Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/* delayacct.c - per-task delay accounting
3 *
4 * Copyright (C) Shailabh Nagar, IBM Corp. 2006
5 */
6
7#include <linux/sched.h>
8#include <linux/sched/task.h>
9#include <linux/sched/cputime.h>
10#include <linux/sched/clock.h>
11#include <linux/slab.h>
12#include <linux/taskstats.h>
13#include <linux/sysctl.h>
14#include <linux/delayacct.h>
15#include <linux/module.h>
16
17DEFINE_STATIC_KEY_FALSE(delayacct_key);
18int delayacct_on __read_mostly; /* Delay accounting turned on/off */
19struct kmem_cache *delayacct_cache;
20
21static void set_delayacct(bool enabled)
22{
23 if (enabled) {
24 static_branch_enable(&delayacct_key);
25 delayacct_on = 1;
26 } else {
27 delayacct_on = 0;
28 static_branch_disable(&delayacct_key);
29 }
30}
31
32static int __init delayacct_setup_enable(char *str)
33{
34 delayacct_on = 1;
35 return 1;
36}
37__setup("delayacct", delayacct_setup_enable);
38
39void delayacct_init(void)
40{
41 delayacct_cache = KMEM_CACHE(task_delay_info, SLAB_PANIC|SLAB_ACCOUNT);
42 delayacct_tsk_init(&init_task);
43 set_delayacct(delayacct_on);
44}
45
46#ifdef CONFIG_PROC_SYSCTL
47int sysctl_delayacct(struct ctl_table *table, int write, void *buffer,
48 size_t *lenp, loff_t *ppos)
49{
50 int state = delayacct_on;
51 struct ctl_table t;
52 int err;
53
54 if (write && !capable(CAP_SYS_ADMIN))
55 return -EPERM;
56
57 t = *table;
58 t.data = &state;
59 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
60 if (err < 0)
61 return err;
62 if (write)
63 set_delayacct(state);
64 return err;
65}
66#endif
67
68void __delayacct_tsk_init(struct task_struct *tsk)
69{
70 tsk->delays = kmem_cache_zalloc(delayacct_cache, GFP_KERNEL);
71 if (tsk->delays)
72 raw_spin_lock_init(&tsk->delays->lock);
73}
74
75/*
76 * Finish delay accounting for a statistic using its timestamps (@start),
77 * accumalator (@total) and @count
78 */
79static void delayacct_end(raw_spinlock_t *lock, u64 *start, u64 *total, u32 *count)
80{
81 s64 ns = local_clock() - *start;
82 unsigned long flags;
83
84 if (ns > 0) {
85 raw_spin_lock_irqsave(lock, flags);
86 *total += ns;
87 (*count)++;
88 raw_spin_unlock_irqrestore(lock, flags);
89 }
90}
91
92void __delayacct_blkio_start(void)
93{
94 current->delays->blkio_start = local_clock();
95}
96
97/*
98 * We cannot rely on the `current` macro, as we haven't yet switched back to
99 * the process being woken.
100 */
101void __delayacct_blkio_end(struct task_struct *p)
102{
103 struct task_delay_info *delays = p->delays;
104 u64 *total;
105 u32 *count;
106
107 if (p->delays->flags & DELAYACCT_PF_SWAPIN) {
108 total = &delays->swapin_delay;
109 count = &delays->swapin_count;
110 } else {
111 total = &delays->blkio_delay;
112 count = &delays->blkio_count;
113 }
114
115 delayacct_end(&delays->lock, &delays->blkio_start, total, count);
116}
117
118int delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk)
119{
120 u64 utime, stime, stimescaled, utimescaled;
121 unsigned long long t2, t3;
122 unsigned long flags, t1;
123 s64 tmp;
124
125 task_cputime(tsk, &utime, &stime);
126 tmp = (s64)d->cpu_run_real_total;
127 tmp += utime + stime;
128 d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp;
129
130 task_cputime_scaled(tsk, &utimescaled, &stimescaled);
131 tmp = (s64)d->cpu_scaled_run_real_total;
132 tmp += utimescaled + stimescaled;
133 d->cpu_scaled_run_real_total =
134 (tmp < (s64)d->cpu_scaled_run_real_total) ? 0 : tmp;
135
136 /*
137 * No locking available for sched_info (and too expensive to add one)
138 * Mitigate by taking snapshot of values
139 */
140 t1 = tsk->sched_info.pcount;
141 t2 = tsk->sched_info.run_delay;
142 t3 = tsk->se.sum_exec_runtime;
143
144 d->cpu_count += t1;
145
146 tmp = (s64)d->cpu_delay_total + t2;
147 d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp;
148
149 tmp = (s64)d->cpu_run_virtual_total + t3;
150 d->cpu_run_virtual_total =
151 (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp;
152
153 if (!tsk->delays)
154 return 0;
155
156 /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */
157
158 raw_spin_lock_irqsave(&tsk->delays->lock, flags);
159 tmp = d->blkio_delay_total + tsk->delays->blkio_delay;
160 d->blkio_delay_total = (tmp < d->blkio_delay_total) ? 0 : tmp;
161 tmp = d->swapin_delay_total + tsk->delays->swapin_delay;
162 d->swapin_delay_total = (tmp < d->swapin_delay_total) ? 0 : tmp;
163 tmp = d->freepages_delay_total + tsk->delays->freepages_delay;
164 d->freepages_delay_total = (tmp < d->freepages_delay_total) ? 0 : tmp;
165 tmp = d->thrashing_delay_total + tsk->delays->thrashing_delay;
166 d->thrashing_delay_total = (tmp < d->thrashing_delay_total) ? 0 : tmp;
167 d->blkio_count += tsk->delays->blkio_count;
168 d->swapin_count += tsk->delays->swapin_count;
169 d->freepages_count += tsk->delays->freepages_count;
170 d->thrashing_count += tsk->delays->thrashing_count;
171 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags);
172
173 return 0;
174}
175
176__u64 __delayacct_blkio_ticks(struct task_struct *tsk)
177{
178 __u64 ret;
179 unsigned long flags;
180
181 raw_spin_lock_irqsave(&tsk->delays->lock, flags);
182 ret = nsec_to_clock_t(tsk->delays->blkio_delay +
183 tsk->delays->swapin_delay);
184 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags);
185 return ret;
186}
187
188void __delayacct_freepages_start(void)
189{
190 current->delays->freepages_start = local_clock();
191}
192
193void __delayacct_freepages_end(void)
194{
195 delayacct_end(¤t->delays->lock,
196 ¤t->delays->freepages_start,
197 ¤t->delays->freepages_delay,
198 ¤t->delays->freepages_count);
199}
200
201void __delayacct_thrashing_start(void)
202{
203 current->delays->thrashing_start = local_clock();
204}
205
206void __delayacct_thrashing_end(void)
207{
208 delayacct_end(¤t->delays->lock,
209 ¤t->delays->thrashing_start,
210 ¤t->delays->thrashing_delay,
211 ¤t->delays->thrashing_count);
212}
1// SPDX-License-Identifier: GPL-2.0-or-later
2/* delayacct.c - per-task delay accounting
3 *
4 * Copyright (C) Shailabh Nagar, IBM Corp. 2006
5 */
6
7#include <linux/sched.h>
8#include <linux/sched/task.h>
9#include <linux/sched/cputime.h>
10#include <linux/slab.h>
11#include <linux/taskstats.h>
12#include <linux/time.h>
13#include <linux/sysctl.h>
14#include <linux/delayacct.h>
15#include <linux/module.h>
16
17int delayacct_on __read_mostly = 1; /* Delay accounting turned on/off */
18EXPORT_SYMBOL_GPL(delayacct_on);
19struct kmem_cache *delayacct_cache;
20
21static int __init delayacct_setup_disable(char *str)
22{
23 delayacct_on = 0;
24 return 1;
25}
26__setup("nodelayacct", delayacct_setup_disable);
27
28void delayacct_init(void)
29{
30 delayacct_cache = KMEM_CACHE(task_delay_info, SLAB_PANIC|SLAB_ACCOUNT);
31 delayacct_tsk_init(&init_task);
32}
33
34void __delayacct_tsk_init(struct task_struct *tsk)
35{
36 tsk->delays = kmem_cache_zalloc(delayacct_cache, GFP_KERNEL);
37 if (tsk->delays)
38 raw_spin_lock_init(&tsk->delays->lock);
39}
40
41/*
42 * Finish delay accounting for a statistic using its timestamps (@start),
43 * accumalator (@total) and @count
44 */
45static void delayacct_end(raw_spinlock_t *lock, u64 *start, u64 *total,
46 u32 *count)
47{
48 s64 ns = ktime_get_ns() - *start;
49 unsigned long flags;
50
51 if (ns > 0) {
52 raw_spin_lock_irqsave(lock, flags);
53 *total += ns;
54 (*count)++;
55 raw_spin_unlock_irqrestore(lock, flags);
56 }
57}
58
59void __delayacct_blkio_start(void)
60{
61 current->delays->blkio_start = ktime_get_ns();
62}
63
64/*
65 * We cannot rely on the `current` macro, as we haven't yet switched back to
66 * the process being woken.
67 */
68void __delayacct_blkio_end(struct task_struct *p)
69{
70 struct task_delay_info *delays = p->delays;
71 u64 *total;
72 u32 *count;
73
74 if (p->delays->flags & DELAYACCT_PF_SWAPIN) {
75 total = &delays->swapin_delay;
76 count = &delays->swapin_count;
77 } else {
78 total = &delays->blkio_delay;
79 count = &delays->blkio_count;
80 }
81
82 delayacct_end(&delays->lock, &delays->blkio_start, total, count);
83}
84
85int __delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk)
86{
87 u64 utime, stime, stimescaled, utimescaled;
88 unsigned long long t2, t3;
89 unsigned long flags, t1;
90 s64 tmp;
91
92 task_cputime(tsk, &utime, &stime);
93 tmp = (s64)d->cpu_run_real_total;
94 tmp += utime + stime;
95 d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp;
96
97 task_cputime_scaled(tsk, &utimescaled, &stimescaled);
98 tmp = (s64)d->cpu_scaled_run_real_total;
99 tmp += utimescaled + stimescaled;
100 d->cpu_scaled_run_real_total =
101 (tmp < (s64)d->cpu_scaled_run_real_total) ? 0 : tmp;
102
103 /*
104 * No locking available for sched_info (and too expensive to add one)
105 * Mitigate by taking snapshot of values
106 */
107 t1 = tsk->sched_info.pcount;
108 t2 = tsk->sched_info.run_delay;
109 t3 = tsk->se.sum_exec_runtime;
110
111 d->cpu_count += t1;
112
113 tmp = (s64)d->cpu_delay_total + t2;
114 d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp;
115
116 tmp = (s64)d->cpu_run_virtual_total + t3;
117 d->cpu_run_virtual_total =
118 (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp;
119
120 /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */
121
122 raw_spin_lock_irqsave(&tsk->delays->lock, flags);
123 tmp = d->blkio_delay_total + tsk->delays->blkio_delay;
124 d->blkio_delay_total = (tmp < d->blkio_delay_total) ? 0 : tmp;
125 tmp = d->swapin_delay_total + tsk->delays->swapin_delay;
126 d->swapin_delay_total = (tmp < d->swapin_delay_total) ? 0 : tmp;
127 tmp = d->freepages_delay_total + tsk->delays->freepages_delay;
128 d->freepages_delay_total = (tmp < d->freepages_delay_total) ? 0 : tmp;
129 tmp = d->thrashing_delay_total + tsk->delays->thrashing_delay;
130 d->thrashing_delay_total = (tmp < d->thrashing_delay_total) ? 0 : tmp;
131 d->blkio_count += tsk->delays->blkio_count;
132 d->swapin_count += tsk->delays->swapin_count;
133 d->freepages_count += tsk->delays->freepages_count;
134 d->thrashing_count += tsk->delays->thrashing_count;
135 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags);
136
137 return 0;
138}
139
140__u64 __delayacct_blkio_ticks(struct task_struct *tsk)
141{
142 __u64 ret;
143 unsigned long flags;
144
145 raw_spin_lock_irqsave(&tsk->delays->lock, flags);
146 ret = nsec_to_clock_t(tsk->delays->blkio_delay +
147 tsk->delays->swapin_delay);
148 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags);
149 return ret;
150}
151
152void __delayacct_freepages_start(void)
153{
154 current->delays->freepages_start = ktime_get_ns();
155}
156
157void __delayacct_freepages_end(void)
158{
159 delayacct_end(
160 ¤t->delays->lock,
161 ¤t->delays->freepages_start,
162 ¤t->delays->freepages_delay,
163 ¤t->delays->freepages_count);
164}
165
166void __delayacct_thrashing_start(void)
167{
168 current->delays->thrashing_start = ktime_get_ns();
169}
170
171void __delayacct_thrashing_end(void)
172{
173 delayacct_end(¤t->delays->lock,
174 ¤t->delays->thrashing_start,
175 ¤t->delays->thrashing_delay,
176 ¤t->delays->thrashing_count);
177}