f014e60883c3a2946ba6c70d38fd2d3a82a03eb8
[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(&td->ru_start.ru_utime,
41                                         &td->ru_end.ru_utime);
42         ts->sys_time += mtime_since(&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 int 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_clz(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 int 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 int **output,
139                                    unsigned int *maxv, unsigned int *minv)
140 {
141         unsigned long sum = 0;
142         unsigned int len, i, j = 0;
143         unsigned int oval_len = 0;
144         unsigned int *ovals = NULL;
145         int is_last;
146
147         *minv = -1U;
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(unsigned int));
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 len, j = 0, minv, maxv;
205         unsigned int *ovals;
206         int is_last, per_line, scale_down;
207         char fmt[32];
208
209         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
210         if (!len)
211                 goto out;
212
213         /*
214          * We default to usecs, but if the value range is such that we
215          * should scale down to msecs, do that.
216          */
217         if (minv > 2000 && maxv > 99999) {
218                 scale_down = 1;
219                 log_buf(out, "    clat percentiles (msec):\n     |");
220         } else {
221                 scale_down = 0;
222                 log_buf(out, "    clat percentiles (usec):\n     |");
223         }
224
225         snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
226         per_line = (80 - 7) / (precision + 14);
227
228         for (j = 0; j < len; j++) {
229                 char fbuf[16], *ptr = fbuf;
230
231                 /* for formatting */
232                 if (j != 0 && (j % per_line) == 0)
233                         log_buf(out, "     |");
234
235                 /* end of the list */
236                 is_last = (j == len - 1);
237
238                 if (plist[j].u.f < 10.0)
239                         ptr += sprintf(fbuf, " ");
240
241                 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
242
243                 if (scale_down)
244                         ovals[j] = (ovals[j] + 999) / 1000;
245
246                 log_buf(out, " %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
247
248                 if (is_last)
249                         break;
250
251                 if ((j % per_line) == per_line - 1)     /* for formatting */
252                         log_buf(out, "\n");
253         }
254
255 out:
256         if (ovals)
257                 free(ovals);
258 }
259
260 bool calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
261               double *mean, double *dev)
262 {
263         double n = (double) is->samples;
264
265         if (n == 0)
266                 return false;
267
268         *min = is->min_val;
269         *max = is->max_val;
270         *mean = is->mean.u.f;
271
272         if (n > 1.0)
273                 *dev = sqrt(is->S.u.f / (n - 1.0));
274         else
275                 *dev = 0;
276
277         return true;
278 }
279
280 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
281 {
282         char *p1, *p2, *p3, *p4;
283         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
284         int i;
285
286         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
287
288         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
289                 const int i2p = is_power_of_2(rs->kb_base);
290
291                 if (!rs->max_run[i])
292                         continue;
293
294                 p1 = num2str(rs->iobytes[i], 6, 1, i2p, 8);
295                 p2 = num2str(rs->agg[i], 6, 1, i2p, rs->unit_base);
296                 p3 = num2str(rs->min_bw[i], 6, 1, i2p, rs->unit_base);
297                 p4 = num2str(rs->max_bw[i], 6, 1, i2p, rs->unit_base);
298
299                 log_buf(out, "%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
300                          " mint=%llumsec, maxt=%llumsec\n",
301                                 rs->unified_rw_rep ? "  MIXED" : str[i],
302                                 p1, p2, p3, p4,
303                                 (unsigned long long) rs->min_run[i],
304                                 (unsigned long long) rs->max_run[i]);
305
306                 free(p1);
307                 free(p2);
308                 free(p3);
309                 free(p4);
310         }
311 }
312
313 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
314 {
315         int i;
316
317         /*
318          * Do depth distribution calculations
319          */
320         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
321                 if (total) {
322                         io_u_dist[i] = (double) map[i] / (double) total;
323                         io_u_dist[i] *= 100.0;
324                         if (io_u_dist[i] < 0.1 && map[i])
325                                 io_u_dist[i] = 0.1;
326                 } else
327                         io_u_dist[i] = 0.0;
328         }
329 }
330
331 static void stat_calc_lat(struct thread_stat *ts, double *dst,
332                           unsigned int *src, int nr)
333 {
334         unsigned long total = ddir_rw_sum(ts->total_io_u);
335         int i;
336
337         /*
338          * Do latency distribution calculations
339          */
340         for (i = 0; i < nr; i++) {
341                 if (total) {
342                         dst[i] = (double) src[i] / (double) total;
343                         dst[i] *= 100.0;
344                         if (dst[i] < 0.01 && src[i])
345                                 dst[i] = 0.01;
346                 } else
347                         dst[i] = 0.0;
348         }
349 }
350
351 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
352 {
353         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
354 }
355
356 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
357 {
358         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
359 }
360
361 static void display_lat(const char *name, unsigned long min, unsigned long max,
362                         double mean, double dev, struct buf_output *out)
363 {
364         const char *base = "(usec)";
365         char *minp, *maxp;
366
367         if (usec_to_msec(&min, &max, &mean, &dev))
368                 base = "(msec)";
369
370         minp = num2str(min, 6, 1, 0, 0);
371         maxp = num2str(max, 6, 1, 0, 0);
372
373         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
374                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
375
376         free(minp);
377         free(maxp);
378 }
379
380 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
381                              int ddir, struct buf_output *out)
382 {
383         const char *str[] = { " read", "write", " trim" };
384         unsigned long min, max, runt;
385         unsigned long long bw, iops;
386         double mean, dev;
387         char *io_p, *bw_p, *iops_p;
388         int i2p;
389
390         assert(ddir_rw(ddir));
391
392         if (!ts->runtime[ddir])
393                 return;
394
395         i2p = is_power_of_2(rs->kb_base);
396         runt = ts->runtime[ddir];
397
398         bw = (1000 * ts->io_bytes[ddir]) / runt;
399         io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
400         bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
401
402         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
403         iops_p = num2str(iops, 6, 1, 0, 0);
404
405         log_buf(out, "  %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
406                                 rs->unified_rw_rep ? "mixed" : str[ddir],
407                                 io_p, bw_p, iops_p,
408                                 (unsigned long long) ts->runtime[ddir]);
409
410         free(io_p);
411         free(bw_p);
412         free(iops_p);
413
414         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
415                 display_lat("slat", min, max, mean, dev, out);
416         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
417                 display_lat("clat", min, max, mean, dev, out);
418         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
419                 display_lat(" lat", min, max, mean, dev, out);
420
421         if (ts->clat_percentiles) {
422                 show_clat_percentiles(ts->io_u_plat[ddir],
423                                         ts->clat_stat[ddir].samples,
424                                         ts->percentile_list,
425                                         ts->percentile_precision, out);
426         }
427         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
428                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
429                 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
430
431                 if (rs->unit_base == 1) {
432                         min *= 8.0;
433                         max *= 8.0;
434                         mean *= 8.0;
435                         dev *= 8.0;
436                 }
437
438                 if (rs->agg[ddir]) {
439                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
440                         if (p_of_agg > 100.0)
441                                 p_of_agg = 100.0;
442                 }
443
444                 if (mean > fkb_base * fkb_base) {
445                         min /= fkb_base;
446                         max /= fkb_base;
447                         mean /= fkb_base;
448                         dev /= fkb_base;
449                         bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
450                 }
451
452                 log_buf(out, "    bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
453                          " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
454                                                         p_of_agg, mean, dev);
455         }
456 }
457
458 static int show_lat(double *io_u_lat, int nr, const char **ranges,
459                     const char *msg, struct buf_output *out)
460 {
461         int new_line = 1, i, line = 0, shown = 0;
462
463         for (i = 0; i < nr; i++) {
464                 if (io_u_lat[i] <= 0.0)
465                         continue;
466                 shown = 1;
467                 if (new_line) {
468                         if (line)
469                                 log_buf(out, "\n");
470                         log_buf(out, "    lat (%s) : ", msg);
471                         new_line = 0;
472                         line = 0;
473                 }
474                 if (line)
475                         log_buf(out, ", ");
476                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
477                 line++;
478                 if (line == 5)
479                         new_line = 1;
480         }
481
482         if (shown)
483                 log_buf(out, "\n");
484
485         return shown;
486 }
487
488 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
489 {
490         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
491                                  "250=", "500=", "750=", "1000=", };
492
493         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
494 }
495
496 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
497 {
498         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
499                                  "250=", "500=", "750=", "1000=", "2000=",
500                                  ">=2000=", };
501
502         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
503 }
504
505 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
506 {
507         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
508         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
509
510         stat_calc_lat_u(ts, io_u_lat_u);
511         stat_calc_lat_m(ts, io_u_lat_m);
512
513         show_lat_u(io_u_lat_u, out);
514         show_lat_m(io_u_lat_m, out);
515 }
516
517 static int block_state_category(int block_state)
518 {
519         switch (block_state) {
520         case BLOCK_STATE_UNINIT:
521                 return 0;
522         case BLOCK_STATE_TRIMMED:
523         case BLOCK_STATE_WRITTEN:
524                 return 1;
525         case BLOCK_STATE_WRITE_FAILURE:
526         case BLOCK_STATE_TRIM_FAILURE:
527                 return 2;
528         default:
529                 /* Silence compile warning on some BSDs and have a return */
530                 assert(0);
531                 return -1;
532         }
533 }
534
535 static int compare_block_infos(const void *bs1, const void *bs2)
536 {
537         uint32_t block1 = *(uint32_t *)bs1;
538         uint32_t block2 = *(uint32_t *)bs2;
539         int state1 = BLOCK_INFO_STATE(block1);
540         int state2 = BLOCK_INFO_STATE(block2);
541         int bscat1 = block_state_category(state1);
542         int bscat2 = block_state_category(state2);
543         int cycles1 = BLOCK_INFO_TRIMS(block1);
544         int cycles2 = BLOCK_INFO_TRIMS(block2);
545
546         if (bscat1 < bscat2)
547                 return -1;
548         if (bscat1 > bscat2)
549                 return 1;
550
551         if (cycles1 < cycles2)
552                 return -1;
553         if (cycles1 > cycles2)
554                 return 1;
555
556         if (state1 < state2)
557                 return -1;
558         if (state1 > state2)
559                 return 1;
560
561         assert(block1 == block2);
562         return 0;
563 }
564
565 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
566                                   fio_fp64_t *plist, unsigned int **percentiles,
567                                   unsigned int *types)
568 {
569         int len = 0;
570         int i, nr_uninit;
571
572         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
573
574         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
575                 len++;
576
577         if (!len)
578                 return 0;
579
580         /*
581          * Sort the percentile list. Note that it may already be sorted if
582          * we are using the default values, but since it's a short list this
583          * isn't a worry. Also note that this does not work for NaN values.
584          */
585         if (len > 1)
586                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
587
588         nr_uninit = 0;
589         /* Start only after the uninit entries end */
590         for (nr_uninit = 0;
591              nr_uninit < nr_block_infos
592                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
593              nr_uninit ++)
594                 ;
595
596         if (nr_uninit == nr_block_infos)
597                 return 0;
598
599         *percentiles = calloc(len, sizeof(**percentiles));
600
601         for (i = 0; i < len; i++) {
602                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
603                                 + nr_uninit;
604                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
605         }
606
607         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
608         for (i = 0; i < nr_block_infos; i++)
609                 types[BLOCK_INFO_STATE(block_infos[i])]++;
610
611         return len;
612 }
613
614 static const char *block_state_names[] = {
615         [BLOCK_STATE_UNINIT] = "unwritten",
616         [BLOCK_STATE_TRIMMED] = "trimmed",
617         [BLOCK_STATE_WRITTEN] = "written",
618         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
619         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
620 };
621
622 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
623                              fio_fp64_t *plist, struct buf_output *out)
624 {
625         int len, pos, i;
626         unsigned int *percentiles = NULL;
627         unsigned int block_state_counts[BLOCK_STATE_COUNT];
628
629         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
630                                      &percentiles, block_state_counts);
631
632         log_buf(out, "  block lifetime percentiles :\n   |");
633         pos = 0;
634         for (i = 0; i < len; i++) {
635                 uint32_t block_info = percentiles[i];
636 #define LINE_LENGTH     75
637                 char str[LINE_LENGTH];
638                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
639                                      plist[i].u.f, block_info,
640                                      i == len - 1 ? '\n' : ',');
641                 assert(strln < LINE_LENGTH);
642                 if (pos + strln > LINE_LENGTH) {
643                         pos = 0;
644                         log_buf(out, "\n   |");
645                 }
646                 log_buf(out, "%s", str);
647                 pos += strln;
648 #undef LINE_LENGTH
649         }
650         if (percentiles)
651                 free(percentiles);
652
653         log_buf(out, "        states               :");
654         for (i = 0; i < BLOCK_STATE_COUNT; i++)
655                 log_buf(out, " %s=%u%c",
656                          block_state_names[i], block_state_counts[i],
657                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
658 }
659
660 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
661 {
662         char *p1, *p2;
663         unsigned long long bw_mean, iops_mean;
664         const int i2p = is_power_of_2(ts->kb_base);
665
666         if (!ts->ss_dur)
667                 return;
668
669         bw_mean = steadystate_bw_mean(ts);
670         iops_mean = steadystate_iops_mean(ts);
671
672         p1 = num2str(bw_mean / ts->kb_base, 6, ts->kb_base, i2p, ts->unit_base);
673         p2 = num2str(iops_mean, 6, 1, 0, 0);
674
675         log_buf(out, "  steadystate  : attained=%s, bw=%s/s, iops=%s, %s%s=%.3f%s\n",
676                 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
677                 p1, p2,
678                 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
679                 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
680                 ts->ss_criterion.u.f,
681                 ts->ss_state & __FIO_SS_PCT ? "%" : "");
682
683         free(p1);
684         free(p2);
685 }
686
687 static void show_thread_status_normal(struct thread_stat *ts,
688                                       struct group_run_stats *rs,
689                                       struct buf_output *out)
690 {
691         double usr_cpu, sys_cpu;
692         unsigned long runtime;
693         double io_u_dist[FIO_IO_U_MAP_NR];
694         time_t time_p;
695         char time_buf[32];
696
697         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
698                 return;
699                 
700         memset(time_buf, 0, sizeof(time_buf));
701
702         time(&time_p);
703         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
704
705         if (!ts->error) {
706                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
707                                         ts->name, ts->groupid, ts->members,
708                                         ts->error, (int) ts->pid, time_buf);
709         } else {
710                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
711                                         ts->name, ts->groupid, ts->members,
712                                         ts->error, ts->verror, (int) ts->pid,
713                                         time_buf);
714         }
715
716         if (strlen(ts->description))
717                 log_buf(out, "  Description  : [%s]\n", ts->description);
718
719         if (ts->io_bytes[DDIR_READ])
720                 show_ddir_status(rs, ts, DDIR_READ, out);
721         if (ts->io_bytes[DDIR_WRITE])
722                 show_ddir_status(rs, ts, DDIR_WRITE, out);
723         if (ts->io_bytes[DDIR_TRIM])
724                 show_ddir_status(rs, ts, DDIR_TRIM, out);
725
726         show_latencies(ts, out);
727
728         runtime = ts->total_run_time;
729         if (runtime) {
730                 double runt = (double) runtime;
731
732                 usr_cpu = (double) ts->usr_time * 100 / runt;
733                 sys_cpu = (double) ts->sys_time * 100 / runt;
734         } else {
735                 usr_cpu = 0;
736                 sys_cpu = 0;
737         }
738
739         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
740                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
741                         (unsigned long long) ts->ctx,
742                         (unsigned long long) ts->majf,
743                         (unsigned long long) ts->minf);
744
745         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
746         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
747                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
748                                         io_u_dist[1], io_u_dist[2],
749                                         io_u_dist[3], io_u_dist[4],
750                                         io_u_dist[5], io_u_dist[6]);
751
752         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
753         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
754                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
755                                         io_u_dist[1], io_u_dist[2],
756                                         io_u_dist[3], io_u_dist[4],
757                                         io_u_dist[5], io_u_dist[6]);
758         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
759         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
760                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
761                                         io_u_dist[1], io_u_dist[2],
762                                         io_u_dist[3], io_u_dist[4],
763                                         io_u_dist[5], io_u_dist[6]);
764         log_buf(out, "     issued    : total=r=%llu/w=%llu/d=%llu,"
765                                  " short=r=%llu/w=%llu/d=%llu,"
766                                  " drop=r=%llu/w=%llu/d=%llu\n",
767                                         (unsigned long long) ts->total_io_u[0],
768                                         (unsigned long long) ts->total_io_u[1],
769                                         (unsigned long long) ts->total_io_u[2],
770                                         (unsigned long long) ts->short_io_u[0],
771                                         (unsigned long long) ts->short_io_u[1],
772                                         (unsigned long long) ts->short_io_u[2],
773                                         (unsigned long long) ts->drop_io_u[0],
774                                         (unsigned long long) ts->drop_io_u[1],
775                                         (unsigned long long) ts->drop_io_u[2]);
776         if (ts->continue_on_error) {
777                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
778                                         (unsigned long long)ts->total_err_count,
779                                         ts->first_error,
780                                         strerror(ts->first_error));
781         }
782         if (ts->latency_depth) {
783                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
784                                         (unsigned long long)ts->latency_target,
785                                         (unsigned long long)ts->latency_window,
786                                         ts->latency_percentile.u.f,
787                                         ts->latency_depth);
788         }
789
790         if (ts->nr_block_infos)
791                 show_block_infos(ts->nr_block_infos, ts->block_infos,
792                                   ts->percentile_list, out);
793
794         if (ts->ss_dur)
795                 show_ss_normal(ts, out);
796 }
797
798 static void show_ddir_status_terse(struct thread_stat *ts,
799                                    struct group_run_stats *rs, int ddir,
800                                    struct buf_output *out)
801 {
802         unsigned long min, max;
803         unsigned long long bw, iops;
804         unsigned int *ovals = NULL;
805         double mean, dev;
806         unsigned int len, minv, maxv;
807         int i;
808
809         assert(ddir_rw(ddir));
810
811         iops = bw = 0;
812         if (ts->runtime[ddir]) {
813                 uint64_t runt = ts->runtime[ddir];
814
815                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
816                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
817         }
818
819         log_buf(out, ";%llu;%llu;%llu;%llu",
820                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
821                                         (unsigned long long) ts->runtime[ddir]);
822
823         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
824                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
825         else
826                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
827
828         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
829                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
830         else
831                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
832
833         if (ts->clat_percentiles) {
834                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
835                                         ts->clat_stat[ddir].samples,
836                                         ts->percentile_list, &ovals, &maxv,
837                                         &minv);
838         } else
839                 len = 0;
840
841         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
842                 if (i >= len) {
843                         log_buf(out, ";0%%=0");
844                         continue;
845                 }
846                 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
847         }
848
849         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
850                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
851         else
852                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
853
854         if (ovals)
855                 free(ovals);
856
857         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
858                 double p_of_agg = 100.0;
859
860                 if (rs->agg[ddir]) {
861                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
862                         if (p_of_agg > 100.0)
863                                 p_of_agg = 100.0;
864                 }
865
866                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
867         } else
868                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
869 }
870
871 static void add_ddir_status_json(struct thread_stat *ts,
872                 struct group_run_stats *rs, int ddir, struct json_object *parent)
873 {
874         unsigned long min, max;
875         unsigned long long bw;
876         unsigned int *ovals = NULL;
877         double mean, dev, iops;
878         unsigned int len, minv, maxv;
879         int i;
880         const char *ddirname[] = {"read", "write", "trim"};
881         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
882         char buf[120];
883         double p_of_agg = 100.0;
884
885         assert(ddir_rw(ddir));
886
887         if (ts->unified_rw_rep && ddir != DDIR_READ)
888                 return;
889
890         dir_object = json_create_object();
891         json_object_add_value_object(parent,
892                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
893
894         bw = 0;
895         iops = 0.0;
896         if (ts->runtime[ddir]) {
897                 uint64_t runt = ts->runtime[ddir];
898
899                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
900                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
901         }
902
903         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
904         json_object_add_value_int(dir_object, "bw", bw);
905         json_object_add_value_float(dir_object, "iops", iops);
906         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
907         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
908         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
909         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
910
911         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
912                 min = max = 0;
913                 mean = dev = 0.0;
914         }
915         tmp_object = json_create_object();
916         json_object_add_value_object(dir_object, "slat", tmp_object);
917         json_object_add_value_int(tmp_object, "min", min);
918         json_object_add_value_int(tmp_object, "max", max);
919         json_object_add_value_float(tmp_object, "mean", mean);
920         json_object_add_value_float(tmp_object, "stddev", dev);
921
922         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
923                 min = max = 0;
924                 mean = dev = 0.0;
925         }
926         tmp_object = json_create_object();
927         json_object_add_value_object(dir_object, "clat", tmp_object);
928         json_object_add_value_int(tmp_object, "min", min);
929         json_object_add_value_int(tmp_object, "max", max);
930         json_object_add_value_float(tmp_object, "mean", mean);
931         json_object_add_value_float(tmp_object, "stddev", dev);
932
933         if (ts->clat_percentiles) {
934                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
935                                         ts->clat_stat[ddir].samples,
936                                         ts->percentile_list, &ovals, &maxv,
937                                         &minv);
938         } else
939                 len = 0;
940
941         percentile_object = json_create_object();
942         json_object_add_value_object(tmp_object, "percentile", percentile_object);
943         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
944                 if (i >= len) {
945                         json_object_add_value_int(percentile_object, "0.00", 0);
946                         continue;
947                 }
948                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
949                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
950         }
951
952         if (output_format & FIO_OUTPUT_JSON_PLUS) {
953                 clat_bins_object = json_create_object();
954                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
955                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
956                         snprintf(buf, sizeof(buf), "%d", i);
957                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
958                 }
959                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS", FIO_IO_U_PLAT_BITS);
960                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL", FIO_IO_U_PLAT_VAL);
961                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR", FIO_IO_U_PLAT_NR);
962         }
963
964         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
965                 min = max = 0;
966                 mean = dev = 0.0;
967         }
968         tmp_object = json_create_object();
969         json_object_add_value_object(dir_object, "lat", tmp_object);
970         json_object_add_value_int(tmp_object, "min", min);
971         json_object_add_value_int(tmp_object, "max", max);
972         json_object_add_value_float(tmp_object, "mean", mean);
973         json_object_add_value_float(tmp_object, "stddev", dev);
974         if (ovals)
975                 free(ovals);
976
977         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
978                 if (rs->agg[ddir]) {
979                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
980                         if (p_of_agg > 100.0)
981                                 p_of_agg = 100.0;
982                 }
983         } else {
984                 min = max = 0;
985                 p_of_agg = mean = dev = 0.0;
986         }
987         json_object_add_value_int(dir_object, "bw_min", min);
988         json_object_add_value_int(dir_object, "bw_max", max);
989         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
990         json_object_add_value_float(dir_object, "bw_mean", mean);
991         json_object_add_value_float(dir_object, "bw_dev", dev);
992 }
993
994 static void show_thread_status_terse_v2(struct thread_stat *ts,
995                                         struct group_run_stats *rs,
996                                         struct buf_output *out)
997 {
998         double io_u_dist[FIO_IO_U_MAP_NR];
999         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1000         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1001         double usr_cpu, sys_cpu;
1002         int i;
1003
1004         /* General Info */
1005         log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1006         /* Log Read Status */
1007         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1008         /* Log Write Status */
1009         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1010         /* Log Trim Status */
1011         show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1012
1013         /* CPU Usage */
1014         if (ts->total_run_time) {
1015                 double runt = (double) ts->total_run_time;
1016
1017                 usr_cpu = (double) ts->usr_time * 100 / runt;
1018                 sys_cpu = (double) ts->sys_time * 100 / runt;
1019         } else {
1020                 usr_cpu = 0;
1021                 sys_cpu = 0;
1022         }
1023
1024         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1025                                                 (unsigned long long) ts->ctx,
1026                                                 (unsigned long long) ts->majf,
1027                                                 (unsigned long long) ts->minf);
1028
1029         /* Calc % distribution of IO depths, usecond, msecond latency */
1030         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1031         stat_calc_lat_u(ts, io_u_lat_u);
1032         stat_calc_lat_m(ts, io_u_lat_m);
1033
1034         /* Only show fixed 7 I/O depth levels*/
1035         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1036                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1037                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1038
1039         /* Microsecond latency */
1040         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1041                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1042         /* Millisecond latency */
1043         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1044                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1045         /* Additional output if continue_on_error set - default off*/
1046         if (ts->continue_on_error)
1047                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1048         log_buf(out, "\n");
1049
1050         /* Additional output if description is set */
1051         if (strlen(ts->description))
1052                 log_buf(out, ";%s", ts->description);
1053
1054         log_buf(out, "\n");
1055 }
1056
1057 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1058                                            struct group_run_stats *rs, int ver,
1059                                            struct buf_output *out)
1060 {
1061         double io_u_dist[FIO_IO_U_MAP_NR];
1062         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1063         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1064         double usr_cpu, sys_cpu;
1065         int i;
1066
1067         /* General Info */
1068         log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1069                                         ts->name, ts->groupid, ts->error);
1070         /* Log Read Status */
1071         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1072         /* Log Write Status */
1073         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1074         /* Log Trim Status */
1075         if (ver == 4)
1076                 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1077
1078         /* CPU Usage */
1079         if (ts->total_run_time) {
1080                 double runt = (double) ts->total_run_time;
1081
1082                 usr_cpu = (double) ts->usr_time * 100 / runt;
1083                 sys_cpu = (double) ts->sys_time * 100 / runt;
1084         } else {
1085                 usr_cpu = 0;
1086                 sys_cpu = 0;
1087         }
1088
1089         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1090                                                 (unsigned long long) ts->ctx,
1091                                                 (unsigned long long) ts->majf,
1092                                                 (unsigned long long) ts->minf);
1093
1094         /* Calc % distribution of IO depths, usecond, msecond latency */
1095         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1096         stat_calc_lat_u(ts, io_u_lat_u);
1097         stat_calc_lat_m(ts, io_u_lat_m);
1098
1099         /* Only show fixed 7 I/O depth levels*/
1100         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1101                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1102                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1103
1104         /* Microsecond latency */
1105         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1106                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1107         /* Millisecond latency */
1108         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1109                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1110
1111         /* disk util stats, if any */
1112         show_disk_util(1, NULL, out);
1113
1114         /* Additional output if continue_on_error set - default off*/
1115         if (ts->continue_on_error)
1116                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1117
1118         /* Additional output if description is set */
1119         if (strlen(ts->description))
1120                 log_buf(out, ";%s", ts->description);
1121
1122         log_buf(out, "\n");
1123 }
1124
1125 static void json_add_job_opts(struct json_object *root, const char *name,
1126                               struct flist_head *opt_list, bool num_jobs)
1127 {
1128         struct json_object *dir_object;
1129         struct flist_head *entry;
1130         struct print_option *p;
1131
1132         if (flist_empty(opt_list))
1133                 return;
1134
1135         dir_object = json_create_object();
1136         json_object_add_value_object(root, name, dir_object);
1137
1138         flist_for_each(entry, opt_list) {
1139                 const char *pos = "";
1140
1141                 p = flist_entry(entry, struct print_option, list);
1142                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1143                         continue;
1144                 if (p->value)
1145                         pos = p->value;
1146                 json_object_add_value_string(dir_object, p->name, pos);
1147         }
1148 }
1149
1150 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1151                                                    struct group_run_stats *rs,
1152                                                    struct flist_head *opt_list)
1153 {
1154         struct json_object *root, *tmp;
1155         struct jobs_eta *je;
1156         double io_u_dist[FIO_IO_U_MAP_NR];
1157         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1158         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1159         double usr_cpu, sys_cpu;
1160         int i;
1161         size_t size;
1162
1163         root = json_create_object();
1164         json_object_add_value_string(root, "jobname", ts->name);
1165         json_object_add_value_int(root, "groupid", ts->groupid);
1166         json_object_add_value_int(root, "error", ts->error);
1167
1168         /* ETA Info */
1169         je = get_jobs_eta(true, &size);
1170         if (je) {
1171                 json_object_add_value_int(root, "eta", je->eta_sec);
1172                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1173         }
1174
1175         if (opt_list)
1176                 json_add_job_opts(root, "job options", opt_list, true);
1177
1178         add_ddir_status_json(ts, rs, DDIR_READ, root);
1179         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1180         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1181
1182         /* CPU Usage */
1183         if (ts->total_run_time) {
1184                 double runt = (double) ts->total_run_time;
1185
1186                 usr_cpu = (double) ts->usr_time * 100 / runt;
1187                 sys_cpu = (double) ts->sys_time * 100 / runt;
1188         } else {
1189                 usr_cpu = 0;
1190                 sys_cpu = 0;
1191         }
1192         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1193         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1194         json_object_add_value_int(root, "ctx", ts->ctx);
1195         json_object_add_value_int(root, "majf", ts->majf);
1196         json_object_add_value_int(root, "minf", ts->minf);
1197
1198
1199         /* Calc % distribution of IO depths, usecond, msecond latency */
1200         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1201         stat_calc_lat_u(ts, io_u_lat_u);
1202         stat_calc_lat_m(ts, io_u_lat_m);
1203
1204         tmp = json_create_object();
1205         json_object_add_value_object(root, "iodepth_level", tmp);
1206         /* Only show fixed 7 I/O depth levels*/
1207         for (i = 0; i < 7; i++) {
1208                 char name[20];
1209                 if (i < 6)
1210                         snprintf(name, 20, "%d", 1 << i);
1211                 else
1212                         snprintf(name, 20, ">=%d", 1 << i);
1213                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1214         }
1215
1216         tmp = json_create_object();
1217         json_object_add_value_object(root, "latency_us", tmp);
1218         /* Microsecond latency */
1219         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1220                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1221                                  "250", "500", "750", "1000", };
1222                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1223         }
1224         /* Millisecond latency */
1225         tmp = json_create_object();
1226         json_object_add_value_object(root, "latency_ms", tmp);
1227         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1228                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1229                                  "250", "500", "750", "1000", "2000",
1230                                  ">=2000", };
1231                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1232         }
1233
1234         /* Additional output if continue_on_error set - default off*/
1235         if (ts->continue_on_error) {
1236                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1237                 json_object_add_value_int(root, "first_error", ts->first_error);
1238         }
1239
1240         if (ts->latency_depth) {
1241                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1242                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1243                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1244                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1245         }
1246
1247         /* Additional output if description is set */
1248         if (strlen(ts->description))
1249                 json_object_add_value_string(root, "desc", ts->description);
1250
1251         if (ts->nr_block_infos) {
1252                 /* Block error histogram and types */
1253                 int len;
1254                 unsigned int *percentiles = NULL;
1255                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1256
1257                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1258                                              ts->percentile_list,
1259                                              &percentiles, block_state_counts);
1260
1261                 if (len) {
1262                         struct json_object *block, *percentile_object, *states;
1263                         int state;
1264                         block = json_create_object();
1265                         json_object_add_value_object(root, "block", block);
1266
1267                         percentile_object = json_create_object();
1268                         json_object_add_value_object(block, "percentiles",
1269                                                      percentile_object);
1270                         for (i = 0; i < len; i++) {
1271                                 char buf[20];
1272                                 snprintf(buf, sizeof(buf), "%f",
1273                                          ts->percentile_list[i].u.f);
1274                                 json_object_add_value_int(percentile_object,
1275                                                           (const char *)buf,
1276                                                           percentiles[i]);
1277                         }
1278
1279                         states = json_create_object();
1280                         json_object_add_value_object(block, "states", states);
1281                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1282                                 json_object_add_value_int(states,
1283                                         block_state_names[state],
1284                                         block_state_counts[state]);
1285                         }
1286                         free(percentiles);
1287                 }
1288         }
1289
1290         if (ts->ss_dur) {
1291                 struct json_object *data;
1292                 struct json_array *iops, *bw;
1293                 int i, j, k;
1294                 char ss_buf[64];
1295
1296                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1297                         ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1298                         ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1299                         (float) ts->ss_limit.u.f,
1300                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1301
1302                 tmp = json_create_object();
1303                 json_object_add_value_object(root, "steadystate", tmp);
1304                 json_object_add_value_string(tmp, "ss", ss_buf);
1305                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1306                 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1307
1308                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1309                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1310                 json_object_add_value_string(tmp, "criterion", ss_buf);
1311                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1312                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1313
1314                 data = json_create_object();
1315                 json_object_add_value_object(tmp, "data", data);
1316                 bw = json_create_array();
1317                 iops = json_create_array();
1318
1319                 /*
1320                 ** if ss was attained or the buffer is not full,
1321                 ** ss->head points to the first element in the list.
1322                 ** otherwise it actually points to the second element
1323                 ** in the list
1324                 */
1325                 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1326                         j = ts->ss_head;
1327                 else
1328                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1329                 for (i = 0; i < ts->ss_dur; i++) {
1330                         k = (j + i) % ts->ss_dur;
1331                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1332                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1333                 }
1334                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1335                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1336                 json_object_add_value_array(data, "iops", iops);
1337                 json_object_add_value_array(data, "bw", bw);
1338         }
1339
1340         return root;
1341 }
1342
1343 static void show_thread_status_terse(struct thread_stat *ts,
1344                                      struct group_run_stats *rs,
1345                                      struct buf_output *out)
1346 {
1347         if (terse_version == 2)
1348                 show_thread_status_terse_v2(ts, rs, out);
1349         else if (terse_version == 3 || terse_version == 4)
1350                 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1351         else
1352                 log_err("fio: bad terse version!? %d\n", terse_version);
1353 }
1354
1355 struct json_object *show_thread_status(struct thread_stat *ts,
1356                                        struct group_run_stats *rs,
1357                                        struct flist_head *opt_list,
1358                                        struct buf_output *out)
1359 {
1360         struct json_object *ret = NULL;
1361
1362         if (output_format & FIO_OUTPUT_TERSE)
1363                 show_thread_status_terse(ts, rs,  out);
1364         if (output_format & FIO_OUTPUT_JSON)
1365                 ret = show_thread_status_json(ts, rs, opt_list);
1366         if (output_format & FIO_OUTPUT_NORMAL)
1367                 show_thread_status_normal(ts, rs,  out);
1368
1369         return ret;
1370 }
1371
1372 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1373 {
1374         double mean, S;
1375
1376         if (src->samples == 0)
1377                 return;
1378
1379         dst->min_val = min(dst->min_val, src->min_val);
1380         dst->max_val = max(dst->max_val, src->max_val);
1381
1382         /*
1383          * Compute new mean and S after the merge
1384          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1385          *  #Parallel_algorithm>
1386          */
1387         if (first) {
1388                 mean = src->mean.u.f;
1389                 S = src->S.u.f;
1390         } else {
1391                 double delta = src->mean.u.f - dst->mean.u.f;
1392
1393                 mean = ((src->mean.u.f * src->samples) +
1394                         (dst->mean.u.f * dst->samples)) /
1395                         (dst->samples + src->samples);
1396
1397                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1398                         (dst->samples * src->samples) /
1399                         (dst->samples + src->samples);
1400         }
1401
1402         dst->samples += src->samples;
1403         dst->mean.u.f = mean;
1404         dst->S.u.f = S;
1405 }
1406
1407 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1408 {
1409         int i;
1410
1411         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1412                 if (dst->max_run[i] < src->max_run[i])
1413                         dst->max_run[i] = src->max_run[i];
1414                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1415                         dst->min_run[i] = src->min_run[i];
1416                 if (dst->max_bw[i] < src->max_bw[i])
1417                         dst->max_bw[i] = src->max_bw[i];
1418                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1419                         dst->min_bw[i] = src->min_bw[i];
1420
1421                 dst->iobytes[i] += src->iobytes[i];
1422                 dst->agg[i] += src->agg[i];
1423         }
1424
1425         if (!dst->kb_base)
1426                 dst->kb_base = src->kb_base;
1427         if (!dst->unit_base)
1428                 dst->unit_base = src->unit_base;
1429 }
1430
1431 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1432                       bool first)
1433 {
1434         int l, k;
1435
1436         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1437                 if (!dst->unified_rw_rep) {
1438                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1439                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1440                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1441                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1442
1443                         dst->io_bytes[l] += src->io_bytes[l];
1444
1445                         if (dst->runtime[l] < src->runtime[l])
1446                                 dst->runtime[l] = src->runtime[l];
1447                 } else {
1448                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1449                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1450                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1451                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1452
1453                         dst->io_bytes[0] += src->io_bytes[l];
1454
1455                         if (dst->runtime[0] < src->runtime[l])
1456                                 dst->runtime[0] = src->runtime[l];
1457
1458                         /*
1459                          * We're summing to the same destination, so override
1460                          * 'first' after the first iteration of the loop
1461                          */
1462                         first = false;
1463                 }
1464         }
1465
1466         dst->usr_time += src->usr_time;
1467         dst->sys_time += src->sys_time;
1468         dst->ctx += src->ctx;
1469         dst->majf += src->majf;
1470         dst->minf += src->minf;
1471
1472         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1473                 dst->io_u_map[k] += src->io_u_map[k];
1474         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1475                 dst->io_u_submit[k] += src->io_u_submit[k];
1476         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1477                 dst->io_u_complete[k] += src->io_u_complete[k];
1478         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1479                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1480         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1481                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1482
1483         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1484                 if (!dst->unified_rw_rep) {
1485                         dst->total_io_u[k] += src->total_io_u[k];
1486                         dst->short_io_u[k] += src->short_io_u[k];
1487                         dst->drop_io_u[k] += src->drop_io_u[k];
1488                 } else {
1489                         dst->total_io_u[0] += src->total_io_u[k];
1490                         dst->short_io_u[0] += src->short_io_u[k];
1491                         dst->drop_io_u[0] += src->drop_io_u[k];
1492                 }
1493         }
1494
1495         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1496                 int m;
1497
1498                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1499                         if (!dst->unified_rw_rep)
1500                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1501                         else
1502                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1503                 }
1504         }
1505
1506         dst->total_run_time += src->total_run_time;
1507         dst->total_submit += src->total_submit;
1508         dst->total_complete += src->total_complete;
1509 }
1510
1511 void init_group_run_stat(struct group_run_stats *gs)
1512 {
1513         int i;
1514         memset(gs, 0, sizeof(*gs));
1515
1516         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1517                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1518 }
1519
1520 void init_thread_stat(struct thread_stat *ts)
1521 {
1522         int j;
1523
1524         memset(ts, 0, sizeof(*ts));
1525
1526         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1527                 ts->lat_stat[j].min_val = -1UL;
1528                 ts->clat_stat[j].min_val = -1UL;
1529                 ts->slat_stat[j].min_val = -1UL;
1530                 ts->bw_stat[j].min_val = -1UL;
1531         }
1532         ts->groupid = -1;
1533 }
1534
1535 void __show_run_stats(void)
1536 {
1537         struct group_run_stats *runstats, *rs;
1538         struct thread_data *td;
1539         struct thread_stat *threadstats, *ts;
1540         int i, j, k, nr_ts, last_ts, idx;
1541         int kb_base_warned = 0;
1542         int unit_base_warned = 0;
1543         struct json_object *root = NULL;
1544         struct json_array *array = NULL;
1545         struct buf_output output[FIO_OUTPUT_NR];
1546         struct flist_head **opt_lists;
1547
1548         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1549
1550         for (i = 0; i < groupid + 1; i++)
1551                 init_group_run_stat(&runstats[i]);
1552
1553         /*
1554          * find out how many threads stats we need. if group reporting isn't
1555          * enabled, it's one-per-td.
1556          */
1557         nr_ts = 0;
1558         last_ts = -1;
1559         for_each_td(td, i) {
1560                 if (!td->o.group_reporting) {
1561                         nr_ts++;
1562                         continue;
1563                 }
1564                 if (last_ts == td->groupid)
1565                         continue;
1566
1567                 last_ts = td->groupid;
1568                 nr_ts++;
1569         }
1570
1571         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1572         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1573
1574         for (i = 0; i < nr_ts; i++) {
1575                 init_thread_stat(&threadstats[i]);
1576                 opt_lists[i] = NULL;
1577         }
1578
1579         j = 0;
1580         last_ts = -1;
1581         idx = 0;
1582         for_each_td(td, i) {
1583                 if (idx && (!td->o.group_reporting ||
1584                     (td->o.group_reporting && last_ts != td->groupid))) {
1585                         idx = 0;
1586                         j++;
1587                 }
1588
1589                 last_ts = td->groupid;
1590
1591                 ts = &threadstats[j];
1592
1593                 ts->clat_percentiles = td->o.clat_percentiles;
1594                 ts->percentile_precision = td->o.percentile_precision;
1595                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1596                 opt_lists[j] = &td->opt_list;
1597
1598                 idx++;
1599                 ts->members++;
1600
1601                 if (ts->groupid == -1) {
1602                         /*
1603                          * These are per-group shared already
1604                          */
1605                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1606                         if (td->o.description)
1607                                 strncpy(ts->description, td->o.description,
1608                                                 FIO_JOBDESC_SIZE - 1);
1609                         else
1610                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1611
1612                         /*
1613                          * If multiple entries in this group, this is
1614                          * the first member.
1615                          */
1616                         ts->thread_number = td->thread_number;
1617                         ts->groupid = td->groupid;
1618
1619                         /*
1620                          * first pid in group, not very useful...
1621                          */
1622                         ts->pid = td->pid;
1623
1624                         ts->kb_base = td->o.kb_base;
1625                         ts->unit_base = td->o.unit_base;
1626                         ts->unified_rw_rep = td->o.unified_rw_rep;
1627                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1628                         log_info("fio: kb_base differs for jobs in group, using"
1629                                  " %u as the base\n", ts->kb_base);
1630                         kb_base_warned = 1;
1631                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1632                         log_info("fio: unit_base differs for jobs in group, using"
1633                                  " %u as the base\n", ts->unit_base);
1634                         unit_base_warned = 1;
1635                 }
1636
1637                 ts->continue_on_error = td->o.continue_on_error;
1638                 ts->total_err_count += td->total_err_count;
1639                 ts->first_error = td->first_error;
1640                 if (!ts->error) {
1641                         if (!td->error && td->o.continue_on_error &&
1642                             td->first_error) {
1643                                 ts->error = td->first_error;
1644                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1645                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1646                         } else  if (td->error) {
1647                                 ts->error = td->error;
1648                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1649                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1650                         }
1651                 }
1652
1653                 ts->latency_depth = td->latency_qd;
1654                 ts->latency_target = td->o.latency_target;
1655                 ts->latency_percentile = td->o.latency_percentile;
1656                 ts->latency_window = td->o.latency_window;
1657
1658                 ts->nr_block_infos = td->ts.nr_block_infos;
1659                 for (k = 0; k < ts->nr_block_infos; k++)
1660                         ts->block_infos[k] = td->ts.block_infos[k];
1661
1662                 sum_thread_stats(ts, &td->ts, idx == 1);
1663
1664                 if (td->o.ss_dur) {
1665                         ts->ss_state = td->ss.state;
1666                         ts->ss_dur = td->ss.dur;
1667                         ts->ss_head = td->ss.head;
1668                         ts->ss_bw_data = td->ss.bw_data;
1669                         ts->ss_iops_data = td->ss.iops_data;
1670                         ts->ss_limit.u.f = td->ss.limit;
1671                         ts->ss_slope.u.f = td->ss.slope;
1672                         ts->ss_deviation.u.f = td->ss.deviation;
1673                         ts->ss_criterion.u.f = td->ss.criterion;
1674                 }
1675                 else
1676                         ts->ss_dur = ts->ss_state = 0;
1677         }
1678
1679         for (i = 0; i < nr_ts; i++) {
1680                 unsigned long long bw;
1681
1682                 ts = &threadstats[i];
1683                 if (ts->groupid == -1)
1684                         continue;
1685                 rs = &runstats[ts->groupid];
1686                 rs->kb_base = ts->kb_base;
1687                 rs->unit_base = ts->unit_base;
1688                 rs->unified_rw_rep += ts->unified_rw_rep;
1689
1690                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1691                         if (!ts->runtime[j])
1692                                 continue;
1693                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1694                                 rs->min_run[j] = ts->runtime[j];
1695                         if (ts->runtime[j] > rs->max_run[j])
1696                                 rs->max_run[j] = ts->runtime[j];
1697
1698                         bw = 0;
1699                         if (ts->runtime[j])
1700                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1701                         if (bw < rs->min_bw[j])
1702                                 rs->min_bw[j] = bw;
1703                         if (bw > rs->max_bw[j])
1704                                 rs->max_bw[j] = bw;
1705
1706                         rs->iobytes[j] += ts->io_bytes[j];
1707                 }
1708         }
1709
1710         for (i = 0; i < groupid + 1; i++) {
1711                 int ddir;
1712
1713                 rs = &runstats[i];
1714
1715                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1716                         if (rs->max_run[ddir])
1717                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1718                                                 rs->max_run[ddir];
1719                 }
1720         }
1721
1722         for (i = 0; i < FIO_OUTPUT_NR; i++)
1723                 buf_output_init(&output[i]);
1724
1725         /*
1726          * don't overwrite last signal output
1727          */
1728         if (output_format & FIO_OUTPUT_NORMAL)
1729                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1730         if (output_format & FIO_OUTPUT_JSON) {
1731                 struct thread_data *global;
1732                 char time_buf[32];
1733                 struct timeval now;
1734                 unsigned long long ms_since_epoch;
1735
1736                 gettimeofday(&now, NULL);
1737                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1738                                  (unsigned long long)(now.tv_usec) / 1000;
1739
1740                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1741                                 sizeof(time_buf));
1742                 if (time_buf[strlen(time_buf) - 1] == '\n')
1743                         time_buf[strlen(time_buf) - 1] = '\0';
1744
1745                 root = json_create_object();
1746                 json_object_add_value_string(root, "fio version", fio_version_string);
1747                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1748                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1749                 json_object_add_value_string(root, "time", time_buf);
1750                 global = get_global_options();
1751                 json_add_job_opts(root, "global options", &global->opt_list, false);
1752                 array = json_create_array();
1753                 json_object_add_value_array(root, "jobs", array);
1754         }
1755
1756         if (is_backend)
1757                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1758
1759         for (i = 0; i < nr_ts; i++) {
1760                 ts = &threadstats[i];
1761                 rs = &runstats[ts->groupid];
1762
1763                 if (is_backend) {
1764                         fio_server_send_job_options(opt_lists[i], i);
1765                         fio_server_send_ts(ts, rs);
1766                 } else {
1767                         if (output_format & FIO_OUTPUT_TERSE)
1768                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1769                         if (output_format & FIO_OUTPUT_JSON) {
1770                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1771                                 json_array_add_value_object(array, tmp);
1772                         }
1773                         if (output_format & FIO_OUTPUT_NORMAL)
1774                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1775                 }
1776         }
1777         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1778                 /* disk util stats, if any */
1779                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1780
1781                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1782
1783                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1784                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1785                 json_free_object(root);
1786         }
1787
1788         for (i = 0; i < groupid + 1; i++) {
1789                 rs = &runstats[i];
1790
1791                 rs->groupid = i;
1792                 if (is_backend)
1793                         fio_server_send_gs(rs);
1794                 else if (output_format & FIO_OUTPUT_NORMAL)
1795                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1796         }
1797
1798         if (is_backend)
1799                 fio_server_send_du();
1800         else if (output_format & FIO_OUTPUT_NORMAL) {
1801                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1802                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1803         }
1804
1805         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1806                 buf_output_flush(&output[i]);
1807                 buf_output_free(&output[i]);
1808         }
1809
1810         log_info_flush();
1811         free(runstats);
1812         free(threadstats);
1813         free(opt_lists);
1814 }
1815
1816 void show_run_stats(void)
1817 {
1818         fio_mutex_down(stat_mutex);
1819         __show_run_stats();
1820         fio_mutex_up(stat_mutex);
1821 }
1822
1823 void __show_running_run_stats(void)
1824 {
1825         struct thread_data *td;
1826         unsigned long long *rt;
1827         struct timeval tv;
1828         int i;
1829
1830         fio_mutex_down(stat_mutex);
1831
1832         rt = malloc(thread_number * sizeof(unsigned long long));
1833         fio_gettime(&tv, NULL);
1834
1835         for_each_td(td, i) {
1836                 td->update_rusage = 1;
1837                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1838                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1839                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1840                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1841
1842                 rt[i] = mtime_since(&td->start, &tv);
1843                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1844                         td->ts.runtime[DDIR_READ] += rt[i];
1845                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1846                         td->ts.runtime[DDIR_WRITE] += rt[i];
1847                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1848                         td->ts.runtime[DDIR_TRIM] += rt[i];
1849         }
1850
1851         for_each_td(td, i) {
1852                 if (td->runstate >= TD_EXITED)
1853                         continue;
1854                 if (td->rusage_sem) {
1855                         td->update_rusage = 1;
1856                         fio_mutex_down(td->rusage_sem);
1857                 }
1858                 td->update_rusage = 0;
1859         }
1860
1861         __show_run_stats();
1862
1863         for_each_td(td, i) {
1864                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1865                         td->ts.runtime[DDIR_READ] -= rt[i];
1866                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1867                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1868                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1869                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1870         }
1871
1872         free(rt);
1873         fio_mutex_up(stat_mutex);
1874 }
1875
1876 static int status_interval_init;
1877 static struct timeval status_time;
1878 static int status_file_disabled;
1879
1880 #define FIO_STATUS_FILE         "fio-dump-status"
1881
1882 static int check_status_file(void)
1883 {
1884         struct stat sb;
1885         const char *temp_dir;
1886         char fio_status_file_path[PATH_MAX];
1887
1888         if (status_file_disabled)
1889                 return 0;
1890
1891         temp_dir = getenv("TMPDIR");
1892         if (temp_dir == NULL) {
1893                 temp_dir = getenv("TEMP");
1894                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1895                         temp_dir = NULL;
1896         }
1897         if (temp_dir == NULL)
1898                 temp_dir = "/tmp";
1899
1900         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1901
1902         if (stat(fio_status_file_path, &sb))
1903                 return 0;
1904
1905         if (unlink(fio_status_file_path) < 0) {
1906                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1907                                                         strerror(errno));
1908                 log_err("fio: disabling status file updates\n");
1909                 status_file_disabled = 1;
1910         }
1911
1912         return 1;
1913 }
1914
1915 void check_for_running_stats(void)
1916 {
1917         if (status_interval) {
1918                 if (!status_interval_init) {
1919                         fio_gettime(&status_time, NULL);
1920                         status_interval_init = 1;
1921                 } else if (mtime_since_now(&status_time) >= status_interval) {
1922                         show_running_run_stats();
1923                         fio_gettime(&status_time, NULL);
1924                         return;
1925                 }
1926         }
1927         if (check_status_file()) {
1928                 show_running_run_stats();
1929                 return;
1930         }
1931 }
1932
1933 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1934 {
1935         double val = data;
1936         double delta;
1937
1938         if (data > is->max_val)
1939                 is->max_val = data;
1940         if (data < is->min_val)
1941                 is->min_val = data;
1942
1943         delta = val - is->mean.u.f;
1944         if (delta) {
1945                 is->mean.u.f += delta / (is->samples + 1.0);
1946                 is->S.u.f += delta * (val - is->mean.u.f);
1947         }
1948
1949         is->samples++;
1950 }
1951
1952 /*
1953  * Return a struct io_logs, which is added to the tail of the log
1954  * list for 'iolog'.
1955  */
1956 static struct io_logs *get_new_log(struct io_log *iolog)
1957 {
1958         size_t new_size, new_samples;
1959         struct io_logs *cur_log;
1960
1961         /*
1962          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
1963          * forever
1964          */
1965         if (!iolog->cur_log_max)
1966                 new_samples = DEF_LOG_ENTRIES;
1967         else {
1968                 new_samples = iolog->cur_log_max * 2;
1969                 if (new_samples > MAX_LOG_ENTRIES)
1970                         new_samples = MAX_LOG_ENTRIES;
1971         }
1972
1973         new_size = new_samples * log_entry_sz(iolog);
1974
1975         cur_log = smalloc(sizeof(*cur_log));
1976         if (cur_log) {
1977                 INIT_FLIST_HEAD(&cur_log->list);
1978                 cur_log->log = malloc(new_size);
1979                 if (cur_log->log) {
1980                         cur_log->nr_samples = 0;
1981                         cur_log->max_samples = new_samples;
1982                         flist_add_tail(&cur_log->list, &iolog->io_logs);
1983                         iolog->cur_log_max = new_samples;
1984                         return cur_log;
1985                 }
1986                 sfree(cur_log);
1987         }
1988
1989         return NULL;
1990 }
1991
1992 /*
1993  * Add and return a new log chunk, or return current log if big enough
1994  */
1995 static struct io_logs *regrow_log(struct io_log *iolog)
1996 {
1997         struct io_logs *cur_log;
1998         int i;
1999
2000         if (!iolog || iolog->disabled)
2001                 goto disable;
2002
2003         cur_log = iolog_cur_log(iolog);
2004         if (!cur_log) {
2005                 cur_log = get_new_log(iolog);
2006                 if (!cur_log)
2007                         return NULL;
2008         }
2009
2010         if (cur_log->nr_samples < cur_log->max_samples)
2011                 return cur_log;
2012
2013         /*
2014          * No room for a new sample. If we're compressing on the fly, flush
2015          * out the current chunk
2016          */
2017         if (iolog->log_gz) {
2018                 if (iolog_cur_flush(iolog, cur_log)) {
2019                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2020                         return NULL;
2021                 }
2022         }
2023
2024         /*
2025          * Get a new log array, and add to our list
2026          */
2027         cur_log = get_new_log(iolog);
2028         if (!cur_log) {
2029                 log_err("fio: failed extending iolog! Will stop logging.\n");
2030                 return NULL;
2031         }
2032
2033         if (!iolog->pending || !iolog->pending->nr_samples)
2034                 return cur_log;
2035
2036         /*
2037          * Flush pending items to new log
2038          */
2039         for (i = 0; i < iolog->pending->nr_samples; i++) {
2040                 struct io_sample *src, *dst;
2041
2042                 src = get_sample(iolog, iolog->pending, i);
2043                 dst = get_sample(iolog, cur_log, i);
2044                 memcpy(dst, src, log_entry_sz(iolog));
2045         }
2046         cur_log->nr_samples = iolog->pending->nr_samples;
2047
2048         iolog->pending->nr_samples = 0;
2049         return cur_log;
2050 disable:
2051         if (iolog)
2052                 iolog->disabled = true;
2053         return NULL;
2054 }
2055
2056 void regrow_logs(struct thread_data *td)
2057 {
2058         regrow_log(td->slat_log);
2059         regrow_log(td->clat_log);
2060         regrow_log(td->clat_hist_log);
2061         regrow_log(td->lat_log);
2062         regrow_log(td->bw_log);
2063         regrow_log(td->iops_log);
2064         td->flags &= ~TD_F_REGROW_LOGS;
2065 }
2066
2067 static struct io_logs *get_cur_log(struct io_log *iolog)
2068 {
2069         struct io_logs *cur_log;
2070
2071         cur_log = iolog_cur_log(iolog);
2072         if (!cur_log) {
2073                 cur_log = get_new_log(iolog);
2074                 if (!cur_log)
2075                         return NULL;
2076         }
2077
2078         if (cur_log->nr_samples < cur_log->max_samples)
2079                 return cur_log;
2080
2081         /*
2082          * Out of space. If we're in IO offload mode, or we're not doing
2083          * per unit logging (hence logging happens outside of the IO thread
2084          * as well), add a new log chunk inline. If we're doing inline
2085          * submissions, flag 'td' as needing a log regrow and we'll take
2086          * care of it on the submission side.
2087          */
2088         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2089             !per_unit_log(iolog))
2090                 return regrow_log(iolog);
2091
2092         iolog->td->flags |= TD_F_REGROW_LOGS;
2093         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2094         return iolog->pending;
2095 }
2096
2097 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2098                              enum fio_ddir ddir, unsigned int bs,
2099                              unsigned long t, uint64_t offset)
2100 {
2101         struct io_logs *cur_log;
2102
2103         if (iolog->disabled)
2104                 return;
2105         if (flist_empty(&iolog->io_logs))
2106                 iolog->avg_last = t;
2107
2108         cur_log = get_cur_log(iolog);
2109         if (cur_log) {
2110                 struct io_sample *s;
2111
2112                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2113
2114                 s->data = data;
2115                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2116                 io_sample_set_ddir(iolog, s, ddir);
2117                 s->bs = bs;
2118
2119                 if (iolog->log_offset) {
2120                         struct io_sample_offset *so = (void *) s;
2121
2122                         so->offset = offset;
2123                 }
2124
2125                 cur_log->nr_samples++;
2126                 return;
2127         }
2128
2129         iolog->disabled = true;
2130 }
2131
2132 static inline void reset_io_stat(struct io_stat *ios)
2133 {
2134         ios->max_val = ios->min_val = ios->samples = 0;
2135         ios->mean.u.f = ios->S.u.f = 0;
2136 }
2137
2138 void reset_io_stats(struct thread_data *td)
2139 {
2140         struct thread_stat *ts = &td->ts;
2141         int i, j;
2142
2143         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2144                 reset_io_stat(&ts->clat_stat[i]);
2145                 reset_io_stat(&ts->slat_stat[i]);
2146                 reset_io_stat(&ts->lat_stat[i]);
2147                 reset_io_stat(&ts->bw_stat[i]);
2148                 reset_io_stat(&ts->iops_stat[i]);
2149
2150                 ts->io_bytes[i] = 0;
2151                 ts->runtime[i] = 0;
2152
2153                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2154                         ts->io_u_plat[i][j] = 0;
2155         }
2156
2157         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2158                 ts->io_u_map[i] = 0;
2159                 ts->io_u_submit[i] = 0;
2160                 ts->io_u_complete[i] = 0;
2161                 ts->io_u_lat_u[i] = 0;
2162                 ts->io_u_lat_m[i] = 0;
2163                 ts->total_submit = 0;
2164                 ts->total_complete = 0;
2165         }
2166
2167         for (i = 0; i < 3; i++) {
2168                 ts->total_io_u[i] = 0;
2169                 ts->short_io_u[i] = 0;
2170                 ts->drop_io_u[i] = 0;
2171         }
2172 }
2173
2174 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2175                               unsigned long elapsed, bool log_max)
2176 {
2177         /*
2178          * Note an entry in the log. Use the mean from the logged samples,
2179          * making sure to properly round up. Only write a log entry if we
2180          * had actual samples done.
2181          */
2182         if (iolog->avg_window[ddir].samples) {
2183                 union io_sample_data data;
2184
2185                 if (log_max)
2186                         data.val = iolog->avg_window[ddir].max_val;
2187                 else
2188                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2189
2190                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2191         }
2192
2193         reset_io_stat(&iolog->avg_window[ddir]);
2194 }
2195
2196 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2197                              bool log_max)
2198 {
2199         int ddir;
2200
2201         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2202                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2203 }
2204
2205 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2206                            union io_sample_data data, enum fio_ddir ddir,
2207                            unsigned int bs, uint64_t offset)
2208 {
2209         unsigned long elapsed, this_window;
2210
2211         if (!ddir_rw(ddir))
2212                 return 0;
2213
2214         elapsed = mtime_since_now(&td->epoch);
2215
2216         /*
2217          * If no time averaging, just add the log sample.
2218          */
2219         if (!iolog->avg_msec) {
2220                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2221                 return 0;
2222         }
2223
2224         /*
2225          * Add the sample. If the time period has passed, then
2226          * add that entry to the log and clear.
2227          */
2228         add_stat_sample(&iolog->avg_window[ddir], data.val);
2229
2230         /*
2231          * If period hasn't passed, adding the above sample is all we
2232          * need to do.
2233          */
2234         this_window = elapsed - iolog->avg_last;
2235         if (elapsed < iolog->avg_last)
2236                 return iolog->avg_last - elapsed;
2237         else if (this_window < iolog->avg_msec) {
2238                 int diff = iolog->avg_msec - this_window;
2239
2240                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2241                         return diff;
2242         }
2243
2244         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2245
2246         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2247         return iolog->avg_msec;
2248 }
2249
2250 void finalize_logs(struct thread_data *td, bool unit_logs)
2251 {
2252         unsigned long elapsed;
2253
2254         elapsed = mtime_since_now(&td->epoch);
2255
2256         if (td->clat_log && unit_logs)
2257                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2258         if (td->slat_log && unit_logs)
2259                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2260         if (td->lat_log && unit_logs)
2261                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2262         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2263                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2264         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2265                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2266 }
2267
2268 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2269 {
2270         struct io_log *iolog;
2271
2272         if (!ddir_rw(ddir))
2273                 return;
2274
2275         iolog = agg_io_log[ddir];
2276         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2277 }
2278
2279 static void add_clat_percentile_sample(struct thread_stat *ts,
2280                                 unsigned long usec, enum fio_ddir ddir)
2281 {
2282         unsigned int idx = plat_val_to_idx(usec);
2283         assert(idx < FIO_IO_U_PLAT_NR);
2284
2285         ts->io_u_plat[ddir][idx]++;
2286 }
2287
2288 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2289                      unsigned long usec, unsigned int bs, uint64_t offset)
2290 {
2291         unsigned long elapsed, this_window;
2292         struct thread_stat *ts = &td->ts;
2293         struct io_log *iolog = td->clat_hist_log;
2294
2295         td_io_u_lock(td);
2296
2297         add_stat_sample(&ts->clat_stat[ddir], usec);
2298
2299         if (td->clat_log)
2300                 add_log_sample(td, td->clat_log, sample_val(usec), ddir, bs,
2301                                offset);
2302
2303         if (ts->clat_percentiles)
2304                 add_clat_percentile_sample(ts, usec, ddir);
2305
2306         if (iolog && iolog->hist_msec) {
2307                 struct io_hist *hw = &iolog->hist_window[ddir];
2308
2309                 hw->samples++;
2310                 elapsed = mtime_since_now(&td->epoch);
2311                 if (!hw->hist_last)
2312                         hw->hist_last = elapsed;
2313                 this_window = elapsed - hw->hist_last;
2314                 
2315                 if (this_window >= iolog->hist_msec) {
2316                         unsigned int *io_u_plat;
2317                         struct io_u_plat_entry *dst;
2318
2319                         /*
2320                          * Make a byte-for-byte copy of the latency histogram
2321                          * stored in td->ts.io_u_plat[ddir], recording it in a
2322                          * log sample. Note that the matching call to free() is
2323                          * located in iolog.c after printing this sample to the
2324                          * log file.
2325                          */
2326                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2327                         dst = malloc(sizeof(struct io_u_plat_entry));
2328                         memcpy(&(dst->io_u_plat), io_u_plat,
2329                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2330                         flist_add(&dst->list, &hw->list);
2331                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2332                                                 elapsed, offset);
2333
2334                         /*
2335                          * Update the last time we recorded as being now, minus
2336                          * any drift in time we encountered before actually
2337                          * making the record.
2338                          */
2339                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2340                         hw->samples = 0;
2341                 }
2342         }
2343
2344         td_io_u_unlock(td);
2345 }
2346
2347 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2348                      unsigned long usec, unsigned int bs, uint64_t offset)
2349 {
2350         struct thread_stat *ts = &td->ts;
2351
2352         if (!ddir_rw(ddir))
2353                 return;
2354
2355         td_io_u_lock(td);
2356
2357         add_stat_sample(&ts->slat_stat[ddir], usec);
2358
2359         if (td->slat_log)
2360                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2361
2362         td_io_u_unlock(td);
2363 }
2364
2365 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2366                     unsigned long usec, unsigned int bs, uint64_t offset)
2367 {
2368         struct thread_stat *ts = &td->ts;
2369
2370         if (!ddir_rw(ddir))
2371                 return;
2372
2373         td_io_u_lock(td);
2374
2375         add_stat_sample(&ts->lat_stat[ddir], usec);
2376
2377         if (td->lat_log)
2378                 add_log_sample(td, td->lat_log, sample_val(usec), ddir, bs,
2379                                offset);
2380
2381         td_io_u_unlock(td);
2382 }
2383
2384 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2385                    unsigned int bytes, unsigned long spent)
2386 {
2387         struct thread_stat *ts = &td->ts;
2388         unsigned long rate;
2389
2390         if (spent)
2391                 rate = bytes * 1000 / spent;
2392         else
2393                 rate = 0;
2394
2395         td_io_u_lock(td);
2396
2397         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2398
2399         if (td->bw_log)
2400                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2401                                bytes, io_u->offset);
2402
2403         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2404         td_io_u_unlock(td);
2405 }
2406
2407 static int add_bw_samples(struct thread_data *td, struct timeval *t)
2408 {
2409         struct thread_stat *ts = &td->ts;
2410         unsigned long spent, rate;
2411         enum fio_ddir ddir;
2412         unsigned int next, next_log;
2413
2414         next_log = td->o.bw_avg_time;
2415
2416         spent = mtime_since(&td->bw_sample_time, t);
2417         if (spent < td->o.bw_avg_time &&
2418             td->o.bw_avg_time - spent >= LOG_MSEC_SLACK)
2419                 return td->o.bw_avg_time - spent;
2420
2421         td_io_u_lock(td);
2422
2423         /*
2424          * Compute both read and write rates for the interval.
2425          */
2426         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2427                 uint64_t delta;
2428
2429                 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2430                 if (!delta)
2431                         continue; /* No entries for interval */
2432
2433                 if (spent)
2434                         rate = delta * 1000 / spent / 1024;
2435                 else
2436                         rate = 0;
2437
2438                 add_stat_sample(&ts->bw_stat[ddir], rate);
2439
2440                 if (td->bw_log) {
2441                         unsigned int bs = 0;
2442
2443                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2444                                 bs = td->o.min_bs[ddir];
2445
2446                         next = add_log_sample(td, td->bw_log, sample_val(rate),
2447                                               ddir, bs, 0);
2448                         next_log = min(next_log, next);
2449                 }
2450
2451                 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2452         }
2453
2454         timeval_add_msec(&td->bw_sample_time, td->o.bw_avg_time);
2455
2456         td_io_u_unlock(td);
2457
2458         if (spent <= td->o.bw_avg_time)
2459                 return min(next_log, td->o.bw_avg_time);
2460
2461         next = td->o.bw_avg_time - (1 + spent - td->o.bw_avg_time);
2462         return min(next, next_log);
2463 }
2464
2465 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2466                      unsigned int bytes)
2467 {
2468         struct thread_stat *ts = &td->ts;
2469
2470         td_io_u_lock(td);
2471
2472         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2473
2474         if (td->iops_log)
2475                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2476                                bytes, io_u->offset);
2477
2478         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2479         td_io_u_unlock(td);
2480 }
2481
2482 static int add_iops_samples(struct thread_data *td, struct timeval *t)
2483 {
2484         struct thread_stat *ts = &td->ts;
2485         unsigned long spent, iops;
2486         enum fio_ddir ddir;
2487         unsigned int next, next_log;
2488
2489         next_log = td->o.iops_avg_time;
2490
2491         spent = mtime_since(&td->iops_sample_time, t);
2492         if (spent < td->o.iops_avg_time &&
2493             td->o.iops_avg_time - spent >= LOG_MSEC_SLACK)
2494                 return td->o.iops_avg_time - spent;
2495
2496         td_io_u_lock(td);
2497
2498         /*
2499          * Compute both read and write rates for the interval.
2500          */
2501         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2502                 uint64_t delta;
2503
2504                 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2505                 if (!delta)
2506                         continue; /* No entries for interval */
2507
2508                 if (spent)
2509                         iops = (delta * 1000) / spent;
2510                 else
2511                         iops = 0;
2512
2513                 add_stat_sample(&ts->iops_stat[ddir], iops);
2514
2515                 if (td->iops_log) {
2516                         unsigned int bs = 0;
2517
2518                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2519                                 bs = td->o.min_bs[ddir];
2520
2521                         next = add_log_sample(td, td->iops_log,
2522                                               sample_val(iops), ddir, bs, 0);
2523                         next_log = min(next_log, next);
2524                 }
2525
2526                 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2527         }
2528
2529         timeval_add_msec(&td->iops_sample_time, td->o.iops_avg_time);
2530
2531         td_io_u_unlock(td);
2532
2533         if (spent <= td->o.iops_avg_time)
2534                 return min(next_log, td->o.iops_avg_time);
2535
2536         next = td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
2537         return min(next, next_log);
2538 }
2539
2540 /*
2541  * Returns msecs to next event
2542  */
2543 int calc_log_samples(void)
2544 {
2545         struct thread_data *td;
2546         unsigned int next = ~0U, tmp;
2547         struct timeval now;
2548         int i;
2549
2550         fio_gettime(&now, NULL);
2551
2552         for_each_td(td, i) {
2553                 if (in_ramp_time(td) ||
2554                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2555                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2556                         continue;
2557                 }
2558                 if (td->bw_log && !per_unit_log(td->bw_log)) {
2559                         tmp = add_bw_samples(td, &now);
2560                         if (tmp < next)
2561                                 next = tmp;
2562                 }
2563                 if (td->iops_log && !per_unit_log(td->iops_log)) {
2564                         tmp = add_iops_samples(td, &now);
2565                         if (tmp < next)
2566                                 next = tmp;
2567                 }
2568         }
2569
2570         return next == ~0U ? 0 : next;
2571 }
2572
2573 void stat_init(void)
2574 {
2575         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2576 }
2577
2578 void stat_exit(void)
2579 {
2580         /*
2581          * When we have the mutex, we know out-of-band access to it
2582          * have ended.
2583          */
2584         fio_mutex_down(stat_mutex);
2585         fio_mutex_remove(stat_mutex);
2586 }
2587
2588 /*
2589  * Called from signal handler. Wake up status thread.
2590  */
2591 void show_running_run_stats(void)
2592 {
2593         helper_do_stat();
2594 }
2595
2596 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2597 {
2598         /* Ignore io_u's which span multiple blocks--they will just get
2599          * inaccurate counts. */
2600         int idx = (io_u->offset - io_u->file->file_offset)
2601                         / td->o.bs[DDIR_TRIM];
2602         uint32_t *info = &td->ts.block_infos[idx];
2603         assert(idx < td->ts.nr_block_infos);
2604         return info;
2605 }