4cc3b7e92fbe802d80cc10f5baee9d4babfb0dee
[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                 if (!strcmp(dir->d_name, "device"))
160                         continue;
161
162                 sprintf(full_path, "%s/%s", path, dir->d_name);
163
164                 if (!strcmp(dir->d_name, "dev")) {
165                         if (!check_dev_match(dev, full_path)) {
166                                 found = 1;
167                                 break;
168                         }
169                 }
170
171                 if (stat(full_path, &st) == -1) {
172                         perror("stat");
173                         break;
174                 }
175
176                 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
177                         continue;
178
179                 found = find_block_dir(dev, full_path);
180                 if (found) {
181                         strcpy(path, full_path);
182                         break;
183                 }
184         }
185
186         closedir(D);
187         return found;
188 }
189
190 void init_disk_util(struct thread_data *td)
191 {
192         struct stat st;
193         char foo[256], tmp[256];
194         dev_t dev;
195         char *p;
196
197         if (!td->do_disk_util)
198                 return;
199
200         if (!stat(td->file_name, &st)) {
201                 if (S_ISBLK(st.st_mode))
202                         dev = st.st_rdev;
203                 else
204                         dev = st.st_dev;
205         } else {
206                 /*
207                  * must be a file, open "." in that path
208                  */
209                 strcpy(foo, td->file_name);
210                 p = dirname(foo);
211                 if (stat(p, &st)) {
212                         perror("disk util stat");
213                         return;
214                 }
215
216                 dev = st.st_dev;
217         }
218
219         if (disk_util_exists(dev))
220                 return;
221                 
222         sprintf(foo, "/sys/block");
223         if (!find_block_dir(dev, foo))
224                 return;
225
226         /*
227          * If there's a ../queue/ directory there, we are inside a partition.
228          * Check if that is the case and jump back. For loop/md/dm etc we
229          * are already in the right spot.
230          */
231         sprintf(tmp, "%s/../queue", foo);
232         if (!stat(tmp, &st)) {
233                 p = dirname(foo);
234                 sprintf(tmp, "%s/queue", p);
235                 if (stat(tmp, &st)) {
236                         log_err("unknown sysfs layout\n");
237                         return;
238                 }
239                 sprintf(foo, "%s", p);
240         }
241
242         td->sysfs_root = strdup(foo);
243         disk_util_add(dev, foo);
244 }
245
246 void disk_util_timer_arm(void)
247 {
248         itimer.it_value.tv_sec = 0;
249         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
250         setitimer(ITIMER_REAL, &itimer, NULL);
251 }
252
253 void update_rusage_stat(struct thread_data *td)
254 {
255         if (!(td->runtime[0] + td->runtime[1]))
256                 return;
257
258         getrusage(RUSAGE_SELF, &td->ru_end);
259
260         td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
261         td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
262         td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
263
264         
265         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
266 }
267
268 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
269                     double *mean, double *dev)
270 {
271         double n;
272
273         if (is->samples == 0)
274                 return 0;
275
276         *min = is->min_val;
277         *max = is->max_val;
278
279         n = (double) is->samples;
280         *mean = (double) is->val / n;
281         *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
282
283         return 1;
284 }
285
286 static void show_group_stats(struct group_run_stats *rs, int id)
287 {
288         fprintf(f_out, "\nRun status group %d (all jobs):\n", id);
289
290         if (rs->max_run[DDIR_READ])
291                 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]);
292         if (rs->max_run[DDIR_WRITE])
293                 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]);
294 }
295
296 static void show_disk_util(void)
297 {
298         struct disk_util_stat *dus;
299         struct list_head *entry;
300         struct disk_util *du;
301         double util;
302
303         fprintf(f_out, "\nDisk stats (read/write):\n");
304
305         list_for_each(entry, &disk_list) {
306                 du = list_entry(entry, struct disk_util, list);
307                 dus = &du->dus;
308
309                 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
310                 if (util > 100.0)
311                         util = 100.0;
312
313                 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);
314         }
315 }
316
317 static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
318                              int ddir)
319 {
320         char *ddir_str[] = { "read ", "write" };
321         unsigned long min, max;
322         unsigned long long bw;
323         double mean, dev;
324
325         if (!td->runtime[ddir])
326                 return;
327
328         bw = td->io_bytes[ddir] / td->runtime[ddir];
329         fprintf(f_out, "  %s: io=%6lluMiB, bw=%6lluKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
330
331         if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
332                 fprintf(f_out, "    slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
333
334         if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
335                 fprintf(f_out, "    clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
336
337         if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
338                 double p_of_agg;
339
340                 p_of_agg = mean * 100 / (double) rs->agg[ddir];
341                 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);
342         }
343 }
344
345 static void show_thread_status(struct thread_data *td,
346                                struct group_run_stats *rs)
347 {
348         double usr_cpu, sys_cpu;
349
350         if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
351                 return;
352
353         fprintf(f_out, "%s: (groupid=%d): err=%2d:\n",td->name, td->groupid, td->error);
354
355         show_ddir_status(td, rs, td->ddir);
356         if (td->io_bytes[td->ddir ^ 1])
357                 show_ddir_status(td, rs, td->ddir ^ 1);
358
359         if (td->runtime[0] + td->runtime[1]) {
360                 double runt = td->runtime[0] + td->runtime[1];
361
362                 usr_cpu = (double) td->usr_time * 100 / runt;
363                 sys_cpu = (double) td->sys_time * 100 / runt;
364         } else {
365                 usr_cpu = 0;
366                 sys_cpu = 0;
367         }
368
369         fprintf(f_out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
370 }
371
372 static void show_ddir_status_terse(struct thread_data *td,
373                                    struct group_run_stats *rs, int ddir)
374 {
375         unsigned long min, max;
376         unsigned long long bw;
377         double mean, dev;
378
379         bw = 0;
380         if (td->runtime[ddir])
381                 bw = td->io_bytes[ddir] / td->runtime[ddir];
382
383         fprintf(f_out, ",%llu,%llu,%lu", td->io_bytes[ddir] >> 10, bw, td->runtime[ddir]);
384
385         if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
386                 fprintf(f_out, ",%lu,%lu,%f,%f", min, max, mean, dev);
387         else
388                 fprintf(f_out, ",%lu,%lu,%f,%f", 0UL, 0UL, 0.0, 0.0);
389
390         if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
391                 fprintf(f_out, ",%lu,%lu,%f,%f", min, max, mean, dev);
392         else
393                 fprintf(f_out, ",%lu,%lu,%f,%f", 0UL, 0UL, 0.0, 0.0);
394
395         if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
396                 double p_of_agg;
397
398                 p_of_agg = mean * 100 / (double) rs->agg[ddir];
399                 fprintf(f_out, ",%lu,%lu,%f%%,%f,%f", min, max, p_of_agg, mean, dev);
400         } else
401                 fprintf(f_out, ",%lu,%lu,%f%%,%f,%f", 0UL, 0UL, 0.0, 0.0, 0.0);
402                 
403 }
404
405
406 static void show_thread_status_terse(struct thread_data *td,
407                                      struct group_run_stats *rs)
408 {
409         double usr_cpu, sys_cpu;
410
411         fprintf(f_out, "%s,%d,%d",td->name, td->groupid, td->error);
412
413         show_ddir_status_terse(td, rs, 0);
414         show_ddir_status_terse(td, rs, 1);
415
416         if (td->runtime[0] + td->runtime[1]) {
417                 double runt = td->runtime[0] + td->runtime[1];
418
419                 usr_cpu = (double) td->usr_time * 100 / runt;
420                 sys_cpu = (double) td->sys_time * 100 / runt;
421         } else {
422                 usr_cpu = 0;
423                 sys_cpu = 0;
424         }
425
426         fprintf(f_out, ",%f%%,%f%%,%lu\n", usr_cpu, sys_cpu, td->ctx);
427 }
428
429 void show_run_stats(void)
430 {
431         struct group_run_stats *runstats, *rs;
432         struct thread_data *td;
433         int i;
434
435         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
436
437         for (i = 0; i < groupid + 1; i++) {
438                 rs = &runstats[i];
439
440                 memset(rs, 0, sizeof(*rs));
441                 rs->min_bw[0] = rs->min_run[0] = ~0UL;
442                 rs->min_bw[1] = rs->min_run[1] = ~0UL;
443         }
444
445         for (i = 0; i < thread_number; i++) {
446                 unsigned long long rbw, wbw;
447
448                 td = &threads[i];
449
450                 if (td->error) {
451                         fprintf(f_out, "%s: %s\n", td->name, td->verror);
452                         continue;
453                 }
454
455                 rs = &runstats[td->groupid];
456
457                 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
458                         rs->min_run[0] = td->runtime[0];
459                 if (td->runtime[0] > rs->max_run[0])
460                         rs->max_run[0] = td->runtime[0];
461                 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
462                         rs->min_run[1] = td->runtime[1];
463                 if (td->runtime[1] > rs->max_run[1])
464                         rs->max_run[1] = td->runtime[1];
465
466                 rbw = wbw = 0;
467                 if (td->runtime[0])
468                         rbw = td->io_bytes[0] / (unsigned long long) td->runtime[0];
469                 if (td->runtime[1])
470                         wbw = td->io_bytes[1] / (unsigned long long) td->runtime[1];
471
472                 if (rbw < rs->min_bw[0])
473                         rs->min_bw[0] = rbw;
474                 if (wbw < rs->min_bw[1])
475                         rs->min_bw[1] = wbw;
476                 if (rbw > rs->max_bw[0])
477                         rs->max_bw[0] = rbw;
478                 if (wbw > rs->max_bw[1])
479                         rs->max_bw[1] = wbw;
480
481                 rs->io_kb[0] += td->io_bytes[0] >> 10;
482                 rs->io_kb[1] += td->io_bytes[1] >> 10;
483         }
484
485         for (i = 0; i < groupid + 1; i++) {
486                 rs = &runstats[i];
487
488                 if (rs->max_run[0])
489                         rs->agg[0] = (rs->io_kb[0]*1024) / rs->max_run[0];
490                 if (rs->max_run[1])
491                         rs->agg[1] = (rs->io_kb[1]*1024) / rs->max_run[1];
492         }
493
494         /*
495          * don't overwrite last signal output
496          */
497         if (!terse_output)
498                 printf("\n");
499
500         for (i = 0; i < thread_number; i++) {
501                 td = &threads[i];
502                 rs = &runstats[td->groupid];
503
504                 if (terse_output)
505                         show_thread_status_terse(td, rs);
506                 else
507                         show_thread_status(td, rs);
508         }
509
510         if (!terse_output) {
511                 for (i = 0; i < groupid + 1; i++)
512                         show_group_stats(&runstats[i], i);
513
514                 show_disk_util();
515         }
516 }
517
518 static inline void add_stat_sample(struct io_stat *is, unsigned long val)
519 {
520         if (val > is->max_val)
521                 is->max_val = val;
522         if (val < is->min_val)
523                 is->min_val = val;
524
525         is->val += val;
526         is->val_sq += val * val;
527         is->samples++;
528 }
529
530 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
531                            unsigned long val, int ddir)
532 {
533         if (iolog->nr_samples == iolog->max_samples) {
534                 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
535
536                 iolog->log = realloc(iolog->log, new_size);
537                 iolog->max_samples <<= 1;
538         }
539
540         iolog->log[iolog->nr_samples].val = val;
541         iolog->log[iolog->nr_samples].time = mtime_since_now(&td->epoch);
542         iolog->log[iolog->nr_samples].ddir = ddir;
543         iolog->nr_samples++;
544 }
545
546 void add_clat_sample(struct thread_data *td, int ddir, unsigned long msec)
547 {
548         add_stat_sample(&td->clat_stat[ddir], msec);
549
550         if (td->clat_log)
551                 add_log_sample(td, td->clat_log, msec, ddir);
552 }
553
554 void add_slat_sample(struct thread_data *td, int ddir, unsigned long msec)
555 {
556         add_stat_sample(&td->slat_stat[ddir], msec);
557
558         if (td->slat_log)
559                 add_log_sample(td, td->slat_log, msec, ddir);
560 }
561
562 void add_bw_sample(struct thread_data *td, int ddir)
563 {
564         unsigned long spent = mtime_since_now(&td->stat_sample_time[ddir]);
565         unsigned long rate;
566
567         if (spent < td->bw_avg_time)
568                 return;
569
570         rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) / spent;
571         add_stat_sample(&td->bw_stat[ddir], rate);
572
573         if (td->bw_log)
574                 add_log_sample(td, td->bw_log, rate, ddir);
575
576         gettimeofday(&td->stat_sample_time[ddir], NULL);
577         td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
578 }
579
580