542         regrow_log(td->slat_log);
2543         regrow_log(td->clat_log);
2544         regrow_log(td->clat_hist_log);
2545         regrow_log(td->lat_log);
2546         regrow_log(td->bw_log);
2547         regrow_log(td->iops_log);
2548         td->flags &= ~TD_F_REGROW_LOGS;
2549 }
2550
2551 static struct io_logs *get_cur_log(struct io_log *iolog)
2552 {
2553         struct io_logs *cur_log;
2554
2555         cur_log = iolog_cur_log(iolog);
2556         if (!cur_log) {
2557                 cur_log = get_new_log(iolog);
2558                 if (!cur_log)
2559                         return NULL;
2560         }
2561
2562         if (cur_log->nr_samples < cur_log->max_samples)
2563                 return cur_log;
2564
2565         /*
2566          * Out of space. If we're in IO offload mode, or we're not doing
2567          * per unit logging (hence logging happens outside of the IO thread
2568          * as well), add a new log chunk inline. If we're doing inline
2569          * submissions, flag 'td' as needing a log regrow and we'll take
2570          * care of it on the submission side.
2571          */
2572         if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2573             !per_unit_log(iolog))
2574                 return regrow_log(iolog);
2575
2576         if (iolog->td)
2577                 iolog->td->flags |= TD_F_REGROW_LOGS;
2578         if (iolog->pending)
2579                 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2580         return iolog->pending;
2581 }
2582
2583 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2584                              enum fio_ddir ddir, unsigned long long bs,
2585                              unsigned long t, uint64_t offset, uint8_t priority_bit)
2586 {
2587         struct io_logs *cur_log;
2588
2589         if (iolog->disabled)
2590                 return;
2591         if (flist_empty(&iolog->io_logs))
2592                 iolog->avg_last[ddir] = t;
2593
2594         cur_log = get_cur_log(iolog);
2595         if (cur_log) {
2596                 struct io_sample *s;
2597
2598                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2599
2600                 s->data = data;
2601                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2602                 io_sample_set_ddir(iolog, s, ddir);
2603                 s->bs = bs;
2604                 s->priority_bit = priority_bit;
2605
2606                 if (iolog->log_offset) {
2607                         struct io_sample_offset *so = (void *) s;
2608
2609                         so->offset = offset;
2610                 }
2611
2612                 cur_log->nr_samples++;
2613                 return;
2614         }
2615
2616         iolog->disabled = true;
2617 }
2618
2619 static inline void reset_io_stat(struct io_stat *ios)
2620 {
2621         ios->min_val = -1ULL;
2622         ios->max_val = ios->samples = 0;
2623         ios->mean.u.f = ios->S.u.f = 0;
2624 }
2625
2626 void reset_io_stats(struct thread_data *td)
2627 {
2628         struct thread_stat *ts = &td->ts;
2629         int i, j, k;
2630
2631         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2632                 reset_io_stat(&ts->clat_high_prio_stat[i]);
2633                 reset_io_stat(&ts->clat_low_prio_stat[i]);
2634                 reset_io_stat(&ts->clat_stat[i]);
2635                 reset_io_stat(&ts->slat_stat[i]);
2636                 reset_io_stat(&ts->lat_stat[i]);
2637                 reset_io_stat(&ts->bw_stat[i]);
2638                 reset_io_stat(&ts->iops_stat[i]);
2639
2640                 ts->io_bytes[i] = 0;
2641                 ts->runtime[i] = 0;
2642                 ts->total_io_u[i] = 0;
2643                 ts->short_io_u[i] = 0;
2644                 ts->drop_io_u[i] = 0;
2645
2646                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2647                         ts->io_u_plat_high_prio[i][j] = 0;
2648                         ts->io_u_plat_low_prio[i][j] = 0;
2649                         if (!i)
2650                                 ts->io_u_sync_plat[j] = 0;
2651                 }
2652         }
2653
2654         for (i = 0; i < FIO_LAT_CNT; i++)
2655                 for (j = 0; j < DDIR_RWDIR_CNT; j++)
2656                         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2657                                 ts->io_u_plat[i][j][k] = 0;
2658
2659         ts->total_io_u[DDIR_SYNC] = 0;
2660
2661         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2662                 ts->io_u_map[i] = 0;
2663                 ts->io_u_submit[i] = 0;
2664                 ts->io_u_complete[i] = 0;
2665         }
2666
2667         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2668                 ts->io_u_lat_n[i] = 0;
2669         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2670                 ts->io_u_lat_u[i] = 0;
2671         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2672                 ts->io_u_lat_m[i] = 0;
2673
2674         ts->total_submit = 0;
2675         ts->total_complete = 0;
2676         ts->nr_zone_resets = 0;
2677         ts->cachehit = ts->cachemiss = 0;
2678 }
2679
2680 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2681                               unsigned long elapsed, bool log_max, uint8_t priority_bit)
2682 {
2683         /*
2684          * Note an entry in the log. Use the mean from the logged samples,
2685          * making sure to properly round up. Only write a log entry if we
2686          * had actual samples done.
2687          */
2688         if (iolog->avg_window[ddir].samples) {
2689                 union io_sample_data data;
2690
2691                 if (log_max)
2692                         data.val = iolog->avg_window[ddir].max_val;
2693                 else
2694                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2695
2696                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, priority_bit);
2697         }
2698
2699         reset_io_stat(&iolog->avg_window[ddir]);
2700 }
2701
2702 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2703                              bool log_max, uint8_t priority_bit)
2704 {
2705         int ddir;
2706
2707         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2708                 __add_stat_to_log(iolog, ddir, elapsed, log_max, priority_bit);
2709 }
2710
2711 static unsigned long add_log_sample(struct thread_data *td,
2712                                     struct io_log *iolog,
2713                                     union io_sample_data data,
2714                                     enum fio_ddir ddir, unsigned long long bs,
2715                                     uint64_t offset, uint8_t priority_bit)
2716 {
2717         unsigned long elapsed, this_window;
2718
2719         if (!ddir_rw(ddir))
2720                 return 0;
2721
2722         elapsed = mtime_since_now(&td->epoch);
2723
2724         /*
2725          * If no time averaging, just add the log sample.
2726          */
2727         if (!iolog->avg_msec) {
2728                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset, priority_bit);
2729                 return 0;
2730         }
2731
2732         /*
2733          * Add the sample. If the time period has passed, then
2734          * add that entry to the log and clear.
2735          */
2736         add_stat_sample(&iolog->avg_window[ddir], data.val);
2737
2738         /*
2739          * If period hasn't passed, adding the above sample is all we
2740          * need to do.
2741          */
2742         this_window = elapsed - iolog->avg_last[ddir];
2743         if (elapsed < iolog->avg_last[ddir])
2744                 return iolog->avg_last[ddir] - elapsed;
2745         else if (this_window < iolog->avg_msec) {
2746                 unsigned long diff = iolog->avg_msec - this_window;
2747
2748                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2749                         return diff;
2750         }
2751
2752         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0, priority_bit);
2753
2754         iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2755         return iolog->avg_msec;
2756 }
2757
2758 void finalize_logs(struct thread_data *td, bool unit_logs)
2759 {
2760         unsigned long elapsed;
2761
2762         elapsed = mtime_since_now(&td->epoch);
2763
2764         if (td->clat_log && unit_logs)
2765                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0, 0);
2766         if (td->slat_log && unit_logs)
2767                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0, 0);
2768         if (td->lat_log && unit_logs)
2769                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0, 0);
2770         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2771                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0, 0);
2772         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2773                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0, 0);
2774 }
2775
2776 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned long long bs,
2777                                         uint8_t priority_bit)
2778 {
2779         struct io_log *iolog;
2780
2781         if (!ddir_rw(ddir))
2782                 return;
2783
2784         iolog = agg_io_log[ddir];
2785         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, priority_bit);
2786 }
2787
2788 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2789 {
2790         unsigned int idx = plat_val_to_idx(nsec);
2791         assert(idx < FIO_IO_U_PLAT_NR);
2792
2793         ts->io_u_sync_plat[idx]++;
2794         add_stat_sample(&ts->sync_stat, nsec);
2795 }
2796
2797 static void add_lat_percentile_sample_noprio(struct thread_stat *ts,
2798                                 unsigned long long nsec, enum fio_ddir ddir, enum fio_lat lat)
2799 {
2800         unsigned int idx = plat_val_to_idx(nsec);
2801         assert(idx < FIO_IO_U_PLAT_NR);
2802
2803         ts->io_u_plat[lat][ddir][idx]++;
2804 }
2805
2806 static void add_lat_percentile_sample(struct thread_stat *ts,
2807                                 unsigned long long nsec, enum fio_ddir ddir, uint8_t priority_bit,
2808                                 enum fio_lat lat)
2809 {
2810         unsigned int idx = plat_val_to_idx(nsec);
2811
2812         add_lat_percentile_sample_noprio(ts, nsec, ddir, lat);
2813
2814         if (!priority_bit)
2815                 ts->io_u_plat_low_prio[ddir][idx]++;
2816         else
2817                 ts->io_u_plat_high_prio[ddir][idx]++;
2818 }
2819
2820 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2821                      unsigned long long nsec, unsigned long long bs,
2822                      uint64_t offset, uint8_t priority_bit)
2823 {
2824         const bool needs_lock = td_async_processing(td);
2825         unsigned long elapsed, this_window;
2826         struct thread_stat *ts = &td->ts;
2827         struct io_log *iolog = td->clat_hist_log;
2828
2829         if (needs_lock)
2830                 __td_io_u_lock(td);
2831
2832         add_stat_sample(&ts->clat_stat[ddir], nsec);
2833
2834         if (!ts->lat_percentiles) {
2835                 if (priority_bit)
2836                         add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
2837                 else
2838                         add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
2839         }
2840
2841         if (td->clat_log)
2842                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2843                                offset, priority_bit);
2844
2845         if (ts->clat_percentiles) {
2846                 if (ts->lat_percentiles)
2847                         add_lat_percentile_sample_noprio(ts, nsec, ddir, FIO_CLAT);
2848                 else
2849                         add_lat_percentile_sample(ts, nsec, ddir, priority_bit, FIO_CLAT);
2850         }
2851
2852         if (iolog && iolog->hist_msec) {
2853                 struct io_hist *hw = &iolog->hist_window[ddir];
2854
2855                 hw->samples++;
2856                 elapsed = mtime_since_now(&td->epoch);
2857                 if (!hw->hist_last)
2858                         hw->hist_last = elapsed;
2859                 this_window = elapsed - hw->hist_last;
2860
2861                 if (this_window >= iolog->hist_msec) {
2862                         uint64_t *io_u_plat;
2863                         struct io_u_plat_entry *dst;
2864
2865                         /*
2866                          * Make a byte-for-byte copy of the latency histogram
2867                          * stored in td->ts.io_u_plat[ddir], recording it in a
2868                          * log sample. Note that the matching call to free() is
2869                          * located in iolog.c after printing this sample to the
2870                          * log file.
2871                          */
2872                         io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
2873                         dst = malloc(sizeof(struct io_u_plat_entry));
2874                         memcpy(&(dst->io_u_plat), io_u_plat,
2875                                 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
2876                         flist_add(&dst->list, &hw->list);
2877                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2878                                                 elapsed, offset, priority_bit);
2879
2880                         /*
2881                          * Update the last time we recorded as being now, minus
2882                          * any drift in time we encountered before actually
2883                          * making the record.
2884                          */
2885                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2886                         hw->samples = 0;
2887                 }
2888         }
2889
2890         if (needs_lock)
2891                 __td_io_u_unlock(td);
2892 }
2893
2894 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2895                         unsigned long long nsec, unsigned long long bs, uint64_t offset,
2896                         uint8_t priority_bit)
2897 {
2898         const bool needs_lock = td_async_processing(td);
2899         struct thread_stat *ts = &td->ts;
2900
2901         if (!ddir_rw(ddir))
2902                 return;
2903
2904         if (needs_lock)
2905                 __td_io_u_lock(td);
2906
2907         add_stat_sample(&ts->slat_stat[ddir], nsec);
2908
2909         if (td->slat_log)
2910                 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs, offset,
2911                         priority_bit);
2912
2913         if (ts->slat_percentiles)
2914                 add_lat_percentile_sample_noprio(ts, nsec, ddir, FIO_SLAT);
2915
2916         if (needs_lock)
2917                 __td_io_u_unlock(td);
2918 }
2919
2920 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2921                     unsigned long long nsec, unsigned long long bs,
2922                     uint64_t offset, uint8_t priority_bit)
2923 {
2924         const bool needs_lock = td_async_processing(td);
2925         struct thread_stat *ts = &td->ts;
2926
2927         if (!ddir_rw(ddir))
2928                 return;
2929
2930         if (needs_lock)
2931                 __td_io_u_lock(td);
2932
2933         add_stat_sample(&ts->lat_stat[ddir], nsec);
2934
2935         if (td->lat_log)
2936                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2937                                offset, priority_bit);
2938
2939         if (ts->lat_percentiles) {
2940                 add_lat_percentile_sample(ts, nsec, ddir, priority_bit, FIO_LAT);
2941                 if (priority_bit)
2942                         add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
2943                 else
2944                         add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
2945
2946         }
2947         if (needs_lock)
2948                 __td_io_u_unlock(td);
2949 }
2950
2951 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2952                    unsigned int bytes, unsigned long long spent)
2953 {
2954         const bool needs_lock = td_async_processing(td);
2955         struct thread_stat *ts = &td->ts;
2956         unsigned long rate;
2957
2958         if (spent)
2959                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2960         else
2961                 rate = 0;
2962
2963         if (needs_lock)
2964                 __td_io_u_lock(td);
2965
2966         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2967
2968         if (td->bw_log)
2969                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2970                                bytes, io_u->offset, io_u_is_prio(io_u));
2971
2972         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2973
2974         if (needs_lock)
2975                 __td_io_u_unlock(td);
2976 }
2977
2978 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2979                          struct timespec *t, unsigned int avg_time,
2980                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2981                          struct io_stat *stat, struct io_log *log,
2982                          bool is_kb)
2983 {
2984         const bool needs_lock = td_async_processing(td);
2985         unsigned long spent, rate;
2986         enum fio_ddir ddir;
2987         unsigned long next, next_log;
2988
2989         next_log = avg_time;
2990
2991         spent = mtime_since(parent_tv, t);
2992         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2993                 return avg_time - spent;
2994
2995         if (needs_lock)
2996                 __td_io_u_lock(td);
2997
2998         /*
2999          * Compute both read and write rates for the interval.
3000          */
3001         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3002                 uint64_t delta;
3003
3004                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3005                 if (!delta)
3006                         continue; /* No entries for interval */
3007
3008                 if (spent) {
3009                         if (is_kb)
3010                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
3011                         else
3012                                 rate = (delta * 1000) / spent;
3013                 } else
3014                         rate = 0;
3015
3016                 add_stat_sample(&stat[ddir], rate);
3017
3018                 if (log) {
3019                         unsigned long long bs = 0;
3020
3021                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3022                                 bs = td->o.min_bs[ddir];
3023
3024                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0, 0);
3025                         next_log = min(next_log, next);
3026                 }
3027
3028                 stat_io_bytes[ddir] = this_io_bytes[ddir];
3029         }
3030
3031         *parent_tv = *t;
3032
3033         if (needs_lock)
3034                 __td_io_u_unlock(td);
3035
3036         if (spent <= avg_time)
3037                 next = avg_time;
3038         else
3039                 next = avg_time - (1 + spent - avg_time);
3040
3041         return min(next, next_log);
3042 }
3043
3044 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3045 {
3046         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3047                                 td->this_io_bytes, td->stat_io_bytes,
3048                                 td->ts.bw_stat, td->bw_log, true);
3049 }
3050
3051 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3052                      unsigned int bytes)
3053 {
3054         const bool needs_lock = td_async_processing(td);
3055         struct thread_stat *ts = &td->ts;
3056
3057         if (needs_lock)
3058                 __td_io_u_lock(td);
3059
3060         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3061
3062         if (td->iops_log)
3063                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3064                                bytes, io_u->offset, io_u_is_prio(io_u));
3065
3066         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3067
3068         if (needs_lock)
3069                 __td_io_u_unlock(td);
3070 }
3071
3072 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3073 {
3074         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3075                                 td->this_io_blocks, td->stat_io_blocks,
3076                                 td->ts.iops_stat, td->iops_log, false);
3077 }
3078
3079 /*
3080  * Returns msecs to next event
3081  */
3082 int calc_log_samples(void)
3083 {
3084         struct thread_data *td;
3085         unsigned int next = ~0U, tmp;
3086         struct timespec now;
3087         int i;
3088
3089         fio_gettime(&now, NULL);
3090
3091         for_each_td(td, i) {
3092                 if (!td->o.stats)
3093                         continue;
3094                 if (in_ramp_time(td) ||
3095                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3096                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3097                         continue;
3098                 }
3099                 if (!td->bw_log ||
3100                         (td->bw_log && !per_unit_log(td->bw_log))) {
3101                         tmp = add_bw_samples(td, &now);
3102                         if (tmp < next)
3103                                 next = tmp;
3104                 }
3105                 if (!td->iops_log ||
3106                         (td->iops_log && !per_unit_log(td->iops_log))) {
3107                         tmp = add_iops_samples(td, &now);
3108                         if (tmp < next)
3109                                 next = tmp;
3110                 }
3111         }
3112
3113         return next == ~0U ? 0 : next;
3114 }
3115
3116 void stat_init(void)
3117 {
3118         stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3119 }
3120
3121 void stat_exit(void)
3122 {
3123         /*
3124          * When we have the mutex, we know out-of-band access to it
3125          * have ended.
3126          */
3127         fio_sem_down(stat_sem);
3128         fio_sem_remove(stat_sem);
3129 }
3130
3131 /*
3132  * Called from signal handler. Wake up status thread.
3133  */
3134 void show_running_run_stats(void)
3135 {
3136         helper_do_stat();
3137 }
3138
3139 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3140 {
3141         /* Ignore io_u's which span multiple blocks--they will just get
3142          * inaccurate counts. */
3143         int idx = (io_u->offset - io_u->file->file_offset)
3144                         / td->o.bs[DDIR_TRIM];
3145         uint32_t *info = &td->ts.block_infos[idx];
3146         assert(idx < td->ts.nr_block_infos);
3147         return info;
3148 }