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 UPDATE_DELAY(type) \ 18 do { \ 19 d->type##_delay_max = tsk->delays->type##_delay_max; \ 20 d->type##_delay_min = tsk->delays->type##_delay_min; \ 21 tmp = d->type##_delay_total + tsk->delays->type##_delay; \ 22 d->type##_delay_total = (tmp < d->type##_delay_total) ? 0 : tmp; \ 23 d->type##_count += tsk->delays->type##_count; \ 24 } while (0) 25 26 DEFINE_STATIC_KEY_FALSE(delayacct_key); 27 int delayacct_on __read_mostly; /* Delay accounting turned on/off */ 28 struct kmem_cache *delayacct_cache; 29 30 static void set_delayacct(bool enabled) 31 { 32 if (enabled) { 33 static_branch_enable(&delayacct_key); 34 delayacct_on = 1; 35 } else { 36 delayacct_on = 0; 37 static_branch_disable(&delayacct_key); 38 } 39 } 40 41 static int __init delayacct_setup_enable(char *str) 42 { 43 delayacct_on = 1; 44 return 1; 45 } 46 __setup("delayacct", delayacct_setup_enable); 47 48 void delayacct_init(void) 49 { 50 delayacct_cache = KMEM_CACHE(task_delay_info, SLAB_PANIC|SLAB_ACCOUNT); 51 delayacct_tsk_init(&init_task); 52 set_delayacct(delayacct_on); 53 } 54 55 #ifdef CONFIG_PROC_SYSCTL 56 static int sysctl_delayacct(const struct ctl_table *table, int write, void *buffer, 57 size_t *lenp, loff_t *ppos) 58 { 59 int state = delayacct_on; 60 struct ctl_table t; 61 int err; 62 63 if (write && !capable(CAP_SYS_ADMIN)) 64 return -EPERM; 65 66 t = *table; 67 t.data = &state; 68 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 69 if (err < 0) 70 return err; 71 if (write) 72 set_delayacct(state); 73 return err; 74 } 75 76 static const struct ctl_table kern_delayacct_table[] = { 77 { 78 .procname = "task_delayacct", 79 .data = NULL, 80 .maxlen = sizeof(unsigned int), 81 .mode = 0644, 82 .proc_handler = sysctl_delayacct, 83 .extra1 = SYSCTL_ZERO, 84 .extra2 = SYSCTL_ONE, 85 }, 86 }; 87 88 static __init int kernel_delayacct_sysctls_init(void) 89 { 90 register_sysctl_init("kernel", kern_delayacct_table); 91 return 0; 92 } 93 late_initcall(kernel_delayacct_sysctls_init); 94 #endif 95 96 void __delayacct_tsk_init(struct task_struct *tsk) 97 { 98 tsk->delays = kmem_cache_zalloc(delayacct_cache, GFP_KERNEL); 99 if (tsk->delays) 100 raw_spin_lock_init(&tsk->delays->lock); 101 } 102 103 /* 104 * Finish delay accounting for a statistic using its timestamps (@start), 105 * accumulator (@total) and @count 106 */ 107 static void delayacct_end(raw_spinlock_t *lock, u64 *start, u64 *total, u32 *count, u64 *max, u64 *min) 108 { 109 s64 ns = local_clock() - *start; 110 unsigned long flags; 111 112 if (ns > 0) { 113 raw_spin_lock_irqsave(lock, flags); 114 *total += ns; 115 (*count)++; 116 if (ns > *max) 117 *max = ns; 118 if (*min == 0 || ns < *min) 119 *min = ns; 120 raw_spin_unlock_irqrestore(lock, flags); 121 } 122 } 123 124 void __delayacct_blkio_start(void) 125 { 126 current->delays->blkio_start = local_clock(); 127 } 128 129 /* 130 * We cannot rely on the `current` macro, as we haven't yet switched back to 131 * the process being woken. 132 */ 133 void __delayacct_blkio_end(struct task_struct *p) 134 { 135 delayacct_end(&p->delays->lock, 136 &p->delays->blkio_start, 137 &p->delays->blkio_delay, 138 &p->delays->blkio_count, 139 &p->delays->blkio_delay_max, 140 &p->delays->blkio_delay_min); 141 } 142 143 int delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk) 144 { 145 u64 utime, stime, stimescaled, utimescaled; 146 unsigned long long t2, t3; 147 unsigned long flags, t1; 148 s64 tmp; 149 150 task_cputime(tsk, &utime, &stime); 151 tmp = (s64)d->cpu_run_real_total; 152 tmp += utime + stime; 153 d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp; 154 155 task_cputime_scaled(tsk, &utimescaled, &stimescaled); 156 tmp = (s64)d->cpu_scaled_run_real_total; 157 tmp += utimescaled + stimescaled; 158 d->cpu_scaled_run_real_total = 159 (tmp < (s64)d->cpu_scaled_run_real_total) ? 0 : tmp; 160 161 /* 162 * No locking available for sched_info (and too expensive to add one) 163 * Mitigate by taking snapshot of values 164 */ 165 t1 = tsk->sched_info.pcount; 166 t2 = tsk->sched_info.run_delay; 167 t3 = tsk->se.sum_exec_runtime; 168 169 d->cpu_count += t1; 170 171 d->cpu_delay_max = tsk->sched_info.max_run_delay; 172 d->cpu_delay_min = tsk->sched_info.min_run_delay; 173 tmp = (s64)d->cpu_delay_total + t2; 174 d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp; 175 tmp = (s64)d->cpu_run_virtual_total + t3; 176 177 d->cpu_run_virtual_total = 178 (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp; 179 180 if (!tsk->delays) 181 return 0; 182 183 /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */ 184 raw_spin_lock_irqsave(&tsk->delays->lock, flags); 185 UPDATE_DELAY(blkio); 186 UPDATE_DELAY(swapin); 187 UPDATE_DELAY(freepages); 188 UPDATE_DELAY(thrashing); 189 UPDATE_DELAY(compact); 190 UPDATE_DELAY(wpcopy); 191 UPDATE_DELAY(irq); 192 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags); 193 194 return 0; 195 } 196 197 __u64 __delayacct_blkio_ticks(struct task_struct *tsk) 198 { 199 __u64 ret; 200 unsigned long flags; 201 202 raw_spin_lock_irqsave(&tsk->delays->lock, flags); 203 ret = nsec_to_clock_t(tsk->delays->blkio_delay); 204 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags); 205 return ret; 206 } 207 208 void __delayacct_freepages_start(void) 209 { 210 current->delays->freepages_start = local_clock(); 211 } 212 213 void __delayacct_freepages_end(void) 214 { 215 delayacct_end(¤t->delays->lock, 216 ¤t->delays->freepages_start, 217 ¤t->delays->freepages_delay, 218 ¤t->delays->freepages_count, 219 ¤t->delays->freepages_delay_max, 220 ¤t->delays->freepages_delay_min); 221 } 222 223 void __delayacct_thrashing_start(bool *in_thrashing) 224 { 225 *in_thrashing = !!current->in_thrashing; 226 if (*in_thrashing) 227 return; 228 229 current->in_thrashing = 1; 230 current->delays->thrashing_start = local_clock(); 231 } 232 233 void __delayacct_thrashing_end(bool *in_thrashing) 234 { 235 if (*in_thrashing) 236 return; 237 238 current->in_thrashing = 0; 239 delayacct_end(¤t->delays->lock, 240 ¤t->delays->thrashing_start, 241 ¤t->delays->thrashing_delay, 242 ¤t->delays->thrashing_count, 243 ¤t->delays->thrashing_delay_max, 244 ¤t->delays->thrashing_delay_min); 245 } 246 247 void __delayacct_swapin_start(void) 248 { 249 current->delays->swapin_start = local_clock(); 250 } 251 252 void __delayacct_swapin_end(void) 253 { 254 delayacct_end(¤t->delays->lock, 255 ¤t->delays->swapin_start, 256 ¤t->delays->swapin_delay, 257 ¤t->delays->swapin_count, 258 ¤t->delays->swapin_delay_max, 259 ¤t->delays->swapin_delay_min); 260 } 261 262 void __delayacct_compact_start(void) 263 { 264 current->delays->compact_start = local_clock(); 265 } 266 267 void __delayacct_compact_end(void) 268 { 269 delayacct_end(¤t->delays->lock, 270 ¤t->delays->compact_start, 271 ¤t->delays->compact_delay, 272 ¤t->delays->compact_count, 273 ¤t->delays->compact_delay_max, 274 ¤t->delays->compact_delay_min); 275 } 276 277 void __delayacct_wpcopy_start(void) 278 { 279 current->delays->wpcopy_start = local_clock(); 280 } 281 282 void __delayacct_wpcopy_end(void) 283 { 284 delayacct_end(¤t->delays->lock, 285 ¤t->delays->wpcopy_start, 286 ¤t->delays->wpcopy_delay, 287 ¤t->delays->wpcopy_count, 288 ¤t->delays->wpcopy_delay_max, 289 ¤t->delays->wpcopy_delay_min); 290 } 291 292 void __delayacct_irq(struct task_struct *task, u32 delta) 293 { 294 unsigned long flags; 295 296 raw_spin_lock_irqsave(&task->delays->lock, flags); 297 task->delays->irq_delay += delta; 298 task->delays->irq_count++; 299 if (delta > task->delays->irq_delay_max) 300 task->delays->irq_delay_max = delta; 301 if (delta && (!task->delays->irq_delay_min || delta < task->delays->irq_delay_min)) 302 task->delays->irq_delay_min = delta; 303 raw_spin_unlock_irqrestore(&task->delays->lock, flags); 304 } 305 306