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/* 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/sched/task.h>
18#include <linux/sched/cputime.h>
19#include <linux/slab.h>
20#include <linux/taskstats.h>
21#include <linux/time.h>
22#include <linux/sysctl.h>
23#include <linux/delayacct.h>
24#include <linux/module.h>
25
26int delayacct_on __read_mostly = 1; /* Delay accounting turned on/off */
27EXPORT_SYMBOL_GPL(delayacct_on);
28struct kmem_cache *delayacct_cache;
29
30static int __init delayacct_setup_disable(char *str)
31{
32 delayacct_on = 0;
33 return 1;
34}
35__setup("nodelayacct", delayacct_setup_disable);
36
37void delayacct_init(void)
38{
39 delayacct_cache = KMEM_CACHE(task_delay_info, SLAB_PANIC|SLAB_ACCOUNT);
40 delayacct_tsk_init(&init_task);
41}
42
43void __delayacct_tsk_init(struct task_struct *tsk)
44{
45 tsk->delays = kmem_cache_zalloc(delayacct_cache, GFP_KERNEL);
46 if (tsk->delays)
47 spin_lock_init(&tsk->delays->lock);
48}
49
50/*
51 * Finish delay accounting for a statistic using its timestamps (@start),
52 * accumalator (@total) and @count
53 */
54static void delayacct_end(spinlock_t *lock, u64 *start, u64 *total, u32 *count)
55{
56 s64 ns = ktime_get_ns() - *start;
57 unsigned long flags;
58
59 if (ns > 0) {
60 spin_lock_irqsave(lock, flags);
61 *total += ns;
62 (*count)++;
63 spin_unlock_irqrestore(lock, flags);
64 }
65}
66
67void __delayacct_blkio_start(void)
68{
69 current->delays->blkio_start = ktime_get_ns();
70}
71
72/*
73 * We cannot rely on the `current` macro, as we haven't yet switched back to
74 * the process being woken.
75 */
76void __delayacct_blkio_end(struct task_struct *p)
77{
78 struct task_delay_info *delays = p->delays;
79 u64 *total;
80 u32 *count;
81
82 if (p->delays->flags & DELAYACCT_PF_SWAPIN) {
83 total = &delays->swapin_delay;
84 count = &delays->swapin_count;
85 } else {
86 total = &delays->blkio_delay;
87 count = &delays->blkio_count;
88 }
89
90 delayacct_end(&delays->lock, &delays->blkio_start, total, count);
91}
92
93int __delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk)
94{
95 u64 utime, stime, stimescaled, utimescaled;
96 unsigned long long t2, t3;
97 unsigned long flags, t1;
98 s64 tmp;
99
100 task_cputime(tsk, &utime, &stime);
101 tmp = (s64)d->cpu_run_real_total;
102 tmp += utime + stime;
103 d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp;
104
105 task_cputime_scaled(tsk, &utimescaled, &stimescaled);
106 tmp = (s64)d->cpu_scaled_run_real_total;
107 tmp += utimescaled + stimescaled;
108 d->cpu_scaled_run_real_total =
109 (tmp < (s64)d->cpu_scaled_run_real_total) ? 0 : tmp;
110
111 /*
112 * No locking available for sched_info (and too expensive to add one)
113 * Mitigate by taking snapshot of values
114 */
115 t1 = tsk->sched_info.pcount;
116 t2 = tsk->sched_info.run_delay;
117 t3 = tsk->se.sum_exec_runtime;
118
119 d->cpu_count += t1;
120
121 tmp = (s64)d->cpu_delay_total + t2;
122 d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp;
123
124 tmp = (s64)d->cpu_run_virtual_total + t3;
125 d->cpu_run_virtual_total =
126 (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp;
127
128 /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */
129
130 spin_lock_irqsave(&tsk->delays->lock, flags);
131 tmp = d->blkio_delay_total + tsk->delays->blkio_delay;
132 d->blkio_delay_total = (tmp < d->blkio_delay_total) ? 0 : tmp;
133 tmp = d->swapin_delay_total + tsk->delays->swapin_delay;
134 d->swapin_delay_total = (tmp < d->swapin_delay_total) ? 0 : tmp;
135 tmp = d->freepages_delay_total + tsk->delays->freepages_delay;
136 d->freepages_delay_total = (tmp < d->freepages_delay_total) ? 0 : tmp;
137 d->blkio_count += tsk->delays->blkio_count;
138 d->swapin_count += tsk->delays->swapin_count;
139 d->freepages_count += tsk->delays->freepages_count;
140 spin_unlock_irqrestore(&tsk->delays->lock, flags);
141
142 return 0;
143}
144
145__u64 __delayacct_blkio_ticks(struct task_struct *tsk)
146{
147 __u64 ret;
148 unsigned long flags;
149
150 spin_lock_irqsave(&tsk->delays->lock, flags);
151 ret = nsec_to_clock_t(tsk->delays->blkio_delay +
152 tsk->delays->swapin_delay);
153 spin_unlock_irqrestore(&tsk->delays->lock, flags);
154 return ret;
155}
156
157void __delayacct_freepages_start(void)
158{
159 current->delays->freepages_start = ktime_get_ns();
160}
161
162void __delayacct_freepages_end(void)
163{
164 delayacct_end(
165 ¤t->delays->lock,
166 ¤t->delays->freepages_start,
167 ¤t->delays->freepages_delay,
168 ¤t->delays->freepages_count);
169}
170