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 static 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 67 static struct ctl_table kern_delayacct_table[] = { 68 { 69 .procname = "task_delayacct", 70 .data = NULL, 71 .maxlen = sizeof(unsigned int), 72 .mode = 0644, 73 .proc_handler = sysctl_delayacct, 74 .extra1 = SYSCTL_ZERO, 75 .extra2 = SYSCTL_ONE, 76 }, 77 { } 78 }; 79 80 static __init int kernel_delayacct_sysctls_init(void) 81 { 82 register_sysctl_init("kernel", kern_delayacct_table); 83 return 0; 84 } 85 late_initcall(kernel_delayacct_sysctls_init); 86 #endif 87 88 void __delayacct_tsk_init(struct task_struct *tsk) 89 { 90 tsk->delays = kmem_cache_zalloc(delayacct_cache, GFP_KERNEL); 91 if (tsk->delays) 92 raw_spin_lock_init(&tsk->delays->lock); 93 } 94 95 /* 96 * Finish delay accounting for a statistic using its timestamps (@start), 97 * accumalator (@total) and @count 98 */ 99 static void delayacct_end(raw_spinlock_t *lock, u64 *start, u64 *total, u32 *count) 100 { 101 s64 ns = local_clock() - *start; 102 unsigned long flags; 103 104 if (ns > 0) { 105 raw_spin_lock_irqsave(lock, flags); 106 *total += ns; 107 (*count)++; 108 raw_spin_unlock_irqrestore(lock, flags); 109 } 110 } 111 112 void __delayacct_blkio_start(void) 113 { 114 current->delays->blkio_start = local_clock(); 115 } 116 117 /* 118 * We cannot rely on the `current` macro, as we haven't yet switched back to 119 * the process being woken. 120 */ 121 void __delayacct_blkio_end(struct task_struct *p) 122 { 123 delayacct_end(&p->delays->lock, 124 &p->delays->blkio_start, 125 &p->delays->blkio_delay, 126 &p->delays->blkio_count); 127 } 128 129 int delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk) 130 { 131 u64 utime, stime, stimescaled, utimescaled; 132 unsigned long long t2, t3; 133 unsigned long flags, t1; 134 s64 tmp; 135 136 task_cputime(tsk, &utime, &stime); 137 tmp = (s64)d->cpu_run_real_total; 138 tmp += utime + stime; 139 d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp; 140 141 task_cputime_scaled(tsk, &utimescaled, &stimescaled); 142 tmp = (s64)d->cpu_scaled_run_real_total; 143 tmp += utimescaled + stimescaled; 144 d->cpu_scaled_run_real_total = 145 (tmp < (s64)d->cpu_scaled_run_real_total) ? 0 : tmp; 146 147 /* 148 * No locking available for sched_info (and too expensive to add one) 149 * Mitigate by taking snapshot of values 150 */ 151 t1 = tsk->sched_info.pcount; 152 t2 = tsk->sched_info.run_delay; 153 t3 = tsk->se.sum_exec_runtime; 154 155 d->cpu_count += t1; 156 157 tmp = (s64)d->cpu_delay_total + t2; 158 d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp; 159 160 tmp = (s64)d->cpu_run_virtual_total + t3; 161 d->cpu_run_virtual_total = 162 (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp; 163 164 if (!tsk->delays) 165 return 0; 166 167 /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */ 168 169 raw_spin_lock_irqsave(&tsk->delays->lock, flags); 170 tmp = d->blkio_delay_total + tsk->delays->blkio_delay; 171 d->blkio_delay_total = (tmp < d->blkio_delay_total) ? 0 : tmp; 172 tmp = d->swapin_delay_total + tsk->delays->swapin_delay; 173 d->swapin_delay_total = (tmp < d->swapin_delay_total) ? 0 : tmp; 174 tmp = d->freepages_delay_total + tsk->delays->freepages_delay; 175 d->freepages_delay_total = (tmp < d->freepages_delay_total) ? 0 : tmp; 176 tmp = d->thrashing_delay_total + tsk->delays->thrashing_delay; 177 d->thrashing_delay_total = (tmp < d->thrashing_delay_total) ? 0 : tmp; 178 tmp = d->compact_delay_total + tsk->delays->compact_delay; 179 d->compact_delay_total = (tmp < d->compact_delay_total) ? 0 : tmp; 180 tmp = d->wpcopy_delay_total + tsk->delays->wpcopy_delay; 181 d->wpcopy_delay_total = (tmp < d->wpcopy_delay_total) ? 0 : tmp; 182 d->blkio_count += tsk->delays->blkio_count; 183 d->swapin_count += tsk->delays->swapin_count; 184 d->freepages_count += tsk->delays->freepages_count; 185 d->thrashing_count += tsk->delays->thrashing_count; 186 d->compact_count += tsk->delays->compact_count; 187 d->wpcopy_count += tsk->delays->wpcopy_count; 188 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags); 189 190 return 0; 191 } 192 193 __u64 __delayacct_blkio_ticks(struct task_struct *tsk) 194 { 195 __u64 ret; 196 unsigned long flags; 197 198 raw_spin_lock_irqsave(&tsk->delays->lock, flags); 199 ret = nsec_to_clock_t(tsk->delays->blkio_delay); 200 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags); 201 return ret; 202 } 203 204 void __delayacct_freepages_start(void) 205 { 206 current->delays->freepages_start = local_clock(); 207 } 208 209 void __delayacct_freepages_end(void) 210 { 211 delayacct_end(¤t->delays->lock, 212 ¤t->delays->freepages_start, 213 ¤t->delays->freepages_delay, 214 ¤t->delays->freepages_count); 215 } 216 217 void __delayacct_thrashing_start(bool *in_thrashing) 218 { 219 *in_thrashing = !!current->in_thrashing; 220 if (*in_thrashing) 221 return; 222 223 current->in_thrashing = 1; 224 current->delays->thrashing_start = local_clock(); 225 } 226 227 void __delayacct_thrashing_end(bool *in_thrashing) 228 { 229 if (*in_thrashing) 230 return; 231 232 current->in_thrashing = 0; 233 delayacct_end(¤t->delays->lock, 234 ¤t->delays->thrashing_start, 235 ¤t->delays->thrashing_delay, 236 ¤t->delays->thrashing_count); 237 } 238 239 void __delayacct_swapin_start(void) 240 { 241 current->delays->swapin_start = local_clock(); 242 } 243 244 void __delayacct_swapin_end(void) 245 { 246 delayacct_end(¤t->delays->lock, 247 ¤t->delays->swapin_start, 248 ¤t->delays->swapin_delay, 249 ¤t->delays->swapin_count); 250 } 251 252 void __delayacct_compact_start(void) 253 { 254 current->delays->compact_start = local_clock(); 255 } 256 257 void __delayacct_compact_end(void) 258 { 259 delayacct_end(¤t->delays->lock, 260 ¤t->delays->compact_start, 261 ¤t->delays->compact_delay, 262 ¤t->delays->compact_count); 263 } 264 265 void __delayacct_wpcopy_start(void) 266 { 267 current->delays->wpcopy_start = local_clock(); 268 } 269 270 void __delayacct_wpcopy_end(void) 271 { 272 delayacct_end(¤t->delays->lock, 273 ¤t->delays->wpcopy_start, 274 ¤t->delays->wpcopy_delay, 275 ¤t->delays->wpcopy_count); 276 } 277