stat: only apply proper stat summing for event timestamps
[fio.git] / stat.c
1 #include <stdio.h>
2 #include <string.h>
3 #include <sys/time.h>
4 #include <sys/stat.h>
5 #include <math.h>
6
7 #include "fio.h"
8 #include "diskutil.h"
9 #include "lib/ieee754.h"
10 #include "json.h"
11 #include "lib/getrusage.h"
12 #include "idletime.h"
13 #include "lib/pow2.h"
14 #include "lib/output_buffer.h"
15 #include "helper_thread.h"
16 #include "smalloc.h"
17 #include "zbd.h"
18
19 #define LOG_MSEC_SLACK  1
20
21 struct fio_sem *stat_sem;
22
23 void clear_rusage_stat(struct thread_data *td)
24 {
25         struct thread_stat *ts = &td->ts;
26
27         fio_getrusage(&td->ru_start);
28         ts->usr_time = ts->sys_time = 0;
29         ts->ctx = 0;
30         ts->minf = ts->majf = 0;
31 }
32
33 void update_rusage_stat(struct thread_data *td)
34 {
35         struct thread_stat *ts = &td->ts;
36
37         fio_getrusage(&td->ru_end);
38         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
39                                         &td->ru_end.ru_utime);
40         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
41                                         &td->ru_end.ru_stime);
42         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
43                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
44         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
45         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
46
47         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
48 }
49
50 /*
51  * Given a latency, return the index of the corresponding bucket in
52  * the structure tracking percentiles.
53  *
54  * (1) find the group (and error bits) that the value (latency)
55  * belongs to by looking at its MSB. (2) find the bucket number in the
56  * group by looking at the index bits.
57  *
58  */
59 static unsigned int plat_val_to_idx(unsigned long long val)
60 {
61         unsigned int msb, error_bits, base, offset, idx;
62
63         /* Find MSB starting from bit 0 */
64         if (val == 0)
65                 msb = 0;
66         else
67                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
68
69         /*
70          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
71          * all bits of the sample as index
72          */
73         if (msb <= FIO_IO_U_PLAT_BITS)
74                 return val;
75
76         /* Compute the number of error bits to discard*/
77         error_bits = msb - FIO_IO_U_PLAT_BITS;
78
79         /* Compute the number of buckets before the group */
80         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
81
82         /*
83          * Discard the error bits and apply the mask to find the
84          * index for the buckets in the group
85          */
86         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
87
88         /* Make sure the index does not exceed (array size - 1) */
89         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
90                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
91
92         return idx;
93 }
94
95 /*
96  * Convert the given index of the bucket array to the value
97  * represented by the bucket
98  */
99 static unsigned long long plat_idx_to_val(unsigned int idx)
100 {
101         unsigned int error_bits;
102         unsigned long long k, base;
103
104         assert(idx < FIO_IO_U_PLAT_NR);
105
106         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
107          * all bits of the sample as index */
108         if (idx < (FIO_IO_U_PLAT_VAL << 1))
109                 return idx;
110
111         /* Find the group and compute the minimum value of that group */
112         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
113         base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
114
115         /* Find its bucket number of the group */
116         k = idx % FIO_IO_U_PLAT_VAL;
117
118         /* Return the mean of the range of the bucket */
119         return base + ((k + 0.5) * (1 << error_bits));
120 }
121
122 static int double_cmp(const void *a, const void *b)
123 {
124         const fio_fp64_t fa = *(const fio_fp64_t *) a;
125         const fio_fp64_t fb = *(const fio_fp64_t *) b;
126         int cmp = 0;
127
128         if (fa.u.f > fb.u.f)
129                 cmp = 1;
130         else if (fa.u.f < fb.u.f)
131                 cmp = -1;
132
133         return cmp;
134 }
135
136 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
137                                    fio_fp64_t *plist, unsigned long long **output,
138                                    unsigned long long *maxv, unsigned long long *minv)
139 {
140         unsigned long long sum = 0;
141         unsigned int len, i, j = 0;
142         unsigned int oval_len = 0;
143         unsigned long long *ovals = NULL;
144         bool is_last;
145
146         *minv = -1ULL;
147         *maxv = 0;
148
149         len = 0;
150         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
151                 len++;
152
153         if (!len)
154                 return 0;
155
156         /*
157          * Sort the percentile list. Note that it may already be sorted if
158          * we are using the default values, but since it's a short list this
159          * isn't a worry. Also note that this does not work for NaN values.
160          */
161         if (len > 1)
162                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
163
164         /*
165          * Calculate bucket values, note down max and min values
166          */
167         is_last = false;
168         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
169                 sum += io_u_plat[i];
170                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
171                         assert(plist[j].u.f <= 100.0);
172
173                         if (j == oval_len) {
174                                 oval_len += 100;
175                                 ovals = realloc(ovals, oval_len * sizeof(*ovals));
176                         }
177
178                         ovals[j] = plat_idx_to_val(i);
179                         if (ovals[j] < *minv)
180                                 *minv = ovals[j];
181                         if (ovals[j] > *maxv)
182                                 *maxv = ovals[j];
183
184                         is_last = (j == len - 1) != 0;
185                         if (is_last)
186                                 break;
187
188                         j++;
189                 }
190         }
191
192         *output = ovals;
193         return len;
194 }
195
196 /*
197  * Find and display the p-th percentile of clat
198  */
199 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
200                                   fio_fp64_t *plist, unsigned int precision,
201                                   const char *pre, struct buf_output *out)
202 {
203         unsigned int divisor, len, i, j = 0;
204         unsigned long long minv, maxv;
205         unsigned long long *ovals;
206         int per_line, scale_down, time_width;
207         bool is_last;
208         char fmt[32];
209
210         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
211         if (!len || !ovals)
212                 goto out;
213
214         /*
215          * We default to nsecs, but if the value range is such that we
216          * should scale down to usecs or msecs, do that.
217          */
218         if (minv > 2000000 && maxv > 99999999ULL) {
219                 scale_down = 2;
220                 divisor = 1000000;
221                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
222         } else if (minv > 2000 && maxv > 99999) {
223                 scale_down = 1;
224                 divisor = 1000;
225                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
226         } else {
227                 scale_down = 0;
228                 divisor = 1;
229                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
230         }
231
232
233         time_width = max(5, (int) (log10(maxv / divisor) + 1));
234         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
235                         precision, time_width);
236         /* fmt will be something like " %5.2fth=[%4llu]%c" */
237         per_line = (80 - 7) / (precision + 10 + time_width);
238
239         for (j = 0; j < len; j++) {
240                 /* for formatting */
241                 if (j != 0 && (j % per_line) == 0)
242                         log_buf(out, "     |");
243
244                 /* end of the list */
245                 is_last = (j == len - 1) != 0;
246
247                 for (i = 0; i < scale_down; i++)
248                         ovals[j] = (ovals[j] + 999) / 1000;
249
250                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
251
252                 if (is_last)
253                         break;
254
255                 if ((j % per_line) == per_line - 1)     /* for formatting */
256                         log_buf(out, "\n");
257         }
258
259 out:
260         if (ovals)
261                 free(ovals);
262 }
263
264 bool calc_lat(struct io_stat *is, unsigned long long *min,
265               unsigned long long *max, double *mean, double *dev)
266 {
267         double n = (double) is->samples;
268
269         if (n == 0)
270                 return false;
271
272         *min = is->min_val;
273         *max = is->max_val;
274         *mean = is->mean.u.f;
275
276         if (n > 1.0)
277                 *dev = sqrt(is->S.u.f / (n - 1.0));
278         else
279                 *dev = 0;
280
281         return true;
282 }
283
284 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
285 {
286         char *io, *agg, *min, *max;
287         char *ioalt, *aggalt, *minalt, *maxalt;
288         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
289         int i;
290
291         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
292
293         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
294                 const int i2p = is_power_of_2(rs->kb_base);
295
296                 if (!rs->max_run[i])
297                         continue;
298
299                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
300                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
301                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
302                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
303                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
304                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
305                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
306                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
307                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
308                                 rs->unified_rw_rep ? "  MIXED" : str[i],
309                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
310                                 (unsigned long long) rs->min_run[i],
311                                 (unsigned long long) rs->max_run[i]);
312
313                 free(io);
314                 free(agg);
315                 free(min);
316                 free(max);
317                 free(ioalt);
318                 free(aggalt);
319                 free(minalt);
320                 free(maxalt);
321         }
322 }
323
324 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
325 {
326         int i;
327
328         /*
329          * Do depth distribution calculations
330          */
331         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
332                 if (total) {
333                         io_u_dist[i] = (double) map[i] / (double) total;
334                         io_u_dist[i] *= 100.0;
335                         if (io_u_dist[i] < 0.1 && map[i])
336                                 io_u_dist[i] = 0.1;
337                 } else
338                         io_u_dist[i] = 0.0;
339         }
340 }
341
342 static void stat_calc_lat(struct thread_stat *ts, double *dst,
343                           uint64_t *src, int nr)
344 {
345         unsigned long total = ddir_rw_sum(ts->total_io_u);
346         int i;
347
348         /*
349          * Do latency distribution calculations
350          */
351         for (i = 0; i < nr; i++) {
352                 if (total) {
353                         dst[i] = (double) src[i] / (double) total;
354                         dst[i] *= 100.0;
355                         if (dst[i] < 0.01 && src[i])
356                                 dst[i] = 0.01;
357                 } else
358                         dst[i] = 0.0;
359         }
360 }
361
362 /*
363  * To keep the terse format unaltered, add all of the ns latency
364  * buckets to the first us latency bucket
365  */
366 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
367 {
368         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
369         int i;
370
371         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
372
373         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
374                 ntotal += ts->io_u_lat_n[i];
375
376         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
377 }
378
379 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
380 {
381         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
382 }
383
384 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
385 {
386         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
387 }
388
389 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
390 {
391         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
392 }
393
394 static void display_lat(const char *name, unsigned long long min,
395                         unsigned long long max, double mean, double dev,
396                         struct buf_output *out)
397 {
398         const char *base = "(nsec)";
399         char *minp, *maxp;
400
401         if (nsec_to_msec(&min, &max, &mean, &dev))
402                 base = "(msec)";
403         else if (nsec_to_usec(&min, &max, &mean, &dev))
404                 base = "(usec)";
405
406         minp = num2str(min, 6, 1, 0, N2S_NONE);
407         maxp = num2str(max, 6, 1, 0, N2S_NONE);
408
409         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
410                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
411
412         free(minp);
413         free(maxp);
414 }
415
416 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
417                              int ddir, struct buf_output *out)
418 {
419         unsigned long runt;
420         unsigned long long min, max, bw, iops;
421         double mean, dev;
422         char *io_p, *bw_p, *bw_p_alt, *iops_p, *zbd_w_st = NULL;
423         int i2p;
424
425         if (ddir_sync(ddir)) {
426                 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
427                         log_buf(out, "  %s:\n", "fsync/fdatasync/sync_file_range");
428                         display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
429                         show_clat_percentiles(ts->io_u_sync_plat,
430                                                 ts->sync_stat.samples,
431                                                 ts->percentile_list,
432                                                 ts->percentile_precision,
433                                                 io_ddir_name(ddir), out);
434                 }
435                 return;
436         }
437
438         assert(ddir_rw(ddir));
439
440         if (!ts->runtime[ddir])
441                 return;
442
443         i2p = is_power_of_2(rs->kb_base);
444         runt = ts->runtime[ddir];
445
446         bw = (1000 * ts->io_bytes[ddir]) / runt;
447         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
448         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
449         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
450
451         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
452         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
453         if (ddir == DDIR_WRITE)
454                 zbd_w_st = zbd_write_status(ts);
455
456         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
457                         rs->unified_rw_rep ? "mixed" : io_ddir_name(ddir),
458                         iops_p, bw_p, bw_p_alt, io_p,
459                         (unsigned long long) ts->runtime[ddir],
460                         zbd_w_st ? : "");
461
462         free(zbd_w_st);
463         free(io_p);
464         free(bw_p);
465         free(bw_p_alt);
466         free(iops_p);
467
468         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
469                 display_lat("slat", min, max, mean, dev, out);
470         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
471                 display_lat("clat", min, max, mean, dev, out);
472         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
473                 display_lat(" lat", min, max, mean, dev, out);
474
475         if (ts->clat_percentiles || ts->lat_percentiles) {
476                 const char *name = ts->clat_percentiles ? "clat" : " lat";
477                 uint64_t samples;
478
479                 if (ts->clat_percentiles)
480                         samples = ts->clat_stat[ddir].samples;
481                 else
482                         samples = ts->lat_stat[ddir].samples;
483
484                 show_clat_percentiles(ts->io_u_plat[ddir],
485                                         samples,
486                                         ts->percentile_list,
487                                         ts->percentile_precision, name, out);
488         }
489         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
490                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
491                 const char *bw_str;
492
493                 if ((rs->unit_base == 1) && i2p)
494                         bw_str = "Kibit";
495                 else if (rs->unit_base == 1)
496                         bw_str = "kbit";
497                 else if (i2p)
498                         bw_str = "KiB";
499                 else
500                         bw_str = "kB";
501
502                 if (rs->agg[ddir]) {
503                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
504                         if (p_of_agg > 100.0)
505                                 p_of_agg = 100.0;
506                 }
507
508                 if (rs->unit_base == 1) {
509                         min *= 8.0;
510                         max *= 8.0;
511                         mean *= 8.0;
512                         dev *= 8.0;
513                 }
514
515                 if (mean > fkb_base * fkb_base) {
516                         min /= fkb_base;
517                         max /= fkb_base;
518                         mean /= fkb_base;
519                         dev /= fkb_base;
520                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
521                 }
522
523                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
524                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
525                         bw_str, min, max, p_of_agg, mean, dev,
526                         (&ts->bw_stat[ddir])->samples);
527         }
528         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
529                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
530                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
531                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
532         }
533 }
534
535 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
536                      const char *msg, struct buf_output *out)
537 {
538         bool new_line = true, shown = false;
539         int i, line = 0;
540
541         for (i = 0; i < nr; i++) {
542                 if (io_u_lat[i] <= 0.0)
543                         continue;
544                 shown = true;
545                 if (new_line) {
546                         if (line)
547                                 log_buf(out, "\n");
548                         log_buf(out, "  lat (%s)   : ", msg);
549                         new_line = false;
550                         line = 0;
551                 }
552                 if (line)
553                         log_buf(out, ", ");
554                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
555                 line++;
556                 if (line == 5)
557                         new_line = true;
558         }
559
560         if (shown)
561                 log_buf(out, "\n");
562
563         return true;
564 }
565
566 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
567 {
568         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
569                                  "250=", "500=", "750=", "1000=", };
570
571         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
572 }
573
574 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
575 {
576         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
577                                  "250=", "500=", "750=", "1000=", };
578
579         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
580 }
581
582 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
583 {
584         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
585                                  "250=", "500=", "750=", "1000=", "2000=",
586                                  ">=2000=", };
587
588         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
589 }
590
591 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
592 {
593         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
594         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
595         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
596
597         stat_calc_lat_n(ts, io_u_lat_n);
598         stat_calc_lat_u(ts, io_u_lat_u);
599         stat_calc_lat_m(ts, io_u_lat_m);
600
601         show_lat_n(io_u_lat_n, out);
602         show_lat_u(io_u_lat_u, out);
603         show_lat_m(io_u_lat_m, out);
604 }
605
606 static int block_state_category(int block_state)
607 {
608         switch (block_state) {
609         case BLOCK_STATE_UNINIT:
610                 return 0;
611         case BLOCK_STATE_TRIMMED:
612         case BLOCK_STATE_WRITTEN:
613                 return 1;
614         case BLOCK_STATE_WRITE_FAILURE:
615         case BLOCK_STATE_TRIM_FAILURE:
616                 return 2;
617         default:
618                 /* Silence compile warning on some BSDs and have a return */
619                 assert(0);
620                 return -1;
621         }
622 }
623
624 static int compare_block_infos(const void *bs1, const void *bs2)
625 {
626         uint64_t block1 = *(uint64_t *)bs1;
627         uint64_t block2 = *(uint64_t *)bs2;
628         int state1 = BLOCK_INFO_STATE(block1);
629         int state2 = BLOCK_INFO_STATE(block2);
630         int bscat1 = block_state_category(state1);
631         int bscat2 = block_state_category(state2);
632         int cycles1 = BLOCK_INFO_TRIMS(block1);
633         int cycles2 = BLOCK_INFO_TRIMS(block2);
634
635         if (bscat1 < bscat2)
636                 return -1;
637         if (bscat1 > bscat2)
638                 return 1;
639
640         if (cycles1 < cycles2)
641                 return -1;
642         if (cycles1 > cycles2)
643                 return 1;
644
645         if (state1 < state2)
646                 return -1;
647         if (state1 > state2)
648                 return 1;
649
650         assert(block1 == block2);
651         return 0;
652 }
653
654 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
655                                   fio_fp64_t *plist, unsigned int **percentiles,
656                                   unsigned int *types)
657 {
658         int len = 0;
659         int i, nr_uninit;
660
661         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
662
663         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
664                 len++;
665
666         if (!len)
667                 return 0;
668
669         /*
670          * Sort the percentile list. Note that it may already be sorted if
671          * we are using the default values, but since it's a short list this
672          * isn't a worry. Also note that this does not work for NaN values.
673          */
674         if (len > 1)
675                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
676
677         /* Start only after the uninit entries end */
678         for (nr_uninit = 0;
679              nr_uninit < nr_block_infos
680                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
681              nr_uninit ++)
682                 ;
683
684         if (nr_uninit == nr_block_infos)
685                 return 0;
686
687         *percentiles = calloc(len, sizeof(**percentiles));
688
689         for (i = 0; i < len; i++) {
690                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
691                                 + nr_uninit;
692                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
693         }
694
695         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
696         for (i = 0; i < nr_block_infos; i++)
697                 types[BLOCK_INFO_STATE(block_infos[i])]++;
698
699         return len;
700 }
701
702 static const char *block_state_names[] = {
703         [BLOCK_STATE_UNINIT] = "unwritten",
704         [BLOCK_STATE_TRIMMED] = "trimmed",
705         [BLOCK_STATE_WRITTEN] = "written",
706         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
707         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
708 };
709
710 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
711                              fio_fp64_t *plist, struct buf_output *out)
712 {
713         int len, pos, i;
714         unsigned int *percentiles = NULL;
715         unsigned int block_state_counts[BLOCK_STATE_COUNT];
716
717         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
718                                      &percentiles, block_state_counts);
719
720         log_buf(out, "  block lifetime percentiles :\n   |");
721         pos = 0;
722         for (i = 0; i < len; i++) {
723                 uint32_t block_info = percentiles[i];
724 #define LINE_LENGTH     75
725                 char str[LINE_LENGTH];
726                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
727                                      plist[i].u.f, block_info,
728                                      i == len - 1 ? '\n' : ',');
729                 assert(strln < LINE_LENGTH);
730                 if (pos + strln > LINE_LENGTH) {
731                         pos = 0;
732                         log_buf(out, "\n   |");
733                 }
734                 log_buf(out, "%s", str);
735                 pos += strln;
736 #undef LINE_LENGTH
737         }
738         if (percentiles)
739                 free(percentiles);
740
741         log_buf(out, "        states               :");
742         for (i = 0; i < BLOCK_STATE_COUNT; i++)
743                 log_buf(out, " %s=%u%c",
744                          block_state_names[i], block_state_counts[i],
745                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
746 }
747
748 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
749 {
750         char *p1, *p1alt, *p2;
751         unsigned long long bw_mean, iops_mean;
752         const int i2p = is_power_of_2(ts->kb_base);
753
754         if (!ts->ss_dur)
755                 return;
756
757         bw_mean = steadystate_bw_mean(ts);
758         iops_mean = steadystate_iops_mean(ts);
759
760         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
761         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
762         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
763
764         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
765                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
766                 p1, p1alt, p2,
767                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
768                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
769                 ts->ss_criterion.u.f,
770                 ts->ss_state & FIO_SS_PCT ? "%" : "");
771
772         free(p1);
773         free(p1alt);
774         free(p2);
775 }
776
777 static void show_thread_status_normal(struct thread_stat *ts,
778                                       struct group_run_stats *rs,
779                                       struct buf_output *out)
780 {
781         double usr_cpu, sys_cpu;
782         unsigned long runtime;
783         double io_u_dist[FIO_IO_U_MAP_NR];
784         time_t time_p;
785         char time_buf[32];
786
787         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
788                 return;
789                 
790         memset(time_buf, 0, sizeof(time_buf));
791
792         time(&time_p);
793         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
794
795         if (!ts->error) {
796                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
797                                         ts->name, ts->groupid, ts->members,
798                                         ts->error, (int) ts->pid, time_buf);
799         } else {
800                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
801                                         ts->name, ts->groupid, ts->members,
802                                         ts->error, ts->verror, (int) ts->pid,
803                                         time_buf);
804         }
805
806         if (strlen(ts->description))
807                 log_buf(out, "  Description  : [%s]\n", ts->description);
808
809         if (ts->io_bytes[DDIR_READ])
810                 show_ddir_status(rs, ts, DDIR_READ, out);
811         if (ts->io_bytes[DDIR_WRITE])
812                 show_ddir_status(rs, ts, DDIR_WRITE, out);
813         if (ts->io_bytes[DDIR_TRIM])
814                 show_ddir_status(rs, ts, DDIR_TRIM, out);
815
816         show_latencies(ts, out);
817
818         if (ts->sync_stat.samples)
819                 show_ddir_status(rs, ts, DDIR_SYNC, out);
820
821         runtime = ts->total_run_time;
822         if (runtime) {
823                 double runt = (double) runtime;
824
825                 usr_cpu = (double) ts->usr_time * 100 / runt;
826                 sys_cpu = (double) ts->sys_time * 100 / runt;
827         } else {
828                 usr_cpu = 0;
829                 sys_cpu = 0;
830         }
831
832         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
833                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
834                         (unsigned long long) ts->ctx,
835                         (unsigned long long) ts->majf,
836                         (unsigned long long) ts->minf);
837
838         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
839         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
840                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
841                                         io_u_dist[1], io_u_dist[2],
842                                         io_u_dist[3], io_u_dist[4],
843                                         io_u_dist[5], io_u_dist[6]);
844
845         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
846         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
847                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
848                                         io_u_dist[1], io_u_dist[2],
849                                         io_u_dist[3], io_u_dist[4],
850                                         io_u_dist[5], io_u_dist[6]);
851         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
852         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
853                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
854                                         io_u_dist[1], io_u_dist[2],
855                                         io_u_dist[3], io_u_dist[4],
856                                         io_u_dist[5], io_u_dist[6]);
857         log_buf(out, "     issued rwts: total=%llu,%llu,%llu,%llu"
858                                  " short=%llu,%llu,%llu,0"
859                                  " dropped=%llu,%llu,%llu,0\n",
860                                         (unsigned long long) ts->total_io_u[0],
861                                         (unsigned long long) ts->total_io_u[1],
862                                         (unsigned long long) ts->total_io_u[2],
863                                         (unsigned long long) ts->total_io_u[3],
864                                         (unsigned long long) ts->short_io_u[0],
865                                         (unsigned long long) ts->short_io_u[1],
866                                         (unsigned long long) ts->short_io_u[2],
867                                         (unsigned long long) ts->drop_io_u[0],
868                                         (unsigned long long) ts->drop_io_u[1],
869                                         (unsigned long long) ts->drop_io_u[2]);
870         if (ts->continue_on_error) {
871                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
872                                         (unsigned long long)ts->total_err_count,
873                                         ts->first_error,
874                                         strerror(ts->first_error));
875         }
876         if (ts->latency_depth) {
877                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
878                                         (unsigned long long)ts->latency_target,
879                                         (unsigned long long)ts->latency_window,
880                                         ts->latency_percentile.u.f,
881                                         ts->latency_depth);
882         }
883
884         if (ts->nr_block_infos)
885                 show_block_infos(ts->nr_block_infos, ts->block_infos,
886                                   ts->percentile_list, out);
887
888         if (ts->ss_dur)
889                 show_ss_normal(ts, out);
890 }
891
892 static void show_ddir_status_terse(struct thread_stat *ts,
893                                    struct group_run_stats *rs, int ddir,
894                                    int ver, struct buf_output *out)
895 {
896         unsigned long long min, max, minv, maxv, bw, iops;
897         unsigned long long *ovals = NULL;
898         double mean, dev;
899         unsigned int len;
900         int i, bw_stat;
901
902         assert(ddir_rw(ddir));
903
904         iops = bw = 0;
905         if (ts->runtime[ddir]) {
906                 uint64_t runt = ts->runtime[ddir];
907
908                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
909                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
910         }
911
912         log_buf(out, ";%llu;%llu;%llu;%llu",
913                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
914                                         (unsigned long long) ts->runtime[ddir]);
915
916         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
917                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
918         else
919                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
920
921         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
922                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
923         else
924                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
925
926         if (ts->clat_percentiles || ts->lat_percentiles) {
927                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
928                                         ts->clat_stat[ddir].samples,
929                                         ts->percentile_list, &ovals, &maxv,
930                                         &minv);
931         } else
932                 len = 0;
933
934         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
935                 if (i >= len) {
936                         log_buf(out, ";0%%=0");
937                         continue;
938                 }
939                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
940         }
941
942         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
943                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
944         else
945                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
946
947         if (ovals)
948                 free(ovals);
949
950         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
951         if (bw_stat) {
952                 double p_of_agg = 100.0;
953
954                 if (rs->agg[ddir]) {
955                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
956                         if (p_of_agg > 100.0)
957                                 p_of_agg = 100.0;
958                 }
959
960                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
961         } else
962                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
963
964         if (ver == 5) {
965                 if (bw_stat)
966                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
967                 else
968                         log_buf(out, ";%lu", 0UL);
969
970                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
971                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
972                                 mean, dev, (&ts->iops_stat[ddir])->samples);
973                 else
974                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
975         }
976 }
977
978 static void add_ddir_status_json(struct thread_stat *ts,
979                 struct group_run_stats *rs, int ddir, struct json_object *parent)
980 {
981         unsigned long long min, max, minv, maxv;
982         unsigned long long bw_bytes, bw;
983         unsigned long long *ovals = NULL;
984         double mean, dev, iops;
985         unsigned int len;
986         int i;
987         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object = NULL;
988         char buf[120];
989         double p_of_agg = 100.0;
990
991         assert(ddir_rw(ddir) || ddir_sync(ddir));
992
993         if (ts->unified_rw_rep && ddir != DDIR_READ)
994                 return;
995
996         dir_object = json_create_object();
997         json_object_add_value_object(parent,
998                 ts->unified_rw_rep ? "mixed" : io_ddir_name(ddir), dir_object);
999
1000         if (ddir_rw(ddir)) {
1001                 bw_bytes = 0;
1002                 bw = 0;
1003                 iops = 0.0;
1004                 if (ts->runtime[ddir]) {
1005                         uint64_t runt = ts->runtime[ddir];
1006
1007                         bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1008                         bw = bw_bytes / 1024; /* KiB/s */
1009                         iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1010                 }
1011
1012                 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1013                 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1014                 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1015                 json_object_add_value_int(dir_object, "bw", bw);
1016                 json_object_add_value_float(dir_object, "iops", iops);
1017                 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1018                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1019                 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1020                 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1021
1022                 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
1023                         min = max = 0;
1024                         mean = dev = 0.0;
1025                 }
1026                 tmp_object = json_create_object();
1027                 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1028                 json_object_add_value_int(tmp_object, "min", min);
1029                 json_object_add_value_int(tmp_object, "max", max);
1030                 json_object_add_value_float(tmp_object, "mean", mean);
1031                 json_object_add_value_float(tmp_object, "stddev", dev);
1032
1033                 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1034                         min = max = 0;
1035                         mean = dev = 0.0;
1036                 }
1037                 tmp_object = json_create_object();
1038                 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1039                 json_object_add_value_int(tmp_object, "min", min);
1040                 json_object_add_value_int(tmp_object, "max", max);
1041                 json_object_add_value_float(tmp_object, "mean", mean);
1042                 json_object_add_value_float(tmp_object, "stddev", dev);
1043         } else {
1044                 if (!calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
1045                         min = max = 0;
1046                         mean = dev = 0.0;
1047                 }
1048
1049                 tmp_object = json_create_object();
1050                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1051                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1052                 json_object_add_value_int(tmp_object, "min", min);
1053                 json_object_add_value_int(tmp_object, "max", max);
1054                 json_object_add_value_float(tmp_object, "mean", mean);
1055                 json_object_add_value_float(tmp_object, "stddev", dev);
1056         }
1057
1058         if (ts->clat_percentiles || ts->lat_percentiles) {
1059                 if (ddir_rw(ddir)) {
1060                         uint64_t samples;
1061
1062                         if (ts->clat_percentiles)
1063                                 samples = ts->clat_stat[ddir].samples;
1064                         else
1065                                 samples = ts->lat_stat[ddir].samples;
1066
1067                         len = calc_clat_percentiles(ts->io_u_plat[ddir],
1068                                         samples, ts->percentile_list, &ovals,
1069                                         &maxv, &minv);
1070                 } else {
1071                         len = calc_clat_percentiles(ts->io_u_sync_plat,
1072                                         ts->sync_stat.samples,
1073                                         ts->percentile_list, &ovals, &maxv,
1074                                         &minv);
1075                 }
1076
1077                 if (len > FIO_IO_U_LIST_MAX_LEN)
1078                         len = FIO_IO_U_LIST_MAX_LEN;
1079         } else
1080                 len = 0;
1081
1082         percentile_object = json_create_object();
1083         json_object_add_value_object(tmp_object, "percentile", percentile_object);
1084         for (i = 0; i < len; i++) {
1085                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1086                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1087         }
1088
1089         if (output_format & FIO_OUTPUT_JSON_PLUS) {
1090                 clat_bins_object = json_create_object();
1091                 if (ts->clat_percentiles)
1092                         json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1093
1094                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1095                         if (ddir_rw(ddir)) {
1096                                 if (ts->io_u_plat[ddir][i]) {
1097                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1098                                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1099                                 }
1100                         } else {
1101                                 if (ts->io_u_sync_plat[i]) {
1102                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1103                                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_sync_plat[i]);
1104                                 }
1105                         }
1106                 }
1107         }
1108
1109         if (!ddir_rw(ddir))
1110                 return;
1111
1112         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1113                 min = max = 0;
1114                 mean = dev = 0.0;
1115         }
1116         tmp_object = json_create_object();
1117         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1118         json_object_add_value_int(tmp_object, "min", min);
1119         json_object_add_value_int(tmp_object, "max", max);
1120         json_object_add_value_float(tmp_object, "mean", mean);
1121         json_object_add_value_float(tmp_object, "stddev", dev);
1122         if (output_format & FIO_OUTPUT_JSON_PLUS && ts->lat_percentiles)
1123                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1124
1125         if (ovals)
1126                 free(ovals);
1127
1128         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1129                 if (rs->agg[ddir]) {
1130                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1131                         if (p_of_agg > 100.0)
1132                                 p_of_agg = 100.0;
1133                 }
1134         } else {
1135                 min = max = 0;
1136                 p_of_agg = mean = dev = 0.0;
1137         }
1138         json_object_add_value_int(dir_object, "bw_min", min);
1139         json_object_add_value_int(dir_object, "bw_max", max);
1140         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1141         json_object_add_value_float(dir_object, "bw_mean", mean);
1142         json_object_add_value_float(dir_object, "bw_dev", dev);
1143         json_object_add_value_int(dir_object, "bw_samples",
1144                                 (&ts->bw_stat[ddir])->samples);
1145
1146         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1147                 min = max = 0;
1148                 mean = dev = 0.0;
1149         }
1150         json_object_add_value_int(dir_object, "iops_min", min);
1151         json_object_add_value_int(dir_object, "iops_max", max);
1152         json_object_add_value_float(dir_object, "iops_mean", mean);
1153         json_object_add_value_float(dir_object, "iops_stddev", dev);
1154         json_object_add_value_int(dir_object, "iops_samples",
1155                                 (&ts->iops_stat[ddir])->samples);
1156 }
1157
1158 static void show_thread_status_terse_all(struct thread_stat *ts,
1159                                          struct group_run_stats *rs, int ver,
1160                                          struct buf_output *out)
1161 {
1162         double io_u_dist[FIO_IO_U_MAP_NR];
1163         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1164         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1165         double usr_cpu, sys_cpu;
1166         int i;
1167
1168         /* General Info */
1169         if (ver == 2)
1170                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1171         else
1172                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1173                         ts->name, ts->groupid, ts->error);
1174
1175         /* Log Read Status */
1176         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1177         /* Log Write Status */
1178         show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1179         /* Log Trim Status */
1180         if (ver == 2 || ver == 4 || ver == 5)
1181                 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1182
1183         /* CPU Usage */
1184         if (ts->total_run_time) {
1185                 double runt = (double) ts->total_run_time;
1186
1187                 usr_cpu = (double) ts->usr_time * 100 / runt;
1188                 sys_cpu = (double) ts->sys_time * 100 / runt;
1189         } else {
1190                 usr_cpu = 0;
1191                 sys_cpu = 0;
1192         }
1193
1194         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1195                                                 (unsigned long long) ts->ctx,
1196                                                 (unsigned long long) ts->majf,
1197                                                 (unsigned long long) ts->minf);
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_nu(ts, io_u_lat_u);
1202         stat_calc_lat_m(ts, io_u_lat_m);
1203
1204         /* Only show fixed 7 I/O depth levels*/
1205         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1206                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1207                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1208
1209         /* Microsecond latency */
1210         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1211                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1212         /* Millisecond latency */
1213         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1214                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1215
1216         /* disk util stats, if any */
1217         if (ver >= 3 && is_running_backend())
1218                 show_disk_util(1, NULL, out);
1219
1220         /* Additional output if continue_on_error set - default off*/
1221         if (ts->continue_on_error)
1222                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1223         if (ver == 2)
1224                 log_buf(out, "\n");
1225
1226         /* Additional output if description is set */
1227         if (strlen(ts->description))
1228                 log_buf(out, ";%s", ts->description);
1229
1230         log_buf(out, "\n");
1231 }
1232
1233 static void json_add_job_opts(struct json_object *root, const char *name,
1234                               struct flist_head *opt_list)
1235 {
1236         struct json_object *dir_object;
1237         struct flist_head *entry;
1238         struct print_option *p;
1239
1240         if (flist_empty(opt_list))
1241                 return;
1242
1243         dir_object = json_create_object();
1244         json_object_add_value_object(root, name, dir_object);
1245
1246         flist_for_each(entry, opt_list) {
1247                 const char *pos = "";
1248
1249                 p = flist_entry(entry, struct print_option, list);
1250                 if (p->value)
1251                         pos = p->value;
1252                 json_object_add_value_string(dir_object, p->name, pos);
1253         }
1254 }
1255
1256 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1257                                                    struct group_run_stats *rs,
1258                                                    struct flist_head *opt_list)
1259 {
1260         struct json_object *root, *tmp;
1261         struct jobs_eta *je;
1262         double io_u_dist[FIO_IO_U_MAP_NR];
1263         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1264         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1265         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1266         double usr_cpu, sys_cpu;
1267         int i;
1268         size_t size;
1269
1270         root = json_create_object();
1271         json_object_add_value_string(root, "jobname", ts->name);
1272         json_object_add_value_int(root, "groupid", ts->groupid);
1273         json_object_add_value_int(root, "error", ts->error);
1274
1275         /* ETA Info */
1276         je = get_jobs_eta(true, &size);
1277         if (je) {
1278                 json_object_add_value_int(root, "eta", je->eta_sec);
1279                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1280         }
1281
1282         if (opt_list)
1283                 json_add_job_opts(root, "job options", opt_list);
1284
1285         add_ddir_status_json(ts, rs, DDIR_READ, root);
1286         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1287         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1288         add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1289
1290         /* CPU Usage */
1291         if (ts->total_run_time) {
1292                 double runt = (double) ts->total_run_time;
1293
1294                 usr_cpu = (double) ts->usr_time * 100 / runt;
1295                 sys_cpu = (double) ts->sys_time * 100 / runt;
1296         } else {
1297                 usr_cpu = 0;
1298                 sys_cpu = 0;
1299         }
1300         json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1301         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1302         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1303         json_object_add_value_int(root, "ctx", ts->ctx);
1304         json_object_add_value_int(root, "majf", ts->majf);
1305         json_object_add_value_int(root, "minf", ts->minf);
1306
1307         /* Calc % distribution of IO depths */
1308         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1309         tmp = json_create_object();
1310         json_object_add_value_object(root, "iodepth_level", tmp);
1311         /* Only show fixed 7 I/O depth levels*/
1312         for (i = 0; i < 7; i++) {
1313                 char name[20];
1314                 if (i < 6)
1315                         snprintf(name, 20, "%d", 1 << i);
1316                 else
1317                         snprintf(name, 20, ">=%d", 1 << i);
1318                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1319         }
1320
1321         /* Calc % distribution of submit IO depths */
1322         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1323         tmp = json_create_object();
1324         json_object_add_value_object(root, "iodepth_submit", tmp);
1325         /* Only show fixed 7 I/O depth levels*/
1326         for (i = 0; i < 7; i++) {
1327                 char name[20];
1328                 if (i == 0)
1329                         snprintf(name, 20, "0");
1330                 else if (i < 6)
1331                         snprintf(name, 20, "%d", 1 << (i+1));
1332                 else
1333                         snprintf(name, 20, ">=%d", 1 << i);
1334                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1335         }
1336
1337         /* Calc % distribution of completion IO depths */
1338         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1339         tmp = json_create_object();
1340         json_object_add_value_object(root, "iodepth_complete", tmp);
1341         /* Only show fixed 7 I/O depth levels*/
1342         for (i = 0; i < 7; i++) {
1343                 char name[20];
1344                 if (i == 0)
1345                         snprintf(name, 20, "0");
1346                 else if (i < 6)
1347                         snprintf(name, 20, "%d", 1 << (i+1));
1348                 else
1349                         snprintf(name, 20, ">=%d", 1 << i);
1350                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1351         }
1352
1353         /* Calc % distribution of nsecond, usecond, msecond latency */
1354         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1355         stat_calc_lat_n(ts, io_u_lat_n);
1356         stat_calc_lat_u(ts, io_u_lat_u);
1357         stat_calc_lat_m(ts, io_u_lat_m);
1358
1359         /* Nanosecond latency */
1360         tmp = json_create_object();
1361         json_object_add_value_object(root, "latency_ns", tmp);
1362         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1363                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1364                                  "250", "500", "750", "1000", };
1365                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1366         }
1367         /* Microsecond latency */
1368         tmp = json_create_object();
1369         json_object_add_value_object(root, "latency_us", tmp);
1370         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1371                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1372                                  "250", "500", "750", "1000", };
1373                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1374         }
1375         /* Millisecond latency */
1376         tmp = json_create_object();
1377         json_object_add_value_object(root, "latency_ms", tmp);
1378         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1379                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1380                                  "250", "500", "750", "1000", "2000",
1381                                  ">=2000", };
1382                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1383         }
1384
1385         /* Additional output if continue_on_error set - default off*/
1386         if (ts->continue_on_error) {
1387                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1388                 json_object_add_value_int(root, "first_error", ts->first_error);
1389         }
1390
1391         if (ts->latency_depth) {
1392                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1393                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1394                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1395                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1396         }
1397
1398         /* Additional output if description is set */
1399         if (strlen(ts->description))
1400                 json_object_add_value_string(root, "desc", ts->description);
1401
1402         if (ts->nr_block_infos) {
1403                 /* Block error histogram and types */
1404                 int len;
1405                 unsigned int *percentiles = NULL;
1406                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1407
1408                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1409                                              ts->percentile_list,
1410                                              &percentiles, block_state_counts);
1411
1412                 if (len) {
1413                         struct json_object *block, *percentile_object, *states;
1414                         int state;
1415                         block = json_create_object();
1416                         json_object_add_value_object(root, "block", block);
1417
1418                         percentile_object = json_create_object();
1419                         json_object_add_value_object(block, "percentiles",
1420                                                      percentile_object);
1421                         for (i = 0; i < len; i++) {
1422                                 char buf[20];
1423                                 snprintf(buf, sizeof(buf), "%f",
1424                                          ts->percentile_list[i].u.f);
1425                                 json_object_add_value_int(percentile_object,
1426                                                           (const char *)buf,
1427                                                           percentiles[i]);
1428                         }
1429
1430                         states = json_create_object();
1431                         json_object_add_value_object(block, "states", states);
1432                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1433                                 json_object_add_value_int(states,
1434                                         block_state_names[state],
1435                                         block_state_counts[state]);
1436                         }
1437                         free(percentiles);
1438                 }
1439         }
1440
1441         if (ts->ss_dur) {
1442                 struct json_object *data;
1443                 struct json_array *iops, *bw;
1444                 int j, k, l;
1445                 char ss_buf[64];
1446
1447                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1448                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1449                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1450                         (float) ts->ss_limit.u.f,
1451                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1452
1453                 tmp = json_create_object();
1454                 json_object_add_value_object(root, "steadystate", tmp);
1455                 json_object_add_value_string(tmp, "ss", ss_buf);
1456                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1457                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1458
1459                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1460                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1461                 json_object_add_value_string(tmp, "criterion", ss_buf);
1462                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1463                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1464
1465                 data = json_create_object();
1466                 json_object_add_value_object(tmp, "data", data);
1467                 bw = json_create_array();
1468                 iops = json_create_array();
1469
1470                 /*
1471                 ** if ss was attained or the buffer is not full,
1472                 ** ss->head points to the first element in the list.
1473                 ** otherwise it actually points to the second element
1474                 ** in the list
1475                 */
1476                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1477                         j = ts->ss_head;
1478                 else
1479                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1480                 for (l = 0; l < ts->ss_dur; l++) {
1481                         k = (j + l) % ts->ss_dur;
1482                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1483                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1484                 }
1485                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1486                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1487                 json_object_add_value_array(data, "iops", iops);
1488                 json_object_add_value_array(data, "bw", bw);
1489         }
1490
1491         return root;
1492 }
1493
1494 static void show_thread_status_terse(struct thread_stat *ts,
1495                                      struct group_run_stats *rs,
1496                                      struct buf_output *out)
1497 {
1498         if (terse_version >= 2 && terse_version <= 5)
1499                 show_thread_status_terse_all(ts, rs, terse_version, out);
1500         else
1501                 log_err("fio: bad terse version!? %d\n", terse_version);
1502 }
1503
1504 struct json_object *show_thread_status(struct thread_stat *ts,
1505                                        struct group_run_stats *rs,
1506                                        struct flist_head *opt_list,
1507                                        struct buf_output *out)
1508 {
1509         struct json_object *ret = NULL;
1510
1511         if (output_format & FIO_OUTPUT_TERSE)
1512                 show_thread_status_terse(ts, rs,  out);
1513         if (output_format & FIO_OUTPUT_JSON)
1514                 ret = show_thread_status_json(ts, rs, opt_list);
1515         if (output_format & FIO_OUTPUT_NORMAL)
1516                 show_thread_status_normal(ts, rs,  out);
1517
1518         return ret;
1519 }
1520
1521 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1522 {
1523         double mean, S;
1524
1525         dst->min_val = min(dst->min_val, src->min_val);
1526         dst->max_val = max(dst->max_val, src->max_val);
1527
1528         /*
1529          * Compute new mean and S after the merge
1530          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1531          *  #Parallel_algorithm>
1532          */
1533         if (first) {
1534                 mean = src->mean.u.f;
1535                 S = src->S.u.f;
1536         } else {
1537                 double delta = src->mean.u.f - dst->mean.u.f;
1538
1539                 mean = ((src->mean.u.f * src->samples) +
1540                         (dst->mean.u.f * dst->samples)) /
1541                         (dst->samples + src->samples);
1542
1543                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1544                         (dst->samples * src->samples) /
1545                         (dst->samples + src->samples);
1546         }
1547
1548         dst->samples += src->samples;
1549         dst->mean.u.f = mean;
1550         dst->S.u.f = S;
1551
1552 }
1553
1554 /*
1555  * We sum two kinds of stats - one that is time based, in which case we
1556  * apply the proper summing technique, and then one that is iops/bw
1557  * numbers. For group_reporting, we should just add those up, not make
1558  * them the mean of everything.
1559  */
1560 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first,
1561                      bool pure_sum)
1562 {
1563         if (src->samples == 0)
1564                 return;
1565
1566         if (!pure_sum) {
1567                 __sum_stat(dst, src, first);
1568                 return;
1569         }
1570
1571         dst->min_val += src->min_val;
1572         dst->max_val += src->max_val;
1573         dst->samples += src->samples;
1574         dst->mean.u.f += src->mean.u.f;
1575         dst->S.u.f += src->S.u.f;
1576 }
1577
1578 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1579 {
1580         int i;
1581
1582         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1583                 if (dst->max_run[i] < src->max_run[i])
1584                         dst->max_run[i] = src->max_run[i];
1585                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1586                         dst->min_run[i] = src->min_run[i];
1587                 if (dst->max_bw[i] < src->max_bw[i])
1588                         dst->max_bw[i] = src->max_bw[i];
1589                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1590                         dst->min_bw[i] = src->min_bw[i];
1591
1592                 dst->iobytes[i] += src->iobytes[i];
1593                 dst->agg[i] += src->agg[i];
1594         }
1595
1596         if (!dst->kb_base)
1597                 dst->kb_base = src->kb_base;
1598         if (!dst->unit_base)
1599                 dst->unit_base = src->unit_base;
1600         if (!dst->sig_figs)
1601                 dst->sig_figs = src->sig_figs;
1602 }
1603
1604 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1605                       bool first)
1606 {
1607         int l, k;
1608
1609         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1610                 if (!dst->unified_rw_rep) {
1611                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first, false);
1612                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first, false);
1613                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first, false);
1614                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first, true);
1615                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first, true);
1616
1617                         dst->io_bytes[l] += src->io_bytes[l];
1618
1619                         if (dst->runtime[l] < src->runtime[l])
1620                                 dst->runtime[l] = src->runtime[l];
1621                 } else {
1622                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first, false);
1623                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first, false);
1624                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first, false);
1625                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first, true);
1626                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first, true);
1627
1628                         dst->io_bytes[0] += src->io_bytes[l];
1629
1630                         if (dst->runtime[0] < src->runtime[l])
1631                                 dst->runtime[0] = src->runtime[l];
1632
1633                         /*
1634                          * We're summing to the same destination, so override
1635                          * 'first' after the first iteration of the loop
1636                          */
1637                         first = false;
1638                 }
1639         }
1640
1641         sum_stat(&dst->sync_stat, &src->sync_stat, first, false);
1642         dst->usr_time += src->usr_time;
1643         dst->sys_time += src->sys_time;
1644         dst->ctx += src->ctx;
1645         dst->majf += src->majf;
1646         dst->minf += src->minf;
1647
1648         for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
1649                 dst->io_u_map[k] += src->io_u_map[k];
1650                 dst->io_u_submit[k] += src->io_u_submit[k];
1651                 dst->io_u_complete[k] += src->io_u_complete[k];
1652         }
1653         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++) {
1654                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1655                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1656                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1657         }
1658         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1659                 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
1660
1661         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1662                 if (!dst->unified_rw_rep) {
1663                         dst->total_io_u[k] += src->total_io_u[k];
1664                         dst->short_io_u[k] += src->short_io_u[k];
1665                         dst->drop_io_u[k] += src->drop_io_u[k];
1666                 } else {
1667                         dst->total_io_u[0] += src->total_io_u[k];
1668                         dst->short_io_u[0] += src->short_io_u[k];
1669                         dst->drop_io_u[0] += src->drop_io_u[k];
1670                 }
1671         }
1672
1673         dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
1674
1675         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1676                 int m;
1677
1678                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1679                         if (!dst->unified_rw_rep)
1680                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1681                         else
1682                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1683                 }
1684         }
1685
1686         dst->total_run_time += src->total_run_time;
1687         dst->total_submit += src->total_submit;
1688         dst->total_complete += src->total_complete;
1689         dst->nr_zone_resets += src->nr_zone_resets;
1690 }
1691
1692 void init_group_run_stat(struct group_run_stats *gs)
1693 {
1694         int i;
1695         memset(gs, 0, sizeof(*gs));
1696
1697         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1698                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1699 }
1700
1701 void init_thread_stat(struct thread_stat *ts)
1702 {
1703         int j;
1704
1705         memset(ts, 0, sizeof(*ts));
1706
1707         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1708                 ts->lat_stat[j].min_val = -1UL;
1709                 ts->clat_stat[j].min_val = -1UL;
1710                 ts->slat_stat[j].min_val = -1UL;
1711                 ts->bw_stat[j].min_val = -1UL;
1712                 ts->iops_stat[j].min_val = -1UL;
1713         }
1714         ts->sync_stat.min_val = -1UL;
1715         ts->groupid = -1;
1716 }
1717
1718 void __show_run_stats(void)
1719 {
1720         struct group_run_stats *runstats, *rs;
1721         struct thread_data *td;
1722         struct thread_stat *threadstats, *ts;
1723         int i, j, k, nr_ts, last_ts, idx;
1724         bool kb_base_warned = false;
1725         bool unit_base_warned = false;
1726         struct json_object *root = NULL;
1727         struct json_array *array = NULL;
1728         struct buf_output output[FIO_OUTPUT_NR];
1729         struct flist_head **opt_lists;
1730
1731         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1732
1733         for (i = 0; i < groupid + 1; i++)
1734                 init_group_run_stat(&runstats[i]);
1735
1736         /*
1737          * find out how many threads stats we need. if group reporting isn't
1738          * enabled, it's one-per-td.
1739          */
1740         nr_ts = 0;
1741         last_ts = -1;
1742         for_each_td(td, i) {
1743                 if (!td->o.group_reporting) {
1744                         nr_ts++;
1745                         continue;
1746                 }
1747                 if (last_ts == td->groupid)
1748                         continue;
1749                 if (!td->o.stats)
1750                         continue;
1751
1752                 last_ts = td->groupid;
1753                 nr_ts++;
1754         }
1755
1756         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1757         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1758
1759         for (i = 0; i < nr_ts; i++) {
1760                 init_thread_stat(&threadstats[i]);
1761                 opt_lists[i] = NULL;
1762         }
1763
1764         j = 0;
1765         last_ts = -1;
1766         idx = 0;
1767         for_each_td(td, i) {
1768                 if (!td->o.stats)
1769                         continue;
1770                 if (idx && (!td->o.group_reporting ||
1771                     (td->o.group_reporting && last_ts != td->groupid))) {
1772                         idx = 0;
1773                         j++;
1774                 }
1775
1776                 last_ts = td->groupid;
1777
1778                 ts = &threadstats[j];
1779
1780                 ts->clat_percentiles = td->o.clat_percentiles;
1781                 ts->lat_percentiles = td->o.lat_percentiles;
1782                 ts->percentile_precision = td->o.percentile_precision;
1783                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1784                 opt_lists[j] = &td->opt_list;
1785
1786                 idx++;
1787                 ts->members++;
1788
1789                 if (ts->groupid == -1) {
1790                         /*
1791                          * These are per-group shared already
1792                          */
1793                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1794                         if (td->o.description)
1795                                 strncpy(ts->description, td->o.description,
1796                                                 FIO_JOBDESC_SIZE - 1);
1797                         else
1798                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1799
1800                         /*
1801                          * If multiple entries in this group, this is
1802                          * the first member.
1803                          */
1804                         ts->thread_number = td->thread_number;
1805                         ts->groupid = td->groupid;
1806
1807                         /*
1808                          * first pid in group, not very useful...
1809                          */
1810                         ts->pid = td->pid;
1811
1812                         ts->kb_base = td->o.kb_base;
1813                         ts->unit_base = td->o.unit_base;
1814                         ts->sig_figs = td->o.sig_figs;
1815                         ts->unified_rw_rep = td->o.unified_rw_rep;
1816                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1817                         log_info("fio: kb_base differs for jobs in group, using"
1818                                  " %u as the base\n", ts->kb_base);
1819                         kb_base_warned = true;
1820                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1821                         log_info("fio: unit_base differs for jobs in group, using"
1822                                  " %u as the base\n", ts->unit_base);
1823                         unit_base_warned = true;
1824                 }
1825
1826                 ts->continue_on_error = td->o.continue_on_error;
1827                 ts->total_err_count += td->total_err_count;
1828                 ts->first_error = td->first_error;
1829                 if (!ts->error) {
1830                         if (!td->error && td->o.continue_on_error &&
1831                             td->first_error) {
1832                                 ts->error = td->first_error;
1833                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1834                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1835                         } else  if (td->error) {
1836                                 ts->error = td->error;
1837                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1838                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1839                         }
1840                 }
1841
1842                 ts->latency_depth = td->latency_qd;
1843                 ts->latency_target = td->o.latency_target;
1844                 ts->latency_percentile = td->o.latency_percentile;
1845                 ts->latency_window = td->o.latency_window;
1846
1847                 ts->nr_block_infos = td->ts.nr_block_infos;
1848                 for (k = 0; k < ts->nr_block_infos; k++)
1849                         ts->block_infos[k] = td->ts.block_infos[k];
1850
1851                 sum_thread_stats(ts, &td->ts, idx == 1);
1852
1853                 if (td->o.ss_dur) {
1854                         ts->ss_state = td->ss.state;
1855                         ts->ss_dur = td->ss.dur;
1856                         ts->ss_head = td->ss.head;
1857                         ts->ss_bw_data = td->ss.bw_data;
1858                         ts->ss_iops_data = td->ss.iops_data;
1859                         ts->ss_limit.u.f = td->ss.limit;
1860                         ts->ss_slope.u.f = td->ss.slope;
1861                         ts->ss_deviation.u.f = td->ss.deviation;
1862                         ts->ss_criterion.u.f = td->ss.criterion;
1863                 }
1864                 else
1865                         ts->ss_dur = ts->ss_state = 0;
1866         }
1867
1868         for (i = 0; i < nr_ts; i++) {
1869                 unsigned long long bw;
1870
1871                 ts = &threadstats[i];
1872                 if (ts->groupid == -1)
1873                         continue;
1874                 rs = &runstats[ts->groupid];
1875                 rs->kb_base = ts->kb_base;
1876                 rs->unit_base = ts->unit_base;
1877                 rs->sig_figs = ts->sig_figs;
1878                 rs->unified_rw_rep += ts->unified_rw_rep;
1879
1880                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1881                         if (!ts->runtime[j])
1882                                 continue;
1883                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1884                                 rs->min_run[j] = ts->runtime[j];
1885                         if (ts->runtime[j] > rs->max_run[j])
1886                                 rs->max_run[j] = ts->runtime[j];
1887
1888                         bw = 0;
1889                         if (ts->runtime[j])
1890                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1891                         if (bw < rs->min_bw[j])
1892                                 rs->min_bw[j] = bw;
1893                         if (bw > rs->max_bw[j])
1894                                 rs->max_bw[j] = bw;
1895
1896                         rs->iobytes[j] += ts->io_bytes[j];
1897                 }
1898         }
1899
1900         for (i = 0; i < groupid + 1; i++) {
1901                 int ddir;
1902
1903                 rs = &runstats[i];
1904
1905                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1906                         if (rs->max_run[ddir])
1907                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1908                                                 rs->max_run[ddir];
1909                 }
1910         }
1911
1912         for (i = 0; i < FIO_OUTPUT_NR; i++)
1913                 buf_output_init(&output[i]);
1914
1915         /*
1916          * don't overwrite last signal output
1917          */
1918         if (output_format & FIO_OUTPUT_NORMAL)
1919                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1920         if (output_format & FIO_OUTPUT_JSON) {
1921                 struct thread_data *global;
1922                 char time_buf[32];
1923                 struct timeval now;
1924                 unsigned long long ms_since_epoch;
1925                 time_t tv_sec;
1926
1927                 gettimeofday(&now, NULL);
1928                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1929                                  (unsigned long long)(now.tv_usec) / 1000;
1930
1931                 tv_sec = now.tv_sec;
1932                 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
1933                 if (time_buf[strlen(time_buf) - 1] == '\n')
1934                         time_buf[strlen(time_buf) - 1] = '\0';
1935
1936                 root = json_create_object();
1937                 json_object_add_value_string(root, "fio version", fio_version_string);
1938                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1939                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1940                 json_object_add_value_string(root, "time", time_buf);
1941                 global = get_global_options();
1942                 json_add_job_opts(root, "global options", &global->opt_list);
1943                 array = json_create_array();
1944                 json_object_add_value_array(root, "jobs", array);
1945         }
1946
1947         if (is_backend)
1948                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1949
1950         for (i = 0; i < nr_ts; i++) {
1951                 ts = &threadstats[i];
1952                 rs = &runstats[ts->groupid];
1953
1954                 if (is_backend) {
1955                         fio_server_send_job_options(opt_lists[i], i);
1956                         fio_server_send_ts(ts, rs);
1957                 } else {
1958                         if (output_format & FIO_OUTPUT_TERSE)
1959                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1960                         if (output_format & FIO_OUTPUT_JSON) {
1961                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1962                                 json_array_add_value_object(array, tmp);
1963                         }
1964                         if (output_format & FIO_OUTPUT_NORMAL)
1965                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1966                 }
1967         }
1968         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1969                 /* disk util stats, if any */
1970                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1971
1972                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1973
1974                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1975                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1976                 json_free_object(root);
1977         }
1978
1979         for (i = 0; i < groupid + 1; i++) {
1980                 rs = &runstats[i];
1981
1982                 rs->groupid = i;
1983                 if (is_backend)
1984                         fio_server_send_gs(rs);
1985                 else if (output_format & FIO_OUTPUT_NORMAL)
1986                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1987         }
1988
1989         if (is_backend)
1990                 fio_server_send_du();
1991         else if (output_format & FIO_OUTPUT_NORMAL) {
1992                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1993                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1994         }
1995
1996         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1997                 struct buf_output *out = &output[i];
1998
1999                 log_info_buf(out->buf, out->buflen);
2000                 buf_output_free(out);
2001         }
2002
2003         fio_idle_prof_cleanup();
2004
2005         log_info_flush();
2006         free(runstats);
2007         free(threadstats);
2008         free(opt_lists);
2009 }
2010
2011 void __show_running_run_stats(void)
2012 {
2013         struct thread_data *td;
2014         unsigned long long *rt;
2015         struct timespec ts;
2016         int i;
2017
2018         fio_sem_down(stat_sem);
2019
2020         rt = malloc(thread_number * sizeof(unsigned long long));
2021         fio_gettime(&ts, NULL);
2022
2023         for_each_td(td, i) {
2024                 td->update_rusage = 1;
2025                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
2026                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
2027                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2028                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2029
2030                 rt[i] = mtime_since(&td->start, &ts);
2031                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2032                         td->ts.runtime[DDIR_READ] += rt[i];
2033                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2034                         td->ts.runtime[DDIR_WRITE] += rt[i];
2035                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2036                         td->ts.runtime[DDIR_TRIM] += rt[i];
2037         }
2038
2039         for_each_td(td, i) {
2040                 if (td->runstate >= TD_EXITED)
2041                         continue;
2042                 if (td->rusage_sem) {
2043                         td->update_rusage = 1;
2044                         fio_sem_down(td->rusage_sem);
2045                 }
2046                 td->update_rusage = 0;
2047         }
2048
2049         __show_run_stats();
2050
2051         for_each_td(td, i) {
2052                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2053                         td->ts.runtime[DDIR_READ] -= rt[i];
2054                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2055                         td->ts.runtime[DDIR_WRITE] -= rt[i];
2056                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2057                         td->ts.runtime[DDIR_TRIM] -= rt[i];
2058         }
2059
2060         free(rt);
2061         fio_sem_up(stat_sem);
2062 }
2063
2064 static bool status_interval_init;
2065 static struct timespec status_time;
2066 static bool status_file_disabled;
2067
2068 #define FIO_STATUS_FILE         "fio-dump-status"
2069
2070 static int check_status_file(void)
2071 {
2072         struct stat sb;
2073         const char *temp_dir;
2074         char fio_status_file_path[PATH_MAX];
2075
2076         if (status_file_disabled)
2077                 return 0;
2078
2079         temp_dir = getenv("TMPDIR");
2080         if (temp_dir == NULL) {
2081                 temp_dir = getenv("TEMP");
2082                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2083                         temp_dir = NULL;
2084         }
2085         if (temp_dir == NULL)
2086                 temp_dir = "/tmp";
2087
2088         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2089
2090         if (stat(fio_status_file_path, &sb))
2091                 return 0;
2092
2093         if (unlink(fio_status_file_path) < 0) {
2094                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2095                                                         strerror(errno));
2096                 log_err("fio: disabling status file updates\n");
2097                 status_file_disabled = true;
2098         }
2099
2100         return 1;
2101 }
2102
2103 void check_for_running_stats(void)
2104 {
2105         if (status_interval) {
2106                 if (!status_interval_init) {
2107                         fio_gettime(&status_time, NULL);
2108                         status_interval_init = true;
2109                 } else if (mtime_since_now(&status_time) >= status_interval) {
2110                         show_running_run_stats();
2111                         fio_gettime(&status_time, NULL);
2112                         return;
2113                 }
2114         }
2115         if (check_status_file()) {
2116                 show_running_run_stats();
2117                 return;
2118         }
2119 }
2120
2121 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2122 {
2123         double val = data;
2124         double delta;
2125
2126         if (data > is->max_val)
2127                 is->max_val = data;
2128         if (data < is->min_val)
2129                 is->min_val = data;
2130
2131         delta = val - is->mean.u.f;
2132         if (delta) {
2133                 is->mean.u.f += delta / (is->samples + 1.0);
2134                 is->S.u.f += delta * (val - is->mean.u.f);
2135         }
2136
2137         is->samples++;
2138 }
2139
2140 /*
2141  * Return a struct io_logs, which is added to the tail of the log
2142  * list for 'iolog'.
2143  */
2144 static struct io_logs *get_new_log(struct io_log *iolog)
2145 {
2146         size_t new_size, new_samples;
2147         struct io_logs *cur_log;
2148
2149         /*
2150          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2151          * forever
2152          */
2153         if (!iolog->cur_log_max)
2154                 new_samples = DEF_LOG_ENTRIES;
2155         else {
2156                 new_samples = iolog->cur_log_max * 2;
2157                 if (new_samples > MAX_LOG_ENTRIES)
2158                         new_samples = MAX_LOG_ENTRIES;
2159         }
2160
2161         new_size = new_samples * log_entry_sz(iolog);
2162
2163         cur_log = smalloc(sizeof(*cur_log));
2164         if (cur_log) {
2165                 INIT_FLIST_HEAD(&cur_log->list);
2166                 cur_log->log = malloc(new_size);
2167                 if (cur_log->log) {
2168                         cur_log->nr_samples = 0;
2169                         cur_log->max_samples = new_samples;
2170                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2171                         iolog->cur_log_max = new_samples;
2172                         return cur_log;
2173                 }
2174                 sfree(cur_log);
2175         }
2176
2177         return NULL;
2178 }
2179
2180 /*
2181  * Add and return a new log chunk, or return current log if big enough
2182  */
2183 static struct io_logs *regrow_log(struct io_log *iolog)
2184 {
2185         struct io_logs *cur_log;
2186         int i;
2187
2188         if (!iolog || iolog->disabled)
2189                 goto disable;
2190
2191         cur_log = iolog_cur_log(iolog);
2192         if (!cur_log) {
2193                 cur_log = get_new_log(iolog);
2194                 if (!cur_log)
2195                         return NULL;
2196         }
2197
2198         if (cur_log->nr_samples < cur_log->max_samples)
2199                 return cur_log;
2200
2201         /*
2202          * No room for a new sample. If we're compressing on the fly, flush
2203          * out the current chunk
2204          */
2205         if (iolog->log_gz) {
2206                 if (iolog_cur_flush(iolog, cur_log)) {
2207                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2208                         return NULL;
2209                 }
2210         }
2211
2212         /*
2213          * Get a new log array, and add to our list
2214          */
2215         cur_log = get_new_log(iolog);
2216         if (!cur_log) {
2217                 log_err("fio: failed extending iolog! Will stop logging.\n");
2218                 return NULL;
2219         }
2220
2221         if (!iolog->pending || !iolog->pending->nr_samples)
2222                 return cur_log;
2223
2224         /*
2225          * Flush pending items to new log
2226          */
2227         for (i = 0; i < iolog->pending->nr_samples; i++) {
2228                 struct io_sample *src, *dst;
2229
2230                 src = get_sample(iolog, iolog->pending, i);
2231                 dst = get_sample(iolog, cur_log, i);
2232                 memcpy(dst, src, log_entry_sz(iolog));
2233         }
2234         cur_log->nr_samples = iolog->pending->nr_samples;
2235
2236         iolog->pending->nr_samples = 0;
2237         return cur_log;
2238 disable:
2239         if (iolog)
2240                 iolog->disabled = true;
2241         return NULL;
2242 }
2243
2244 void regrow_logs(struct thread_data *td)
2245 {
2246         regrow_log(td->slat_log);
2247         regrow_log(td->clat_log);
2248         regrow_log(td->clat_hist_log);
2249         regrow_log(td->lat_log);
2250         regrow_log(td->bw_log);
2251         regrow_log(td->iops_log);
2252         td->flags &= ~TD_F_REGROW_LOGS;
2253 }
2254
2255 static struct io_logs *get_cur_log(struct io_log *iolog)
2256 {
2257         struct io_logs *cur_log;
2258
2259         cur_log = iolog_cur_log(iolog);
2260         if (!cur_log) {
2261                 cur_log = get_new_log(iolog);
2262                 if (!cur_log)
2263                         return NULL;
2264         }
2265
2266         if (cur_log->nr_samples < cur_log->max_samples)
2267                 return cur_log;
2268
2269         /*
2270          * Out of space. If we're in IO offload mode, or we're not doing
2271          * per unit logging (hence logging happens outside of the IO thread
2272          * as well), add a new log chunk inline. If we're doing inline
2273          * submissions, flag 'td' as needing a log regrow and we'll take
2274          * care of it on the submission side.
2275          */
2276         if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2277             !per_unit_log(iolog))
2278                 return regrow_log(iolog);
2279
2280         if (iolog->td)
2281                 iolog->td->flags |= TD_F_REGROW_LOGS;
2282         if (iolog->pending)
2283                 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2284         return iolog->pending;
2285 }
2286
2287 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2288                              enum fio_ddir ddir, unsigned long long bs,
2289                              unsigned long t, uint64_t offset)
2290 {
2291         struct io_logs *cur_log;
2292
2293         if (iolog->disabled)
2294                 return;
2295         if (flist_empty(&iolog->io_logs))
2296                 iolog->avg_last[ddir] = t;
2297
2298         cur_log = get_cur_log(iolog);
2299         if (cur_log) {
2300                 struct io_sample *s;
2301
2302                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2303
2304                 s->data = data;
2305                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2306                 io_sample_set_ddir(iolog, s, ddir);
2307                 s->bs = bs;
2308
2309                 if (iolog->log_offset) {
2310                         struct io_sample_offset *so = (void *) s;
2311
2312                         so->offset = offset;
2313                 }
2314
2315                 cur_log->nr_samples++;
2316                 return;
2317         }
2318
2319         iolog->disabled = true;
2320 }
2321
2322 static inline void reset_io_stat(struct io_stat *ios)
2323 {
2324         ios->max_val = ios->min_val = ios->samples = 0;
2325         ios->mean.u.f = ios->S.u.f = 0;
2326 }
2327
2328 void reset_io_stats(struct thread_data *td)
2329 {
2330         struct thread_stat *ts = &td->ts;
2331         int i, j;
2332
2333         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2334                 reset_io_stat(&ts->clat_stat[i]);
2335                 reset_io_stat(&ts->slat_stat[i]);
2336                 reset_io_stat(&ts->lat_stat[i]);
2337                 reset_io_stat(&ts->bw_stat[i]);
2338                 reset_io_stat(&ts->iops_stat[i]);
2339
2340                 ts->io_bytes[i] = 0;
2341                 ts->runtime[i] = 0;
2342                 ts->total_io_u[i] = 0;
2343                 ts->short_io_u[i] = 0;
2344                 ts->drop_io_u[i] = 0;
2345
2346                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2347                         ts->io_u_plat[i][j] = 0;
2348                         if (!i)
2349                                 ts->io_u_sync_plat[j] = 0;
2350                 }
2351         }
2352
2353         ts->total_io_u[DDIR_SYNC] = 0;
2354
2355         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2356                 ts->io_u_map[i] = 0;
2357                 ts->io_u_submit[i] = 0;
2358                 ts->io_u_complete[i] = 0;
2359         }
2360
2361         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2362                 ts->io_u_lat_n[i] = 0;
2363         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2364                 ts->io_u_lat_u[i] = 0;
2365         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2366                 ts->io_u_lat_m[i] = 0;
2367
2368         ts->total_submit = 0;
2369         ts->total_complete = 0;
2370         ts->nr_zone_resets = 0;
2371 }
2372
2373 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2374                               unsigned long elapsed, bool log_max)
2375 {
2376         /*
2377          * Note an entry in the log. Use the mean from the logged samples,
2378          * making sure to properly round up. Only write a log entry if we
2379          * had actual samples done.
2380          */
2381         if (iolog->avg_window[ddir].samples) {
2382                 union io_sample_data data;
2383
2384                 if (log_max)
2385                         data.val = iolog->avg_window[ddir].max_val;
2386                 else
2387                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2388
2389                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2390         }
2391
2392         reset_io_stat(&iolog->avg_window[ddir]);
2393 }
2394
2395 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2396                              bool log_max)
2397 {
2398         int ddir;
2399
2400         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2401                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2402 }
2403
2404 static unsigned long add_log_sample(struct thread_data *td,
2405                                     struct io_log *iolog,
2406                                     union io_sample_data data,
2407                                     enum fio_ddir ddir, unsigned long long bs,
2408                                     uint64_t offset)
2409 {
2410         unsigned long elapsed, this_window;
2411
2412         if (!ddir_rw(ddir))
2413                 return 0;
2414
2415         elapsed = mtime_since_now(&td->epoch);
2416
2417         /*
2418          * If no time averaging, just add the log sample.
2419          */
2420         if (!iolog->avg_msec) {
2421                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2422                 return 0;
2423         }
2424
2425         /*
2426          * Add the sample. If the time period has passed, then
2427          * add that entry to the log and clear.
2428          */
2429         add_stat_sample(&iolog->avg_window[ddir], data.val);
2430
2431         /*
2432          * If period hasn't passed, adding the above sample is all we
2433          * need to do.
2434          */
2435         this_window = elapsed - iolog->avg_last[ddir];
2436         if (elapsed < iolog->avg_last[ddir])
2437                 return iolog->avg_last[ddir] - elapsed;
2438         else if (this_window < iolog->avg_msec) {
2439                 unsigned long diff = iolog->avg_msec - this_window;
2440
2441                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2442                         return diff;
2443         }
2444
2445         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2446
2447         iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2448         return iolog->avg_msec;
2449 }
2450
2451 void finalize_logs(struct thread_data *td, bool unit_logs)
2452 {
2453         unsigned long elapsed;
2454
2455         elapsed = mtime_since_now(&td->epoch);
2456
2457         if (td->clat_log && unit_logs)
2458                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2459         if (td->slat_log && unit_logs)
2460                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2461         if (td->lat_log && unit_logs)
2462                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2463         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2464                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2465         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2466                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2467 }
2468
2469 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned long long bs)
2470 {
2471         struct io_log *iolog;
2472
2473         if (!ddir_rw(ddir))
2474                 return;
2475
2476         iolog = agg_io_log[ddir];
2477         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2478 }
2479
2480 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2481 {
2482         unsigned int idx = plat_val_to_idx(nsec);
2483         assert(idx < FIO_IO_U_PLAT_NR);
2484
2485         ts->io_u_sync_plat[idx]++;
2486         add_stat_sample(&ts->sync_stat, nsec);
2487 }
2488
2489 static void add_clat_percentile_sample(struct thread_stat *ts,
2490                                 unsigned long long nsec, enum fio_ddir ddir)
2491 {
2492         unsigned int idx = plat_val_to_idx(nsec);
2493         assert(idx < FIO_IO_U_PLAT_NR);
2494
2495         ts->io_u_plat[ddir][idx]++;
2496 }
2497
2498 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2499                      unsigned long long nsec, unsigned long long bs,
2500                      uint64_t offset)
2501 {
2502         const bool needs_lock = td_async_processing(td);
2503         unsigned long elapsed, this_window;
2504         struct thread_stat *ts = &td->ts;
2505         struct io_log *iolog = td->clat_hist_log;
2506
2507         if (needs_lock)
2508                 __td_io_u_lock(td);
2509
2510         add_stat_sample(&ts->clat_stat[ddir], nsec);
2511
2512         if (td->clat_log)
2513                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2514                                offset);
2515
2516         if (ts->clat_percentiles)
2517                 add_clat_percentile_sample(ts, nsec, ddir);
2518
2519         if (iolog && iolog->hist_msec) {
2520                 struct io_hist *hw = &iolog->hist_window[ddir];
2521
2522                 hw->samples++;
2523                 elapsed = mtime_since_now(&td->epoch);
2524                 if (!hw->hist_last)
2525                         hw->hist_last = elapsed;
2526                 this_window = elapsed - hw->hist_last;
2527                 
2528                 if (this_window >= iolog->hist_msec) {
2529                         uint64_t *io_u_plat;
2530                         struct io_u_plat_entry *dst;
2531
2532                         /*
2533                          * Make a byte-for-byte copy of the latency histogram
2534                          * stored in td->ts.io_u_plat[ddir], recording it in a
2535                          * log sample. Note that the matching call to free() is
2536                          * located in iolog.c after printing this sample to the
2537                          * log file.
2538                          */
2539                         io_u_plat = (uint64_t *) td->ts.io_u_plat[ddir];
2540                         dst = malloc(sizeof(struct io_u_plat_entry));
2541                         memcpy(&(dst->io_u_plat), io_u_plat,
2542                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2543                         flist_add(&dst->list, &hw->list);
2544                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2545                                                 elapsed, offset);
2546
2547                         /*
2548                          * Update the last time we recorded as being now, minus
2549                          * any drift in time we encountered before actually
2550                          * making the record.
2551                          */
2552                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2553                         hw->samples = 0;
2554                 }
2555         }
2556
2557         if (needs_lock)
2558                 __td_io_u_unlock(td);
2559 }
2560
2561 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2562                      unsigned long usec, unsigned long long bs, uint64_t offset)
2563 {
2564         const bool needs_lock = td_async_processing(td);
2565         struct thread_stat *ts = &td->ts;
2566
2567         if (!ddir_rw(ddir))
2568                 return;
2569
2570         if (needs_lock)
2571                 __td_io_u_lock(td);
2572
2573         add_stat_sample(&ts->slat_stat[ddir], usec);
2574
2575         if (td->slat_log)
2576                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2577
2578         if (needs_lock)
2579                 __td_io_u_unlock(td);
2580 }
2581
2582 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2583                     unsigned long long nsec, unsigned long long bs,
2584                     uint64_t offset)
2585 {
2586         const bool needs_lock = td_async_processing(td);
2587         struct thread_stat *ts = &td->ts;
2588
2589         if (!ddir_rw(ddir))
2590                 return;
2591
2592         if (needs_lock)
2593                 __td_io_u_lock(td);
2594
2595         add_stat_sample(&ts->lat_stat[ddir], nsec);
2596
2597         if (td->lat_log)
2598                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2599                                offset);
2600
2601         if (ts->lat_percentiles)
2602                 add_clat_percentile_sample(ts, nsec, ddir);
2603
2604         if (needs_lock)
2605                 __td_io_u_unlock(td);
2606 }
2607
2608 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2609                    unsigned int bytes, unsigned long long spent)
2610 {
2611         const bool needs_lock = td_async_processing(td);
2612         struct thread_stat *ts = &td->ts;
2613         unsigned long rate;
2614
2615         if (spent)
2616                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2617         else
2618                 rate = 0;
2619
2620         if (needs_lock)
2621                 __td_io_u_lock(td);
2622
2623         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2624
2625         if (td->bw_log)
2626                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2627                                bytes, io_u->offset);
2628
2629         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2630
2631         if (needs_lock)
2632                 __td_io_u_unlock(td);
2633 }
2634
2635 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2636                          struct timespec *t, unsigned int avg_time,
2637                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2638                          struct io_stat *stat, struct io_log *log,
2639                          bool is_kb)
2640 {
2641         const bool needs_lock = td_async_processing(td);
2642         unsigned long spent, rate;
2643         enum fio_ddir ddir;
2644         unsigned long next, next_log;
2645
2646         next_log = avg_time;
2647
2648         spent = mtime_since(parent_tv, t);
2649         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2650                 return avg_time - spent;
2651
2652         if (needs_lock)
2653                 __td_io_u_lock(td);
2654
2655         /*
2656          * Compute both read and write rates for the interval.
2657          */
2658         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2659                 uint64_t delta;
2660
2661                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2662                 if (!delta)
2663                         continue; /* No entries for interval */
2664
2665                 if (spent) {
2666                         if (is_kb)
2667                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2668                         else
2669                                 rate = (delta * 1000) / spent;
2670                 } else
2671                         rate = 0;
2672
2673                 add_stat_sample(&stat[ddir], rate);
2674
2675                 if (log) {
2676                         unsigned long long bs = 0;
2677
2678                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2679                                 bs = td->o.min_bs[ddir];
2680
2681                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2682                         next_log = min(next_log, next);
2683                 }
2684
2685                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2686         }
2687
2688         timespec_add_msec(parent_tv, avg_time);
2689
2690         if (needs_lock)
2691                 __td_io_u_unlock(td);
2692
2693         if (spent <= avg_time)
2694                 next = avg_time;
2695         else
2696                 next = avg_time - (1 + spent - avg_time);
2697
2698         return min(next, next_log);
2699 }
2700
2701 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2702 {
2703         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2704                                 td->this_io_bytes, td->stat_io_bytes,
2705                                 td->ts.bw_stat, td->bw_log, true);
2706 }
2707
2708 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2709                      unsigned int bytes)
2710 {
2711         const bool needs_lock = td_async_processing(td);
2712         struct thread_stat *ts = &td->ts;
2713
2714         if (needs_lock)
2715                 __td_io_u_lock(td);
2716
2717         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2718
2719         if (td->iops_log)
2720                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2721                                bytes, io_u->offset);
2722
2723         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2724
2725         if (needs_lock)
2726                 __td_io_u_unlock(td);
2727 }
2728
2729 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2730 {
2731         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2732                                 td->this_io_blocks, td->stat_io_blocks,
2733                                 td->ts.iops_stat, td->iops_log, false);
2734 }
2735
2736 /*
2737  * Returns msecs to next event
2738  */
2739 int calc_log_samples(void)
2740 {
2741         struct thread_data *td;
2742         unsigned int next = ~0U, tmp;
2743         struct timespec now;
2744         int i;
2745
2746         fio_gettime(&now, NULL);
2747
2748         for_each_td(td, i) {
2749                 if (!td->o.stats)
2750                         continue;
2751                 if (in_ramp_time(td) ||
2752                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2753                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2754                         continue;
2755                 }
2756                 if (!td->bw_log ||
2757                         (td->bw_log && !per_unit_log(td->bw_log))) {
2758                         tmp = add_bw_samples(td, &now);
2759                         if (tmp < next)
2760                                 next = tmp;
2761                 }
2762                 if (!td->iops_log ||
2763                         (td->iops_log && !per_unit_log(td->iops_log))) {
2764                         tmp = add_iops_samples(td, &now);
2765                         if (tmp < next)
2766                                 next = tmp;
2767                 }
2768         }
2769
2770         return next == ~0U ? 0 : next;
2771 }
2772
2773 void stat_init(void)
2774 {
2775         stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
2776 }
2777
2778 void stat_exit(void)
2779 {
2780         /*
2781          * When we have the mutex, we know out-of-band access to it
2782          * have ended.
2783          */
2784         fio_sem_down(stat_sem);
2785         fio_sem_remove(stat_sem);
2786 }
2787
2788 /*
2789  * Called from signal handler. Wake up status thread.
2790  */
2791 void show_running_run_stats(void)
2792 {
2793         helper_do_stat();
2794 }
2795
2796 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2797 {
2798         /* Ignore io_u's which span multiple blocks--they will just get
2799          * inaccurate counts. */
2800         int idx = (io_u->offset - io_u->file->file_offset)
2801                         / td->o.bs[DDIR_TRIM];
2802         uint32_t *info = &td->ts.block_infos[idx];
2803         assert(idx < td->ts.nr_block_infos);
2804         return info;
2805 }