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 struct task_delay_info *delays = p->delays; 104 u64 *total; 105 u32 *count; 106 107 if (p->delays->flags & DELAYACCT_PF_SWAPIN) { 108 total = &delays->swapin_delay; 109 count = &delays->swapin_count; 110 } else { 111 total = &delays->blkio_delay; 112 count = &delays->blkio_count; 113 } 114 115 delayacct_end(&delays->lock, &delays->blkio_start, total, count); 116 } 117 118 int delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk) 119 { 120 u64 utime, stime, stimescaled, utimescaled; 121 unsigned long long t2, t3; 122 unsigned long flags, t1; 123 s64 tmp; 124 125 task_cputime(tsk, &utime, &stime); 126 tmp = (s64)d->cpu_run_real_total; 127 tmp += utime + stime; 128 d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp; 129 130 task_cputime_scaled(tsk, &utimescaled, &stimescaled); 131 tmp = (s64)d->cpu_scaled_run_real_total; 132 tmp += utimescaled + stimescaled; 133 d->cpu_scaled_run_real_total = 134 (tmp < (s64)d->cpu_scaled_run_real_total) ? 0 : tmp; 135 136 /* 137 * No locking available for sched_info (and too expensive to add one) 138 * Mitigate by taking snapshot of values 139 */ 140 t1 = tsk->sched_info.pcount; 141 t2 = tsk->sched_info.run_delay; 142 t3 = tsk->se.sum_exec_runtime; 143 144 d->cpu_count += t1; 145 146 tmp = (s64)d->cpu_delay_total + t2; 147 d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp; 148 149 tmp = (s64)d->cpu_run_virtual_total + t3; 150 d->cpu_run_virtual_total = 151 (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp; 152 153 if (!tsk->delays) 154 return 0; 155 156 /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */ 157 158 raw_spin_lock_irqsave(&tsk->delays->lock, flags); 159 tmp = d->blkio_delay_total + tsk->delays->blkio_delay; 160 d->blkio_delay_total = (tmp < d->blkio_delay_total) ? 0 : tmp; 161 tmp = d->swapin_delay_total + tsk->delays->swapin_delay; 162 d->swapin_delay_total = (tmp < d->swapin_delay_total) ? 0 : tmp; 163 tmp = d->freepages_delay_total + tsk->delays->freepages_delay; 164 d->freepages_delay_total = (tmp < d->freepages_delay_total) ? 0 : tmp; 165 tmp = d->thrashing_delay_total + tsk->delays->thrashing_delay; 166 d->thrashing_delay_total = (tmp < d->thrashing_delay_total) ? 0 : tmp; 167 d->blkio_count += tsk->delays->blkio_count; 168 d->swapin_count += tsk->delays->swapin_count; 169 d->freepages_count += tsk->delays->freepages_count; 170 d->thrashing_count += tsk->delays->thrashing_count; 171 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags); 172 173 return 0; 174 } 175 176 __u64 __delayacct_blkio_ticks(struct task_struct *tsk) 177 { 178 __u64 ret; 179 unsigned long flags; 180 181 raw_spin_lock_irqsave(&tsk->delays->lock, flags); 182 ret = nsec_to_clock_t(tsk->delays->blkio_delay + 183 tsk->delays->swapin_delay); 184 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags); 185 return ret; 186 } 187 188 void __delayacct_freepages_start(void) 189 { 190 current->delays->freepages_start = local_clock(); 191 } 192 193 void __delayacct_freepages_end(void) 194 { 195 delayacct_end(¤t->delays->lock, 196 ¤t->delays->freepages_start, 197 ¤t->delays->freepages_delay, 198 ¤t->delays->freepages_count); 199 } 200 201 void __delayacct_thrashing_start(void) 202 { 203 current->delays->thrashing_start = local_clock(); 204 } 205 206 void __delayacct_thrashing_end(void) 207 { 208 delayacct_end(¤t->delays->lock, 209 ¤t->delays->thrashing_start, 210 ¤t->delays->thrashing_delay, 211 ¤t->delays->thrashing_count); 212 } 213