treat error in addr conversion to string as non-fatal
[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
18 struct fio_mutex *stat_mutex;
19
20 void update_rusage_stat(struct thread_data *td)
21 {
22         struct thread_stat *ts = &td->ts;
23
24         fio_getrusage(&td->ru_end);
25         ts->usr_time += mtime_since(&td->ru_start.ru_utime,
26                                         &td->ru_end.ru_utime);
27         ts->sys_time += mtime_since(&td->ru_start.ru_stime,
28                                         &td->ru_end.ru_stime);
29         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
30                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
31         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
32         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
33
34         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
35 }
36
37 /*
38  * Given a latency, return the index of the corresponding bucket in
39  * the structure tracking percentiles.
40  *
41  * (1) find the group (and error bits) that the value (latency)
42  * belongs to by looking at its MSB. (2) find the bucket number in the
43  * group by looking at the index bits.
44  *
45  */
46 static unsigned int plat_val_to_idx(unsigned int val)
47 {
48         unsigned int msb, error_bits, base, offset, idx;
49
50         /* Find MSB starting from bit 0 */
51         if (val == 0)
52                 msb = 0;
53         else
54                 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
55
56         /*
57          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
58          * all bits of the sample as index
59          */
60         if (msb <= FIO_IO_U_PLAT_BITS)
61                 return val;
62
63         /* Compute the number of error bits to discard*/
64         error_bits = msb - FIO_IO_U_PLAT_BITS;
65
66         /* Compute the number of buckets before the group */
67         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
68
69         /*
70          * Discard the error bits and apply the mask to find the
71          * index for the buckets in the group
72          */
73         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
74
75         /* Make sure the index does not exceed (array size - 1) */
76         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
77                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
78
79         return idx;
80 }
81
82 /*
83  * Convert the given index of the bucket array to the value
84  * represented by the bucket
85  */
86 static unsigned int plat_idx_to_val(unsigned int idx)
87 {
88         unsigned int error_bits, k, base;
89
90         assert(idx < FIO_IO_U_PLAT_NR);
91
92         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
93          * all bits of the sample as index */
94         if (idx < (FIO_IO_U_PLAT_VAL << 1))
95                 return idx;
96
97         /* Find the group and compute the minimum value of that group */
98         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
99         base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
100
101         /* Find its bucket number of the group */
102         k = idx % FIO_IO_U_PLAT_VAL;
103
104         /* Return the mean of the range of the bucket */
105         return base + ((k + 0.5) * (1 << error_bits));
106 }
107
108 static int double_cmp(const void *a, const void *b)
109 {
110         const fio_fp64_t fa = *(const fio_fp64_t *) a;
111         const fio_fp64_t fb = *(const fio_fp64_t *) b;
112         int cmp = 0;
113
114         if (fa.u.f > fb.u.f)
115                 cmp = 1;
116         else if (fa.u.f < fb.u.f)
117                 cmp = -1;
118
119         return cmp;
120 }
121
122 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
123                                    fio_fp64_t *plist, unsigned int **output,
124                                    unsigned int *maxv, unsigned int *minv)
125 {
126         unsigned long sum = 0;
127         unsigned int len, i, j = 0;
128         unsigned int oval_len = 0;
129         unsigned int *ovals = NULL;
130         int is_last;
131
132         *minv = -1U;
133         *maxv = 0;
134
135         len = 0;
136         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
137                 len++;
138
139         if (!len)
140                 return 0;
141
142         /*
143          * Sort the percentile list. Note that it may already be sorted if
144          * we are using the default values, but since it's a short list this
145          * isn't a worry. Also note that this does not work for NaN values.
146          */
147         if (len > 1)
148                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
149
150         /*
151          * Calculate bucket values, note down max and min values
152          */
153         is_last = 0;
154         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
155                 sum += io_u_plat[i];
156                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
157                         assert(plist[j].u.f <= 100.0);
158
159                         if (j == oval_len) {
160                                 oval_len += 100;
161                                 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
162                         }
163
164                         ovals[j] = plat_idx_to_val(i);
165                         if (ovals[j] < *minv)
166                                 *minv = ovals[j];
167                         if (ovals[j] > *maxv)
168                                 *maxv = ovals[j];
169
170                         is_last = (j == len - 1);
171                         if (is_last)
172                                 break;
173
174                         j++;
175                 }
176         }
177
178         *output = ovals;
179         return len;
180 }
181
182 /*
183  * Find and display the p-th percentile of clat
184  */
185 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
186                                   fio_fp64_t *plist, unsigned int precision)
187 {
188         unsigned int len, j = 0, minv, maxv;
189         unsigned int *ovals;
190         int is_last, per_line, scale_down;
191         char fmt[32];
192
193         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
194         if (!len)
195                 goto out;
196
197         /*
198          * We default to usecs, but if the value range is such that we
199          * should scale down to msecs, do that.
200          */
201         if (minv > 2000 && maxv > 99999) {
202                 scale_down = 1;
203                 log_info("    clat percentiles (msec):\n     |");
204         } else {
205                 scale_down = 0;
206                 log_info("    clat percentiles (usec):\n     |");
207         }
208
209         snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
210         per_line = (80 - 7) / (precision + 14);
211
212         for (j = 0; j < len; j++) {
213                 char fbuf[16], *ptr = fbuf;
214
215                 /* for formatting */
216                 if (j != 0 && (j % per_line) == 0)
217                         log_info("     |");
218
219                 /* end of the list */
220                 is_last = (j == len - 1);
221
222                 if (plist[j].u.f < 10.0)
223                         ptr += sprintf(fbuf, " ");
224
225                 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
226
227                 if (scale_down)
228                         ovals[j] = (ovals[j] + 999) / 1000;
229
230                 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
231
232                 if (is_last)
233                         break;
234
235                 if ((j % per_line) == per_line - 1)     /* for formatting */
236                         log_info("\n");
237         }
238
239 out:
240         if (ovals)
241                 free(ovals);
242 }
243
244 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
245              double *mean, double *dev)
246 {
247         double n = (double) is->samples;
248
249         if (n == 0)
250                 return 0;
251
252         *min = is->min_val;
253         *max = is->max_val;
254         *mean = is->mean.u.f;
255
256         if (n > 1.0)
257                 *dev = sqrt(is->S.u.f / (n - 1.0));
258         else
259                 *dev = 0;
260
261         return 1;
262 }
263
264 void show_group_stats(struct group_run_stats *rs)
265 {
266         char *p1, *p2, *p3, *p4;
267         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
268         int i;
269
270         log_info("\nRun status group %d (all jobs):\n", rs->groupid);
271
272         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
273                 const int i2p = is_power_of_2(rs->kb_base);
274
275                 if (!rs->max_run[i])
276                         continue;
277
278                 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
279                 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
280                 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
281                 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
282
283                 log_info("%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
284                          " mint=%llumsec, maxt=%llumsec\n",
285                                 rs->unified_rw_rep ? "  MIXED" : str[i],
286                                 p1, p2, p3, p4,
287                                 (unsigned long long) rs->min_run[i],
288                                 (unsigned long long) rs->max_run[i]);
289
290                 free(p1);
291                 free(p2);
292                 free(p3);
293                 free(p4);
294         }
295 }
296
297 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
298 {
299         int i;
300
301         /*
302          * Do depth distribution calculations
303          */
304         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
305                 if (total) {
306                         io_u_dist[i] = (double) map[i] / (double) total;
307                         io_u_dist[i] *= 100.0;
308                         if (io_u_dist[i] < 0.1 && map[i])
309                                 io_u_dist[i] = 0.1;
310                 } else
311                         io_u_dist[i] = 0.0;
312         }
313 }
314
315 static void stat_calc_lat(struct thread_stat *ts, double *dst,
316                           unsigned int *src, int nr)
317 {
318         unsigned long total = ddir_rw_sum(ts->total_io_u);
319         int i;
320
321         /*
322          * Do latency distribution calculations
323          */
324         for (i = 0; i < nr; i++) {
325                 if (total) {
326                         dst[i] = (double) src[i] / (double) total;
327                         dst[i] *= 100.0;
328                         if (dst[i] < 0.01 && src[i])
329                                 dst[i] = 0.01;
330                 } else
331                         dst[i] = 0.0;
332         }
333 }
334
335 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
336 {
337         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
338 }
339
340 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
341 {
342         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
343 }
344
345 static void display_lat(const char *name, unsigned long min, unsigned long max,
346                         double mean, double dev)
347 {
348         const char *base = "(usec)";
349         char *minp, *maxp;
350
351         if (!usec_to_msec(&min, &max, &mean, &dev))
352                 base = "(msec)";
353
354         minp = num2str(min, 6, 1, 0, 0);
355         maxp = num2str(max, 6, 1, 0, 0);
356
357         log_info("    %s %s: min=%s, max=%s, avg=%5.02f,"
358                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
359
360         free(minp);
361         free(maxp);
362 }
363
364 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
365                              int ddir)
366 {
367         const char *str[] = { "read ", "write", "trim" };
368         unsigned long min, max, runt;
369         unsigned long long bw, iops;
370         double mean, dev;
371         char *io_p, *bw_p, *iops_p;
372         int i2p;
373
374         assert(ddir_rw(ddir));
375
376         if (!ts->runtime[ddir])
377                 return;
378
379         i2p = is_power_of_2(rs->kb_base);
380         runt = ts->runtime[ddir];
381
382         bw = (1000 * ts->io_bytes[ddir]) / runt;
383         io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
384         bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
385
386         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
387         iops_p = num2str(iops, 6, 1, 0, 0);
388
389         log_info("  %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
390                                 rs->unified_rw_rep ? "mixed" : str[ddir],
391                                 io_p, bw_p, iops_p,
392                                 (unsigned long long) ts->runtime[ddir]);
393
394         free(io_p);
395         free(bw_p);
396         free(iops_p);
397
398         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
399                 display_lat("slat", min, max, mean, dev);
400         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
401                 display_lat("clat", min, max, mean, dev);
402         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
403                 display_lat(" lat", min, max, mean, dev);
404
405         if (ts->clat_percentiles) {
406                 show_clat_percentiles(ts->io_u_plat[ddir],
407                                         ts->clat_stat[ddir].samples,
408                                         ts->percentile_list,
409                                         ts->percentile_precision);
410         }
411         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
412                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
413                 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
414
415                 if (rs->unit_base == 1) {
416                         min *= 8.0;
417                         max *= 8.0;
418                         mean *= 8.0;
419                         dev *= 8.0;
420                 }
421
422                 if (rs->agg[ddir]) {
423                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
424                         if (p_of_agg > 100.0)
425                                 p_of_agg = 100.0;
426                 }
427
428                 if (mean > fkb_base * fkb_base) {
429                         min /= fkb_base;
430                         max /= fkb_base;
431                         mean /= fkb_base;
432                         dev /= fkb_base;
433                         bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
434                 }
435
436                 log_info("    bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
437                          " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
438                                                         p_of_agg, mean, dev);
439         }
440 }
441
442 static int show_lat(double *io_u_lat, int nr, const char **ranges,
443                     const char *msg)
444 {
445         int new_line = 1, i, line = 0, shown = 0;
446
447         for (i = 0; i < nr; i++) {
448                 if (io_u_lat[i] <= 0.0)
449                         continue;
450                 shown = 1;
451                 if (new_line) {
452                         if (line)
453                                 log_info("\n");
454                         log_info("    lat (%s) : ", msg);
455                         new_line = 0;
456                         line = 0;
457                 }
458                 if (line)
459                         log_info(", ");
460                 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
461                 line++;
462                 if (line == 5)
463                         new_line = 1;
464         }
465
466         if (shown)
467                 log_info("\n");
468
469         return shown;
470 }
471
472 static void show_lat_u(double *io_u_lat_u)
473 {
474         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
475                                  "250=", "500=", "750=", "1000=", };
476
477         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
478 }
479
480 static void show_lat_m(double *io_u_lat_m)
481 {
482         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
483                                  "250=", "500=", "750=", "1000=", "2000=",
484                                  ">=2000=", };
485
486         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
487 }
488
489 static void show_latencies(struct thread_stat *ts)
490 {
491         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
492         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
493
494         stat_calc_lat_u(ts, io_u_lat_u);
495         stat_calc_lat_m(ts, io_u_lat_m);
496
497         show_lat_u(io_u_lat_u);
498         show_lat_m(io_u_lat_m);
499 }
500
501 static int block_state_category(int block_state)
502 {
503         switch (block_state) {
504         case BLOCK_STATE_UNINIT:
505                 return 0;
506         case BLOCK_STATE_TRIMMED:
507         case BLOCK_STATE_WRITTEN:
508                 return 1;
509         case BLOCK_STATE_WRITE_FAILURE:
510         case BLOCK_STATE_TRIM_FAILURE:
511                 return 2;
512         default:
513                 /* Silence compile warning on some BSDs and have a return */
514                 assert(0);
515                 return -1;
516         }
517 }
518
519 static int compare_block_infos(const void *bs1, const void *bs2)
520 {
521         uint32_t block1 = *(uint32_t *)bs1;
522         uint32_t block2 = *(uint32_t *)bs2;
523         int state1 = BLOCK_INFO_STATE(block1);
524         int state2 = BLOCK_INFO_STATE(block2);
525         int bscat1 = block_state_category(state1);
526         int bscat2 = block_state_category(state2);
527         int cycles1 = BLOCK_INFO_TRIMS(block1);
528         int cycles2 = BLOCK_INFO_TRIMS(block2);
529
530         if (bscat1 < bscat2)
531                 return -1;
532         if (bscat1 > bscat2)
533                 return 1;
534
535         if (cycles1 < cycles2)
536                 return -1;
537         if (cycles1 > cycles2)
538                 return 1;
539
540         if (state1 < state2)
541                 return -1;
542         if (state1 > state2)
543                 return 1;
544
545         assert(block1 == block2);
546         return 0;
547 }
548
549 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
550                                   fio_fp64_t *plist, unsigned int **percentiles,
551                                   unsigned int *types)
552 {
553         int len = 0;
554         int i, nr_uninit;
555
556         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
557
558         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
559                 len++;
560
561         if (!len)
562                 return 0;
563
564         /*
565          * Sort the percentile list. Note that it may already be sorted if
566          * we are using the default values, but since it's a short list this
567          * isn't a worry. Also note that this does not work for NaN values.
568          */
569         if (len > 1)
570                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
571
572         nr_uninit = 0;
573         /* Start only after the uninit entries end */
574         for (nr_uninit = 0;
575              nr_uninit < nr_block_infos
576                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
577              nr_uninit ++)
578                 ;
579
580         if (nr_uninit == nr_block_infos)
581                 return 0;
582
583         *percentiles = calloc(len, sizeof(**percentiles));
584
585         for (i = 0; i < len; i++) {
586                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
587                                 + nr_uninit;
588                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
589         }
590
591         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
592         for (i = 0; i < nr_block_infos; i++)
593                 types[BLOCK_INFO_STATE(block_infos[i])]++;
594
595         return len;
596 }
597
598 static const char *block_state_names[] = {
599         [BLOCK_STATE_UNINIT] = "unwritten",
600         [BLOCK_STATE_TRIMMED] = "trimmed",
601         [BLOCK_STATE_WRITTEN] = "written",
602         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
603         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
604 };
605
606 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
607                              fio_fp64_t *plist)
608 {
609         int len, pos, i;
610         unsigned int *percentiles = NULL;
611         unsigned int block_state_counts[BLOCK_STATE_COUNT];
612
613         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
614                                      &percentiles, block_state_counts);
615
616         log_info("  block lifetime percentiles :\n   |");
617         pos = 0;
618         for (i = 0; i < len; i++) {
619                 uint32_t block_info = percentiles[i];
620 #define LINE_LENGTH     75
621                 char str[LINE_LENGTH];
622                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
623                                      plist[i].u.f, block_info,
624                                      i == len - 1 ? '\n' : ',');
625                 assert(strln < LINE_LENGTH);
626                 if (pos + strln > LINE_LENGTH) {
627                         pos = 0;
628                         log_info("\n   |");
629                 }
630                 log_info("%s", str);
631                 pos += strln;
632 #undef LINE_LENGTH
633         }
634         if (percentiles)
635                 free(percentiles);
636
637         log_info("        states               :");
638         for (i = 0; i < BLOCK_STATE_COUNT; i++)
639                 log_info(" %s=%u%c",
640                          block_state_names[i], block_state_counts[i],
641                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
642 }
643
644 static void show_thread_status_normal(struct thread_stat *ts,
645                                       struct group_run_stats *rs)
646 {
647         double usr_cpu, sys_cpu;
648         unsigned long runtime;
649         double io_u_dist[FIO_IO_U_MAP_NR];
650         time_t time_p;
651         char time_buf[32];
652
653         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
654                 return;
655
656         time(&time_p);
657         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
658
659         if (!ts->error) {
660                 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
661                                         ts->name, ts->groupid, ts->members,
662                                         ts->error, (int) ts->pid, time_buf);
663         } else {
664                 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
665                                         ts->name, ts->groupid, ts->members,
666                                         ts->error, ts->verror, (int) ts->pid,
667                                         time_buf);
668         }
669
670         if (strlen(ts->description))
671                 log_info("  Description  : [%s]\n", ts->description);
672
673         if (ts->io_bytes[DDIR_READ])
674                 show_ddir_status(rs, ts, DDIR_READ);
675         if (ts->io_bytes[DDIR_WRITE])
676                 show_ddir_status(rs, ts, DDIR_WRITE);
677         if (ts->io_bytes[DDIR_TRIM])
678                 show_ddir_status(rs, ts, DDIR_TRIM);
679
680         show_latencies(ts);
681
682         runtime = ts->total_run_time;
683         if (runtime) {
684                 double runt = (double) runtime;
685
686                 usr_cpu = (double) ts->usr_time * 100 / runt;
687                 sys_cpu = (double) ts->sys_time * 100 / runt;
688         } else {
689                 usr_cpu = 0;
690                 sys_cpu = 0;
691         }
692
693         log_info("  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
694                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
695                         (unsigned long long) ts->ctx,
696                         (unsigned long long) ts->majf,
697                         (unsigned long long) ts->minf);
698
699         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
700         log_info("  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
701                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
702                                         io_u_dist[1], io_u_dist[2],
703                                         io_u_dist[3], io_u_dist[4],
704                                         io_u_dist[5], io_u_dist[6]);
705
706         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
707         log_info("     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
708                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
709                                         io_u_dist[1], io_u_dist[2],
710                                         io_u_dist[3], io_u_dist[4],
711                                         io_u_dist[5], io_u_dist[6]);
712         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
713         log_info("     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
714                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
715                                         io_u_dist[1], io_u_dist[2],
716                                         io_u_dist[3], io_u_dist[4],
717                                         io_u_dist[5], io_u_dist[6]);
718         log_info("     issued    : total=r=%llu/w=%llu/d=%llu,"
719                                  " short=r=%llu/w=%llu/d=%llu,"
720                                  " drop=r=%llu/w=%llu/d=%llu\n",
721                                         (unsigned long long) ts->total_io_u[0],
722                                         (unsigned long long) ts->total_io_u[1],
723                                         (unsigned long long) ts->total_io_u[2],
724                                         (unsigned long long) ts->short_io_u[0],
725                                         (unsigned long long) ts->short_io_u[1],
726                                         (unsigned long long) ts->short_io_u[2],
727                                         (unsigned long long) ts->drop_io_u[0],
728                                         (unsigned long long) ts->drop_io_u[1],
729                                         (unsigned long long) ts->drop_io_u[2]);
730         if (ts->continue_on_error) {
731                 log_info("     errors    : total=%llu, first_error=%d/<%s>\n",
732                                         (unsigned long long)ts->total_err_count,
733                                         ts->first_error,
734                                         strerror(ts->first_error));
735         }
736         if (ts->latency_depth) {
737                 log_info("     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
738                                         (unsigned long long)ts->latency_target,
739                                         (unsigned long long)ts->latency_window,
740                                         ts->latency_percentile.u.f,
741                                         ts->latency_depth);
742         }
743
744         if (ts->nr_block_infos)
745                 show_block_infos(ts->nr_block_infos, ts->block_infos,
746                                   ts->percentile_list);
747 }
748
749 static void show_ddir_status_terse(struct thread_stat *ts,
750                                    struct group_run_stats *rs, int ddir)
751 {
752         unsigned long min, max;
753         unsigned long long bw, iops;
754         unsigned int *ovals = NULL;
755         double mean, dev;
756         unsigned int len, minv, maxv;
757         int i;
758
759         assert(ddir_rw(ddir));
760
761         iops = bw = 0;
762         if (ts->runtime[ddir]) {
763                 uint64_t runt = ts->runtime[ddir];
764
765                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
766                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
767         }
768
769         log_info(";%llu;%llu;%llu;%llu",
770                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
771                                         (unsigned long long) ts->runtime[ddir]);
772
773         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
774                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
775         else
776                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
777
778         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
779                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
780         else
781                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
782
783         if (ts->clat_percentiles) {
784                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
785                                         ts->clat_stat[ddir].samples,
786                                         ts->percentile_list, &ovals, &maxv,
787                                         &minv);
788         } else
789                 len = 0;
790
791         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
792                 if (i >= len) {
793                         log_info(";0%%=0");
794                         continue;
795                 }
796                 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
797         }
798
799         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
800                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
801         else
802                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
803
804         if (ovals)
805                 free(ovals);
806
807         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
808                 double p_of_agg = 100.0;
809
810                 if (rs->agg[ddir]) {
811                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
812                         if (p_of_agg > 100.0)
813                                 p_of_agg = 100.0;
814                 }
815
816                 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
817         } else
818                 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
819 }
820
821 static void add_ddir_status_json(struct thread_stat *ts,
822                 struct group_run_stats *rs, int ddir, struct json_object *parent)
823 {
824         unsigned long min, max;
825         unsigned long long bw;
826         unsigned int *ovals = NULL;
827         double mean, dev, iops;
828         unsigned int len, minv, maxv;
829         int i;
830         const char *ddirname[] = {"read", "write", "trim"};
831         struct json_object *dir_object, *tmp_object, *percentile_object;
832         char buf[120];
833         double p_of_agg = 100.0;
834
835         assert(ddir_rw(ddir));
836
837         if (ts->unified_rw_rep && ddir != DDIR_READ)
838                 return;
839
840         dir_object = json_create_object();
841         json_object_add_value_object(parent,
842                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
843
844         bw = 0;
845         iops = 0.0;
846         if (ts->runtime[ddir]) {
847                 uint64_t runt = ts->runtime[ddir];
848
849                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
850                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
851         }
852
853         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
854         json_object_add_value_int(dir_object, "bw", bw);
855         json_object_add_value_float(dir_object, "iops", iops);
856         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
857         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
858         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
859         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
860
861         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
862                 min = max = 0;
863                 mean = dev = 0.0;
864         }
865         tmp_object = json_create_object();
866         json_object_add_value_object(dir_object, "slat", tmp_object);
867         json_object_add_value_int(tmp_object, "min", min);
868         json_object_add_value_int(tmp_object, "max", max);
869         json_object_add_value_float(tmp_object, "mean", mean);
870         json_object_add_value_float(tmp_object, "stddev", dev);
871
872         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
873                 min = max = 0;
874                 mean = dev = 0.0;
875         }
876         tmp_object = json_create_object();
877         json_object_add_value_object(dir_object, "clat", tmp_object);
878         json_object_add_value_int(tmp_object, "min", min);
879         json_object_add_value_int(tmp_object, "max", max);
880         json_object_add_value_float(tmp_object, "mean", mean);
881         json_object_add_value_float(tmp_object, "stddev", dev);
882
883         if (ts->clat_percentiles) {
884                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
885                                         ts->clat_stat[ddir].samples,
886                                         ts->percentile_list, &ovals, &maxv,
887                                         &minv);
888         } else
889                 len = 0;
890
891         percentile_object = json_create_object();
892         json_object_add_value_object(tmp_object, "percentile", percentile_object);
893         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
894                 if (i >= len) {
895                         json_object_add_value_int(percentile_object, "0.00", 0);
896                         continue;
897                 }
898                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
899                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
900         }
901
902         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
903                 min = max = 0;
904                 mean = dev = 0.0;
905         }
906         tmp_object = json_create_object();
907         json_object_add_value_object(dir_object, "lat", tmp_object);
908         json_object_add_value_int(tmp_object, "min", min);
909         json_object_add_value_int(tmp_object, "max", max);
910         json_object_add_value_float(tmp_object, "mean", mean);
911         json_object_add_value_float(tmp_object, "stddev", dev);
912         if (ovals)
913                 free(ovals);
914
915         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
916                 if (rs->agg[ddir]) {
917                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
918                         if (p_of_agg > 100.0)
919                                 p_of_agg = 100.0;
920                 }
921         } else {
922                 min = max = 0;
923                 p_of_agg = mean = dev = 0.0;
924         }
925         json_object_add_value_int(dir_object, "bw_min", min);
926         json_object_add_value_int(dir_object, "bw_max", max);
927         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
928         json_object_add_value_float(dir_object, "bw_mean", mean);
929         json_object_add_value_float(dir_object, "bw_dev", dev);
930 }
931
932 static void show_thread_status_terse_v2(struct thread_stat *ts,
933                                         struct group_run_stats *rs)
934 {
935         double io_u_dist[FIO_IO_U_MAP_NR];
936         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
937         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
938         double usr_cpu, sys_cpu;
939         int i;
940
941         /* General Info */
942         log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
943         /* Log Read Status */
944         show_ddir_status_terse(ts, rs, DDIR_READ);
945         /* Log Write Status */
946         show_ddir_status_terse(ts, rs, DDIR_WRITE);
947         /* Log Trim Status */
948         show_ddir_status_terse(ts, rs, DDIR_TRIM);
949
950         /* CPU Usage */
951         if (ts->total_run_time) {
952                 double runt = (double) ts->total_run_time;
953
954                 usr_cpu = (double) ts->usr_time * 100 / runt;
955                 sys_cpu = (double) ts->sys_time * 100 / runt;
956         } else {
957                 usr_cpu = 0;
958                 sys_cpu = 0;
959         }
960
961         log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
962                                                 (unsigned long long) ts->ctx,
963                                                 (unsigned long long) ts->majf,
964                                                 (unsigned long long) ts->minf);
965
966         /* Calc % distribution of IO depths, usecond, msecond latency */
967         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
968         stat_calc_lat_u(ts, io_u_lat_u);
969         stat_calc_lat_m(ts, io_u_lat_m);
970
971         /* Only show fixed 7 I/O depth levels*/
972         log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
973                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
974                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
975
976         /* Microsecond latency */
977         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
978                 log_info(";%3.2f%%", io_u_lat_u[i]);
979         /* Millisecond latency */
980         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
981                 log_info(";%3.2f%%", io_u_lat_m[i]);
982         /* Additional output if continue_on_error set - default off*/
983         if (ts->continue_on_error)
984                 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
985         log_info("\n");
986
987         /* Additional output if description is set */
988         if (strlen(ts->description))
989                 log_info(";%s", ts->description);
990
991         log_info("\n");
992 }
993
994 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
995                                            struct group_run_stats *rs, int ver)
996 {
997         double io_u_dist[FIO_IO_U_MAP_NR];
998         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
999         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1000         double usr_cpu, sys_cpu;
1001         int i;
1002
1003         /* General Info */
1004         log_info("%d;%s;%s;%d;%d", ver, fio_version_string,
1005                                         ts->name, ts->groupid, ts->error);
1006         /* Log Read Status */
1007         show_ddir_status_terse(ts, rs, DDIR_READ);
1008         /* Log Write Status */
1009         show_ddir_status_terse(ts, rs, DDIR_WRITE);
1010         /* Log Trim Status */
1011         if (ver == 4)
1012                 show_ddir_status_terse(ts, rs, DDIR_TRIM);
1013
1014         /* CPU Usage */
1015         if (ts->total_run_time) {
1016                 double runt = (double) ts->total_run_time;
1017
1018                 usr_cpu = (double) ts->usr_time * 100 / runt;
1019                 sys_cpu = (double) ts->sys_time * 100 / runt;
1020         } else {
1021                 usr_cpu = 0;
1022                 sys_cpu = 0;
1023         }
1024
1025         log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1026                                                 (unsigned long long) ts->ctx,
1027                                                 (unsigned long long) ts->majf,
1028                                                 (unsigned long long) ts->minf);
1029
1030         /* Calc % distribution of IO depths, usecond, msecond latency */
1031         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1032         stat_calc_lat_u(ts, io_u_lat_u);
1033         stat_calc_lat_m(ts, io_u_lat_m);
1034
1035         /* Only show fixed 7 I/O depth levels*/
1036         log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1037                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1038                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1039
1040         /* Microsecond latency */
1041         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1042                 log_info(";%3.2f%%", io_u_lat_u[i]);
1043         /* Millisecond latency */
1044         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1045                 log_info(";%3.2f%%", io_u_lat_m[i]);
1046
1047         /* disk util stats, if any */
1048         show_disk_util(1, NULL);
1049
1050         /* Additional output if continue_on_error set - default off*/
1051         if (ts->continue_on_error)
1052                 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1053
1054         /* Additional output if description is set */
1055         if (strlen(ts->description))
1056                 log_info(";%s", ts->description);
1057
1058         log_info("\n");
1059 }
1060
1061 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1062                                     struct group_run_stats *rs)
1063 {
1064         struct json_object *root, *tmp;
1065         double io_u_dist[FIO_IO_U_MAP_NR];
1066         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1067         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1068         double usr_cpu, sys_cpu;
1069         int i;
1070
1071         root = json_create_object();
1072         json_object_add_value_string(root, "jobname", ts->name);
1073         json_object_add_value_int(root, "groupid", ts->groupid);
1074         json_object_add_value_int(root, "error", ts->error);
1075
1076         add_ddir_status_json(ts, rs, DDIR_READ, root);
1077         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1078         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1079
1080         /* CPU Usage */
1081         if (ts->total_run_time) {
1082                 double runt = (double) ts->total_run_time;
1083
1084                 usr_cpu = (double) ts->usr_time * 100 / runt;
1085                 sys_cpu = (double) ts->sys_time * 100 / runt;
1086         } else {
1087                 usr_cpu = 0;
1088                 sys_cpu = 0;
1089         }
1090         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1091         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1092         json_object_add_value_int(root, "ctx", ts->ctx);
1093         json_object_add_value_int(root, "majf", ts->majf);
1094         json_object_add_value_int(root, "minf", ts->minf);
1095
1096
1097         /* Calc % distribution of IO depths, usecond, msecond latency */
1098         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1099         stat_calc_lat_u(ts, io_u_lat_u);
1100         stat_calc_lat_m(ts, io_u_lat_m);
1101
1102         tmp = json_create_object();
1103         json_object_add_value_object(root, "iodepth_level", tmp);
1104         /* Only show fixed 7 I/O depth levels*/
1105         for (i = 0; i < 7; i++) {
1106                 char name[20];
1107                 if (i < 6)
1108                         snprintf(name, 20, "%d", 1 << i);
1109                 else
1110                         snprintf(name, 20, ">=%d", 1 << i);
1111                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1112         }
1113
1114         tmp = json_create_object();
1115         json_object_add_value_object(root, "latency_us", tmp);
1116         /* Microsecond latency */
1117         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1118                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1119                                  "250", "500", "750", "1000", };
1120                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1121         }
1122         /* Millisecond latency */
1123         tmp = json_create_object();
1124         json_object_add_value_object(root, "latency_ms", tmp);
1125         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1126                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1127                                  "250", "500", "750", "1000", "2000",
1128                                  ">=2000", };
1129                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1130         }
1131
1132         /* Additional output if continue_on_error set - default off*/
1133         if (ts->continue_on_error) {
1134                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1135                 json_object_add_value_int(root, "first_error", ts->first_error);
1136         }
1137
1138         if (ts->latency_depth) {
1139                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1140                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1141                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1142                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1143         }
1144
1145         /* Additional output if description is set */
1146         if (strlen(ts->description))
1147                 json_object_add_value_string(root, "desc", ts->description);
1148
1149         if (ts->nr_block_infos) {
1150                 /* Block error histogram and types */
1151                 int len;
1152                 unsigned int *percentiles = NULL;
1153                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1154
1155                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1156                                              ts->percentile_list,
1157                                              &percentiles, block_state_counts);
1158
1159                 if (len) {
1160                         struct json_object *block, *percentile_object, *states;
1161                         int state, i;
1162                         block = json_create_object();
1163                         json_object_add_value_object(root, "block", block);
1164
1165                         percentile_object = json_create_object();
1166                         json_object_add_value_object(block, "percentiles",
1167                                                      percentile_object);
1168                         for (i = 0; i < len; i++) {
1169                                 char buf[20];
1170                                 snprintf(buf, sizeof(buf), "%f",
1171                                          ts->percentile_list[i].u.f);
1172                                 json_object_add_value_int(percentile_object,
1173                                                           (const char *)buf,
1174                                                           percentiles[i]);
1175                         }
1176
1177                         states = json_create_object();
1178                         json_object_add_value_object(block, "states", states);
1179                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1180                                 json_object_add_value_int(states,
1181                                         block_state_names[state],
1182                                         block_state_counts[state]);
1183                         }
1184                         free(percentiles);
1185                 }
1186         }
1187
1188         return root;
1189 }
1190
1191 static void show_thread_status_terse(struct thread_stat *ts,
1192                                      struct group_run_stats *rs)
1193 {
1194         if (terse_version == 2)
1195                 show_thread_status_terse_v2(ts, rs);
1196         else if (terse_version == 3 || terse_version == 4)
1197                 show_thread_status_terse_v3_v4(ts, rs, terse_version);
1198         else
1199                 log_err("fio: bad terse version!? %d\n", terse_version);
1200 }
1201
1202 struct json_object *show_thread_status(struct thread_stat *ts,
1203                                        struct group_run_stats *rs)
1204 {
1205         if (output_format == FIO_OUTPUT_TERSE)
1206                 show_thread_status_terse(ts, rs);
1207         else if (output_format == FIO_OUTPUT_JSON)
1208                 return show_thread_status_json(ts, rs);
1209         else
1210                 show_thread_status_normal(ts, rs);
1211         return NULL;
1212 }
1213
1214 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
1215 {
1216         double mean, S;
1217
1218         if (src->samples == 0)
1219                 return;
1220
1221         dst->min_val = min(dst->min_val, src->min_val);
1222         dst->max_val = max(dst->max_val, src->max_val);
1223
1224         /*
1225          * Compute new mean and S after the merge
1226          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1227          *  #Parallel_algorithm>
1228          */
1229         if (nr == 1) {
1230                 mean = src->mean.u.f;
1231                 S = src->S.u.f;
1232         } else {
1233                 double delta = src->mean.u.f - dst->mean.u.f;
1234
1235                 mean = ((src->mean.u.f * src->samples) +
1236                         (dst->mean.u.f * dst->samples)) /
1237                         (dst->samples + src->samples);
1238
1239                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1240                         (dst->samples * src->samples) /
1241                         (dst->samples + src->samples);
1242         }
1243
1244         dst->samples += src->samples;
1245         dst->mean.u.f = mean;
1246         dst->S.u.f = S;
1247 }
1248
1249 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1250 {
1251         int i;
1252
1253         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1254                 if (dst->max_run[i] < src->max_run[i])
1255                         dst->max_run[i] = src->max_run[i];
1256                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1257                         dst->min_run[i] = src->min_run[i];
1258                 if (dst->max_bw[i] < src->max_bw[i])
1259                         dst->max_bw[i] = src->max_bw[i];
1260                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1261                         dst->min_bw[i] = src->min_bw[i];
1262
1263                 dst->io_kb[i] += src->io_kb[i];
1264                 dst->agg[i] += src->agg[i];
1265         }
1266
1267         if (!dst->kb_base)
1268                 dst->kb_base = src->kb_base;
1269         if (!dst->unit_base)
1270                 dst->unit_base = src->unit_base;
1271 }
1272
1273 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1274 {
1275         int l, k;
1276
1277         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1278                 if (!dst->unified_rw_rep) {
1279                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1280                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1281                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1282                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1283
1284                         dst->io_bytes[l] += src->io_bytes[l];
1285
1286                         if (dst->runtime[l] < src->runtime[l])
1287                                 dst->runtime[l] = src->runtime[l];
1288                 } else {
1289                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1290                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1291                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1292                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1293
1294                         dst->io_bytes[0] += src->io_bytes[l];
1295
1296                         if (dst->runtime[0] < src->runtime[l])
1297                                 dst->runtime[0] = src->runtime[l];
1298                 }
1299         }
1300
1301         dst->usr_time += src->usr_time;
1302         dst->sys_time += src->sys_time;
1303         dst->ctx += src->ctx;
1304         dst->majf += src->majf;
1305         dst->minf += src->minf;
1306
1307         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1308                 dst->io_u_map[k] += src->io_u_map[k];
1309         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1310                 dst->io_u_submit[k] += src->io_u_submit[k];
1311         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1312                 dst->io_u_complete[k] += src->io_u_complete[k];
1313         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1314                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1315         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1316                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1317
1318         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1319                 if (!dst->unified_rw_rep) {
1320                         dst->total_io_u[k] += src->total_io_u[k];
1321                         dst->short_io_u[k] += src->short_io_u[k];
1322                         dst->drop_io_u[k] += src->drop_io_u[k];
1323                 } else {
1324                         dst->total_io_u[0] += src->total_io_u[k];
1325                         dst->short_io_u[0] += src->short_io_u[k];
1326                         dst->drop_io_u[0] += src->drop_io_u[k];
1327                 }
1328         }
1329
1330         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1331                 int m;
1332
1333                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1334                         if (!dst->unified_rw_rep)
1335                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1336                         else
1337                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1338                 }
1339         }
1340
1341         dst->total_run_time += src->total_run_time;
1342         dst->total_submit += src->total_submit;
1343         dst->total_complete += src->total_complete;
1344 }
1345
1346 void init_group_run_stat(struct group_run_stats *gs)
1347 {
1348         int i;
1349         memset(gs, 0, sizeof(*gs));
1350
1351         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1352                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1353 }
1354
1355 void init_thread_stat(struct thread_stat *ts)
1356 {
1357         int j;
1358
1359         memset(ts, 0, sizeof(*ts));
1360
1361         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1362                 ts->lat_stat[j].min_val = -1UL;
1363                 ts->clat_stat[j].min_val = -1UL;
1364                 ts->slat_stat[j].min_val = -1UL;
1365                 ts->bw_stat[j].min_val = -1UL;
1366         }
1367         ts->groupid = -1;
1368 }
1369
1370 void __show_run_stats(void)
1371 {
1372         struct group_run_stats *runstats, *rs;
1373         struct thread_data *td;
1374         struct thread_stat *threadstats, *ts;
1375         int i, j, k, nr_ts, last_ts, idx;
1376         int kb_base_warned = 0;
1377         int unit_base_warned = 0;
1378         struct json_object *root = NULL;
1379         struct json_array *array = NULL;
1380         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1381
1382         for (i = 0; i < groupid + 1; i++)
1383                 init_group_run_stat(&runstats[i]);
1384
1385         /*
1386          * find out how many threads stats we need. if group reporting isn't
1387          * enabled, it's one-per-td.
1388          */
1389         nr_ts = 0;
1390         last_ts = -1;
1391         for_each_td(td, i) {
1392                 if (!td->o.group_reporting) {
1393                         nr_ts++;
1394                         continue;
1395                 }
1396                 if (last_ts == td->groupid)
1397                         continue;
1398
1399                 last_ts = td->groupid;
1400                 nr_ts++;
1401         }
1402
1403         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1404
1405         for (i = 0; i < nr_ts; i++)
1406                 init_thread_stat(&threadstats[i]);
1407
1408         j = 0;
1409         last_ts = -1;
1410         idx = 0;
1411         for_each_td(td, i) {
1412                 if (idx && (!td->o.group_reporting ||
1413                     (td->o.group_reporting && last_ts != td->groupid))) {
1414                         idx = 0;
1415                         j++;
1416                 }
1417
1418                 last_ts = td->groupid;
1419
1420                 ts = &threadstats[j];
1421
1422                 ts->clat_percentiles = td->o.clat_percentiles;
1423                 ts->percentile_precision = td->o.percentile_precision;
1424                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1425
1426                 idx++;
1427                 ts->members++;
1428
1429                 if (ts->groupid == -1) {
1430                         /*
1431                          * These are per-group shared already
1432                          */
1433                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1434                         if (td->o.description)
1435                                 strncpy(ts->description, td->o.description,
1436                                                 FIO_JOBDESC_SIZE - 1);
1437                         else
1438                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1439
1440                         /*
1441                          * If multiple entries in this group, this is
1442                          * the first member.
1443                          */
1444                         ts->thread_number = td->thread_number;
1445                         ts->groupid = td->groupid;
1446
1447                         /*
1448                          * first pid in group, not very useful...
1449                          */
1450                         ts->pid = td->pid;
1451
1452                         ts->kb_base = td->o.kb_base;
1453                         ts->unit_base = td->o.unit_base;
1454                         ts->unified_rw_rep = td->o.unified_rw_rep;
1455                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1456                         log_info("fio: kb_base differs for jobs in group, using"
1457                                  " %u as the base\n", ts->kb_base);
1458                         kb_base_warned = 1;
1459                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1460                         log_info("fio: unit_base differs for jobs in group, using"
1461                                  " %u as the base\n", ts->unit_base);
1462                         unit_base_warned = 1;
1463                 }
1464
1465                 ts->continue_on_error = td->o.continue_on_error;
1466                 ts->total_err_count += td->total_err_count;
1467                 ts->first_error = td->first_error;
1468                 if (!ts->error) {
1469                         if (!td->error && td->o.continue_on_error &&
1470                             td->first_error) {
1471                                 ts->error = td->first_error;
1472                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1473                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1474                         } else  if (td->error) {
1475                                 ts->error = td->error;
1476                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1477                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1478                         }
1479                 }
1480
1481                 ts->latency_depth = td->latency_qd;
1482                 ts->latency_target = td->o.latency_target;
1483                 ts->latency_percentile = td->o.latency_percentile;
1484                 ts->latency_window = td->o.latency_window;
1485
1486                 ts->nr_block_infos = td->ts.nr_block_infos;
1487                 for (k = 0; k < ts->nr_block_infos; k++)
1488                         ts->block_infos[k] = td->ts.block_infos[k];
1489
1490                 sum_thread_stats(ts, &td->ts, idx);
1491         }
1492
1493         for (i = 0; i < nr_ts; i++) {
1494                 unsigned long long bw;
1495
1496                 ts = &threadstats[i];
1497                 rs = &runstats[ts->groupid];
1498                 rs->kb_base = ts->kb_base;
1499                 rs->unit_base = ts->unit_base;
1500                 rs->unified_rw_rep += ts->unified_rw_rep;
1501
1502                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1503                         if (!ts->runtime[j])
1504                                 continue;
1505                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1506                                 rs->min_run[j] = ts->runtime[j];
1507                         if (ts->runtime[j] > rs->max_run[j])
1508                                 rs->max_run[j] = ts->runtime[j];
1509
1510                         bw = 0;
1511                         if (ts->runtime[j]) {
1512                                 unsigned long runt = ts->runtime[j];
1513                                 unsigned long long kb;
1514
1515                                 kb = ts->io_bytes[j] / rs->kb_base;
1516                                 bw = kb * 1000 / runt;
1517                         }
1518                         if (bw < rs->min_bw[j])
1519                                 rs->min_bw[j] = bw;
1520                         if (bw > rs->max_bw[j])
1521                                 rs->max_bw[j] = bw;
1522
1523                         rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1524                 }
1525         }
1526
1527         for (i = 0; i < groupid + 1; i++) {
1528                 int ddir;
1529
1530                 rs = &runstats[i];
1531
1532                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1533                         if (rs->max_run[ddir])
1534                                 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1535                                                 rs->max_run[ddir];
1536                 }
1537         }
1538
1539         /*
1540          * don't overwrite last signal output
1541          */
1542         if (output_format == FIO_OUTPUT_NORMAL)
1543                 log_info("\n");
1544         else if (output_format == FIO_OUTPUT_JSON) {
1545                 char time_buf[32];
1546                 time_t time_p;
1547
1548                 time(&time_p);
1549                 os_ctime_r((const time_t *) &time_p, time_buf,
1550                                 sizeof(time_buf));
1551                 time_buf[strlen(time_buf) - 1] = '\0';
1552
1553                 root = json_create_object();
1554                 json_object_add_value_string(root, "fio version", fio_version_string);
1555                 json_object_add_value_int(root, "timestamp", time_p);
1556                 json_object_add_value_string(root, "time", time_buf);
1557                 array = json_create_array();
1558                 json_object_add_value_array(root, "jobs", array);
1559         }
1560
1561         for (i = 0; i < nr_ts; i++) {
1562                 ts = &threadstats[i];
1563                 rs = &runstats[ts->groupid];
1564
1565                 if (is_backend)
1566                         fio_server_send_ts(ts, rs);
1567                 else if (output_format == FIO_OUTPUT_TERSE)
1568                         show_thread_status_terse(ts, rs);
1569                 else if (output_format == FIO_OUTPUT_JSON) {
1570                         struct json_object *tmp = show_thread_status_json(ts, rs);
1571                         json_array_add_value_object(array, tmp);
1572                 } else
1573                         show_thread_status_normal(ts, rs);
1574         }
1575         if (output_format == FIO_OUTPUT_JSON) {
1576                 /* disk util stats, if any */
1577                 show_disk_util(1, root);
1578
1579                 show_idle_prof_stats(FIO_OUTPUT_JSON, root);
1580
1581                 json_print_object(root);
1582                 log_info("\n");
1583                 json_free_object(root);
1584         }
1585
1586         for (i = 0; i < groupid + 1; i++) {
1587                 rs = &runstats[i];
1588
1589                 rs->groupid = i;
1590                 if (is_backend)
1591                         fio_server_send_gs(rs);
1592                 else if (output_format == FIO_OUTPUT_NORMAL)
1593                         show_group_stats(rs);
1594         }
1595
1596         if (is_backend)
1597                 fio_server_send_du();
1598         else if (output_format == FIO_OUTPUT_NORMAL) {
1599                 show_disk_util(0, NULL);
1600                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
1601         }
1602
1603         if ( !(output_format == FIO_OUTPUT_TERSE) && append_terse_output) {
1604                 log_info("\nAdditional Terse Output:\n");
1605
1606                 for (i = 0; i < nr_ts; i++) {
1607                         ts = &threadstats[i];
1608                         rs = &runstats[ts->groupid];
1609                         show_thread_status_terse(ts, rs);
1610                 }
1611         }
1612
1613         log_info_flush();
1614         free(runstats);
1615         free(threadstats);
1616 }
1617
1618 void show_run_stats(void)
1619 {
1620         fio_mutex_down(stat_mutex);
1621         __show_run_stats();
1622         fio_mutex_up(stat_mutex);
1623 }
1624
1625 void __show_running_run_stats(void)
1626 {
1627         struct thread_data *td;
1628         unsigned long long *rt;
1629         struct timeval tv;
1630         int i;
1631
1632         fio_mutex_down(stat_mutex);
1633
1634         rt = malloc(thread_number * sizeof(unsigned long long));
1635         fio_gettime(&tv, NULL);
1636
1637         for_each_td(td, i) {
1638                 rt[i] = mtime_since(&td->start, &tv);
1639                 if (td_read(td) && td->io_bytes[DDIR_READ])
1640                         td->ts.runtime[DDIR_READ] += rt[i];
1641                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1642                         td->ts.runtime[DDIR_WRITE] += rt[i];
1643                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1644                         td->ts.runtime[DDIR_TRIM] += rt[i];
1645
1646                 td->update_rusage = 1;
1647                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1648                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1649                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1650                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1651         }
1652
1653         for_each_td(td, i) {
1654                 if (td->runstate >= TD_EXITED)
1655                         continue;
1656                 if (td->rusage_sem) {
1657                         td->update_rusage = 1;
1658                         fio_mutex_down(td->rusage_sem);
1659                 }
1660                 td->update_rusage = 0;
1661         }
1662
1663         __show_run_stats();
1664
1665         for_each_td(td, i) {
1666                 if (td_read(td) && td->io_bytes[DDIR_READ])
1667                         td->ts.runtime[DDIR_READ] -= rt[i];
1668                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1669                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1670                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1671                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1672         }
1673
1674         free(rt);
1675         fio_mutex_up(stat_mutex);
1676 }
1677
1678 static int status_interval_init;
1679 static struct timeval status_time;
1680 static int status_file_disabled;
1681
1682 #define FIO_STATUS_FILE         "fio-dump-status"
1683
1684 static int check_status_file(void)
1685 {
1686         struct stat sb;
1687         const char *temp_dir;
1688         char fio_status_file_path[PATH_MAX];
1689
1690         if (status_file_disabled)
1691                 return 0;
1692
1693         temp_dir = getenv("TMPDIR");
1694         if (temp_dir == NULL) {
1695                 temp_dir = getenv("TEMP");
1696                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1697                         temp_dir = NULL;
1698         }
1699         if (temp_dir == NULL)
1700                 temp_dir = "/tmp";
1701
1702         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1703
1704         if (stat(fio_status_file_path, &sb))
1705                 return 0;
1706
1707         if (unlink(fio_status_file_path) < 0) {
1708                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1709                                                         strerror(errno));
1710                 log_err("fio: disabling status file updates\n");
1711                 status_file_disabled = 1;
1712         }
1713
1714         return 1;
1715 }
1716
1717 void check_for_running_stats(void)
1718 {
1719         if (status_interval) {
1720                 if (!status_interval_init) {
1721                         fio_gettime(&status_time, NULL);
1722                         status_interval_init = 1;
1723                 } else if (mtime_since_now(&status_time) >= status_interval) {
1724                         show_running_run_stats();
1725                         fio_gettime(&status_time, NULL);
1726                         return;
1727                 }
1728         }
1729         if (check_status_file()) {
1730                 show_running_run_stats();
1731                 return;
1732         }
1733 }
1734
1735 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1736 {
1737         double val = data;
1738         double delta;
1739
1740         if (data > is->max_val)
1741                 is->max_val = data;
1742         if (data < is->min_val)
1743                 is->min_val = data;
1744
1745         delta = val - is->mean.u.f;
1746         if (delta) {
1747                 is->mean.u.f += delta / (is->samples + 1.0);
1748                 is->S.u.f += delta * (val - is->mean.u.f);
1749         }
1750
1751         is->samples++;
1752 }
1753
1754 static void __add_log_sample(struct io_log *iolog, unsigned long val,
1755                              enum fio_ddir ddir, unsigned int bs,
1756                              unsigned long t, uint64_t offset)
1757 {
1758         uint64_t nr_samples = iolog->nr_samples;
1759         struct io_sample *s;
1760
1761         if (iolog->disabled)
1762                 return;
1763
1764         if (!iolog->nr_samples)
1765                 iolog->avg_last = t;
1766
1767         if (iolog->nr_samples == iolog->max_samples) {
1768                 size_t new_size;
1769                 void *new_log;
1770
1771                 new_size = 2 * iolog->max_samples * log_entry_sz(iolog);
1772
1773                 if (iolog->log_gz && (new_size > iolog->log_gz)) {
1774                         if (iolog_flush(iolog, 0)) {
1775                                 log_err("fio: failed flushing iolog! Will stop logging.\n");
1776                                 iolog->disabled = 1;
1777                                 return;
1778                         }
1779                         nr_samples = iolog->nr_samples;
1780                 } else {
1781                         new_log = realloc(iolog->log, new_size);
1782                         if (!new_log) {
1783                                 log_err("fio: failed extending iolog! Will stop logging.\n");
1784                                 iolog->disabled = 1;
1785                                 return;
1786                         }
1787                         iolog->log = new_log;
1788                         iolog->max_samples <<= 1;
1789                 }
1790         }
1791
1792         s = get_sample(iolog, nr_samples);
1793
1794         s->val = val;
1795         s->time = t;
1796         io_sample_set_ddir(iolog, s, ddir);
1797         s->bs = bs;
1798
1799         if (iolog->log_offset) {
1800                 struct io_sample_offset *so = (void *) s;
1801
1802                 so->offset = offset;
1803         }
1804
1805         iolog->nr_samples++;
1806 }
1807
1808 static inline void reset_io_stat(struct io_stat *ios)
1809 {
1810         ios->max_val = ios->min_val = ios->samples = 0;
1811         ios->mean.u.f = ios->S.u.f = 0;
1812 }
1813
1814 void reset_io_stats(struct thread_data *td)
1815 {
1816         struct thread_stat *ts = &td->ts;
1817         int i, j;
1818
1819         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1820                 reset_io_stat(&ts->clat_stat[i]);
1821                 reset_io_stat(&ts->slat_stat[i]);
1822                 reset_io_stat(&ts->lat_stat[i]);
1823                 reset_io_stat(&ts->bw_stat[i]);
1824                 reset_io_stat(&ts->iops_stat[i]);
1825
1826                 ts->io_bytes[i] = 0;
1827                 ts->runtime[i] = 0;
1828
1829                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1830                         ts->io_u_plat[i][j] = 0;
1831         }
1832
1833         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1834                 ts->io_u_map[i] = 0;
1835                 ts->io_u_submit[i] = 0;
1836                 ts->io_u_complete[i] = 0;
1837                 ts->io_u_lat_u[i] = 0;
1838                 ts->io_u_lat_m[i] = 0;
1839                 ts->total_submit = 0;
1840                 ts->total_complete = 0;
1841         }
1842
1843         for (i = 0; i < 3; i++) {
1844                 ts->total_io_u[i] = 0;
1845                 ts->short_io_u[i] = 0;
1846                 ts->drop_io_u[i] = 0;
1847         }
1848 }
1849
1850 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
1851 {
1852         /*
1853          * Note an entry in the log. Use the mean from the logged samples,
1854          * making sure to properly round up. Only write a log entry if we
1855          * had actual samples done.
1856          */
1857         if (iolog->avg_window[DDIR_READ].samples) {
1858                 unsigned long mr;
1859
1860                 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1861                 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed, 0);
1862         }
1863         if (iolog->avg_window[DDIR_WRITE].samples) {
1864                 unsigned long mw;
1865
1866                 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1867                 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed, 0);
1868         }
1869         if (iolog->avg_window[DDIR_TRIM].samples) {
1870                 unsigned long mw;
1871
1872                 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1873                 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed, 0);
1874         }
1875
1876         reset_io_stat(&iolog->avg_window[DDIR_READ]);
1877         reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1878         reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1879 }
1880
1881 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1882                            unsigned long val, enum fio_ddir ddir,
1883                            unsigned int bs, uint64_t offset)
1884 {
1885         unsigned long elapsed, this_window;
1886
1887         if (!ddir_rw(ddir))
1888                 return;
1889
1890         elapsed = mtime_since_now(&td->epoch);
1891
1892         /*
1893          * If no time averaging, just add the log sample.
1894          */
1895         if (!iolog->avg_msec) {
1896                 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
1897                 return;
1898         }
1899
1900         /*
1901          * Add the sample. If the time period has passed, then
1902          * add that entry to the log and clear.
1903          */
1904         add_stat_sample(&iolog->avg_window[ddir], val);
1905
1906         /*
1907          * If period hasn't passed, adding the above sample is all we
1908          * need to do.
1909          */
1910         this_window = elapsed - iolog->avg_last;
1911         if (this_window < iolog->avg_msec)
1912                 return;
1913
1914         _add_stat_to_log(iolog, elapsed);
1915
1916         iolog->avg_last = elapsed;
1917 }
1918
1919 void finalize_logs(struct thread_data *td)
1920 {
1921         unsigned long elapsed;
1922
1923         elapsed = mtime_since_now(&td->epoch);
1924
1925         if (td->clat_log)
1926                 _add_stat_to_log(td->clat_log, elapsed);
1927         if (td->slat_log)
1928                 _add_stat_to_log(td->slat_log, elapsed);
1929         if (td->lat_log)
1930                 _add_stat_to_log(td->lat_log, elapsed);
1931         if (td->bw_log)
1932                 _add_stat_to_log(td->bw_log, elapsed);
1933         if (td->iops_log)
1934                 _add_stat_to_log(td->iops_log, elapsed);
1935 }
1936
1937 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1938 {
1939         struct io_log *iolog;
1940
1941         if (!ddir_rw(ddir))
1942                 return;
1943
1944         iolog = agg_io_log[ddir];
1945         __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
1946 }
1947
1948 static void add_clat_percentile_sample(struct thread_stat *ts,
1949                                 unsigned long usec, enum fio_ddir ddir)
1950 {
1951         unsigned int idx = plat_val_to_idx(usec);
1952         assert(idx < FIO_IO_U_PLAT_NR);
1953
1954         ts->io_u_plat[ddir][idx]++;
1955 }
1956
1957 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1958                      unsigned long usec, unsigned int bs, uint64_t offset)
1959 {
1960         struct thread_stat *ts = &td->ts;
1961
1962         if (!ddir_rw(ddir))
1963                 return;
1964
1965         add_stat_sample(&ts->clat_stat[ddir], usec);
1966
1967         if (td->clat_log)
1968                 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
1969
1970         if (ts->clat_percentiles)
1971                 add_clat_percentile_sample(ts, usec, ddir);
1972 }
1973
1974 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1975                      unsigned long usec, unsigned int bs, uint64_t offset)
1976 {
1977         struct thread_stat *ts = &td->ts;
1978
1979         if (!ddir_rw(ddir))
1980                 return;
1981
1982         add_stat_sample(&ts->slat_stat[ddir], usec);
1983
1984         if (td->slat_log)
1985                 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
1986 }
1987
1988 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1989                     unsigned long usec, unsigned int bs, uint64_t offset)
1990 {
1991         struct thread_stat *ts = &td->ts;
1992
1993         if (!ddir_rw(ddir))
1994                 return;
1995
1996         add_stat_sample(&ts->lat_stat[ddir], usec);
1997
1998         if (td->lat_log)
1999                 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
2000 }
2001
2002 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
2003                    struct timeval *t)
2004 {
2005         struct thread_stat *ts = &td->ts;
2006         unsigned long spent, rate;
2007
2008         if (!ddir_rw(ddir))
2009                 return;
2010
2011         spent = mtime_since(&td->bw_sample_time, t);
2012         if (spent < td->o.bw_avg_time)
2013                 return;
2014
2015         td_io_u_lock(td);
2016
2017         /*
2018          * Compute both read and write rates for the interval.
2019          */
2020         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2021                 uint64_t delta;
2022
2023                 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2024                 if (!delta)
2025                         continue; /* No entries for interval */
2026
2027                 if (spent)
2028                         rate = delta * 1000 / spent / 1024;
2029                 else
2030                         rate = 0;
2031
2032                 add_stat_sample(&ts->bw_stat[ddir], rate);
2033
2034                 if (td->bw_log)
2035                         add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2036
2037                 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2038         }
2039
2040         fio_gettime(&td->bw_sample_time, NULL);
2041         td_io_u_unlock(td);
2042 }
2043
2044 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
2045                      struct timeval *t)
2046 {
2047         struct thread_stat *ts = &td->ts;
2048         unsigned long spent, iops;
2049
2050         if (!ddir_rw(ddir))
2051                 return;
2052
2053         spent = mtime_since(&td->iops_sample_time, t);
2054         if (spent < td->o.iops_avg_time)
2055                 return;
2056
2057         td_io_u_lock(td);
2058
2059         /*
2060          * Compute both read and write rates for the interval.
2061          */
2062         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2063                 uint64_t delta;
2064
2065                 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2066                 if (!delta)
2067                         continue; /* No entries for interval */
2068
2069                 if (spent)
2070                         iops = (delta * 1000) / spent;
2071                 else
2072                         iops = 0;
2073
2074                 add_stat_sample(&ts->iops_stat[ddir], iops);
2075
2076                 if (td->iops_log)
2077                         add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2078
2079                 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2080         }
2081
2082         fio_gettime(&td->iops_sample_time, NULL);
2083         td_io_u_unlock(td);
2084 }
2085
2086 void stat_init(void)
2087 {
2088         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2089 }
2090
2091 void stat_exit(void)
2092 {
2093         /*
2094          * When we have the mutex, we know out-of-band access to it
2095          * have ended.
2096          */
2097         fio_mutex_down(stat_mutex);
2098         fio_mutex_remove(stat_mutex);
2099 }
2100
2101 /*
2102  * Called from signal handler. Wake up status thread.
2103  */
2104 void show_running_run_stats(void)
2105 {
2106         helper_do_stat = 1;
2107         pthread_cond_signal(&helper_cond);
2108 }
2109
2110 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2111 {
2112         /* Ignore io_u's which span multiple blocks--they will just get
2113          * inaccurate counts. */
2114         int idx = (io_u->offset - io_u->file->file_offset)
2115                         / td->o.bs[DDIR_TRIM];
2116         uint32_t *info = &td->ts.block_infos[idx];
2117         assert(idx < td->ts.nr_block_infos);
2118         return info;
2119 }