Commit | Line | Data |
---|---|---|
7170066e | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
ca74e92b SN |
2 | /* delayacct.c - per-task delay accounting |
3 | * | |
4 | * Copyright (C) Shailabh Nagar, IBM Corp. 2006 | |
ca74e92b SN |
5 | */ |
6 | ||
7 | #include <linux/sched.h> | |
9164bb4a | 8 | #include <linux/sched/task.h> |
32ef5517 | 9 | #include <linux/sched/cputime.h> |
4b7a08a0 | 10 | #include <linux/sched/clock.h> |
ca74e92b | 11 | #include <linux/slab.h> |
6952b61d | 12 | #include <linux/taskstats.h> |
ca74e92b SN |
13 | #include <linux/sysctl.h> |
14 | #include <linux/delayacct.h> | |
c9aaa895 | 15 | #include <linux/module.h> |
ca74e92b | 16 | |
e4042ad4 PZ |
17 | DEFINE_STATIC_KEY_FALSE(delayacct_key); |
18 | int delayacct_on __read_mostly; /* Delay accounting turned on/off */ | |
e18b890b | 19 | struct kmem_cache *delayacct_cache; |
ca74e92b | 20 | |
0cd7c741 PZ |
21 | static 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 | ||
e4042ad4 | 32 | static int __init delayacct_setup_enable(char *str) |
ca74e92b | 33 | { |
e4042ad4 | 34 | delayacct_on = 1; |
ca74e92b SN |
35 | return 1; |
36 | } | |
e4042ad4 | 37 | __setup("delayacct", delayacct_setup_enable); |
ca74e92b SN |
38 | |
39 | void delayacct_init(void) | |
40 | { | |
5d097056 | 41 | delayacct_cache = KMEM_CACHE(task_delay_info, SLAB_PANIC|SLAB_ACCOUNT); |
ca74e92b | 42 | delayacct_tsk_init(&init_task); |
0cd7c741 PZ |
43 | set_delayacct(delayacct_on); |
44 | } | |
45 | ||
46 | #ifdef CONFIG_PROC_SYSCTL | |
1186618a | 47 | static int sysctl_delayacct(struct ctl_table *table, int write, void *buffer, |
0cd7c741 PZ |
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; | |
ca74e92b | 65 | } |
1186618a | 66 | |
67 | static 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 | ||
80 | static __init int kernel_delayacct_sysctls_init(void) | |
81 | { | |
82 | register_sysctl_init("kernel", kern_delayacct_table); | |
83 | return 0; | |
84 | } | |
85 | late_initcall(kernel_delayacct_sysctls_init); | |
0cd7c741 | 86 | #endif |
ca74e92b SN |
87 | |
88 | void __delayacct_tsk_init(struct task_struct *tsk) | |
89 | { | |
e94b1766 | 90 | tsk->delays = kmem_cache_zalloc(delayacct_cache, GFP_KERNEL); |
ca74e92b | 91 | if (tsk->delays) |
02acc80d | 92 | raw_spin_lock_init(&tsk->delays->lock); |
ca74e92b SN |
93 | } |
94 | ||
ca74e92b | 95 | /* |
9667a23d TG |
96 | * Finish delay accounting for a statistic using its timestamps (@start), |
97 | * accumalator (@total) and @count | |
ca74e92b | 98 | */ |
4b7a08a0 | 99 | static void delayacct_end(raw_spinlock_t *lock, u64 *start, u64 *total, u32 *count) |
ca74e92b | 100 | { |
4b7a08a0 | 101 | s64 ns = local_clock() - *start; |
64efade1 | 102 | unsigned long flags; |
ca74e92b | 103 | |
9667a23d | 104 | if (ns > 0) { |
02acc80d | 105 | raw_spin_lock_irqsave(lock, flags); |
9667a23d TG |
106 | *total += ns; |
107 | (*count)++; | |
02acc80d | 108 | raw_spin_unlock_irqrestore(lock, flags); |
9667a23d | 109 | } |
ca74e92b SN |
110 | } |
111 | ||
0ff92245 SN |
112 | void __delayacct_blkio_start(void) |
113 | { | |
4b7a08a0 | 114 | current->delays->blkio_start = local_clock(); |
0ff92245 SN |
115 | } |
116 | ||
c96f5471 JS |
117 | /* |
118 | * We cannot rely on the `current` macro, as we haven't yet switched back to | |
119 | * the process being woken. | |
120 | */ | |
121 | void __delayacct_blkio_end(struct task_struct *p) | |
0ff92245 | 122 | { |
a3d5dc90 YY |
123 | delayacct_end(&p->delays->lock, |
124 | &p->delays->blkio_start, | |
125 | &p->delays->blkio_delay, | |
126 | &p->delays->blkio_count); | |
0ff92245 | 127 | } |
6f44993f | 128 | |
e4042ad4 | 129 | int delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk) |
6f44993f | 130 | { |
dbf3da1c | 131 | u64 utime, stime, stimescaled, utimescaled; |
68f6783d TG |
132 | unsigned long long t2, t3; |
133 | unsigned long flags, t1; | |
134 | s64 tmp; | |
6f44993f | 135 | |
dbf3da1c | 136 | task_cputime(tsk, &utime, &stime); |
68f6783d | 137 | tmp = (s64)d->cpu_run_real_total; |
dbf3da1c | 138 | tmp += utime + stime; |
6f44993f SN |
139 | d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp; |
140 | ||
dbf3da1c | 141 | task_cputime_scaled(tsk, &utimescaled, &stimescaled); |
68f6783d | 142 | tmp = (s64)d->cpu_scaled_run_real_total; |
dbf3da1c | 143 | tmp += utimescaled + stimescaled; |
c66f08be MN |
144 | d->cpu_scaled_run_real_total = |
145 | (tmp < (s64)d->cpu_scaled_run_real_total) ? 0 : tmp; | |
146 | ||
6f44993f SN |
147 | /* |
148 | * No locking available for sched_info (and too expensive to add one) | |
149 | * Mitigate by taking snapshot of values | |
150 | */ | |
2d72376b | 151 | t1 = tsk->sched_info.pcount; |
6f44993f | 152 | t2 = tsk->sched_info.run_delay; |
9c2c4802 | 153 | t3 = tsk->se.sum_exec_runtime; |
6f44993f SN |
154 | |
155 | d->cpu_count += t1; | |
156 | ||
172ba844 | 157 | tmp = (s64)d->cpu_delay_total + t2; |
6f44993f SN |
158 | d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp; |
159 | ||
172ba844 | 160 | tmp = (s64)d->cpu_run_virtual_total + t3; |
6f44993f SN |
161 | d->cpu_run_virtual_total = |
162 | (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp; | |
163 | ||
e4042ad4 PZ |
164 | if (!tsk->delays) |
165 | return 0; | |
166 | ||
6f44993f SN |
167 | /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */ |
168 | ||
02acc80d | 169 | raw_spin_lock_irqsave(&tsk->delays->lock, flags); |
6f44993f SN |
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; | |
016ae219 KK |
174 | tmp = d->freepages_delay_total + tsk->delays->freepages_delay; |
175 | d->freepages_delay_total = (tmp < d->freepages_delay_total) ? 0 : tmp; | |
b1d29ba8 JW |
176 | tmp = d->thrashing_delay_total + tsk->delays->thrashing_delay; |
177 | d->thrashing_delay_total = (tmp < d->thrashing_delay_total) ? 0 : tmp; | |
5bf18281 | 178 | tmp = d->compact_delay_total + tsk->delays->compact_delay; |
179 | d->compact_delay_total = (tmp < d->compact_delay_total) ? 0 : tmp; | |
662ce1dc YY |
180 | tmp = d->wpcopy_delay_total + tsk->delays->wpcopy_delay; |
181 | d->wpcopy_delay_total = (tmp < d->wpcopy_delay_total) ? 0 : tmp; | |
a3b2aeac YY |
182 | tmp = d->irq_delay_total + tsk->delays->irq_delay; |
183 | d->irq_delay_total = (tmp < d->irq_delay_total) ? 0 : tmp; | |
6f44993f SN |
184 | d->blkio_count += tsk->delays->blkio_count; |
185 | d->swapin_count += tsk->delays->swapin_count; | |
016ae219 | 186 | d->freepages_count += tsk->delays->freepages_count; |
b1d29ba8 | 187 | d->thrashing_count += tsk->delays->thrashing_count; |
5bf18281 | 188 | d->compact_count += tsk->delays->compact_count; |
662ce1dc | 189 | d->wpcopy_count += tsk->delays->wpcopy_count; |
a3b2aeac | 190 | d->irq_count += tsk->delays->irq_count; |
02acc80d | 191 | raw_spin_unlock_irqrestore(&tsk->delays->lock, flags); |
6f44993f | 192 | |
6f44993f SN |
193 | return 0; |
194 | } | |
25890454 SN |
195 | |
196 | __u64 __delayacct_blkio_ticks(struct task_struct *tsk) | |
197 | { | |
198 | __u64 ret; | |
64efade1 | 199 | unsigned long flags; |
25890454 | 200 | |
02acc80d | 201 | raw_spin_lock_irqsave(&tsk->delays->lock, flags); |
a3d5dc90 | 202 | ret = nsec_to_clock_t(tsk->delays->blkio_delay); |
02acc80d | 203 | raw_spin_unlock_irqrestore(&tsk->delays->lock, flags); |
25890454 SN |
204 | return ret; |
205 | } | |
206 | ||
873b4771 KK |
207 | void __delayacct_freepages_start(void) |
208 | { | |
4b7a08a0 | 209 | current->delays->freepages_start = local_clock(); |
873b4771 KK |
210 | } |
211 | ||
212 | void __delayacct_freepages_end(void) | |
213 | { | |
4b7a08a0 PZ |
214 | delayacct_end(¤t->delays->lock, |
215 | ¤t->delays->freepages_start, | |
216 | ¤t->delays->freepages_delay, | |
217 | ¤t->delays->freepages_count); | |
873b4771 KK |
218 | } |
219 | ||
aa1cf99b | 220 | void __delayacct_thrashing_start(bool *in_thrashing) |
b1d29ba8 | 221 | { |
aa1cf99b YY |
222 | *in_thrashing = !!current->in_thrashing; |
223 | if (*in_thrashing) | |
224 | return; | |
225 | ||
226 | current->in_thrashing = 1; | |
4b7a08a0 | 227 | current->delays->thrashing_start = local_clock(); |
b1d29ba8 JW |
228 | } |
229 | ||
aa1cf99b | 230 | void __delayacct_thrashing_end(bool *in_thrashing) |
b1d29ba8 | 231 | { |
aa1cf99b YY |
232 | if (*in_thrashing) |
233 | return; | |
234 | ||
235 | current->in_thrashing = 0; | |
b1d29ba8 JW |
236 | delayacct_end(¤t->delays->lock, |
237 | ¤t->delays->thrashing_start, | |
238 | ¤t->delays->thrashing_delay, | |
239 | ¤t->delays->thrashing_count); | |
240 | } | |
a3d5dc90 YY |
241 | |
242 | void __delayacct_swapin_start(void) | |
243 | { | |
244 | current->delays->swapin_start = local_clock(); | |
245 | } | |
246 | ||
247 | void __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 | } | |
5bf18281 | 254 | |
255 | void __delayacct_compact_start(void) | |
256 | { | |
257 | current->delays->compact_start = local_clock(); | |
258 | } | |
259 | ||
260 | void __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 | } | |
662ce1dc YY |
267 | |
268 | void __delayacct_wpcopy_start(void) | |
269 | { | |
270 | current->delays->wpcopy_start = local_clock(); | |
271 | } | |
272 | ||
273 | void __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 | } | |
a3b2aeac YY |
280 | |
281 | void __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 |