kernel/delayacct.c

   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
  17 DEFINE_STATIC_KEY_FALSE(delayacct_key);
  18 int delayacct_on __read_mostly; /* Delay accounting turned on/off */
  19 struct kmem_cache *delayacct_cache;
  20
  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
  32 static int __init delayacct_setup_enable(char *str)
  33 {
  34         delayacct_on = 1;
  35         return 1;
  36 }
  37 __setup("delayacct", delayacct_setup_enable);
  38
  39 void 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
  47 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 #endif
  67
  68 void __delayacct_tsk_init(struct task_struct *tsk)
  69 {
  70         tsk->delays = kmem_cache_zalloc(delayacct_cache, GFP_KERNEL);
  71         if (tsk->delays)
  72                 raw_spin_lock_init(&tsk->delays->lock);
  73 }
  74
  75 /*
  76  * Finish delay accounting for a statistic using its timestamps (@start),
  77  * accumalator (@total) and @count
  78  */
  79 static void delayacct_end(raw_spinlock_t *lock, u64 *start, u64 *total, u32 *count)
  80 {
  81         s64 ns = local_clock() - *start;
  82         unsigned long flags;
  83
  84         if (ns > 0) {
  85                 raw_spin_lock_irqsave(lock, flags);
  86                 *total += ns;
  87                 (*count)++;
  88                 raw_spin_unlock_irqrestore(lock, flags);
  89         }
  90 }
  91
  92 void __delayacct_blkio_start(void)
  93 {
  94         current->delays->blkio_start = local_clock();
  95 }
  96
  97 /*
  98  * We cannot rely on the `current` macro, as we haven't yet switched back to
  99  * the process being woken.
 100  */
 101 void __delayacct_blkio_end(struct task_struct *p)
 102 {
 103         delayacct_end(&p->delays->lock,
 104                       &p->delays->blkio_start,
 105                       &p->delays->blkio_delay,
 106                       &p->delays->blkio_count);
 107 }
 108
 109 int delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk)
 110 {
 111         u64 utime, stime, stimescaled, utimescaled;
 112         unsigned long long t2, t3;
 113         unsigned long flags, t1;
 114         s64 tmp;
 115
 116         task_cputime(tsk, &utime, &stime);
 117         tmp = (s64)d->cpu_run_real_total;
 118         tmp += utime + stime;
 119         d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp;
 120
 121         task_cputime_scaled(tsk, &utimescaled, &stimescaled);
 122         tmp = (s64)d->cpu_scaled_run_real_total;
 123         tmp += utimescaled + stimescaled;
 124         d->cpu_scaled_run_real_total =
 125                 (tmp < (s64)d->cpu_scaled_run_real_total) ? 0 : tmp;
 126
 127         /*
 128          * No locking available for sched_info (and too expensive to add one)
 129          * Mitigate by taking snapshot of values
 130          */
 131         t1 = tsk->sched_info.pcount;
 132         t2 = tsk->sched_info.run_delay;
 133         t3 = tsk->se.sum_exec_runtime;
 134
 135         d->cpu_count += t1;
 136
 137         tmp = (s64)d->cpu_delay_total + t2;
 138         d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp;
 139
 140         tmp = (s64)d->cpu_run_virtual_total + t3;
 141         d->cpu_run_virtual_total =
 142                 (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp;
 143
 144         if (!tsk->delays)
 145                 return 0;
 146
 147         /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */
 148
 149         raw_spin_lock_irqsave(&tsk->delays->lock, flags);
 150         tmp = d->blkio_delay_total + tsk->delays->blkio_delay;
 151         d->blkio_delay_total = (tmp < d->blkio_delay_total) ? 0 : tmp;
 152         tmp = d->swapin_delay_total + tsk->delays->swapin_delay;
 153         d->swapin_delay_total = (tmp < d->swapin_delay_total) ? 0 : tmp;
 154         tmp = d->freepages_delay_total + tsk->delays->freepages_delay;
 155         d->freepages_delay_total = (tmp < d->freepages_delay_total) ? 0 : tmp;
 156         tmp = d->thrashing_delay_total + tsk->delays->thrashing_delay;
 157         d->thrashing_delay_total = (tmp < d->thrashing_delay_total) ? 0 : tmp;
 158         tmp = d->compact_delay_total + tsk->delays->compact_delay;
 159         d->compact_delay_total = (tmp < d->compact_delay_total) ? 0 : tmp;
 160         d->blkio_count += tsk->delays->blkio_count;
 161         d->swapin_count += tsk->delays->swapin_count;
 162         d->freepages_count += tsk->delays->freepages_count;
 163         d->thrashing_count += tsk->delays->thrashing_count;
 164         d->compact_count += tsk->delays->compact_count;
 165         raw_spin_unlock_irqrestore(&tsk->delays->lock, flags);
 166
 167         return 0;
 168 }
 169
 170 __u64 __delayacct_blkio_ticks(struct task_struct *tsk)
 171 {
 172         __u64 ret;
 173         unsigned long flags;
 174
 175         raw_spin_lock_irqsave(&tsk->delays->lock, flags);
 176         ret = nsec_to_clock_t(tsk->delays->blkio_delay);
 177         raw_spin_unlock_irqrestore(&tsk->delays->lock, flags);
 178         return ret;
 179 }
 180
 181 void __delayacct_freepages_start(void)
 182 {
 183         current->delays->freepages_start = local_clock();
 184 }
 185
 186 void __delayacct_freepages_end(void)
 187 {
 188         delayacct_end(&current->delays->lock,
 189                       &current->delays->freepages_start,
 190                       &current->delays->freepages_delay,
 191                       &current->delays->freepages_count);
 192 }
 193
 194 void __delayacct_thrashing_start(void)
 195 {
 196         current->delays->thrashing_start = local_clock();
 197 }
 198
 199 void __delayacct_thrashing_end(void)
 200 {
 201         delayacct_end(&current->delays->lock,
 202                       &current->delays->thrashing_start,
 203                       &current->delays->thrashing_delay,
 204                       &current->delays->thrashing_count);
 205 }
 206
 207 void __delayacct_swapin_start(void)
 208 {
 209         current->delays->swapin_start = local_clock();
 210 }
 211
 212 void __delayacct_swapin_end(void)
 213 {
 214         delayacct_end(&current->delays->lock,
 215                       &current->delays->swapin_start,
 216                       &current->delays->swapin_delay,
 217                       &current->delays->swapin_count);
 218 }
 219
 220 void __delayacct_compact_start(void)
 221 {
 222         current->delays->compact_start = local_clock();
 223 }
 224
 225 void __delayacct_compact_end(void)
 226 {
 227         delayacct_end(&current->delays->lock,
 228                       &current->delays->compact_start,
 229                       &current->delays->compact_delay,
 230                       &current->delays->compact_count);
 231 }