[PATCH] Add bw/lat log as job options
[fio.git] / stat.c
1 #include <stdio.h>
2 #include <string.h>
3 #include <sys/time.h>
4 #include <sys/types.h>
5 #include <sys/stat.h>
6 #include <dirent.h>
7 #include <libgen.h>
8 #include <math.h>
9
10 #include "fio.h"
11
12 static struct itimerval itimer;
13 static struct list_head disk_list = LIST_HEAD_INIT(disk_list);
14
15 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
16 {
17         unsigned in_flight;
18         char line[256];
19         FILE *f;
20         char *p;
21
22         f = fopen(du->path, "r");
23         if (!f)
24                 return 1;
25
26         p = fgets(line, sizeof(line), f);
27         if (!p) {
28                 fclose(f);
29                 return 1;
30         }
31
32         if (sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0], &dus->merges[0], &dus->sectors[0], &dus->ticks[0], &dus->ios[1], &dus->merges[1], &dus->sectors[1], &dus->ticks[1], &in_flight, &dus->io_ticks, &dus->time_in_queue) != 11) {
33                 fclose(f);
34                 return 1;
35         }
36
37         fclose(f);
38         return 0;
39 }
40
41 static void update_io_tick_disk(struct disk_util *du)
42 {
43         struct disk_util_stat __dus, *dus, *ldus;
44         struct timeval t;
45
46         if (get_io_ticks(du, &__dus))
47                 return;
48
49         dus = &du->dus;
50         ldus = &du->last_dus;
51
52         dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
53         dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
54         dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
55         dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
56         dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
57         dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
58         dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
59         dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
60         dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
61         dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
62
63         gettimeofday(&t, NULL);
64         du->msec += mtime_since(&du->time, &t);
65         memcpy(&du->time, &t, sizeof(t));
66         memcpy(ldus, &__dus, sizeof(__dus));
67 }
68
69 void update_io_ticks(void)
70 {
71         struct list_head *entry;
72         struct disk_util *du;
73
74         list_for_each(entry, &disk_list) {
75                 du = list_entry(entry, struct disk_util, list);
76                 update_io_tick_disk(du);
77         }
78 }
79
80 static int disk_util_exists(dev_t dev)
81 {
82         struct list_head *entry;
83         struct disk_util *du;
84
85         list_for_each(entry, &disk_list) {
86                 du = list_entry(entry, struct disk_util, list);
87
88                 if (du->dev == dev)
89                         return 1;
90         }
91
92         return 0;
93 }
94
95 static void disk_util_add(dev_t dev, char *path)
96 {
97         struct disk_util *du = malloc(sizeof(*du));
98
99         memset(du, 0, sizeof(*du));
100         INIT_LIST_HEAD(&du->list);
101         sprintf(du->path, "%s/stat", path);
102         du->name = strdup(basename(path));
103         du->dev = dev;
104
105         gettimeofday(&du->time, NULL);
106         get_io_ticks(du, &du->last_dus);
107
108         list_add_tail(&du->list, &disk_list);
109 }
110
111 static int check_dev_match(dev_t dev, char *path)
112 {
113         unsigned int major, minor;
114         char line[256], *p;
115         FILE *f;
116
117         f = fopen(path, "r");
118         if (!f) {
119                 perror("open path");
120                 return 1;
121         }
122
123         p = fgets(line, sizeof(line), f);
124         if (!p) {
125                 fclose(f);
126                 return 1;
127         }
128
129         if (sscanf(p, "%u:%u", &major, &minor) != 2) {
130                 fclose(f);
131                 return 1;
132         }
133
134         if (((major << 8) | minor) == dev) {
135                 fclose(f);
136                 return 0;
137         }
138
139         fclose(f);
140         return 1;
141 }
142
143 static int find_block_dir(dev_t dev, char *path)
144 {
145         struct dirent *dir;
146         struct stat st;
147         int found = 0;
148         DIR *D;
149
150         D = opendir(path);
151         if (!D)
152                 return 0;
153
154         while ((dir = readdir(D)) != NULL) {
155                 char full_path[256];
156
157                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
158                         continue;
159
160                 sprintf(full_path, "%s/%s", path, dir->d_name);
161
162                 if (!strcmp(dir->d_name, "dev")) {
163                         if (!check_dev_match(dev, full_path)) {
164                                 found = 1;
165                                 break;
166                         }
167                 }
168
169                 if (lstat(full_path, &st) == -1) {
170                         perror("stat");
171                         break;
172                 }
173
174                 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
175                         continue;
176
177                 found = find_block_dir(dev, full_path);
178                 if (found) {
179                         strcpy(path, full_path);
180                         break;
181                 }
182         }
183
184         closedir(D);
185         return found;
186 }
187
188 void init_disk_util(struct thread_data *td)
189 {
190         struct fio_file *f;
191         struct stat st;
192         char foo[256], tmp[256];
193         dev_t dev;
194         char *p;
195
196         if (!td->do_disk_util)
197                 return;
198
199         /*
200          * Just use the same file, they are on the same device.
201          */
202         f = &td->files[0];
203         if (!stat(f->file_name, &st)) {
204                 if (S_ISBLK(st.st_mode))
205                         dev = st.st_rdev;
206                 else
207                         dev = st.st_dev;
208         } else {
209                 /*
210                  * must be a file, open "." in that path
211                  */
212                 strcpy(foo, f->file_name);
213                 p = dirname(foo);
214                 if (stat(p, &st)) {
215                         perror("disk util stat");
216                         return;
217                 }
218
219                 dev = st.st_dev;
220         }
221
222         if (disk_util_exists(dev))
223                 return;
224                 
225         sprintf(foo, "/sys/block");
226         if (!find_block_dir(dev, foo))
227                 return;
228
229         /*
230          * If there's a ../queue/ directory there, we are inside a partition.
231          * Check if that is the case and jump back. For loop/md/dm etc we
232          * are already in the right spot.
233          */
234         sprintf(tmp, "%s/../queue", foo);
235         if (!stat(tmp, &st)) {
236                 p = dirname(foo);
237                 sprintf(tmp, "%s/queue", p);
238                 if (stat(tmp, &st)) {
239                         log_err("unknown sysfs layout\n");
240                         return;
241                 }
242                 strcpy(tmp, p);
243                 sprintf(foo, "%s", tmp);
244         }
245
246         td->sysfs_root = strdup(foo);
247         disk_util_add(dev, foo);
248 }
249
250 void disk_util_timer_arm(void)
251 {
252         itimer.it_value.tv_sec = 0;
253         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
254         setitimer(ITIMER_REAL, &itimer, NULL);
255 }
256
257 void update_rusage_stat(struct thread_data *td)
258 {
259         if (!(td->runtime[0] + td->runtime[1]))
260                 return;
261
262         getrusage(RUSAGE_SELF, &td->ru_end);
263
264         td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
265         td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
266         td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
267
268         
269         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
270 }
271
272 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
273                     double *mean, double *dev)
274 {
275         double n;
276
277         if (is->samples == 0)
278                 return 0;
279
280         *min = is->min_val;
281         *max = is->max_val;
282
283         n = (double) is->samples;
284         *mean = (double) is->val / n;
285         *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
286
287         return 1;
288 }
289
290 static void show_group_stats(struct group_run_stats *rs, int id)
291 {
292         fprintf(f_out, "\nRun status group %d (all jobs):\n", id);
293
294         if (rs->max_run[DDIR_READ])
295                 fprintf(f_out, "   READ: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_kb[0] >> 10, rs->agg[0], rs->min_bw[0], rs->max_bw[0], rs->min_run[0], rs->max_run[0]);
296         if (rs->max_run[DDIR_WRITE])
297                 fprintf(f_out, "  WRITE: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_kb[1] >> 10, rs->agg[1], rs->min_bw[1], rs->max_bw[1], rs->min_run[1], rs->max_run[1]);
298 }
299
300 static void show_disk_util(void)
301 {
302         struct disk_util_stat *dus;
303         struct list_head *entry;
304         struct disk_util *du;
305         double util;
306
307         fprintf(f_out, "\nDisk stats (read/write):\n");
308
309         list_for_each(entry, &disk_list) {
310                 du = list_entry(entry, struct disk_util, list);
311                 dus = &du->dus;
312
313                 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
314                 if (util > 100.0)
315                         util = 100.0;
316
317                 fprintf(f_out, "  %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, in_queue=%u, util=%3.2f%%\n", du->name, dus->ios[0], dus->ios[1], dus->merges[0], dus->merges[1], dus->ticks[0], dus->ticks[1], dus->time_in_queue, util);
318         }
319 }
320
321 static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
322                              int ddir)
323 {
324         char *ddir_str[] = { "read ", "write" };
325         unsigned long min, max;
326         unsigned long long bw;
327         double mean, dev;
328
329         if (!td->runtime[ddir])
330                 return;
331
332         bw = td->io_bytes[ddir] / td->runtime[ddir];
333         fprintf(f_out, "  %s: io=%6lluMiB, bw=%6lluKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
334
335         if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
336                 fprintf(f_out, "    slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
337
338         if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
339                 fprintf(f_out, "    clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
340
341         if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
342                 double p_of_agg;
343
344                 p_of_agg = mean * 100 / (double) rs->agg[ddir];
345                 fprintf(f_out, "    bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, dev=%5.02f\n", min, max, p_of_agg, mean, dev);
346         }
347 }
348
349 static void show_thread_status(struct thread_data *td,
350                                struct group_run_stats *rs)
351 {
352         double usr_cpu, sys_cpu;
353
354         if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
355                 return;
356
357         fprintf(f_out, "%s: (groupid=%d): err=%2d:\n",td->name, td->groupid, td->error);
358
359         show_ddir_status(td, rs, td->ddir);
360         if (td->io_bytes[td->ddir ^ 1])
361                 show_ddir_status(td, rs, td->ddir ^ 1);
362
363         if (td->runtime[0] + td->runtime[1]) {
364                 double runt = td->runtime[0] + td->runtime[1];
365
366                 usr_cpu = (double) td->usr_time * 100 / runt;
367                 sys_cpu = (double) td->sys_time * 100 / runt;
368         } else {
369                 usr_cpu = 0;
370                 sys_cpu = 0;
371         }
372
373         fprintf(f_out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
374 }
375
376 static void show_ddir_status_terse(struct thread_data *td,
377                                    struct group_run_stats *rs, int ddir)
378 {
379         unsigned long min, max;
380         unsigned long long bw;
381         double mean, dev;
382
383         bw = 0;
384         if (td->runtime[ddir])
385                 bw = td->io_bytes[ddir] / td->runtime[ddir];
386
387         fprintf(f_out, ",%llu,%llu,%lu", td->io_bytes[ddir] >> 10, bw, td->runtime[ddir]);
388
389         if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
390                 fprintf(f_out, ",%lu,%lu,%f,%f", min, max, mean, dev);
391         else
392                 fprintf(f_out, ",%lu,%lu,%f,%f", 0UL, 0UL, 0.0, 0.0);
393
394         if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
395                 fprintf(f_out, ",%lu,%lu,%f,%f", min, max, mean, dev);
396         else
397                 fprintf(f_out, ",%lu,%lu,%f,%f", 0UL, 0UL, 0.0, 0.0);
398
399         if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
400                 double p_of_agg;
401
402                 p_of_agg = mean * 100 / (double) rs->agg[ddir];
403                 fprintf(f_out, ",%lu,%lu,%f%%,%f,%f", min, max, p_of_agg, mean, dev);
404         } else
405                 fprintf(f_out, ",%lu,%lu,%f%%,%f,%f", 0UL, 0UL, 0.0, 0.0, 0.0);
406                 
407 }
408
409
410 static void show_thread_status_terse(struct thread_data *td,
411                                      struct group_run_stats *rs)
412 {
413         double usr_cpu, sys_cpu;
414
415         fprintf(f_out, "%s,%d,%d",td->name, td->groupid, td->error);
416
417         show_ddir_status_terse(td, rs, 0);
418         show_ddir_status_terse(td, rs, 1);
419
420         if (td->runtime[0] + td->runtime[1]) {
421                 double runt = td->runtime[0] + td->runtime[1];
422
423                 usr_cpu = (double) td->usr_time * 100 / runt;
424                 sys_cpu = (double) td->sys_time * 100 / runt;
425         } else {
426                 usr_cpu = 0;
427                 sys_cpu = 0;
428         }
429
430         fprintf(f_out, ",%f%%,%f%%,%lu\n", usr_cpu, sys_cpu, td->ctx);
431 }
432
433 void show_run_stats(void)
434 {
435         struct group_run_stats *runstats, *rs;
436         struct thread_data *td;
437         int i;
438
439         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
440
441         for (i = 0; i < groupid + 1; i++) {
442                 rs = &runstats[i];
443
444                 memset(rs, 0, sizeof(*rs));
445                 rs->min_bw[0] = rs->min_run[0] = ~0UL;
446                 rs->min_bw[1] = rs->min_run[1] = ~0UL;
447         }
448
449         for_each_td(td, i) {
450                 unsigned long long rbw, wbw;
451
452                 if (td->error) {
453                         fprintf(f_out, "%s: %s\n", td->name, td->verror);
454                         continue;
455                 }
456
457                 rs = &runstats[td->groupid];
458
459                 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
460                         rs->min_run[0] = td->runtime[0];
461                 if (td->runtime[0] > rs->max_run[0])
462                         rs->max_run[0] = td->runtime[0];
463                 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
464                         rs->min_run[1] = td->runtime[1];
465                 if (td->runtime[1] > rs->max_run[1])
466                         rs->max_run[1] = td->runtime[1];
467
468                 rbw = wbw = 0;
469                 if (td->runtime[0])
470                         rbw = td->io_bytes[0] / (unsigned long long) td->runtime[0];
471                 if (td->runtime[1])
472                         wbw = td->io_bytes[1] / (unsigned long long) td->runtime[1];
473
474                 if (rbw < rs->min_bw[0])
475                         rs->min_bw[0] = rbw;
476                 if (wbw < rs->min_bw[1])
477                         rs->min_bw[1] = wbw;
478                 if (rbw > rs->max_bw[0])
479                         rs->max_bw[0] = rbw;
480                 if (wbw > rs->max_bw[1])
481                         rs->max_bw[1] = wbw;
482
483                 rs->io_kb[0] += td->io_bytes[0] >> 10;
484                 rs->io_kb[1] += td->io_bytes[1] >> 10;
485         }
486
487         for (i = 0; i < groupid + 1; i++) {
488                 rs = &runstats[i];
489
490                 if (rs->max_run[0])
491                         rs->agg[0] = (rs->io_kb[0]*1024) / rs->max_run[0];
492                 if (rs->max_run[1])
493                         rs->agg[1] = (rs->io_kb[1]*1024) / rs->max_run[1];
494         }
495
496         /*
497          * don't overwrite last signal output
498          */
499         if (!terse_output)
500                 printf("\n");
501
502         for_each_td(td, i) {
503                 rs = &runstats[td->groupid];
504
505                 if (terse_output)
506                         show_thread_status_terse(td, rs);
507                 else
508                         show_thread_status(td, rs);
509         }
510
511         if (!terse_output) {
512                 for (i = 0; i < groupid + 1; i++)
513                         show_group_stats(&runstats[i], i);
514
515                 show_disk_util();
516         }
517 }
518
519 static inline void add_stat_sample(struct io_stat *is, unsigned long val)
520 {
521         if (val > is->max_val)
522                 is->max_val = val;
523         if (val < is->min_val)
524                 is->min_val = val;
525
526         is->val += val;
527         is->val_sq += val * val;
528         is->samples++;
529 }
530
531 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
532                            unsigned long val, int ddir)
533 {
534         if (iolog->nr_samples == iolog->max_samples) {
535                 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
536
537                 iolog->log = realloc(iolog->log, new_size);
538                 iolog->max_samples <<= 1;
539         }
540
541         iolog->log[iolog->nr_samples].val = val;
542         iolog->log[iolog->nr_samples].time = mtime_since_now(&td->epoch);
543         iolog->log[iolog->nr_samples].ddir = ddir;
544         iolog->nr_samples++;
545 }
546
547 void add_clat_sample(struct thread_data *td, int ddir, unsigned long msec)
548 {
549         add_stat_sample(&td->clat_stat[ddir], msec);
550
551         if (td->clat_log)
552                 add_log_sample(td, td->clat_log, msec, ddir);
553 }
554
555 void add_slat_sample(struct thread_data *td, int ddir, unsigned long msec)
556 {
557         add_stat_sample(&td->slat_stat[ddir], msec);
558
559         if (td->slat_log)
560                 add_log_sample(td, td->slat_log, msec, ddir);
561 }
562
563 void add_bw_sample(struct thread_data *td, int ddir)
564 {
565         unsigned long spent = mtime_since_now(&td->stat_sample_time[ddir]);
566         unsigned long rate;
567
568         if (spent < td->bw_avg_time)
569                 return;
570
571         rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) / spent;
572         add_stat_sample(&td->bw_stat[ddir], rate);
573
574         if (td->bw_log)
575                 add_log_sample(td, td->bw_log, rate, ddir);
576
577         gettimeofday(&td->stat_sample_time[ddir], NULL);
578         td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
579 }
580
581