jbd commit: fix transaction dropping
[linux-2.6] / kernel / delayacct.c
1 /* delayacct.c - per-task delay accounting
2  *
3  * Copyright (C) Shailabh Nagar, IBM Corp. 2006
4  *
5  * This program is free software;  you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License as published by
7  * the Free Software Foundation; either version 2 of the License, or
8  * (at your option) any later version.
9  *
10  * This program is distributed in the hope that it would be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See
13  * the GNU General Public License for more details.
14  */
15
16 #include <linux/sched.h>
17 #include <linux/slab.h>
18 #include <linux/time.h>
19 #include <linux/sysctl.h>
20 #include <linux/delayacct.h>
21
22 int delayacct_on __read_mostly = 1;     /* Delay accounting turned on/off */
23 struct kmem_cache *delayacct_cache;
24
25 static int __init delayacct_setup_disable(char *str)
26 {
27         delayacct_on = 0;
28         return 1;
29 }
30 __setup("nodelayacct", delayacct_setup_disable);
31
32 void delayacct_init(void)
33 {
34         delayacct_cache = KMEM_CACHE(task_delay_info, SLAB_PANIC);
35         delayacct_tsk_init(&init_task);
36 }
37
38 void __delayacct_tsk_init(struct task_struct *tsk)
39 {
40         tsk->delays = kmem_cache_zalloc(delayacct_cache, GFP_KERNEL);
41         if (tsk->delays)
42                 spin_lock_init(&tsk->delays->lock);
43 }
44
45 /*
46  * Start accounting for a delay statistic using
47  * its starting timestamp (@start)
48  */
49
50 static inline void delayacct_start(struct timespec *start)
51 {
52         do_posix_clock_monotonic_gettime(start);
53 }
54
55 /*
56  * Finish delay accounting for a statistic using
57  * its timestamps (@start, @end), accumalator (@total) and @count
58  */
59
60 static void delayacct_end(struct timespec *start, struct timespec *end,
61                                 u64 *total, u32 *count)
62 {
63         struct timespec ts;
64         s64 ns;
65         unsigned long flags;
66
67         do_posix_clock_monotonic_gettime(end);
68         ts = timespec_sub(*end, *start);
69         ns = timespec_to_ns(&ts);
70         if (ns < 0)
71                 return;
72
73         spin_lock_irqsave(&current->delays->lock, flags);
74         *total += ns;
75         (*count)++;
76         spin_unlock_irqrestore(&current->delays->lock, flags);
77 }
78
79 void __delayacct_blkio_start(void)
80 {
81         delayacct_start(&current->delays->blkio_start);
82 }
83
84 void __delayacct_blkio_end(void)
85 {
86         if (current->delays->flags & DELAYACCT_PF_SWAPIN)
87                 /* Swapin block I/O */
88                 delayacct_end(&current->delays->blkio_start,
89                         &current->delays->blkio_end,
90                         &current->delays->swapin_delay,
91                         &current->delays->swapin_count);
92         else    /* Other block I/O */
93                 delayacct_end(&current->delays->blkio_start,
94                         &current->delays->blkio_end,
95                         &current->delays->blkio_delay,
96                         &current->delays->blkio_count);
97 }
98
99 int __delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk)
100 {
101         s64 tmp;
102         unsigned long t1;
103         unsigned long long t2, t3;
104         unsigned long flags;
105         struct timespec ts;
106
107         /* Though tsk->delays accessed later, early exit avoids
108          * unnecessary returning of other data
109          */
110         if (!tsk->delays)
111                 goto done;
112
113         tmp = (s64)d->cpu_run_real_total;
114         cputime_to_timespec(tsk->utime + tsk->stime, &ts);
115         tmp += timespec_to_ns(&ts);
116         d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp;
117
118         /*
119          * No locking available for sched_info (and too expensive to add one)
120          * Mitigate by taking snapshot of values
121          */
122         t1 = tsk->sched_info.pcnt;
123         t2 = tsk->sched_info.run_delay;
124         t3 = tsk->sched_info.cpu_time;
125
126         d->cpu_count += t1;
127
128         tmp = (s64)d->cpu_delay_total + t2;
129         d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp;
130
131         tmp = (s64)d->cpu_run_virtual_total + t3;
132         d->cpu_run_virtual_total =
133                 (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp;
134
135         /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */
136
137         spin_lock_irqsave(&tsk->delays->lock, flags);
138         tmp = d->blkio_delay_total + tsk->delays->blkio_delay;
139         d->blkio_delay_total = (tmp < d->blkio_delay_total) ? 0 : tmp;
140         tmp = d->swapin_delay_total + tsk->delays->swapin_delay;
141         d->swapin_delay_total = (tmp < d->swapin_delay_total) ? 0 : tmp;
142         d->blkio_count += tsk->delays->blkio_count;
143         d->swapin_count += tsk->delays->swapin_count;
144         spin_unlock_irqrestore(&tsk->delays->lock, flags);
145
146 done:
147         return 0;
148 }
149
150 __u64 __delayacct_blkio_ticks(struct task_struct *tsk)
151 {
152         __u64 ret;
153         unsigned long flags;
154
155         spin_lock_irqsave(&tsk->delays->lock, flags);
156         ret = nsec_to_clock_t(tsk->delays->blkio_delay +
157                                 tsk->delays->swapin_delay);
158         spin_unlock_irqrestore(&tsk->delays->lock, flags);
159         return ret;
160 }
161