kernel/sched/stats.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * /proc/schedstat implementation
   4  */
   5 #include "sched.h"
   6
   7 void __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
  21 void __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(p, 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
  48 void __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(p, delta >> 10, 1);
  70                         trace_sched_stat_sleep(p, 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(p, delta);
  92                         }
  93
  94                         trace_sched_stat_blocked(p, delta);
  95
  96                         /*
  97                          * Blocking time is in units of nanosecs, so shift by
  98                          * 20 to get a milliseconds-range estimation of the
  99                          * amount of time that the task spent sleeping:
 100                          */
 101                         if (unlikely(prof_on == SLEEP_PROFILING)) {
 102                                 profile_hits(SLEEP_PROFILING,
 103                                              (void *)get_wchan(p),
 104                                              delta >> 20);
 105                         }
 106                         account_scheduler_latency(p, delta >> 10, 0);
 107                 }
 108         }
 109 }
 110
 111 /*
 112  * Current schedstat API version.
 113  *
 114  * Bump this up when changing the output format or the meaning of an existing
 115  * format, so that tools can adapt (or abort)
 116  */
 117 #define SCHEDSTAT_VERSION 15
 118
 119 static int show_schedstat(struct seq_file *seq, void *v)
 120 {
 121         int cpu;
 122
 123         if (v == (void *)1) {
 124                 seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
 125                 seq_printf(seq, "timestamp %lu\n", jiffies);
 126         } else {
 127                 struct rq *rq;
 128 #ifdef CONFIG_SMP
 129                 struct sched_domain *sd;
 130                 int dcount = 0;
 131 #endif
 132                 cpu = (unsigned long)(v - 2);
 133                 rq = cpu_rq(cpu);
 134
 135                 /* runqueue-specific stats */
 136                 seq_printf(seq,
 137                     "cpu%d %u 0 %u %u %u %u %llu %llu %lu",
 138                     cpu, rq->yld_count,
 139                     rq->sched_count, rq->sched_goidle,
 140                     rq->ttwu_count, rq->ttwu_local,
 141                     rq->rq_cpu_time,
 142                     rq->rq_sched_info.run_delay, rq->rq_sched_info.pcount);
 143
 144                 seq_printf(seq, "\n");
 145
 146 #ifdef CONFIG_SMP
 147                 /* domain-specific stats */
 148                 rcu_read_lock();
 149                 for_each_domain(cpu, sd) {
 150                         enum cpu_idle_type itype;
 151
 152                         seq_printf(seq, "domain%d %*pb", dcount++,
 153                                    cpumask_pr_args(sched_domain_span(sd)));
 154                         for (itype = CPU_IDLE; itype < CPU_MAX_IDLE_TYPES;
 155                                         itype++) {
 156                                 seq_printf(seq, " %u %u %u %u %u %u %u %u",
 157                                     sd->lb_count[itype],
 158                                     sd->lb_balanced[itype],
 159                                     sd->lb_failed[itype],
 160                                     sd->lb_imbalance[itype],
 161                                     sd->lb_gained[itype],
 162                                     sd->lb_hot_gained[itype],
 163                                     sd->lb_nobusyq[itype],
 164                                     sd->lb_nobusyg[itype]);
 165                         }
 166                         seq_printf(seq,
 167                                    " %u %u %u %u %u %u %u %u %u %u %u %u\n",
 168                             sd->alb_count, sd->alb_failed, sd->alb_pushed,
 169                             sd->sbe_count, sd->sbe_balanced, sd->sbe_pushed,
 170                             sd->sbf_count, sd->sbf_balanced, sd->sbf_pushed,
 171                             sd->ttwu_wake_remote, sd->ttwu_move_affine,
 172                             sd->ttwu_move_balance);
 173                 }
 174                 rcu_read_unlock();
 175 #endif
 176         }
 177         return 0;
 178 }
 179
 180 /*
 181  * This iterator needs some explanation.
 182  * It returns 1 for the header position.
 183  * This means 2 is cpu 0.
 184  * In a hotplugged system some CPUs, including cpu 0, may be missing so we have
 185  * to use cpumask_* to iterate over the CPUs.
 186  */
 187 static void *schedstat_start(struct seq_file *file, loff_t *offset)
 188 {
 189         unsigned long n = *offset;
 190
 191         if (n == 0)
 192                 return (void *) 1;
 193
 194         n--;
 195
 196         if (n > 0)
 197                 n = cpumask_next(n - 1, cpu_online_mask);
 198         else
 199                 n = cpumask_first(cpu_online_mask);
 200
 201         *offset = n + 1;
 202
 203         if (n < nr_cpu_ids)
 204                 return (void *)(unsigned long)(n + 2);
 205
 206         return NULL;
 207 }
 208
 209 static void *schedstat_next(struct seq_file *file, void *data, loff_t *offset)
 210 {
 211         (*offset)++;
 212
 213         return schedstat_start(file, offset);
 214 }
 215
 216 static void schedstat_stop(struct seq_file *file, void *data)
 217 {
 218 }
 219
 220 static const struct seq_operations schedstat_sops = {
 221         .start = schedstat_start,
 222         .next  = schedstat_next,
 223         .stop  = schedstat_stop,
 224         .show  = show_schedstat,
 225 };
 226
 227 static int __init proc_schedstat_init(void)
 228 {
 229         proc_create_seq("schedstat", 0, NULL, &schedstat_sops);
 230         return 0;
 231 }
 232 subsys_initcall(proc_schedstat_init);