Loading...
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/taskstats.h>
19#include <linux/time.h>
20#include <linux/sysctl.h>
21#include <linux/delayacct.h>
22#include <linux/module.h>
23
24int delayacct_on __read_mostly = 1; /* Delay accounting turned on/off */
25EXPORT_SYMBOL_GPL(delayacct_on);
26struct kmem_cache *delayacct_cache;
27
28static int __init delayacct_setup_disable(char *str)
29{
30 delayacct_on = 0;
31 return 1;
32}
33__setup("nodelayacct", delayacct_setup_disable);
34
35void delayacct_init(void)
36{
37 delayacct_cache = KMEM_CACHE(task_delay_info, SLAB_PANIC|SLAB_ACCOUNT);
38 delayacct_tsk_init(&init_task);
39}
40
41void __delayacct_tsk_init(struct task_struct *tsk)
42{
43 tsk->delays = kmem_cache_zalloc(delayacct_cache, GFP_KERNEL);
44 if (tsk->delays)
45 spin_lock_init(&tsk->delays->lock);
46}
47
48/*
49 * Finish delay accounting for a statistic using its timestamps (@start),
50 * accumalator (@total) and @count
51 */
52static void delayacct_end(u64 *start, u64 *total, u32 *count)
53{
54 s64 ns = ktime_get_ns() - *start;
55 unsigned long flags;
56
57 if (ns > 0) {
58 spin_lock_irqsave(¤t->delays->lock, flags);
59 *total += ns;
60 (*count)++;
61 spin_unlock_irqrestore(¤t->delays->lock, flags);
62 }
63}
64
65void __delayacct_blkio_start(void)
66{
67 current->delays->blkio_start = ktime_get_ns();
68}
69
70void __delayacct_blkio_end(void)
71{
72 if (current->delays->flags & DELAYACCT_PF_SWAPIN)
73 /* Swapin block I/O */
74 delayacct_end(¤t->delays->blkio_start,
75 ¤t->delays->swapin_delay,
76 ¤t->delays->swapin_count);
77 else /* Other block I/O */
78 delayacct_end(¤t->delays->blkio_start,
79 ¤t->delays->blkio_delay,
80 ¤t->delays->blkio_count);
81}
82
83int __delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk)
84{
85 cputime_t utime, stime, stimescaled, utimescaled;
86 unsigned long long t2, t3;
87 unsigned long flags, t1;
88 s64 tmp;
89
90 task_cputime(tsk, &utime, &stime);
91 tmp = (s64)d->cpu_run_real_total;
92 tmp += cputime_to_nsecs(utime + stime);
93 d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp;
94
95 task_cputime_scaled(tsk, &utimescaled, &stimescaled);
96 tmp = (s64)d->cpu_scaled_run_real_total;
97 tmp += cputime_to_nsecs(utimescaled + stimescaled);
98 d->cpu_scaled_run_real_total =
99 (tmp < (s64)d->cpu_scaled_run_real_total) ? 0 : tmp;
100
101 /*
102 * No locking available for sched_info (and too expensive to add one)
103 * Mitigate by taking snapshot of values
104 */
105 t1 = tsk->sched_info.pcount;
106 t2 = tsk->sched_info.run_delay;
107 t3 = tsk->se.sum_exec_runtime;
108
109 d->cpu_count += t1;
110
111 tmp = (s64)d->cpu_delay_total + t2;
112 d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp;
113
114 tmp = (s64)d->cpu_run_virtual_total + t3;
115 d->cpu_run_virtual_total =
116 (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp;
117
118 /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */
119
120 spin_lock_irqsave(&tsk->delays->lock, flags);
121 tmp = d->blkio_delay_total + tsk->delays->blkio_delay;
122 d->blkio_delay_total = (tmp < d->blkio_delay_total) ? 0 : tmp;
123 tmp = d->swapin_delay_total + tsk->delays->swapin_delay;
124 d->swapin_delay_total = (tmp < d->swapin_delay_total) ? 0 : tmp;
125 tmp = d->freepages_delay_total + tsk->delays->freepages_delay;
126 d->freepages_delay_total = (tmp < d->freepages_delay_total) ? 0 : tmp;
127 d->blkio_count += tsk->delays->blkio_count;
128 d->swapin_count += tsk->delays->swapin_count;
129 d->freepages_count += tsk->delays->freepages_count;
130 spin_unlock_irqrestore(&tsk->delays->lock, flags);
131
132 return 0;
133}
134
135__u64 __delayacct_blkio_ticks(struct task_struct *tsk)
136{
137 __u64 ret;
138 unsigned long flags;
139
140 spin_lock_irqsave(&tsk->delays->lock, flags);
141 ret = nsec_to_clock_t(tsk->delays->blkio_delay +
142 tsk->delays->swapin_delay);
143 spin_unlock_irqrestore(&tsk->delays->lock, flags);
144 return ret;
145}
146
147void __delayacct_freepages_start(void)
148{
149 current->delays->freepages_start = ktime_get_ns();
150}
151
152void __delayacct_freepages_end(void)
153{
154 delayacct_end(¤t->delays->freepages_start,
155 ¤t->delays->freepages_delay,
156 ¤t->delays->freepages_count);
157}
158
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
47static int 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
67static struct ctl_table kern_delayacct_table[] = {
68 {
69 .procname = "task_delayacct",
70 .data = NULL,
71 .maxlen = sizeof(unsigned int),
72 .mode = 0644,
73 .proc_handler = sysctl_delayacct,
74 .extra1 = SYSCTL_ZERO,
75 .extra2 = SYSCTL_ONE,
76 },
77 { }
78};
79
80static __init int kernel_delayacct_sysctls_init(void)
81{
82 register_sysctl_init("kernel", kern_delayacct_table);
83 return 0;
84}
85late_initcall(kernel_delayacct_sysctls_init);
86#endif
87
88void __delayacct_tsk_init(struct task_struct *tsk)
89{
90 tsk->delays = kmem_cache_zalloc(delayacct_cache, GFP_KERNEL);
91 if (tsk->delays)
92 raw_spin_lock_init(&tsk->delays->lock);
93}
94
95/*
96 * Finish delay accounting for a statistic using its timestamps (@start),
97 * accumalator (@total) and @count
98 */
99static void delayacct_end(raw_spinlock_t *lock, u64 *start, u64 *total, u32 *count)
100{
101 s64 ns = local_clock() - *start;
102 unsigned long flags;
103
104 if (ns > 0) {
105 raw_spin_lock_irqsave(lock, flags);
106 *total += ns;
107 (*count)++;
108 raw_spin_unlock_irqrestore(lock, flags);
109 }
110}
111
112void __delayacct_blkio_start(void)
113{
114 current->delays->blkio_start = local_clock();
115}
116
117/*
118 * We cannot rely on the `current` macro, as we haven't yet switched back to
119 * the process being woken.
120 */
121void __delayacct_blkio_end(struct task_struct *p)
122{
123 delayacct_end(&p->delays->lock,
124 &p->delays->blkio_start,
125 &p->delays->blkio_delay,
126 &p->delays->blkio_count);
127}
128
129int delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk)
130{
131 u64 utime, stime, stimescaled, utimescaled;
132 unsigned long long t2, t3;
133 unsigned long flags, t1;
134 s64 tmp;
135
136 task_cputime(tsk, &utime, &stime);
137 tmp = (s64)d->cpu_run_real_total;
138 tmp += utime + stime;
139 d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp;
140
141 task_cputime_scaled(tsk, &utimescaled, &stimescaled);
142 tmp = (s64)d->cpu_scaled_run_real_total;
143 tmp += utimescaled + stimescaled;
144 d->cpu_scaled_run_real_total =
145 (tmp < (s64)d->cpu_scaled_run_real_total) ? 0 : tmp;
146
147 /*
148 * No locking available for sched_info (and too expensive to add one)
149 * Mitigate by taking snapshot of values
150 */
151 t1 = tsk->sched_info.pcount;
152 t2 = tsk->sched_info.run_delay;
153 t3 = tsk->se.sum_exec_runtime;
154
155 d->cpu_count += t1;
156
157 tmp = (s64)d->cpu_delay_total + t2;
158 d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp;
159
160 tmp = (s64)d->cpu_run_virtual_total + t3;
161 d->cpu_run_virtual_total =
162 (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp;
163
164 if (!tsk->delays)
165 return 0;
166
167 /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */
168
169 raw_spin_lock_irqsave(&tsk->delays->lock, flags);
170 tmp = d->blkio_delay_total + tsk->delays->blkio_delay;
171 d->blkio_delay_total = (tmp < d->blkio_delay_total) ? 0 : tmp;
172 tmp = d->swapin_delay_total + tsk->delays->swapin_delay;
173 d->swapin_delay_total = (tmp < d->swapin_delay_total) ? 0 : tmp;
174 tmp = d->freepages_delay_total + tsk->delays->freepages_delay;
175 d->freepages_delay_total = (tmp < d->freepages_delay_total) ? 0 : tmp;
176 tmp = d->thrashing_delay_total + tsk->delays->thrashing_delay;
177 d->thrashing_delay_total = (tmp < d->thrashing_delay_total) ? 0 : tmp;
178 tmp = d->compact_delay_total + tsk->delays->compact_delay;
179 d->compact_delay_total = (tmp < d->compact_delay_total) ? 0 : tmp;
180 tmp = d->wpcopy_delay_total + tsk->delays->wpcopy_delay;
181 d->wpcopy_delay_total = (tmp < d->wpcopy_delay_total) ? 0 : tmp;
182 tmp = d->irq_delay_total + tsk->delays->irq_delay;
183 d->irq_delay_total = (tmp < d->irq_delay_total) ? 0 : tmp;
184 d->blkio_count += tsk->delays->blkio_count;
185 d->swapin_count += tsk->delays->swapin_count;
186 d->freepages_count += tsk->delays->freepages_count;
187 d->thrashing_count += tsk->delays->thrashing_count;
188 d->compact_count += tsk->delays->compact_count;
189 d->wpcopy_count += tsk->delays->wpcopy_count;
190 d->irq_count += tsk->delays->irq_count;
191 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags);
192
193 return 0;
194}
195
196__u64 __delayacct_blkio_ticks(struct task_struct *tsk)
197{
198 __u64 ret;
199 unsigned long flags;
200
201 raw_spin_lock_irqsave(&tsk->delays->lock, flags);
202 ret = nsec_to_clock_t(tsk->delays->blkio_delay);
203 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags);
204 return ret;
205}
206
207void __delayacct_freepages_start(void)
208{
209 current->delays->freepages_start = local_clock();
210}
211
212void __delayacct_freepages_end(void)
213{
214 delayacct_end(¤t->delays->lock,
215 ¤t->delays->freepages_start,
216 ¤t->delays->freepages_delay,
217 ¤t->delays->freepages_count);
218}
219
220void __delayacct_thrashing_start(bool *in_thrashing)
221{
222 *in_thrashing = !!current->in_thrashing;
223 if (*in_thrashing)
224 return;
225
226 current->in_thrashing = 1;
227 current->delays->thrashing_start = local_clock();
228}
229
230void __delayacct_thrashing_end(bool *in_thrashing)
231{
232 if (*in_thrashing)
233 return;
234
235 current->in_thrashing = 0;
236 delayacct_end(¤t->delays->lock,
237 ¤t->delays->thrashing_start,
238 ¤t->delays->thrashing_delay,
239 ¤t->delays->thrashing_count);
240}
241
242void __delayacct_swapin_start(void)
243{
244 current->delays->swapin_start = local_clock();
245}
246
247void __delayacct_swapin_end(void)
248{
249 delayacct_end(¤t->delays->lock,
250 ¤t->delays->swapin_start,
251 ¤t->delays->swapin_delay,
252 ¤t->delays->swapin_count);
253}
254
255void __delayacct_compact_start(void)
256{
257 current->delays->compact_start = local_clock();
258}
259
260void __delayacct_compact_end(void)
261{
262 delayacct_end(¤t->delays->lock,
263 ¤t->delays->compact_start,
264 ¤t->delays->compact_delay,
265 ¤t->delays->compact_count);
266}
267
268void __delayacct_wpcopy_start(void)
269{
270 current->delays->wpcopy_start = local_clock();
271}
272
273void __delayacct_wpcopy_end(void)
274{
275 delayacct_end(¤t->delays->lock,
276 ¤t->delays->wpcopy_start,
277 ¤t->delays->wpcopy_delay,
278 ¤t->delays->wpcopy_count);
279}
280
281void __delayacct_irq(struct task_struct *task, u32 delta)
282{
283 unsigned long flags;
284
285 raw_spin_lock_irqsave(&task->delays->lock, flags);
286 task->delays->irq_delay += delta;
287 task->delays->irq_count++;
288 raw_spin_unlock_irqrestore(&task->delays->lock, flags);
289}
290