1// SPDX-License-Identifier: GPL-2.0
2/*
3 * /proc/schedstat implementation
4 */
5#include "sched.h"
6
7void __update_stats_wait_start(struct rq *rq, struct task_struct *p,
8 struct sched_statistics *stats)
9{
10 u64 wait_start, prev_wait_start;
11
12 wait_start = rq_clock(rq);
13 prev_wait_start = schedstat_val(stats->wait_start);
14
15 if (p && likely(wait_start > prev_wait_start))
16 wait_start -= prev_wait_start;
17
18 __schedstat_set(stats->wait_start, wait_start);
19}
20
21void __update_stats_wait_end(struct rq *rq, struct task_struct *p,
22 struct sched_statistics *stats)
23{
24 u64 delta = rq_clock(rq) - schedstat_val(stats->wait_start);
25
26 if (p) {
27 if (task_on_rq_migrating(p)) {
28 /*
29 * Preserve migrating task's wait time so wait_start
30 * time stamp can be adjusted to accumulate wait time
31 * prior to migration.
32 */
33 __schedstat_set(stats->wait_start, delta);
34
35 return;
36 }
37
38 trace_sched_stat_wait(tsk: p, delay: delta);
39 }
40
41 __schedstat_set(stats->wait_max,
42 max(schedstat_val(stats->wait_max), delta));
43 __schedstat_inc(stats->wait_count);
44 __schedstat_add(stats->wait_sum, delta);
45 __schedstat_set(stats->wait_start, 0);
46}
47
48void __update_stats_enqueue_sleeper(struct rq *rq, struct task_struct *p,
49 struct sched_statistics *stats)
50{
51 u64 sleep_start, block_start;
52
53 sleep_start = schedstat_val(stats->sleep_start);
54 block_start = schedstat_val(stats->block_start);
55
56 if (sleep_start) {
57 u64 delta = rq_clock(rq) - sleep_start;
58
59 if ((s64)delta < 0)
60 delta = 0;
61
62 if (unlikely(delta > schedstat_val(stats->sleep_max)))
63 __schedstat_set(stats->sleep_max, delta);
64
65 __schedstat_set(stats->sleep_start, 0);
66 __schedstat_add(stats->sum_sleep_runtime, delta);
67
68 if (p) {
69 account_scheduler_latency(task: p, usecs: delta >> 10, inter: 1);
70 trace_sched_stat_sleep(tsk: p, delay: delta);
71 }
72 }
73
74 if (block_start) {
75 u64 delta = rq_clock(rq) - block_start;
76
77 if ((s64)delta < 0)
78 delta = 0;
79
80 if (unlikely(delta > schedstat_val(stats->block_max)))
81 __schedstat_set(stats->block_max, delta);
82
83 __schedstat_set(stats->block_start, 0);
84 __schedstat_add(stats->sum_sleep_runtime, delta);
85 __schedstat_add(stats->sum_block_runtime, delta);
86
87 if (p) {
88 if (p->in_iowait) {
89 __schedstat_add(stats->iowait_sum, delta);
90 __schedstat_inc(stats->iowait_count);
91 trace_sched_stat_iowait(tsk: p, delay: delta);
92 }
93
94 trace_sched_stat_blocked(tsk: p, delay: delta);
95
96 account_scheduler_latency(task: p, usecs: delta >> 10, inter: 0);
97 }
98 }
99}
100
101/*
102 * Current schedstat API version.
103 *
104 * Bump this up when changing the output format or the meaning of an existing
105 * format, so that tools can adapt (or abort)
106 */
107#define SCHEDSTAT_VERSION 17
108
109static int show_schedstat(struct seq_file *seq, void *v)
110{
111 int cpu;
112
113 if (v == (void *)1) {
114 seq_printf(m: seq, fmt: "version %d\n", SCHEDSTAT_VERSION);
115 seq_printf(m: seq, fmt: "timestamp %lu\n", jiffies);
116 } else {
117 struct rq *rq;
118 struct sched_domain *sd;
119 int dcount = 0;
120 cpu = (unsigned long)(v - 2);
121 rq = cpu_rq(cpu);
122
123 /* runqueue-specific stats */
124 seq_printf(m: seq,
125 fmt: "cpu%d %u 0 %u %u %u %u %llu %llu %lu",
126 cpu, rq->yld_count,
127 rq->sched_count, rq->sched_goidle,
128 rq->ttwu_count, rq->ttwu_local,
129 rq->rq_cpu_time,
130 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcount);
131
132 seq_printf(m: seq, fmt: "\n");
133
134 /* domain-specific stats */
135 rcu_read_lock();
136 for_each_domain(cpu, sd) {
137 enum cpu_idle_type itype;
138
139 seq_printf(m: seq, fmt: "domain%d %s %*pb", dcount++, sd->name,
140 cpumask_pr_args(sched_domain_span(sd)));
141 for (itype = 0; itype < CPU_MAX_IDLE_TYPES; itype++) {
142 seq_printf(m: seq, fmt: " %u %u %u %u %u %u %u %u %u %u %u",
143 sd->lb_count[itype],
144 sd->lb_balanced[itype],
145 sd->lb_failed[itype],
146 sd->lb_imbalance_load[itype],
147 sd->lb_imbalance_util[itype],
148 sd->lb_imbalance_task[itype],
149 sd->lb_imbalance_misfit[itype],
150 sd->lb_gained[itype],
151 sd->lb_hot_gained[itype],
152 sd->lb_nobusyq[itype],
153 sd->lb_nobusyg[itype]);
154 }
155 seq_printf(m: seq,
156 fmt: " %u %u %u %u %u %u %u %u %u %u %u %u\n",
157 sd->alb_count, sd->alb_failed, sd->alb_pushed,
158 sd->sbe_count, sd->sbe_balanced, sd->sbe_pushed,
159 sd->sbf_count, sd->sbf_balanced, sd->sbf_pushed,
160 sd->ttwu_wake_remote, sd->ttwu_move_affine,
161 sd->ttwu_move_balance);
162 }
163 rcu_read_unlock();
164 }
165 return 0;
166}
167
168/*
169 * This iterator needs some explanation.
170 * It returns 1 for the header position.
171 * This means 2 is cpu 0.
172 * In a hotplugged system some CPUs, including cpu 0, may be missing so we have
173 * to use cpumask_* to iterate over the CPUs.
174 */
175static void *schedstat_start(struct seq_file *file, loff_t *offset)
176{
177 unsigned long n = *offset;
178
179 if (n == 0)
180 return (void *) 1;
181
182 n--;
183
184 if (n > 0)
185 n = cpumask_next(n: n - 1, cpu_online_mask);
186 else
187 n = cpumask_first(cpu_online_mask);
188
189 *offset = n + 1;
190
191 if (n < nr_cpu_ids)
192 return (void *)(unsigned long)(n + 2);
193
194 return NULL;
195}
196
197static void *schedstat_next(struct seq_file *file, void *data, loff_t *offset)
198{
199 (*offset)++;
200
201 return schedstat_start(file, offset);
202}
203
204static void schedstat_stop(struct seq_file *file, void *data)
205{
206}
207
208static const struct seq_operations schedstat_sops = {
209 .start = schedstat_start,
210 .next = schedstat_next,
211 .stop = schedstat_stop,
212 .show = show_schedstat,
213};
214
215static int __init proc_schedstat_init(void)
216{
217 proc_create_seq("schedstat", 0, NULL, &schedstat_sops);
218 return 0;
219}
220subsys_initcall(proc_schedstat_init);
221