beec574d8bd4e80767e683f5631e838d6df853d1
[fio.git] / stat.c
1 #include <stdio.h>
2 #include <string.h>
3 #include <sys/time.h>
4 #include <sys/types.h>
5 #include <sys/stat.h>
6 #include <dirent.h>
7 #include <libgen.h>
8 #include <math.h>
9
10 #include "fio.h"
11 #include "diskutil.h"
12 #include "lib/ieee754.h"
13 #include "json.h"
14 #include "lib/getrusage.h"
15 #include "idletime.h"
16 #include "lib/pow2.h"
17 #include "lib/output_buffer.h"
18 #include "helper_thread.h"
19 #include "smalloc.h"
20
21 #define LOG_MSEC_SLACK  10
22
23 struct fio_mutex *stat_mutex;
24
25 void clear_rusage_stat(struct thread_data *td)
26 {
27         struct thread_stat *ts = &td->ts;
28
29         fio_getrusage(&td->ru_start);
30         ts->usr_time = ts->sys_time = 0;
31         ts->ctx = 0;
32         ts->minf = ts->majf = 0;
33 }
34
35 void update_rusage_stat(struct thread_data *td)
36 {
37         struct thread_stat *ts = &td->ts;
38
39         fio_getrusage(&td->ru_end);
40         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
41                                         &td->ru_end.ru_utime);
42         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
43                                         &td->ru_end.ru_stime);
44         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
45                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
46         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
47         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
48
49         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
50 }
51
52 /*
53  * Given a latency, return the index of the corresponding bucket in
54  * the structure tracking percentiles.
55  *
56  * (1) find the group (and error bits) that the value (latency)
57  * belongs to by looking at its MSB. (2) find the bucket number in the
58  * group by looking at the index bits.
59  *
60  */
61 static unsigned int plat_val_to_idx(unsigned long long val)
62 {
63         unsigned int msb, error_bits, base, offset, idx;
64
65         /* Find MSB starting from bit 0 */
66         if (val == 0)
67                 msb = 0;
68         else
69                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
70
71         /*
72          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
73          * all bits of the sample as index
74          */
75         if (msb <= FIO_IO_U_PLAT_BITS)
76                 return val;
77
78         /* Compute the number of error bits to discard*/
79         error_bits = msb - FIO_IO_U_PLAT_BITS;
80
81         /* Compute the number of buckets before the group */
82         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
83
84         /*
85          * Discard the error bits and apply the mask to find the
86          * index for the buckets in the group
87          */
88         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
89
90         /* Make sure the index does not exceed (array size - 1) */
91         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
92                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
93
94         return idx;
95 }
96
97 /*
98  * Convert the given index of the bucket array to the value
99  * represented by the bucket
100  */
101 static unsigned long long plat_idx_to_val(unsigned int idx)
102 {
103         unsigned int error_bits, 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 = 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(unsigned int *io_u_plat, unsigned long nr,
138                                    fio_fp64_t *plist, unsigned long long **output,
139                                    unsigned long long *maxv, unsigned long long *minv)
140 {
141         unsigned long sum = 0;
142         unsigned int len, i, j = 0;
143         unsigned int oval_len = 0;
144         unsigned long long *ovals = NULL;
145         int is_last;
146
147         *minv = -1ULL;
148         *maxv = 0;
149
150         len = 0;
151         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
152                 len++;
153
154         if (!len)
155                 return 0;
156
157         /*
158          * Sort the percentile list. Note that it may already be sorted if
159          * we are using the default values, but since it's a short list this
160          * isn't a worry. Also note that this does not work for NaN values.
161          */
162         if (len > 1)
163                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
164
165         /*
166          * Calculate bucket values, note down max and min values
167          */
168         is_last = 0;
169         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
170                 sum += io_u_plat[i];
171                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
172                         assert(plist[j].u.f <= 100.0);
173
174                         if (j == oval_len) {
175                                 oval_len += 100;
176                                 ovals = realloc(ovals, oval_len * sizeof(*ovals));
177                         }
178
179                         ovals[j] = plat_idx_to_val(i);
180                         if (ovals[j] < *minv)
181                                 *minv = ovals[j];
182                         if (ovals[j] > *maxv)
183                                 *maxv = ovals[j];
184
185                         is_last = (j == len - 1);
186                         if (is_last)
187                                 break;
188
189                         j++;
190                 }
191         }
192
193         *output = ovals;
194         return len;
195 }
196
197 /*
198  * Find and display the p-th percentile of clat
199  */
200 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
201                                   fio_fp64_t *plist, unsigned int precision,
202                                   struct buf_output *out)
203 {
204         unsigned int divisor, len, i, j = 0;
205         unsigned long long minv, maxv;
206         unsigned long long *ovals;
207         int is_last, per_line, scale_down, time_width;
208         char fmt[32];
209
210         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
211         if (!len)
212                 goto out;
213
214         /*
215          * We default to nsecs, but if the value range is such that we
216          * should scale down to usecs or msecs, do that.
217          */
218         if (minv > 2000000 && maxv > 99999999ULL) {
219                 scale_down = 2;
220                 divisor = 1000000;
221                 log_buf(out, "    clat percentiles (msec):\n     |");
222         } else if (minv > 2000 && maxv > 99999) {
223                 scale_down = 1;
224                 divisor = 1000;
225                 log_buf(out, "    clat percentiles (usec):\n     |");
226         } else {
227                 scale_down = 0;
228                 divisor = 1;
229                 log_buf(out, "    clat percentiles (nsec):\n     |");
230         }
231
232
233         time_width = max(5, (int) (log10(maxv / divisor) + 1));
234         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
235                         precision, time_width);
236         /* fmt will be something like " %5.2fth=[%4llu]%c" */
237         per_line = (80 - 7) / (precision + 10 + time_width);
238
239         for (j = 0; j < len; j++) {
240                 /* for formatting */
241                 if (j != 0 && (j % per_line) == 0)
242                         log_buf(out, "     |");
243
244                 /* end of the list */
245                 is_last = (j == len - 1);
246
247                 for (i = 0; i < scale_down; i++)
248                         ovals[j] = (ovals[j] + 999) / 1000;
249
250                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
251
252                 if (is_last)
253                         break;
254
255                 if ((j % per_line) == per_line - 1)     /* for formatting */
256                         log_buf(out, "\n");
257         }
258
259 out:
260         if (ovals)
261                 free(ovals);
262 }
263
264 bool calc_lat(struct io_stat *is, unsigned long long *min,
265               unsigned long long *max, double *mean, double *dev)
266 {
267         double n = (double) is->samples;
268
269         if (n == 0)
270                 return false;
271
272         *min = is->min_val;
273         *max = is->max_val;
274         *mean = is->mean.u.f;
275
276         if (n > 1.0)
277                 *dev = sqrt(is->S.u.f / (n - 1.0));
278         else
279                 *dev = 0;
280
281         return true;
282 }
283
284 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
285 {
286         char *io, *agg, *min, *max;
287         char *ioalt, *aggalt, *minalt, *maxalt;
288         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
289         int i;
290
291         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
292
293         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
294                 const int i2p = is_power_of_2(rs->kb_base);
295
296                 if (!rs->max_run[i])
297                         continue;
298
299                 io = num2str(rs->iobytes[i], 4, 1, i2p, N2S_BYTE);
300                 ioalt = num2str(rs->iobytes[i], 4, 1, !i2p, N2S_BYTE);
301                 agg = num2str(rs->agg[i], 4, 1, i2p, rs->unit_base);
302                 aggalt = num2str(rs->agg[i], 4, 1, !i2p, rs->unit_base);
303                 min = num2str(rs->min_bw[i], 4, 1, i2p, rs->unit_base);
304                 minalt = num2str(rs->min_bw[i], 4, 1, !i2p, rs->unit_base);
305                 max = num2str(rs->max_bw[i], 4, 1, i2p, rs->unit_base);
306                 maxalt = num2str(rs->max_bw[i], 4, 1, !i2p, rs->unit_base);
307                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
308                                 rs->unified_rw_rep ? "  MIXED" : str[i],
309                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
310                                 (unsigned long long) rs->min_run[i],
311                                 (unsigned long long) rs->max_run[i]);
312
313                 free(io);
314                 free(agg);
315                 free(min);
316                 free(max);
317                 free(ioalt);
318                 free(aggalt);
319                 free(minalt);
320                 free(maxalt);
321         }
322 }
323
324 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
325 {
326         int i;
327
328         /*
329          * Do depth distribution calculations
330          */
331         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
332                 if (total) {
333                         io_u_dist[i] = (double) map[i] / (double) total;
334                         io_u_dist[i] *= 100.0;
335                         if (io_u_dist[i] < 0.1 && map[i])
336                                 io_u_dist[i] = 0.1;
337                 } else
338                         io_u_dist[i] = 0.0;
339         }
340 }
341
342 static void stat_calc_lat(struct thread_stat *ts, double *dst,
343                           unsigned int *src, int nr)
344 {
345         unsigned long total = ddir_rw_sum(ts->total_io_u);
346         int i;
347
348         /*
349          * Do latency distribution calculations
350          */
351         for (i = 0; i < nr; i++) {
352                 if (total) {
353                         dst[i] = (double) src[i] / (double) total;
354                         dst[i] *= 100.0;
355                         if (dst[i] < 0.01 && src[i])
356                                 dst[i] = 0.01;
357                 } else
358                         dst[i] = 0.0;
359         }
360 }
361
362 /*
363  * To keep the terse format unaltered, add all of the ns latency
364  * buckets to the first us latency bucket
365  */
366 void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
367 {
368         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
369         int i;
370
371         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
372
373         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
374                 ntotal += ts->io_u_lat_n[i];
375
376         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
377 }
378
379 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
380 {
381         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
382 }
383
384 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
385 {
386         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
387 }
388
389 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
390 {
391         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
392 }
393
394 static void display_lat(const char *name, unsigned long long min,
395                         unsigned long long max, double mean, double dev,
396                         struct buf_output *out)
397 {
398         const char *base = "(nsec)";
399         char *minp, *maxp;
400
401         if (nsec_to_msec(&min, &max, &mean, &dev))
402                 base = "(msec)";
403         else if (nsec_to_usec(&min, &max, &mean, &dev))
404                 base = "(usec)";
405
406         minp = num2str(min, 6, 1, 0, N2S_NONE);
407         maxp = num2str(max, 6, 1, 0, N2S_NONE);
408
409         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
410                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
411
412         free(minp);
413         free(maxp);
414 }
415
416 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
417                              int ddir, struct buf_output *out)
418 {
419         const char *str[] = { " read", "write", " trim" };
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;
424         int i2p;
425
426         assert(ddir_rw(ddir));
427
428         if (!ts->runtime[ddir])
429                 return;
430
431         i2p = is_power_of_2(rs->kb_base);
432         runt = ts->runtime[ddir];
433
434         bw = (1000 * ts->io_bytes[ddir]) / runt;
435         io_p = num2str(ts->io_bytes[ddir], 4, 1, i2p, N2S_BYTE);
436         bw_p = num2str(bw, 4, 1, i2p, ts->unit_base);
437         bw_p_alt = num2str(bw, 4, 1, !i2p, ts->unit_base);
438
439         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
440         iops_p = num2str(iops, 4, 1, 0, N2S_NONE);
441
442         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
443                         rs->unified_rw_rep ? "mixed" : str[ddir],
444                         iops_p, bw_p, bw_p_alt, io_p,
445                         (unsigned long long) ts->runtime[ddir]);
446
447         free(io_p);
448         free(bw_p);
449         free(bw_p_alt);
450         free(iops_p);
451
452         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
453                 display_lat("slat", min, max, mean, dev, out);
454         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
455                 display_lat("clat", min, max, mean, dev, out);
456         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
457                 display_lat(" lat", min, max, mean, dev, out);
458
459         if (ts->clat_percentiles) {
460                 show_clat_percentiles(ts->io_u_plat[ddir],
461                                         ts->clat_stat[ddir].samples,
462                                         ts->percentile_list,
463                                         ts->percentile_precision, out);
464         }
465         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
466                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
467                 const char *bw_str;
468
469                 if ((rs->unit_base == 1) && i2p)
470                         bw_str = "Kibit";
471                 else if (rs->unit_base == 1)
472                         bw_str = "kbit";
473                 else if (i2p)
474                         bw_str = "KiB";
475                 else
476                         bw_str = "kB";
477
478                 if (rs->unit_base == 1) {
479                         min *= 8.0;
480                         max *= 8.0;
481                         mean *= 8.0;
482                         dev *= 8.0;
483                 }
484
485                 if (rs->agg[ddir]) {
486                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
487                         if (p_of_agg > 100.0)
488                                 p_of_agg = 100.0;
489                 }
490
491                 if (mean > fkb_base * fkb_base) {
492                         min /= fkb_base;
493                         max /= fkb_base;
494                         mean /= fkb_base;
495                         dev /= fkb_base;
496                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
497                 }
498
499                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
500                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
501                         bw_str, min, max, p_of_agg, mean, dev,
502                         (&ts->bw_stat[ddir])->samples);
503         }
504         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
505                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
506                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
507                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
508         }
509 }
510
511 static int show_lat(double *io_u_lat, int nr, const char **ranges,
512                     const char *msg, struct buf_output *out)
513 {
514         int new_line = 1, i, line = 0, shown = 0;
515
516         for (i = 0; i < nr; i++) {
517                 if (io_u_lat[i] <= 0.0)
518                         continue;
519                 shown = 1;
520                 if (new_line) {
521                         if (line)
522                                 log_buf(out, "\n");
523                         log_buf(out, "    lat (%s) : ", msg);
524                         new_line = 0;
525                         line = 0;
526                 }
527                 if (line)
528                         log_buf(out, ", ");
529                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
530                 line++;
531                 if (line == 5)
532                         new_line = 1;
533         }
534
535         if (shown)
536                 log_buf(out, "\n");
537
538         return shown;
539 }
540
541 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
542 {
543         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
544                                  "250=", "500=", "750=", "1000=", };
545
546         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
547 }
548
549 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
550 {
551         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
552                                  "250=", "500=", "750=", "1000=", };
553
554         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
555 }
556
557 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
558 {
559         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
560                                  "250=", "500=", "750=", "1000=", "2000=",
561                                  ">=2000=", };
562
563         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
564 }
565
566 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
567 {
568         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
569         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
570         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
571
572         stat_calc_lat_n(ts, io_u_lat_n);
573         stat_calc_lat_u(ts, io_u_lat_u);
574         stat_calc_lat_m(ts, io_u_lat_m);
575
576         show_lat_n(io_u_lat_n, out);
577         show_lat_u(io_u_lat_u, out);
578         show_lat_m(io_u_lat_m, out);
579 }
580
581 static int block_state_category(int block_state)
582 {
583         switch (block_state) {
584         case BLOCK_STATE_UNINIT:
585                 return 0;
586         case BLOCK_STATE_TRIMMED:
587         case BLOCK_STATE_WRITTEN:
588                 return 1;
589         case BLOCK_STATE_WRITE_FAILURE:
590         case BLOCK_STATE_TRIM_FAILURE:
591                 return 2;
592         default:
593                 /* Silence compile warning on some BSDs and have a return */
594                 assert(0);
595                 return -1;
596         }
597 }
598
599 static int compare_block_infos(const void *bs1, const void *bs2)
600 {
601         uint32_t block1 = *(uint32_t *)bs1;
602         uint32_t block2 = *(uint32_t *)bs2;
603         int state1 = BLOCK_INFO_STATE(block1);
604         int state2 = BLOCK_INFO_STATE(block2);
605         int bscat1 = block_state_category(state1);
606         int bscat2 = block_state_category(state2);
607         int cycles1 = BLOCK_INFO_TRIMS(block1);
608         int cycles2 = BLOCK_INFO_TRIMS(block2);
609
610         if (bscat1 < bscat2)
611                 return -1;
612         if (bscat1 > bscat2)
613                 return 1;
614
615         if (cycles1 < cycles2)
616                 return -1;
617         if (cycles1 > cycles2)
618                 return 1;
619
620         if (state1 < state2)
621                 return -1;
622         if (state1 > state2)
623                 return 1;
624
625         assert(block1 == block2);
626         return 0;
627 }
628
629 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
630                                   fio_fp64_t *plist, unsigned int **percentiles,
631                                   unsigned int *types)
632 {
633         int len = 0;
634         int i, nr_uninit;
635
636         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
637
638         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
639                 len++;
640
641         if (!len)
642                 return 0;
643
644         /*
645          * Sort the percentile list. Note that it may already be sorted if
646          * we are using the default values, but since it's a short list this
647          * isn't a worry. Also note that this does not work for NaN values.
648          */
649         if (len > 1)
650                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
651
652         nr_uninit = 0;
653         /* Start only after the uninit entries end */
654         for (nr_uninit = 0;
655              nr_uninit < nr_block_infos
656                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
657              nr_uninit ++)
658                 ;
659
660         if (nr_uninit == nr_block_infos)
661                 return 0;
662
663         *percentiles = calloc(len, sizeof(**percentiles));
664
665         for (i = 0; i < len; i++) {
666                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
667                                 + nr_uninit;
668                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
669         }
670
671         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
672         for (i = 0; i < nr_block_infos; i++)
673                 types[BLOCK_INFO_STATE(block_infos[i])]++;
674
675         return len;
676 }
677
678 static const char *block_state_names[] = {
679         [BLOCK_STATE_UNINIT] = "unwritten",
680         [BLOCK_STATE_TRIMMED] = "trimmed",
681         [BLOCK_STATE_WRITTEN] = "written",
682         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
683         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
684 };
685
686 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
687                              fio_fp64_t *plist, struct buf_output *out)
688 {
689         int len, pos, i;
690         unsigned int *percentiles = NULL;
691         unsigned int block_state_counts[BLOCK_STATE_COUNT];
692
693         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
694                                      &percentiles, block_state_counts);
695
696         log_buf(out, "  block lifetime percentiles :\n   |");
697         pos = 0;
698         for (i = 0; i < len; i++) {
699                 uint32_t block_info = percentiles[i];
700 #define LINE_LENGTH     75
701                 char str[LINE_LENGTH];
702                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
703                                      plist[i].u.f, block_info,
704                                      i == len - 1 ? '\n' : ',');
705                 assert(strln < LINE_LENGTH);
706                 if (pos + strln > LINE_LENGTH) {
707                         pos = 0;
708                         log_buf(out, "\n   |");
709                 }
710                 log_buf(out, "%s", str);
711                 pos += strln;
712 #undef LINE_LENGTH
713         }
714         if (percentiles)
715                 free(percentiles);
716
717         log_buf(out, "        states               :");
718         for (i = 0; i < BLOCK_STATE_COUNT; i++)
719                 log_buf(out, " %s=%u%c",
720                          block_state_names[i], block_state_counts[i],
721                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
722 }
723
724 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
725 {
726         char *p1, *p1alt, *p2;
727         unsigned long long bw_mean, iops_mean;
728         const int i2p = is_power_of_2(ts->kb_base);
729
730         if (!ts->ss_dur)
731                 return;
732
733         bw_mean = steadystate_bw_mean(ts);
734         iops_mean = steadystate_iops_mean(ts);
735
736         p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
737         p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
738         p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
739
740         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
741                 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
742                 p1, p1alt, p2,
743                 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
744                 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
745                 ts->ss_criterion.u.f,
746                 ts->ss_state & __FIO_SS_PCT ? "%" : "");
747
748         free(p1);
749         free(p1alt);
750         free(p2);
751 }
752
753 static void show_thread_status_normal(struct thread_stat *ts,
754                                       struct group_run_stats *rs,
755                                       struct buf_output *out)
756 {
757         double usr_cpu, sys_cpu;
758         unsigned long runtime;
759         double io_u_dist[FIO_IO_U_MAP_NR];
760         time_t time_p;
761         char time_buf[32];
762
763         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
764                 return;
765                 
766         memset(time_buf, 0, sizeof(time_buf));
767
768         time(&time_p);
769         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
770
771         if (!ts->error) {
772                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
773                                         ts->name, ts->groupid, ts->members,
774                                         ts->error, (int) ts->pid, time_buf);
775         } else {
776                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
777                                         ts->name, ts->groupid, ts->members,
778                                         ts->error, ts->verror, (int) ts->pid,
779                                         time_buf);
780         }
781
782         if (strlen(ts->description))
783                 log_buf(out, "  Description  : [%s]\n", ts->description);
784
785         if (ts->io_bytes[DDIR_READ])
786                 show_ddir_status(rs, ts, DDIR_READ, out);
787         if (ts->io_bytes[DDIR_WRITE])
788                 show_ddir_status(rs, ts, DDIR_WRITE, out);
789         if (ts->io_bytes[DDIR_TRIM])
790                 show_ddir_status(rs, ts, DDIR_TRIM, out);
791
792         show_latencies(ts, out);
793
794         runtime = ts->total_run_time;
795         if (runtime) {
796                 double runt = (double) runtime;
797
798                 usr_cpu = (double) ts->usr_time * 100 / runt;
799                 sys_cpu = (double) ts->sys_time * 100 / runt;
800         } else {
801                 usr_cpu = 0;
802                 sys_cpu = 0;
803         }
804
805         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
806                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
807                         (unsigned long long) ts->ctx,
808                         (unsigned long long) ts->majf,
809                         (unsigned long long) ts->minf);
810
811         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
812         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
813                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
814                                         io_u_dist[1], io_u_dist[2],
815                                         io_u_dist[3], io_u_dist[4],
816                                         io_u_dist[5], io_u_dist[6]);
817
818         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
819         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
820                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
821                                         io_u_dist[1], io_u_dist[2],
822                                         io_u_dist[3], io_u_dist[4],
823                                         io_u_dist[5], io_u_dist[6]);
824         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
825         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
826                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
827                                         io_u_dist[1], io_u_dist[2],
828                                         io_u_dist[3], io_u_dist[4],
829                                         io_u_dist[5], io_u_dist[6]);
830         log_buf(out, "     issued rwt: total=%llu,%llu,%llu,"
831                                  " short=%llu,%llu,%llu,"
832                                  " dropped=%llu,%llu,%llu\n",
833                                         (unsigned long long) ts->total_io_u[0],
834                                         (unsigned long long) ts->total_io_u[1],
835                                         (unsigned long long) ts->total_io_u[2],
836                                         (unsigned long long) ts->short_io_u[0],
837                                         (unsigned long long) ts->short_io_u[1],
838                                         (unsigned long long) ts->short_io_u[2],
839                                         (unsigned long long) ts->drop_io_u[0],
840                                         (unsigned long long) ts->drop_io_u[1],
841                                         (unsigned long long) ts->drop_io_u[2]);
842         if (ts->continue_on_error) {
843                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
844                                         (unsigned long long)ts->total_err_count,
845                                         ts->first_error,
846                                         strerror(ts->first_error));
847         }
848         if (ts->latency_depth) {
849                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
850                                         (unsigned long long)ts->latency_target,
851                                         (unsigned long long)ts->latency_window,
852                                         ts->latency_percentile.u.f,
853                                         ts->latency_depth);
854         }
855
856         if (ts->nr_block_infos)
857                 show_block_infos(ts->nr_block_infos, ts->block_infos,
858                                   ts->percentile_list, out);
859
860         if (ts->ss_dur)
861                 show_ss_normal(ts, out);
862 }
863
864 static void show_ddir_status_terse(struct thread_stat *ts,
865                                    struct group_run_stats *rs, int ddir,
866                                    int ver, struct buf_output *out)
867 {
868         unsigned long long min, max, minv, maxv, bw, iops;
869         unsigned long long *ovals = NULL;
870         double mean, dev;
871         unsigned int len;
872         int i, bw_stat;
873
874         assert(ddir_rw(ddir));
875
876         iops = bw = 0;
877         if (ts->runtime[ddir]) {
878                 uint64_t runt = ts->runtime[ddir];
879
880                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
881                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
882         }
883
884         log_buf(out, ";%llu;%llu;%llu;%llu",
885                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
886                                         (unsigned long long) ts->runtime[ddir]);
887
888         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
889                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
890         else
891                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
892
893         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
894                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
895         else
896                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
897
898         if (ts->clat_percentiles) {
899                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
900                                         ts->clat_stat[ddir].samples,
901                                         ts->percentile_list, &ovals, &maxv,
902                                         &minv);
903         } else
904                 len = 0;
905
906         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
907                 if (i >= len) {
908                         log_buf(out, ";0%%=0");
909                         continue;
910                 }
911                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
912         }
913
914         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
915                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
916         else
917                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
918
919         if (ovals)
920                 free(ovals);
921
922         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
923         if (bw_stat) {
924                 double p_of_agg = 100.0;
925
926                 if (rs->agg[ddir]) {
927                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
928                         if (p_of_agg > 100.0)
929                                 p_of_agg = 100.0;
930                 }
931
932                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
933         } else
934                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
935
936         if (ver == 5) {
937                 if (bw_stat)
938                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
939                 else
940                         log_buf(out, ";%lu", 0UL);
941
942                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
943                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
944                                 mean, dev, (&ts->iops_stat[ddir])->samples);
945                 else
946                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
947         }
948 }
949
950 static void add_ddir_status_json(struct thread_stat *ts,
951                 struct group_run_stats *rs, int ddir, struct json_object *parent)
952 {
953         unsigned long long min, max, minv, maxv;
954         unsigned long long bw;
955         unsigned long long *ovals = NULL;
956         double mean, dev, iops;
957         unsigned int len;
958         int i;
959         const char *ddirname[] = {"read", "write", "trim"};
960         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
961         char buf[120];
962         double p_of_agg = 100.0;
963
964         assert(ddir_rw(ddir));
965
966         if (ts->unified_rw_rep && ddir != DDIR_READ)
967                 return;
968
969         dir_object = json_create_object();
970         json_object_add_value_object(parent,
971                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
972
973         bw = 0;
974         iops = 0.0;
975         if (ts->runtime[ddir]) {
976                 uint64_t runt = ts->runtime[ddir];
977
978                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
979                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
980         }
981
982         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
983         json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
984         json_object_add_value_int(dir_object, "bw", bw);
985         json_object_add_value_float(dir_object, "iops", iops);
986         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
987         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
988         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
989         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
990
991         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
992                 min = max = 0;
993                 mean = dev = 0.0;
994         }
995         tmp_object = json_create_object();
996         json_object_add_value_object(dir_object, "slat_ns", tmp_object);
997         json_object_add_value_int(tmp_object, "min", min);
998         json_object_add_value_int(tmp_object, "max", max);
999         json_object_add_value_float(tmp_object, "mean", mean);
1000         json_object_add_value_float(tmp_object, "stddev", dev);
1001
1002         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1003                 min = max = 0;
1004                 mean = dev = 0.0;
1005         }
1006         tmp_object = json_create_object();
1007         json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1008         json_object_add_value_int(tmp_object, "min", min);
1009         json_object_add_value_int(tmp_object, "max", max);
1010         json_object_add_value_float(tmp_object, "mean", mean);
1011         json_object_add_value_float(tmp_object, "stddev", dev);
1012
1013         if (ts->clat_percentiles) {
1014                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
1015                                         ts->clat_stat[ddir].samples,
1016                                         ts->percentile_list, &ovals, &maxv,
1017                                         &minv);
1018         } else
1019                 len = 0;
1020
1021         percentile_object = json_create_object();
1022         json_object_add_value_object(tmp_object, "percentile", percentile_object);
1023         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1024                 if (i >= len) {
1025                         json_object_add_value_int(percentile_object, "0.00", 0);
1026                         continue;
1027                 }
1028                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1029                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1030         }
1031
1032         if (output_format & FIO_OUTPUT_JSON_PLUS) {
1033                 clat_bins_object = json_create_object();
1034                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1035                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1036                         if (ts->io_u_plat[ddir][i]) {
1037                                 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1038                                 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1039                         }
1040                 }
1041         }
1042
1043         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1044                 min = max = 0;
1045                 mean = dev = 0.0;
1046         }
1047         tmp_object = json_create_object();
1048         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1049         json_object_add_value_int(tmp_object, "min", min);
1050         json_object_add_value_int(tmp_object, "max", max);
1051         json_object_add_value_float(tmp_object, "mean", mean);
1052         json_object_add_value_float(tmp_object, "stddev", dev);
1053         if (ovals)
1054                 free(ovals);
1055
1056         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1057                 if (rs->agg[ddir]) {
1058                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
1059                         if (p_of_agg > 100.0)
1060                                 p_of_agg = 100.0;
1061                 }
1062         } else {
1063                 min = max = 0;
1064                 p_of_agg = mean = dev = 0.0;
1065         }
1066         json_object_add_value_int(dir_object, "bw_min", min);
1067         json_object_add_value_int(dir_object, "bw_max", max);
1068         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1069         json_object_add_value_float(dir_object, "bw_mean", mean);
1070         json_object_add_value_float(dir_object, "bw_dev", dev);
1071         json_object_add_value_int(dir_object, "bw_samples",
1072                                 (&ts->bw_stat[ddir])->samples);
1073
1074         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1075                 min = max = 0;
1076                 mean = dev = 0.0;
1077         }
1078         json_object_add_value_int(dir_object, "iops_min", min);
1079         json_object_add_value_int(dir_object, "iops_max", max);
1080         json_object_add_value_float(dir_object, "iops_mean", mean);
1081         json_object_add_value_float(dir_object, "iops_stddev", dev);
1082         json_object_add_value_int(dir_object, "iops_samples",
1083                                 (&ts->iops_stat[ddir])->samples);
1084 }
1085
1086 static void show_thread_status_terse_all(struct thread_stat *ts,
1087                                          struct group_run_stats *rs, int ver,
1088                                          struct buf_output *out)
1089 {
1090         double io_u_dist[FIO_IO_U_MAP_NR];
1091         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1092         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1093         double usr_cpu, sys_cpu;
1094         int i;
1095
1096         /* General Info */
1097         if (ver == 2)
1098                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1099         else
1100                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1101                         ts->name, ts->groupid, ts->error);
1102
1103         /* Log Read Status */
1104         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1105         /* Log Write Status */
1106         show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1107         /* Log Trim Status */
1108         if (ver == 2 || ver == 4 || ver == 5)
1109                 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1110
1111         /* CPU Usage */
1112         if (ts->total_run_time) {
1113                 double runt = (double) ts->total_run_time;
1114
1115                 usr_cpu = (double) ts->usr_time * 100 / runt;
1116                 sys_cpu = (double) ts->sys_time * 100 / runt;
1117         } else {
1118                 usr_cpu = 0;
1119                 sys_cpu = 0;
1120         }
1121
1122         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1123                                                 (unsigned long long) ts->ctx,
1124                                                 (unsigned long long) ts->majf,
1125                                                 (unsigned long long) ts->minf);
1126
1127         /* Calc % distribution of IO depths, usecond, msecond latency */
1128         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1129         stat_calc_lat_nu(ts, io_u_lat_u);
1130         stat_calc_lat_m(ts, io_u_lat_m);
1131
1132         /* Only show fixed 7 I/O depth levels*/
1133         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1134                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1135                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1136
1137         /* Microsecond latency */
1138         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1139                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1140         /* Millisecond latency */
1141         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1142                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1143
1144         /* disk util stats, if any */
1145         if (ver >= 3)
1146                 show_disk_util(1, NULL, out);
1147
1148         /* Additional output if continue_on_error set - default off*/
1149         if (ts->continue_on_error)
1150                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1151         if (ver == 2)
1152                 log_buf(out, "\n");
1153
1154         /* Additional output if description is set */
1155         if (strlen(ts->description))
1156                 log_buf(out, ";%s", ts->description);
1157
1158         log_buf(out, "\n");
1159 }
1160
1161 static void json_add_job_opts(struct json_object *root, const char *name,
1162                               struct flist_head *opt_list, bool num_jobs)
1163 {
1164         struct json_object *dir_object;
1165         struct flist_head *entry;
1166         struct print_option *p;
1167
1168         if (flist_empty(opt_list))
1169                 return;
1170
1171         dir_object = json_create_object();
1172         json_object_add_value_object(root, name, dir_object);
1173
1174         flist_for_each(entry, opt_list) {
1175                 const char *pos = "";
1176
1177                 p = flist_entry(entry, struct print_option, list);
1178                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1179                         continue;
1180                 if (p->value)
1181                         pos = p->value;
1182                 json_object_add_value_string(dir_object, p->name, pos);
1183         }
1184 }
1185
1186 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1187                                                    struct group_run_stats *rs,
1188                                                    struct flist_head *opt_list)
1189 {
1190         struct json_object *root, *tmp;
1191         struct jobs_eta *je;
1192         double io_u_dist[FIO_IO_U_MAP_NR];
1193         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1194         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1195         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1196         double usr_cpu, sys_cpu;
1197         int i;
1198         size_t size;
1199
1200         root = json_create_object();
1201         json_object_add_value_string(root, "jobname", ts->name);
1202         json_object_add_value_int(root, "groupid", ts->groupid);
1203         json_object_add_value_int(root, "error", ts->error);
1204
1205         /* ETA Info */
1206         je = get_jobs_eta(true, &size);
1207         if (je) {
1208                 json_object_add_value_int(root, "eta", je->eta_sec);
1209                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1210         }
1211
1212         if (opt_list)
1213                 json_add_job_opts(root, "job options", opt_list, true);
1214
1215         add_ddir_status_json(ts, rs, DDIR_READ, root);
1216         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1217         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1218
1219         /* CPU Usage */
1220         if (ts->total_run_time) {
1221                 double runt = (double) ts->total_run_time;
1222
1223                 usr_cpu = (double) ts->usr_time * 100 / runt;
1224                 sys_cpu = (double) ts->sys_time * 100 / runt;
1225         } else {
1226                 usr_cpu = 0;
1227                 sys_cpu = 0;
1228         }
1229         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1230         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1231         json_object_add_value_int(root, "ctx", ts->ctx);
1232         json_object_add_value_int(root, "majf", ts->majf);
1233         json_object_add_value_int(root, "minf", ts->minf);
1234
1235
1236         /* Calc % distribution of IO depths, usecond, msecond latency */
1237         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1238         stat_calc_lat_n(ts, io_u_lat_n);
1239         stat_calc_lat_u(ts, io_u_lat_u);
1240         stat_calc_lat_m(ts, io_u_lat_m);
1241
1242         tmp = json_create_object();
1243         json_object_add_value_object(root, "iodepth_level", tmp);
1244         /* Only show fixed 7 I/O depth levels*/
1245         for (i = 0; i < 7; i++) {
1246                 char name[20];
1247                 if (i < 6)
1248                         snprintf(name, 20, "%d", 1 << i);
1249                 else
1250                         snprintf(name, 20, ">=%d", 1 << i);
1251                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1252         }
1253
1254         /* Nanosecond latency */
1255         tmp = json_create_object();
1256         json_object_add_value_object(root, "latency_ns", tmp);
1257         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1258                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1259                                  "250", "500", "750", "1000", };
1260                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1261         }
1262         /* Microsecond latency */
1263         tmp = json_create_object();
1264         json_object_add_value_object(root, "latency_us", tmp);
1265         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1266                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1267                                  "250", "500", "750", "1000", };
1268                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1269         }
1270         /* Millisecond latency */
1271         tmp = json_create_object();
1272         json_object_add_value_object(root, "latency_ms", tmp);
1273         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1274                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1275                                  "250", "500", "750", "1000", "2000",
1276                                  ">=2000", };
1277                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1278         }
1279
1280         /* Additional output if continue_on_error set - default off*/
1281         if (ts->continue_on_error) {
1282                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1283                 json_object_add_value_int(root, "first_error", ts->first_error);
1284         }
1285
1286         if (ts->latency_depth) {
1287                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1288                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1289                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1290                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1291         }
1292
1293         /* Additional output if description is set */
1294         if (strlen(ts->description))
1295                 json_object_add_value_string(root, "desc", ts->description);
1296
1297         if (ts->nr_block_infos) {
1298                 /* Block error histogram and types */
1299                 int len;
1300                 unsigned int *percentiles = NULL;
1301                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1302
1303                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1304                                              ts->percentile_list,
1305                                              &percentiles, block_state_counts);
1306
1307                 if (len) {
1308                         struct json_object *block, *percentile_object, *states;
1309                         int state;
1310                         block = json_create_object();
1311                         json_object_add_value_object(root, "block", block);
1312
1313                         percentile_object = json_create_object();
1314                         json_object_add_value_object(block, "percentiles",
1315                                                      percentile_object);
1316                         for (i = 0; i < len; i++) {
1317                                 char buf[20];
1318                                 snprintf(buf, sizeof(buf), "%f",
1319                                          ts->percentile_list[i].u.f);
1320                                 json_object_add_value_int(percentile_object,
1321                                                           (const char *)buf,
1322                                                           percentiles[i]);
1323                         }
1324
1325                         states = json_create_object();
1326                         json_object_add_value_object(block, "states", states);
1327                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1328                                 json_object_add_value_int(states,
1329                                         block_state_names[state],
1330                                         block_state_counts[state]);
1331                         }
1332                         free(percentiles);
1333                 }
1334         }
1335
1336         if (ts->ss_dur) {
1337                 struct json_object *data;
1338                 struct json_array *iops, *bw;
1339                 int i, j, k;
1340                 char ss_buf[64];
1341
1342                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1343                         ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1344                         ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1345                         (float) ts->ss_limit.u.f,
1346                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1347
1348                 tmp = json_create_object();
1349                 json_object_add_value_object(root, "steadystate", tmp);
1350                 json_object_add_value_string(tmp, "ss", ss_buf);
1351                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1352                 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1353
1354                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1355                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1356                 json_object_add_value_string(tmp, "criterion", ss_buf);
1357                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1358                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1359
1360                 data = json_create_object();
1361                 json_object_add_value_object(tmp, "data", data);
1362                 bw = json_create_array();
1363                 iops = json_create_array();
1364
1365                 /*
1366                 ** if ss was attained or the buffer is not full,
1367                 ** ss->head points to the first element in the list.
1368                 ** otherwise it actually points to the second element
1369                 ** in the list
1370                 */
1371                 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1372                         j = ts->ss_head;
1373                 else
1374                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1375                 for (i = 0; i < ts->ss_dur; i++) {
1376                         k = (j + i) % ts->ss_dur;
1377                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1378                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1379                 }
1380                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1381                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1382                 json_object_add_value_array(data, "iops", iops);
1383                 json_object_add_value_array(data, "bw", bw);
1384         }
1385
1386         return root;
1387 }
1388
1389 static void show_thread_status_terse(struct thread_stat *ts,
1390                                      struct group_run_stats *rs,
1391                                      struct buf_output *out)
1392 {
1393         if (terse_version >= 2 && terse_version <= 5)
1394                 show_thread_status_terse_all(ts, rs, terse_version, out);
1395         else
1396                 log_err("fio: bad terse version!? %d\n", terse_version);
1397 }
1398
1399 struct json_object *show_thread_status(struct thread_stat *ts,
1400                                        struct group_run_stats *rs,
1401                                        struct flist_head *opt_list,
1402                                        struct buf_output *out)
1403 {
1404         struct json_object *ret = NULL;
1405
1406         if (output_format & FIO_OUTPUT_TERSE)
1407                 show_thread_status_terse(ts, rs,  out);
1408         if (output_format & FIO_OUTPUT_JSON)
1409                 ret = show_thread_status_json(ts, rs, opt_list);
1410         if (output_format & FIO_OUTPUT_NORMAL)
1411                 show_thread_status_normal(ts, rs,  out);
1412
1413         return ret;
1414 }
1415
1416 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1417 {
1418         double mean, S;
1419
1420         if (src->samples == 0)
1421                 return;
1422
1423         dst->min_val = min(dst->min_val, src->min_val);
1424         dst->max_val = max(dst->max_val, src->max_val);
1425
1426         /*
1427          * Compute new mean and S after the merge
1428          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1429          *  #Parallel_algorithm>
1430          */
1431         if (first) {
1432                 mean = src->mean.u.f;
1433                 S = src->S.u.f;
1434         } else {
1435                 double delta = src->mean.u.f - dst->mean.u.f;
1436
1437                 mean = ((src->mean.u.f * src->samples) +
1438                         (dst->mean.u.f * dst->samples)) /
1439                         (dst->samples + src->samples);
1440
1441                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1442                         (dst->samples * src->samples) /
1443                         (dst->samples + src->samples);
1444         }
1445
1446         dst->samples += src->samples;
1447         dst->mean.u.f = mean;
1448         dst->S.u.f = S;
1449 }
1450
1451 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1452 {
1453         int i;
1454
1455         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1456                 if (dst->max_run[i] < src->max_run[i])
1457                         dst->max_run[i] = src->max_run[i];
1458                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1459                         dst->min_run[i] = src->min_run[i];
1460                 if (dst->max_bw[i] < src->max_bw[i])
1461                         dst->max_bw[i] = src->max_bw[i];
1462                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1463                         dst->min_bw[i] = src->min_bw[i];
1464
1465                 dst->iobytes[i] += src->iobytes[i];
1466                 dst->agg[i] += src->agg[i];
1467         }
1468
1469         if (!dst->kb_base)
1470                 dst->kb_base = src->kb_base;
1471         if (!dst->unit_base)
1472                 dst->unit_base = src->unit_base;
1473 }
1474
1475 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1476                       bool first)
1477 {
1478         int l, k;
1479
1480         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1481                 if (!dst->unified_rw_rep) {
1482                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1483                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1484                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1485                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1486                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first);
1487
1488                         dst->io_bytes[l] += src->io_bytes[l];
1489
1490                         if (dst->runtime[l] < src->runtime[l])
1491                                 dst->runtime[l] = src->runtime[l];
1492                 } else {
1493                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1494                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1495                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1496                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1497                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first);
1498
1499                         dst->io_bytes[0] += src->io_bytes[l];
1500
1501                         if (dst->runtime[0] < src->runtime[l])
1502                                 dst->runtime[0] = src->runtime[l];
1503
1504                         /*
1505                          * We're summing to the same destination, so override
1506                          * 'first' after the first iteration of the loop
1507                          */
1508                         first = false;
1509                 }
1510         }
1511
1512         dst->usr_time += src->usr_time;
1513         dst->sys_time += src->sys_time;
1514         dst->ctx += src->ctx;
1515         dst->majf += src->majf;
1516         dst->minf += src->minf;
1517
1518         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1519                 dst->io_u_map[k] += src->io_u_map[k];
1520         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1521                 dst->io_u_submit[k] += src->io_u_submit[k];
1522         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1523                 dst->io_u_complete[k] += src->io_u_complete[k];
1524         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1525                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1526         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1527                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1528         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1529                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1530
1531         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1532                 if (!dst->unified_rw_rep) {
1533                         dst->total_io_u[k] += src->total_io_u[k];
1534                         dst->short_io_u[k] += src->short_io_u[k];
1535                         dst->drop_io_u[k] += src->drop_io_u[k];
1536                 } else {
1537                         dst->total_io_u[0] += src->total_io_u[k];
1538                         dst->short_io_u[0] += src->short_io_u[k];
1539                         dst->drop_io_u[0] += src->drop_io_u[k];
1540                 }
1541         }
1542
1543         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1544                 int m;
1545
1546                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1547                         if (!dst->unified_rw_rep)
1548                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1549                         else
1550                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1551                 }
1552         }
1553
1554         dst->total_run_time += src->total_run_time;
1555         dst->total_submit += src->total_submit;
1556         dst->total_complete += src->total_complete;
1557 }
1558
1559 void init_group_run_stat(struct group_run_stats *gs)
1560 {
1561         int i;
1562         memset(gs, 0, sizeof(*gs));
1563
1564         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1565                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1566 }
1567
1568 void init_thread_stat(struct thread_stat *ts)
1569 {
1570         int j;
1571
1572         memset(ts, 0, sizeof(*ts));
1573
1574         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1575                 ts->lat_stat[j].min_val = -1UL;
1576                 ts->clat_stat[j].min_val = -1UL;
1577                 ts->slat_stat[j].min_val = -1UL;
1578                 ts->bw_stat[j].min_val = -1UL;
1579                 ts->iops_stat[j].min_val = -1UL;
1580         }
1581         ts->groupid = -1;
1582 }
1583
1584 void __show_run_stats(void)
1585 {
1586         struct group_run_stats *runstats, *rs;
1587         struct thread_data *td;
1588         struct thread_stat *threadstats, *ts;
1589         int i, j, k, nr_ts, last_ts, idx;
1590         int kb_base_warned = 0;
1591         int unit_base_warned = 0;
1592         struct json_object *root = NULL;
1593         struct json_array *array = NULL;
1594         struct buf_output output[FIO_OUTPUT_NR];
1595         struct flist_head **opt_lists;
1596
1597         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1598
1599         for (i = 0; i < groupid + 1; i++)
1600                 init_group_run_stat(&runstats[i]);
1601
1602         /*
1603          * find out how many threads stats we need. if group reporting isn't
1604          * enabled, it's one-per-td.
1605          */
1606         nr_ts = 0;
1607         last_ts = -1;
1608         for_each_td(td, i) {
1609                 if (!td->o.group_reporting) {
1610                         nr_ts++;
1611                         continue;
1612                 }
1613                 if (last_ts == td->groupid)
1614                         continue;
1615                 if (!td->o.stats)
1616                         continue;
1617
1618                 last_ts = td->groupid;
1619                 nr_ts++;
1620         }
1621
1622         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1623         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1624
1625         for (i = 0; i < nr_ts; i++) {
1626                 init_thread_stat(&threadstats[i]);
1627                 opt_lists[i] = NULL;
1628         }
1629
1630         j = 0;
1631         last_ts = -1;
1632         idx = 0;
1633         for_each_td(td, i) {
1634                 if (!td->o.stats)
1635                         continue;
1636                 if (idx && (!td->o.group_reporting ||
1637                     (td->o.group_reporting && last_ts != td->groupid))) {
1638                         idx = 0;
1639                         j++;
1640                 }
1641
1642                 last_ts = td->groupid;
1643
1644                 ts = &threadstats[j];
1645
1646                 ts->clat_percentiles = td->o.clat_percentiles;
1647                 ts->percentile_precision = td->o.percentile_precision;
1648                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1649                 opt_lists[j] = &td->opt_list;
1650
1651                 idx++;
1652                 ts->members++;
1653
1654                 if (ts->groupid == -1) {
1655                         /*
1656                          * These are per-group shared already
1657                          */
1658                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1659                         if (td->o.description)
1660                                 strncpy(ts->description, td->o.description,
1661                                                 FIO_JOBDESC_SIZE - 1);
1662                         else
1663                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1664
1665                         /*
1666                          * If multiple entries in this group, this is
1667                          * the first member.
1668                          */
1669                         ts->thread_number = td->thread_number;
1670                         ts->groupid = td->groupid;
1671
1672                         /*
1673                          * first pid in group, not very useful...
1674                          */
1675                         ts->pid = td->pid;
1676
1677                         ts->kb_base = td->o.kb_base;
1678                         ts->unit_base = td->o.unit_base;
1679                         ts->unified_rw_rep = td->o.unified_rw_rep;
1680                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1681                         log_info("fio: kb_base differs for jobs in group, using"
1682                                  " %u as the base\n", ts->kb_base);
1683                         kb_base_warned = 1;
1684                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1685                         log_info("fio: unit_base differs for jobs in group, using"
1686                                  " %u as the base\n", ts->unit_base);
1687                         unit_base_warned = 1;
1688                 }
1689
1690                 ts->continue_on_error = td->o.continue_on_error;
1691                 ts->total_err_count += td->total_err_count;
1692                 ts->first_error = td->first_error;
1693                 if (!ts->error) {
1694                         if (!td->error && td->o.continue_on_error &&
1695                             td->first_error) {
1696                                 ts->error = td->first_error;
1697                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1698                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1699                         } else  if (td->error) {
1700                                 ts->error = td->error;
1701                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1702                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1703                         }
1704                 }
1705
1706                 ts->latency_depth = td->latency_qd;
1707                 ts->latency_target = td->o.latency_target;
1708                 ts->latency_percentile = td->o.latency_percentile;
1709                 ts->latency_window = td->o.latency_window;
1710
1711                 ts->nr_block_infos = td->ts.nr_block_infos;
1712                 for (k = 0; k < ts->nr_block_infos; k++)
1713                         ts->block_infos[k] = td->ts.block_infos[k];
1714
1715                 sum_thread_stats(ts, &td->ts, idx == 1);
1716
1717                 if (td->o.ss_dur) {
1718                         ts->ss_state = td->ss.state;
1719                         ts->ss_dur = td->ss.dur;
1720                         ts->ss_head = td->ss.head;
1721                         ts->ss_bw_data = td->ss.bw_data;
1722                         ts->ss_iops_data = td->ss.iops_data;
1723                         ts->ss_limit.u.f = td->ss.limit;
1724                         ts->ss_slope.u.f = td->ss.slope;
1725                         ts->ss_deviation.u.f = td->ss.deviation;
1726                         ts->ss_criterion.u.f = td->ss.criterion;
1727                 }
1728                 else
1729                         ts->ss_dur = ts->ss_state = 0;
1730         }
1731
1732         for (i = 0; i < nr_ts; i++) {
1733                 unsigned long long bw;
1734
1735                 ts = &threadstats[i];
1736                 if (ts->groupid == -1)
1737                         continue;
1738                 rs = &runstats[ts->groupid];
1739                 rs->kb_base = ts->kb_base;
1740                 rs->unit_base = ts->unit_base;
1741                 rs->unified_rw_rep += ts->unified_rw_rep;
1742
1743                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1744                         if (!ts->runtime[j])
1745                                 continue;
1746                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1747                                 rs->min_run[j] = ts->runtime[j];
1748                         if (ts->runtime[j] > rs->max_run[j])
1749                                 rs->max_run[j] = ts->runtime[j];
1750
1751                         bw = 0;
1752                         if (ts->runtime[j])
1753                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1754                         if (bw < rs->min_bw[j])
1755                                 rs->min_bw[j] = bw;
1756                         if (bw > rs->max_bw[j])
1757                                 rs->max_bw[j] = bw;
1758
1759                         rs->iobytes[j] += ts->io_bytes[j];
1760                 }
1761         }
1762
1763         for (i = 0; i < groupid + 1; i++) {
1764                 int ddir;
1765
1766                 rs = &runstats[i];
1767
1768                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1769                         if (rs->max_run[ddir])
1770                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1771                                                 rs->max_run[ddir];
1772                 }
1773         }
1774
1775         for (i = 0; i < FIO_OUTPUT_NR; i++)
1776                 buf_output_init(&output[i]);
1777
1778         /*
1779          * don't overwrite last signal output
1780          */
1781         if (output_format & FIO_OUTPUT_NORMAL)
1782                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1783         if (output_format & FIO_OUTPUT_JSON) {
1784                 struct thread_data *global;
1785                 char time_buf[32];
1786                 struct timeval now;
1787                 unsigned long long ms_since_epoch;
1788
1789                 gettimeofday(&now, NULL);
1790                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1791                                  (unsigned long long)(now.tv_usec) / 1000;
1792
1793                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1794                                 sizeof(time_buf));
1795                 if (time_buf[strlen(time_buf) - 1] == '\n')
1796                         time_buf[strlen(time_buf) - 1] = '\0';
1797
1798                 root = json_create_object();
1799                 json_object_add_value_string(root, "fio version", fio_version_string);
1800                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1801                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1802                 json_object_add_value_string(root, "time", time_buf);
1803                 global = get_global_options();
1804                 json_add_job_opts(root, "global options", &global->opt_list, false);
1805                 array = json_create_array();
1806                 json_object_add_value_array(root, "jobs", array);
1807         }
1808
1809         if (is_backend)
1810                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1811
1812         for (i = 0; i < nr_ts; i++) {
1813                 ts = &threadstats[i];
1814                 rs = &runstats[ts->groupid];
1815
1816                 if (is_backend) {
1817                         fio_server_send_job_options(opt_lists[i], i);
1818                         fio_server_send_ts(ts, rs);
1819                 } else {
1820                         if (output_format & FIO_OUTPUT_TERSE)
1821                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1822                         if (output_format & FIO_OUTPUT_JSON) {
1823                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1824                                 json_array_add_value_object(array, tmp);
1825                         }
1826                         if (output_format & FIO_OUTPUT_NORMAL)
1827                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1828                 }
1829         }
1830         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1831                 /* disk util stats, if any */
1832                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1833
1834                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1835
1836                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1837                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1838                 json_free_object(root);
1839         }
1840
1841         for (i = 0; i < groupid + 1; i++) {
1842                 rs = &runstats[i];
1843
1844                 rs->groupid = i;
1845                 if (is_backend)
1846                         fio_server_send_gs(rs);
1847                 else if (output_format & FIO_OUTPUT_NORMAL)
1848                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1849         }
1850
1851         if (is_backend)
1852                 fio_server_send_du();
1853         else if (output_format & FIO_OUTPUT_NORMAL) {
1854                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1855                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1856         }
1857
1858         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1859                 struct buf_output *out = &output[i];
1860
1861                 log_info_buf(out->buf, out->buflen);
1862                 buf_output_free(out);
1863         }
1864
1865         log_info_flush();
1866         free(runstats);
1867         free(threadstats);
1868         free(opt_lists);
1869 }
1870
1871 void show_run_stats(void)
1872 {
1873         fio_mutex_down(stat_mutex);
1874         __show_run_stats();
1875         fio_mutex_up(stat_mutex);
1876 }
1877
1878 void __show_running_run_stats(void)
1879 {
1880         struct thread_data *td;
1881         unsigned long long *rt;
1882         struct timespec ts;
1883         int i;
1884
1885         fio_mutex_down(stat_mutex);
1886
1887         rt = malloc(thread_number * sizeof(unsigned long long));
1888         fio_gettime(&ts, NULL);
1889
1890         for_each_td(td, i) {
1891                 td->update_rusage = 1;
1892                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1893                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1894                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1895                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1896
1897                 rt[i] = mtime_since(&td->start, &ts);
1898                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1899                         td->ts.runtime[DDIR_READ] += rt[i];
1900                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1901                         td->ts.runtime[DDIR_WRITE] += rt[i];
1902                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1903                         td->ts.runtime[DDIR_TRIM] += rt[i];
1904         }
1905
1906         for_each_td(td, i) {
1907                 if (td->runstate >= TD_EXITED)
1908                         continue;
1909                 if (td->rusage_sem) {
1910                         td->update_rusage = 1;
1911                         fio_mutex_down(td->rusage_sem);
1912                 }
1913                 td->update_rusage = 0;
1914         }
1915
1916         __show_run_stats();
1917
1918         for_each_td(td, i) {
1919                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1920                         td->ts.runtime[DDIR_READ] -= rt[i];
1921                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1922                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1923                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1924                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1925         }
1926
1927         free(rt);
1928         fio_mutex_up(stat_mutex);
1929 }
1930
1931 static int status_interval_init;
1932 static struct timespec status_time;
1933 static int status_file_disabled;
1934
1935 #define FIO_STATUS_FILE         "fio-dump-status"
1936
1937 static int check_status_file(void)
1938 {
1939         struct stat sb;
1940         const char *temp_dir;
1941         char fio_status_file_path[PATH_MAX];
1942
1943         if (status_file_disabled)
1944                 return 0;
1945
1946         temp_dir = getenv("TMPDIR");
1947         if (temp_dir == NULL) {
1948                 temp_dir = getenv("TEMP");
1949                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1950                         temp_dir = NULL;
1951         }
1952         if (temp_dir == NULL)
1953                 temp_dir = "/tmp";
1954
1955         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1956
1957         if (stat(fio_status_file_path, &sb))
1958                 return 0;
1959
1960         if (unlink(fio_status_file_path) < 0) {
1961                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1962                                                         strerror(errno));
1963                 log_err("fio: disabling status file updates\n");
1964                 status_file_disabled = 1;
1965         }
1966
1967         return 1;
1968 }
1969
1970 void check_for_running_stats(void)
1971 {
1972         if (status_interval) {
1973                 if (!status_interval_init) {
1974                         fio_gettime(&status_time, NULL);
1975                         status_interval_init = 1;
1976                 } else if (mtime_since_now(&status_time) >= status_interval) {
1977                         show_running_run_stats();
1978                         fio_gettime(&status_time, NULL);
1979                         return;
1980                 }
1981         }
1982         if (check_status_file()) {
1983                 show_running_run_stats();
1984                 return;
1985         }
1986 }
1987
1988 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
1989 {
1990         double val = data;
1991         double delta;
1992
1993         if (data > is->max_val)
1994                 is->max_val = data;
1995         if (data < is->min_val)
1996                 is->min_val = data;
1997
1998         delta = val - is->mean.u.f;
1999         if (delta) {
2000                 is->mean.u.f += delta / (is->samples + 1.0);
2001                 is->S.u.f += delta * (val - is->mean.u.f);
2002         }
2003
2004         is->samples++;
2005 }
2006
2007 /*
2008  * Return a struct io_logs, which is added to the tail of the log
2009  * list for 'iolog'.
2010  */
2011 static struct io_logs *get_new_log(struct io_log *iolog)
2012 {
2013         size_t new_size, new_samples;
2014         struct io_logs *cur_log;
2015
2016         /*
2017          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2018          * forever
2019          */
2020         if (!iolog->cur_log_max)
2021                 new_samples = DEF_LOG_ENTRIES;
2022         else {
2023                 new_samples = iolog->cur_log_max * 2;
2024                 if (new_samples > MAX_LOG_ENTRIES)
2025                         new_samples = MAX_LOG_ENTRIES;
2026         }
2027
2028         new_size = new_samples * log_entry_sz(iolog);
2029
2030         cur_log = smalloc(sizeof(*cur_log));
2031         if (cur_log) {
2032                 INIT_FLIST_HEAD(&cur_log->list);
2033                 cur_log->log = malloc(new_size);
2034                 if (cur_log->log) {
2035                         cur_log->nr_samples = 0;
2036                         cur_log->max_samples = new_samples;
2037                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2038                         iolog->cur_log_max = new_samples;
2039                         return cur_log;
2040                 }
2041                 sfree(cur_log);
2042         }
2043
2044         return NULL;
2045 }
2046
2047 /*
2048  * Add and return a new log chunk, or return current log if big enough
2049  */
2050 static struct io_logs *regrow_log(struct io_log *iolog)
2051 {
2052         struct io_logs *cur_log;
2053         int i;
2054
2055         if (!iolog || iolog->disabled)
2056                 goto disable;
2057
2058         cur_log = iolog_cur_log(iolog);
2059         if (!cur_log) {
2060                 cur_log = get_new_log(iolog);
2061                 if (!cur_log)
2062                         return NULL;
2063         }
2064
2065         if (cur_log->nr_samples < cur_log->max_samples)
2066                 return cur_log;
2067
2068         /*
2069          * No room for a new sample. If we're compressing on the fly, flush
2070          * out the current chunk
2071          */
2072         if (iolog->log_gz) {
2073                 if (iolog_cur_flush(iolog, cur_log)) {
2074                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2075                         return NULL;
2076                 }
2077         }
2078
2079         /*
2080          * Get a new log array, and add to our list
2081          */
2082         cur_log = get_new_log(iolog);
2083         if (!cur_log) {
2084                 log_err("fio: failed extending iolog! Will stop logging.\n");
2085                 return NULL;
2086         }
2087
2088         if (!iolog->pending || !iolog->pending->nr_samples)
2089                 return cur_log;
2090
2091         /*
2092          * Flush pending items to new log
2093          */
2094         for (i = 0; i < iolog->pending->nr_samples; i++) {
2095                 struct io_sample *src, *dst;
2096
2097                 src = get_sample(iolog, iolog->pending, i);
2098                 dst = get_sample(iolog, cur_log, i);
2099                 memcpy(dst, src, log_entry_sz(iolog));
2100         }
2101         cur_log->nr_samples = iolog->pending->nr_samples;
2102
2103         iolog->pending->nr_samples = 0;
2104         return cur_log;
2105 disable:
2106         if (iolog)
2107                 iolog->disabled = true;
2108         return NULL;
2109 }
2110
2111 void regrow_logs(struct thread_data *td)
2112 {
2113         regrow_log(td->slat_log);
2114         regrow_log(td->clat_log);
2115         regrow_log(td->clat_hist_log);
2116         regrow_log(td->lat_log);
2117         regrow_log(td->bw_log);
2118         regrow_log(td->iops_log);
2119         td->flags &= ~TD_F_REGROW_LOGS;
2120 }
2121
2122 static struct io_logs *get_cur_log(struct io_log *iolog)
2123 {
2124         struct io_logs *cur_log;
2125
2126         cur_log = iolog_cur_log(iolog);
2127         if (!cur_log) {
2128                 cur_log = get_new_log(iolog);
2129                 if (!cur_log)
2130                         return NULL;
2131         }
2132
2133         if (cur_log->nr_samples < cur_log->max_samples)
2134                 return cur_log;
2135
2136         /*
2137          * Out of space. If we're in IO offload mode, or we're not doing
2138          * per unit logging (hence logging happens outside of the IO thread
2139          * as well), add a new log chunk inline. If we're doing inline
2140          * submissions, flag 'td' as needing a log regrow and we'll take
2141          * care of it on the submission side.
2142          */
2143         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2144             !per_unit_log(iolog))
2145                 return regrow_log(iolog);
2146
2147         iolog->td->flags |= TD_F_REGROW_LOGS;
2148         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2149         return iolog->pending;
2150 }
2151
2152 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2153                              enum fio_ddir ddir, unsigned int bs,
2154                              unsigned long t, uint64_t offset)
2155 {
2156         struct io_logs *cur_log;
2157
2158         if (iolog->disabled)
2159                 return;
2160         if (flist_empty(&iolog->io_logs))
2161                 iolog->avg_last = t;
2162
2163         cur_log = get_cur_log(iolog);
2164         if (cur_log) {
2165                 struct io_sample *s;
2166
2167                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2168
2169                 s->data = data;
2170                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2171                 io_sample_set_ddir(iolog, s, ddir);
2172                 s->bs = bs;
2173
2174                 if (iolog->log_offset) {
2175                         struct io_sample_offset *so = (void *) s;
2176
2177                         so->offset = offset;
2178                 }
2179
2180                 cur_log->nr_samples++;
2181                 return;
2182         }
2183
2184         iolog->disabled = true;
2185 }
2186
2187 static inline void reset_io_stat(struct io_stat *ios)
2188 {
2189         ios->max_val = ios->min_val = ios->samples = 0;
2190         ios->mean.u.f = ios->S.u.f = 0;
2191 }
2192
2193 void reset_io_stats(struct thread_data *td)
2194 {
2195         struct thread_stat *ts = &td->ts;
2196         int i, j;
2197
2198         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2199                 reset_io_stat(&ts->clat_stat[i]);
2200                 reset_io_stat(&ts->slat_stat[i]);
2201                 reset_io_stat(&ts->lat_stat[i]);
2202                 reset_io_stat(&ts->bw_stat[i]);
2203                 reset_io_stat(&ts->iops_stat[i]);
2204
2205                 ts->io_bytes[i] = 0;
2206                 ts->runtime[i] = 0;
2207                 ts->total_io_u[i] = 0;
2208                 ts->short_io_u[i] = 0;
2209                 ts->drop_io_u[i] = 0;
2210
2211                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2212                         ts->io_u_plat[i][j] = 0;
2213         }
2214
2215         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2216                 ts->io_u_map[i] = 0;
2217                 ts->io_u_submit[i] = 0;
2218                 ts->io_u_complete[i] = 0;
2219         }
2220
2221         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2222                 ts->io_u_lat_n[i] = 0;
2223         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2224                 ts->io_u_lat_u[i] = 0;
2225         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2226                 ts->io_u_lat_m[i] = 0;
2227
2228         ts->total_submit = 0;
2229         ts->total_complete = 0;
2230 }
2231
2232 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2233                               unsigned long elapsed, bool log_max)
2234 {
2235         /*
2236          * Note an entry in the log. Use the mean from the logged samples,
2237          * making sure to properly round up. Only write a log entry if we
2238          * had actual samples done.
2239          */
2240         if (iolog->avg_window[ddir].samples) {
2241                 union io_sample_data data;
2242
2243                 if (log_max)
2244                         data.val = iolog->avg_window[ddir].max_val;
2245                 else
2246                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2247
2248                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2249         }
2250
2251         reset_io_stat(&iolog->avg_window[ddir]);
2252 }
2253
2254 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2255                              bool log_max)
2256 {
2257         int ddir;
2258
2259         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2260                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2261 }
2262
2263 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2264                            union io_sample_data data, enum fio_ddir ddir,
2265                            unsigned int bs, uint64_t offset)
2266 {
2267         unsigned long elapsed, this_window;
2268
2269         if (!ddir_rw(ddir))
2270                 return 0;
2271
2272         elapsed = mtime_since_now(&td->epoch);
2273
2274         /*
2275          * If no time averaging, just add the log sample.
2276          */
2277         if (!iolog->avg_msec) {
2278                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2279                 return 0;
2280         }
2281
2282         /*
2283          * Add the sample. If the time period has passed, then
2284          * add that entry to the log and clear.
2285          */
2286         add_stat_sample(&iolog->avg_window[ddir], data.val);
2287
2288         /*
2289          * If period hasn't passed, adding the above sample is all we
2290          * need to do.
2291          */
2292         this_window = elapsed - iolog->avg_last;
2293         if (elapsed < iolog->avg_last)
2294                 return iolog->avg_last - elapsed;
2295         else if (this_window < iolog->avg_msec) {
2296                 int diff = iolog->avg_msec - this_window;
2297
2298                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2299                         return diff;
2300         }
2301
2302         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2303
2304         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2305         return iolog->avg_msec;
2306 }
2307
2308 void finalize_logs(struct thread_data *td, bool unit_logs)
2309 {
2310         unsigned long elapsed;
2311
2312         elapsed = mtime_since_now(&td->epoch);
2313
2314         if (td->clat_log && unit_logs)
2315                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2316         if (td->slat_log && unit_logs)
2317                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2318         if (td->lat_log && unit_logs)
2319                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2320         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2321                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2322         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2323                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2324 }
2325
2326 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2327 {
2328         struct io_log *iolog;
2329
2330         if (!ddir_rw(ddir))
2331                 return;
2332
2333         iolog = agg_io_log[ddir];
2334         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2335 }
2336
2337 static void add_clat_percentile_sample(struct thread_stat *ts,
2338                                 unsigned long long nsec, enum fio_ddir ddir)
2339 {
2340         unsigned int idx = plat_val_to_idx(nsec);
2341         assert(idx < FIO_IO_U_PLAT_NR);
2342
2343         ts->io_u_plat[ddir][idx]++;
2344 }
2345
2346 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2347                      unsigned long long nsec, unsigned int bs, uint64_t offset)
2348 {
2349         unsigned long elapsed, this_window;
2350         struct thread_stat *ts = &td->ts;
2351         struct io_log *iolog = td->clat_hist_log;
2352
2353         td_io_u_lock(td);
2354
2355         add_stat_sample(&ts->clat_stat[ddir], nsec);
2356
2357         if (td->clat_log)
2358                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2359                                offset);
2360
2361         if (ts->clat_percentiles)
2362                 add_clat_percentile_sample(ts, nsec, ddir);
2363
2364         if (iolog && iolog->hist_msec) {
2365                 struct io_hist *hw = &iolog->hist_window[ddir];
2366
2367                 hw->samples++;
2368                 elapsed = mtime_since_now(&td->epoch);
2369                 if (!hw->hist_last)
2370                         hw->hist_last = elapsed;
2371                 this_window = elapsed - hw->hist_last;
2372                 
2373                 if (this_window >= iolog->hist_msec) {
2374                         unsigned int *io_u_plat;
2375                         struct io_u_plat_entry *dst;
2376
2377                         /*
2378                          * Make a byte-for-byte copy of the latency histogram
2379                          * stored in td->ts.io_u_plat[ddir], recording it in a
2380                          * log sample. Note that the matching call to free() is
2381                          * located in iolog.c after printing this sample to the
2382                          * log file.
2383                          */
2384                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2385                         dst = malloc(sizeof(struct io_u_plat_entry));
2386                         memcpy(&(dst->io_u_plat), io_u_plat,
2387                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2388                         flist_add(&dst->list, &hw->list);
2389                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2390                                                 elapsed, offset);
2391
2392                         /*
2393                          * Update the last time we recorded as being now, minus
2394                          * any drift in time we encountered before actually
2395                          * making the record.
2396                          */
2397                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2398                         hw->samples = 0;
2399                 }
2400         }
2401
2402         td_io_u_unlock(td);
2403 }
2404
2405 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2406                      unsigned long usec, unsigned int bs, uint64_t offset)
2407 {
2408         struct thread_stat *ts = &td->ts;
2409
2410         if (!ddir_rw(ddir))
2411                 return;
2412
2413         td_io_u_lock(td);
2414
2415         add_stat_sample(&ts->slat_stat[ddir], usec);
2416
2417         if (td->slat_log)
2418                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2419
2420         td_io_u_unlock(td);
2421 }
2422
2423 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2424                     unsigned long long nsec, unsigned int bs, uint64_t offset)
2425 {
2426         struct thread_stat *ts = &td->ts;
2427
2428         if (!ddir_rw(ddir))
2429                 return;
2430
2431         td_io_u_lock(td);
2432
2433         add_stat_sample(&ts->lat_stat[ddir], nsec);
2434
2435         if (td->lat_log)
2436                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2437                                offset);
2438
2439         td_io_u_unlock(td);
2440 }
2441
2442 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2443                    unsigned int bytes, unsigned long long spent)
2444 {
2445         struct thread_stat *ts = &td->ts;
2446         unsigned long rate;
2447
2448         if (spent)
2449                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2450         else
2451                 rate = 0;
2452
2453         td_io_u_lock(td);
2454
2455         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2456
2457         if (td->bw_log)
2458                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2459                                bytes, io_u->offset);
2460
2461         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2462         td_io_u_unlock(td);
2463 }
2464
2465 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2466                          struct timespec *t, unsigned int avg_time,
2467                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2468                          struct io_stat *stat, struct io_log *log,
2469                          bool is_kb)
2470 {
2471         unsigned long spent, rate;
2472         enum fio_ddir ddir;
2473         unsigned int next, next_log;
2474
2475         next_log = avg_time;
2476
2477         spent = mtime_since(parent_tv, t);
2478         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2479                 return avg_time - spent;
2480
2481         td_io_u_lock(td);
2482
2483         /*
2484          * Compute both read and write rates for the interval.
2485          */
2486         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2487                 uint64_t delta;
2488
2489                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2490                 if (!delta)
2491                         continue; /* No entries for interval */
2492
2493                 if (spent) {
2494                         if (is_kb)
2495                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2496                         else
2497                                 rate = (delta * 1000) / spent;
2498                 } else
2499                         rate = 0;
2500
2501                 add_stat_sample(&stat[ddir], rate);
2502
2503                 if (log) {
2504                         unsigned int bs = 0;
2505
2506                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2507                                 bs = td->o.min_bs[ddir];
2508
2509                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2510                         next_log = min(next_log, next);
2511                 }
2512
2513                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2514         }
2515
2516         timespec_add_msec(parent_tv, avg_time);
2517
2518         td_io_u_unlock(td);
2519
2520         if (spent <= avg_time)
2521                 next = avg_time;
2522         else
2523                 next = avg_time - (1 + spent - avg_time);
2524
2525         return min(next, next_log);
2526 }
2527
2528 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2529 {
2530         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2531                                 td->this_io_bytes, td->stat_io_bytes,
2532                                 td->ts.bw_stat, td->bw_log, true);
2533 }
2534
2535 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2536                      unsigned int bytes)
2537 {
2538         struct thread_stat *ts = &td->ts;
2539
2540         td_io_u_lock(td);
2541
2542         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2543
2544         if (td->iops_log)
2545                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2546                                bytes, io_u->offset);
2547
2548         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2549         td_io_u_unlock(td);
2550 }
2551
2552 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2553 {
2554         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2555                                 td->this_io_blocks, td->stat_io_blocks,
2556                                 td->ts.iops_stat, td->iops_log, false);
2557 }
2558
2559 /*
2560  * Returns msecs to next event
2561  */
2562 int calc_log_samples(void)
2563 {
2564         struct thread_data *td;
2565         unsigned int next = ~0U, tmp;
2566         struct timespec now;
2567         int i;
2568
2569         fio_gettime(&now, NULL);
2570
2571         for_each_td(td, i) {
2572                 if (!td->o.stats)
2573                         continue;
2574                 if (in_ramp_time(td) ||
2575                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2576                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2577                         continue;
2578                 }
2579                 if (!td->bw_log ||
2580                         (td->bw_log && !per_unit_log(td->bw_log))) {
2581                         tmp = add_bw_samples(td, &now);
2582                         if (tmp < next)
2583                                 next = tmp;
2584                 }
2585                 if (!td->iops_log ||
2586                         (td->iops_log && !per_unit_log(td->iops_log))) {
2587                         tmp = add_iops_samples(td, &now);
2588                         if (tmp < next)
2589                                 next = tmp;
2590                 }
2591         }
2592
2593         return next == ~0U ? 0 : next;
2594 }
2595
2596 void stat_init(void)
2597 {
2598         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2599 }
2600
2601 void stat_exit(void)
2602 {
2603         /*
2604          * When we have the mutex, we know out-of-band access to it
2605          * have ended.
2606          */
2607         fio_mutex_down(stat_mutex);
2608         fio_mutex_remove(stat_mutex);
2609 }
2610
2611 /*
2612  * Called from signal handler. Wake up status thread.
2613  */
2614 void show_running_run_stats(void)
2615 {
2616         helper_do_stat();
2617 }
2618
2619 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2620 {
2621         /* Ignore io_u's which span multiple blocks--they will just get
2622          * inaccurate counts. */
2623         int idx = (io_u->offset - io_u->file->file_offset)
2624                         / td->o.bs[DDIR_TRIM];
2625         uint32_t *info = &td->ts.block_infos[idx];
2626         assert(idx < td->ts.nr_block_infos);
2627         return info;
2628 }