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(const 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 const 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 static __init int kernel_delayacct_sysctls_init(void) 80 { 81 register_sysctl_init("kernel", kern_delayacct_table); 82 return 0; 83 } 84 late_initcall(kernel_delayacct_sysctls_init); 85 #endif 86 87 void __delayacct_tsk_init(struct task_struct *tsk) 88 { 89 tsk->delays = kmem_cache_zalloc(delayacct_cache, GFP_KERNEL); 90 if (tsk->delays) 91 raw_spin_lock_init(&tsk->delays->lock); 92 } 93 94 /* 95 * Finish delay accounting for a statistic using its timestamps (@start), 96 * accumulator (@total) and @count 97 */ 98 static void delayacct_end(raw_spinlock_t *lock, u64 *start, u64 *total, u32 *count, u64 *max, u64 *min) 99 { 100 s64 ns = local_clock() - *start; 101 unsigned long flags; 102 103 if (ns > 0) { 104 raw_spin_lock_irqsave(lock, flags); 105 *total += ns; 106 (*count)++; 107 if (ns > *max) 108 *max = ns; 109 if (*min == 0 || ns < *min) 110 *min = ns; 111 raw_spin_unlock_irqrestore(lock, flags); 112 } 113 } 114 115 void __delayacct_blkio_start(void) 116 { 117 current->delays->blkio_start = local_clock(); 118 } 119 120 /* 121 * We cannot rely on the `current` macro, as we haven't yet switched back to 122 * the process being woken. 123 */ 124 void __delayacct_blkio_end(struct task_struct *p) 125 { 126 delayacct_end(&p->delays->lock, 127 &p->delays->blkio_start, 128 &p->delays->blkio_delay, 129 &p->delays->blkio_count, 130 &p->delays->blkio_delay_max, 131 &p->delays->blkio_delay_min); 132 } 133 134 int delayacct_add_tsk(struct taskstats *d, struct task_struct *tsk) 135 { 136 u64 utime, stime, stimescaled, utimescaled; 137 unsigned long long t2, t3; 138 unsigned long flags, t1; 139 s64 tmp; 140 141 task_cputime(tsk, &utime, &stime); 142 tmp = (s64)d->cpu_run_real_total; 143 tmp += utime + stime; 144 d->cpu_run_real_total = (tmp < (s64)d->cpu_run_real_total) ? 0 : tmp; 145 146 task_cputime_scaled(tsk, &utimescaled, &stimescaled); 147 tmp = (s64)d->cpu_scaled_run_real_total; 148 tmp += utimescaled + stimescaled; 149 d->cpu_scaled_run_real_total = 150 (tmp < (s64)d->cpu_scaled_run_real_total) ? 0 : tmp; 151 152 /* 153 * No locking available for sched_info (and too expensive to add one) 154 * Mitigate by taking snapshot of values 155 */ 156 t1 = tsk->sched_info.pcount; 157 t2 = tsk->sched_info.run_delay; 158 t3 = tsk->se.sum_exec_runtime; 159 160 d->cpu_count += t1; 161 162 d->cpu_delay_max = tsk->sched_info.max_run_delay; 163 d->cpu_delay_min = tsk->sched_info.min_run_delay; 164 tmp = (s64)d->cpu_delay_total + t2; 165 d->cpu_delay_total = (tmp < (s64)d->cpu_delay_total) ? 0 : tmp; 166 tmp = (s64)d->cpu_run_virtual_total + t3; 167 168 d->cpu_run_virtual_total = 169 (tmp < (s64)d->cpu_run_virtual_total) ? 0 : tmp; 170 171 if (!tsk->delays) 172 return 0; 173 174 /* zero XXX_total, non-zero XXX_count implies XXX stat overflowed */ 175 raw_spin_lock_irqsave(&tsk->delays->lock, flags); 176 d->blkio_delay_max = tsk->delays->blkio_delay_max; 177 d->blkio_delay_min = tsk->delays->blkio_delay_min; 178 tmp = d->blkio_delay_total + tsk->delays->blkio_delay; 179 d->blkio_delay_total = (tmp < d->blkio_delay_total) ? 0 : tmp; 180 d->swapin_delay_max = tsk->delays->swapin_delay_max; 181 d->swapin_delay_min = tsk->delays->swapin_delay_min; 182 tmp = d->swapin_delay_total + tsk->delays->swapin_delay; 183 d->swapin_delay_total = (tmp < d->swapin_delay_total) ? 0 : tmp; 184 d->freepages_delay_max = tsk->delays->freepages_delay_max; 185 d->freepages_delay_min = tsk->delays->freepages_delay_min; 186 tmp = d->freepages_delay_total + tsk->delays->freepages_delay; 187 d->freepages_delay_total = (tmp < d->freepages_delay_total) ? 0 : tmp; 188 d->thrashing_delay_max = tsk->delays->thrashing_delay_max; 189 d->thrashing_delay_min = tsk->delays->thrashing_delay_min; 190 tmp = d->thrashing_delay_total + tsk->delays->thrashing_delay; 191 d->thrashing_delay_total = (tmp < d->thrashing_delay_total) ? 0 : tmp; 192 d->compact_delay_max = tsk->delays->compact_delay_max; 193 d->compact_delay_min = tsk->delays->compact_delay_min; 194 tmp = d->compact_delay_total + tsk->delays->compact_delay; 195 d->compact_delay_total = (tmp < d->compact_delay_total) ? 0 : tmp; 196 d->wpcopy_delay_max = tsk->delays->wpcopy_delay_max; 197 d->wpcopy_delay_min = tsk->delays->wpcopy_delay_min; 198 tmp = d->wpcopy_delay_total + tsk->delays->wpcopy_delay; 199 d->wpcopy_delay_total = (tmp < d->wpcopy_delay_total) ? 0 : tmp; 200 d->irq_delay_max = tsk->delays->irq_delay_max; 201 d->irq_delay_min = tsk->delays->irq_delay_min; 202 tmp = d->irq_delay_total + tsk->delays->irq_delay; 203 d->irq_delay_total = (tmp < d->irq_delay_total) ? 0 : tmp; 204 d->blkio_count += tsk->delays->blkio_count; 205 d->swapin_count += tsk->delays->swapin_count; 206 d->freepages_count += tsk->delays->freepages_count; 207 d->thrashing_count += tsk->delays->thrashing_count; 208 d->compact_count += tsk->delays->compact_count; 209 d->wpcopy_count += tsk->delays->wpcopy_count; 210 d->irq_count += tsk->delays->irq_count; 211 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags); 212 213 return 0; 214 } 215 216 __u64 __delayacct_blkio_ticks(struct task_struct *tsk) 217 { 218 __u64 ret; 219 unsigned long flags; 220 221 raw_spin_lock_irqsave(&tsk->delays->lock, flags); 222 ret = nsec_to_clock_t(tsk->delays->blkio_delay); 223 raw_spin_unlock_irqrestore(&tsk->delays->lock, flags); 224 return ret; 225 } 226 227 void __delayacct_freepages_start(void) 228 { 229 current->delays->freepages_start = local_clock(); 230 } 231 232 void __delayacct_freepages_end(void) 233 { 234 delayacct_end(¤t->delays->lock, 235 ¤t->delays->freepages_start, 236 ¤t->delays->freepages_delay, 237 ¤t->delays->freepages_count, 238 ¤t->delays->freepages_delay_max, 239 ¤t->delays->freepages_delay_min); 240 } 241 242 void __delayacct_thrashing_start(bool *in_thrashing) 243 { 244 *in_thrashing = !!current->in_thrashing; 245 if (*in_thrashing) 246 return; 247 248 current->in_thrashing = 1; 249 current->delays->thrashing_start = local_clock(); 250 } 251 252 void __delayacct_thrashing_end(bool *in_thrashing) 253 { 254 if (*in_thrashing) 255 return; 256 257 current->in_thrashing = 0; 258 delayacct_end(¤t->delays->lock, 259 ¤t->delays->thrashing_start, 260 ¤t->delays->thrashing_delay, 261 ¤t->delays->thrashing_count, 262 ¤t->delays->thrashing_delay_max, 263 ¤t->delays->thrashing_delay_min); 264 } 265 266 void __delayacct_swapin_start(void) 267 { 268 current->delays->swapin_start = local_clock(); 269 } 270 271 void __delayacct_swapin_end(void) 272 { 273 delayacct_end(¤t->delays->lock, 274 ¤t->delays->swapin_start, 275 ¤t->delays->swapin_delay, 276 ¤t->delays->swapin_count, 277 ¤t->delays->swapin_delay_max, 278 ¤t->delays->swapin_delay_min); 279 } 280 281 void __delayacct_compact_start(void) 282 { 283 current->delays->compact_start = local_clock(); 284 } 285 286 void __delayacct_compact_end(void) 287 { 288 delayacct_end(¤t->delays->lock, 289 ¤t->delays->compact_start, 290 ¤t->delays->compact_delay, 291 ¤t->delays->compact_count, 292 ¤t->delays->compact_delay_max, 293 ¤t->delays->compact_delay_min); 294 } 295 296 void __delayacct_wpcopy_start(void) 297 { 298 current->delays->wpcopy_start = local_clock(); 299 } 300 301 void __delayacct_wpcopy_end(void) 302 { 303 delayacct_end(¤t->delays->lock, 304 ¤t->delays->wpcopy_start, 305 ¤t->delays->wpcopy_delay, 306 ¤t->delays->wpcopy_count, 307 ¤t->delays->wpcopy_delay_max, 308 ¤t->delays->wpcopy_delay_min); 309 } 310 311 void __delayacct_irq(struct task_struct *task, u32 delta) 312 { 313 unsigned long flags; 314 315 raw_spin_lock_irqsave(&task->delays->lock, flags); 316 task->delays->irq_delay += delta; 317 task->delays->irq_count++; 318 if (delta > task->delays->irq_delay_max) 319 task->delays->irq_delay_max = delta; 320 if (delta && (!task->delays->irq_delay_min || delta < task->delays->irq_delay_min)) 321 task->delays->irq_delay_min = delta; 322 raw_spin_unlock_irqrestore(&task->delays->lock, flags); 323 } 324 325