1// SPDX-License-Identifier: GPL-2.0
2/*
3 * /proc/schedstat implementation
4 */
5
6void __update_stats_wait_start(struct rq *rq, struct task_struct *p,
7 struct sched_statistics *stats)
8{
9 u64 wait_start, prev_wait_start;
10
11 wait_start = rq_clock(rq);
12 prev_wait_start = schedstat_val(stats->wait_start);
13
14 if (p && likely(wait_start > prev_wait_start))
15 wait_start -= prev_wait_start;
16
17 __schedstat_set(stats->wait_start, wait_start);
18}
19
20void __update_stats_wait_end(struct rq *rq, struct task_struct *p,
21 struct sched_statistics *stats)
22{
23 u64 delta = rq_clock(rq) - schedstat_val(stats->wait_start);
24
25 if (p) {
26 if (task_on_rq_migrating(p)) {
27 /*
28 * Preserve migrating task's wait time so wait_start
29 * time stamp can be adjusted to accumulate wait time
30 * prior to migration.
31 */
32 __schedstat_set(stats->wait_start, delta);
33
34 return;
35 }
36
37 trace_sched_stat_wait(tsk: p, delay: delta);
38 }
39
40 __schedstat_set(stats->wait_max,
41 max(schedstat_val(stats->wait_max), delta));
42 __schedstat_inc(stats->wait_count);
43 __schedstat_add(stats->wait_sum, delta);
44 __schedstat_set(stats->wait_start, 0);
45}
46
47void __update_stats_enqueue_sleeper(struct rq *rq, struct task_struct *p,
48 struct sched_statistics *stats)
49{
50 u64 sleep_start, block_start;
51
52 sleep_start = schedstat_val(stats->sleep_start);
53 block_start = schedstat_val(stats->block_start);
54
55 if (sleep_start) {
56 u64 delta = rq_clock(rq) - sleep_start;
57
58 if ((s64)delta < 0)
59 delta = 0;
60
61 if (unlikely(delta > schedstat_val(stats->sleep_max)))
62 __schedstat_set(stats->sleep_max, delta);
63
64 __schedstat_set(stats->sleep_start, 0);
65 __schedstat_add(stats->sum_sleep_runtime, delta);
66
67 if (p) {
68 account_scheduler_latency(task: p, usecs: delta >> 10, inter: 1);
69 trace_sched_stat_sleep(tsk: p, delay: delta);
70 }
71 }
72
73 if (block_start) {
74 u64 delta = rq_clock(rq) - block_start;
75
76 if ((s64)delta < 0)
77 delta = 0;
78
79 if (unlikely(delta > schedstat_val(stats->block_max)))
80 __schedstat_set(stats->block_max, delta);
81
82 __schedstat_set(stats->block_start, 0);
83 __schedstat_add(stats->sum_sleep_runtime, delta);
84 __schedstat_add(stats->sum_block_runtime, delta);
85
86 if (p) {
87 if (p->in_iowait) {
88 __schedstat_add(stats->iowait_sum, delta);
89 __schedstat_inc(stats->iowait_count);
90 trace_sched_stat_iowait(tsk: p, delay: delta);
91 }
92
93 trace_sched_stat_blocked(tsk: p, delay: delta);
94
95 /*
96 * Blocking time is in units of nanosecs, so shift by
97 * 20 to get a milliseconds-range estimation of the
98 * amount of time that the task spent sleeping:
99 */
100 if (unlikely(prof_on == SLEEP_PROFILING)) {
101 profile_hits(SLEEP_PROFILING,
102 ip: (void *)get_wchan(p),
103 nr_hits: delta >> 20);
104 }
105 account_scheduler_latency(task: p, usecs: delta >> 10, inter: 0);
106 }
107 }
108}
109
110/*
111 * Current schedstat API version.
112 *
113 * Bump this up when changing the output format or the meaning of an existing
114 * format, so that tools can adapt (or abort)
115 */
116#define SCHEDSTAT_VERSION 15
117
118static int show_schedstat(struct seq_file *seq, void *v)
119{
120 int cpu;
121
122 if (v == (void *)1) {
123 seq_printf(m: seq, fmt: "version %d\n", SCHEDSTAT_VERSION);
124 seq_printf(m: seq, fmt: "timestamp %lu\n", jiffies);
125 } else {
126 struct rq *rq;
127#ifdef CONFIG_SMP
128 struct sched_domain *sd;
129 int dcount = 0;
130#endif
131 cpu = (unsigned long)(v - 2);
132 rq = cpu_rq(cpu);
133
134 /* runqueue-specific stats */
135 seq_printf(m: seq,
136 fmt: "cpu%d %u 0 %u %u %u %u %llu %llu %lu",
137 cpu, rq->yld_count,
138 rq->sched_count, rq->sched_goidle,
139 rq->ttwu_count, rq->ttwu_local,
140 rq->rq_cpu_time,
141 rq->rq_sched_info.run_delay, rq->rq_sched_info.pcount);
142
143 seq_printf(m: seq, fmt: "\n");
144
145#ifdef CONFIG_SMP
146 /* domain-specific stats */
147 rcu_read_lock();
148 for_each_domain(cpu, sd) {
149 enum cpu_idle_type itype;
150
151 seq_printf(m: seq, fmt: "domain%d %*pb", dcount++,
152 cpumask_pr_args(sched_domain_span(sd)));
153 for (itype = CPU_IDLE; itype < CPU_MAX_IDLE_TYPES;
154 itype++) {
155 seq_printf(m: seq, fmt: " %u %u %u %u %u %u %u %u",
156 sd->lb_count[itype],
157 sd->lb_balanced[itype],
158 sd->lb_failed[itype],
159 sd->lb_imbalance[itype],
160 sd->lb_gained[itype],
161 sd->lb_hot_gained[itype],
162 sd->lb_nobusyq[itype],
163 sd->lb_nobusyg[itype]);
164 }
165 seq_printf(m: seq,
166 fmt: " %u %u %u %u %u %u %u %u %u %u %u %u\n",
167 sd->alb_count, sd->alb_failed, sd->alb_pushed,
168 sd->sbe_count, sd->sbe_balanced, sd->sbe_pushed,
169 sd->sbf_count, sd->sbf_balanced, sd->sbf_pushed,
170 sd->ttwu_wake_remote, sd->ttwu_move_affine,
171 sd->ttwu_move_balance);
172 }
173 rcu_read_unlock();
174#endif
175 }
176 return 0;
177}
178
179/*
180 * This iterator needs some explanation.
181 * It returns 1 for the header position.
182 * This means 2 is cpu 0.
183 * In a hotplugged system some CPUs, including cpu 0, may be missing so we have
184 * to use cpumask_* to iterate over the CPUs.
185 */
186static void *schedstat_start(struct seq_file *file, loff_t *offset)
187{
188 unsigned long n = *offset;
189
190 if (n == 0)
191 return (void *) 1;
192
193 n--;
194
195 if (n > 0)
196 n = cpumask_next(n: n - 1, cpu_online_mask);
197 else
198 n = cpumask_first(cpu_online_mask);
199
200 *offset = n + 1;
201
202 if (n < nr_cpu_ids)
203 return (void *)(unsigned long)(n + 2);
204
205 return NULL;
206}
207
208static void *schedstat_next(struct seq_file *file, void *data, loff_t *offset)
209{
210 (*offset)++;
211
212 return schedstat_start(file, offset);
213}
214
215static void schedstat_stop(struct seq_file *file, void *data)
216{
217}
218
219static const struct seq_operations schedstat_sops = {
220 .start = schedstat_start,
221 .next = schedstat_next,
222 .stop = schedstat_stop,
223 .show = show_schedstat,
224};
225
226static int __init proc_schedstat_init(void)
227{
228 proc_create_seq("schedstat", 0, NULL, &schedstat_sops);
229 return 0;
230}
231subsys_initcall(proc_schedstat_init);
232

source code of linux/kernel/sched/stats.c