1 #include <linux/export.h> 2 #include <linux/sched.h> 3 #include <linux/tsacct_kern.h> 4 #include <linux/kernel_stat.h> 5 #include <linux/static_key.h> 6 #include "sched.h" 7 8 9 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 10 11 /* 12 * There are no locks covering percpu hardirq/softirq time. 13 * They are only modified in vtime_account, on corresponding CPU 14 * with interrupts disabled. So, writes are safe. 15 * They are read and saved off onto struct rq in update_rq_clock(). 16 * This may result in other CPU reading this CPU's irq time and can 17 * race with irq/vtime_account on this CPU. We would either get old 18 * or new value with a side effect of accounting a slice of irq time to wrong 19 * task when irq is in progress while we read rq->clock. That is a worthy 20 * compromise in place of having locks on each irq in account_system_time. 21 */ 22 DEFINE_PER_CPU(u64, cpu_hardirq_time); 23 DEFINE_PER_CPU(u64, cpu_softirq_time); 24 25 static DEFINE_PER_CPU(u64, irq_start_time); 26 static int sched_clock_irqtime; 27 28 void enable_sched_clock_irqtime(void) 29 { 30 sched_clock_irqtime = 1; 31 } 32 33 void disable_sched_clock_irqtime(void) 34 { 35 sched_clock_irqtime = 0; 36 } 37 38 #ifndef CONFIG_64BIT 39 DEFINE_PER_CPU(seqcount_t, irq_time_seq); 40 #endif /* CONFIG_64BIT */ 41 42 /* 43 * Called before incrementing preempt_count on {soft,}irq_enter 44 * and before decrementing preempt_count on {soft,}irq_exit. 45 */ 46 void irqtime_account_irq(struct task_struct *curr) 47 { 48 unsigned long flags; 49 s64 delta; 50 int cpu; 51 52 if (!sched_clock_irqtime) 53 return; 54 55 local_irq_save(flags); 56 57 cpu = smp_processor_id(); 58 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time); 59 __this_cpu_add(irq_start_time, delta); 60 61 irq_time_write_begin(); 62 /* 63 * We do not account for softirq time from ksoftirqd here. 64 * We want to continue accounting softirq time to ksoftirqd thread 65 * in that case, so as not to confuse scheduler with a special task 66 * that do not consume any time, but still wants to run. 67 */ 68 if (hardirq_count()) 69 __this_cpu_add(cpu_hardirq_time, delta); 70 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd()) 71 __this_cpu_add(cpu_softirq_time, delta); 72 73 irq_time_write_end(); 74 local_irq_restore(flags); 75 } 76 EXPORT_SYMBOL_GPL(irqtime_account_irq); 77 78 static int irqtime_account_hi_update(void) 79 { 80 u64 *cpustat = kcpustat_this_cpu->cpustat; 81 unsigned long flags; 82 u64 latest_ns; 83 int ret = 0; 84 85 local_irq_save(flags); 86 latest_ns = this_cpu_read(cpu_hardirq_time); 87 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ]) 88 ret = 1; 89 local_irq_restore(flags); 90 return ret; 91 } 92 93 static int irqtime_account_si_update(void) 94 { 95 u64 *cpustat = kcpustat_this_cpu->cpustat; 96 unsigned long flags; 97 u64 latest_ns; 98 int ret = 0; 99 100 local_irq_save(flags); 101 latest_ns = this_cpu_read(cpu_softirq_time); 102 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ]) 103 ret = 1; 104 local_irq_restore(flags); 105 return ret; 106 } 107 108 #else /* CONFIG_IRQ_TIME_ACCOUNTING */ 109 110 #define sched_clock_irqtime (0) 111 112 #endif /* !CONFIG_IRQ_TIME_ACCOUNTING */ 113 114 static inline void task_group_account_field(struct task_struct *p, int index, 115 u64 tmp) 116 { 117 #ifdef CONFIG_CGROUP_CPUACCT 118 struct kernel_cpustat *kcpustat; 119 struct cpuacct *ca; 120 #endif 121 /* 122 * Since all updates are sure to touch the root cgroup, we 123 * get ourselves ahead and touch it first. If the root cgroup 124 * is the only cgroup, then nothing else should be necessary. 125 * 126 */ 127 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp; 128 129 #ifdef CONFIG_CGROUP_CPUACCT 130 if (unlikely(!cpuacct_subsys.active)) 131 return; 132 133 rcu_read_lock(); 134 ca = task_ca(p); 135 while (ca && (ca != &root_cpuacct)) { 136 kcpustat = this_cpu_ptr(ca->cpustat); 137 kcpustat->cpustat[index] += tmp; 138 ca = parent_ca(ca); 139 } 140 rcu_read_unlock(); 141 #endif 142 } 143 144 /* 145 * Account user cpu time to a process. 146 * @p: the process that the cpu time gets accounted to 147 * @cputime: the cpu time spent in user space since the last update 148 * @cputime_scaled: cputime scaled by cpu frequency 149 */ 150 void account_user_time(struct task_struct *p, cputime_t cputime, 151 cputime_t cputime_scaled) 152 { 153 int index; 154 155 /* Add user time to process. */ 156 p->utime += cputime; 157 p->utimescaled += cputime_scaled; 158 account_group_user_time(p, cputime); 159 160 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER; 161 162 /* Add user time to cpustat. */ 163 task_group_account_field(p, index, (__force u64) cputime); 164 165 /* Account for user time used */ 166 acct_update_integrals(p); 167 } 168 169 /* 170 * Account guest cpu time to a process. 171 * @p: the process that the cpu time gets accounted to 172 * @cputime: the cpu time spent in virtual machine since the last update 173 * @cputime_scaled: cputime scaled by cpu frequency 174 */ 175 static void account_guest_time(struct task_struct *p, cputime_t cputime, 176 cputime_t cputime_scaled) 177 { 178 u64 *cpustat = kcpustat_this_cpu->cpustat; 179 180 /* Add guest time to process. */ 181 p->utime += cputime; 182 p->utimescaled += cputime_scaled; 183 account_group_user_time(p, cputime); 184 p->gtime += cputime; 185 186 /* Add guest time to cpustat. */ 187 if (TASK_NICE(p) > 0) { 188 cpustat[CPUTIME_NICE] += (__force u64) cputime; 189 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime; 190 } else { 191 cpustat[CPUTIME_USER] += (__force u64) cputime; 192 cpustat[CPUTIME_GUEST] += (__force u64) cputime; 193 } 194 } 195 196 /* 197 * Account system cpu time to a process and desired cpustat field 198 * @p: the process that the cpu time gets accounted to 199 * @cputime: the cpu time spent in kernel space since the last update 200 * @cputime_scaled: cputime scaled by cpu frequency 201 * @target_cputime64: pointer to cpustat field that has to be updated 202 */ 203 static inline 204 void __account_system_time(struct task_struct *p, cputime_t cputime, 205 cputime_t cputime_scaled, int index) 206 { 207 /* Add system time to process. */ 208 p->stime += cputime; 209 p->stimescaled += cputime_scaled; 210 account_group_system_time(p, cputime); 211 212 /* Add system time to cpustat. */ 213 task_group_account_field(p, index, (__force u64) cputime); 214 215 /* Account for system time used */ 216 acct_update_integrals(p); 217 } 218 219 /* 220 * Account system cpu time to a process. 221 * @p: the process that the cpu time gets accounted to 222 * @hardirq_offset: the offset to subtract from hardirq_count() 223 * @cputime: the cpu time spent in kernel space since the last update 224 * @cputime_scaled: cputime scaled by cpu frequency 225 */ 226 void account_system_time(struct task_struct *p, int hardirq_offset, 227 cputime_t cputime, cputime_t cputime_scaled) 228 { 229 int index; 230 231 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { 232 account_guest_time(p, cputime, cputime_scaled); 233 return; 234 } 235 236 if (hardirq_count() - hardirq_offset) 237 index = CPUTIME_IRQ; 238 else if (in_serving_softirq()) 239 index = CPUTIME_SOFTIRQ; 240 else 241 index = CPUTIME_SYSTEM; 242 243 __account_system_time(p, cputime, cputime_scaled, index); 244 } 245 246 /* 247 * Account for involuntary wait time. 248 * @cputime: the cpu time spent in involuntary wait 249 */ 250 void account_steal_time(cputime_t cputime) 251 { 252 u64 *cpustat = kcpustat_this_cpu->cpustat; 253 254 cpustat[CPUTIME_STEAL] += (__force u64) cputime; 255 } 256 257 /* 258 * Account for idle time. 259 * @cputime: the cpu time spent in idle wait 260 */ 261 void account_idle_time(cputime_t cputime) 262 { 263 u64 *cpustat = kcpustat_this_cpu->cpustat; 264 struct rq *rq = this_rq(); 265 266 if (atomic_read(&rq->nr_iowait) > 0) 267 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime; 268 else 269 cpustat[CPUTIME_IDLE] += (__force u64) cputime; 270 } 271 272 static __always_inline bool steal_account_process_tick(void) 273 { 274 #ifdef CONFIG_PARAVIRT 275 if (static_key_false(¶virt_steal_enabled)) { 276 u64 steal, st = 0; 277 278 steal = paravirt_steal_clock(smp_processor_id()); 279 steal -= this_rq()->prev_steal_time; 280 281 st = steal_ticks(steal); 282 this_rq()->prev_steal_time += st * TICK_NSEC; 283 284 account_steal_time(st); 285 return st; 286 } 287 #endif 288 return false; 289 } 290 291 /* 292 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live 293 * tasks (sum on group iteration) belonging to @tsk's group. 294 */ 295 void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times) 296 { 297 struct signal_struct *sig = tsk->signal; 298 struct task_struct *t; 299 300 times->utime = sig->utime; 301 times->stime = sig->stime; 302 times->sum_exec_runtime = sig->sum_sched_runtime; 303 304 rcu_read_lock(); 305 /* make sure we can trust tsk->thread_group list */ 306 if (!likely(pid_alive(tsk))) 307 goto out; 308 309 t = tsk; 310 do { 311 times->utime += t->utime; 312 times->stime += t->stime; 313 times->sum_exec_runtime += task_sched_runtime(t); 314 } while_each_thread(tsk, t); 315 out: 316 rcu_read_unlock(); 317 } 318 319 #ifndef CONFIG_VIRT_CPU_ACCOUNTING 320 321 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 322 /* 323 * Account a tick to a process and cpustat 324 * @p: the process that the cpu time gets accounted to 325 * @user_tick: is the tick from userspace 326 * @rq: the pointer to rq 327 * 328 * Tick demultiplexing follows the order 329 * - pending hardirq update 330 * - pending softirq update 331 * - user_time 332 * - idle_time 333 * - system time 334 * - check for guest_time 335 * - else account as system_time 336 * 337 * Check for hardirq is done both for system and user time as there is 338 * no timer going off while we are on hardirq and hence we may never get an 339 * opportunity to update it solely in system time. 340 * p->stime and friends are only updated on system time and not on irq 341 * softirq as those do not count in task exec_runtime any more. 342 */ 343 static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 344 struct rq *rq) 345 { 346 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 347 u64 *cpustat = kcpustat_this_cpu->cpustat; 348 349 if (steal_account_process_tick()) 350 return; 351 352 if (irqtime_account_hi_update()) { 353 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy; 354 } else if (irqtime_account_si_update()) { 355 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy; 356 } else if (this_cpu_ksoftirqd() == p) { 357 /* 358 * ksoftirqd time do not get accounted in cpu_softirq_time. 359 * So, we have to handle it separately here. 360 * Also, p->stime needs to be updated for ksoftirqd. 361 */ 362 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 363 CPUTIME_SOFTIRQ); 364 } else if (user_tick) { 365 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 366 } else if (p == rq->idle) { 367 account_idle_time(cputime_one_jiffy); 368 } else if (p->flags & PF_VCPU) { /* System time or guest time */ 369 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled); 370 } else { 371 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 372 CPUTIME_SYSTEM); 373 } 374 } 375 376 static void irqtime_account_idle_ticks(int ticks) 377 { 378 int i; 379 struct rq *rq = this_rq(); 380 381 for (i = 0; i < ticks; i++) 382 irqtime_account_process_tick(current, 0, rq); 383 } 384 #else /* CONFIG_IRQ_TIME_ACCOUNTING */ 385 static void irqtime_account_idle_ticks(int ticks) {} 386 static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 387 struct rq *rq) {} 388 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 389 390 /* 391 * Account a single tick of cpu time. 392 * @p: the process that the cpu time gets accounted to 393 * @user_tick: indicates if the tick is a user or a system tick 394 */ 395 void account_process_tick(struct task_struct *p, int user_tick) 396 { 397 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 398 struct rq *rq = this_rq(); 399 400 if (sched_clock_irqtime) { 401 irqtime_account_process_tick(p, user_tick, rq); 402 return; 403 } 404 405 if (steal_account_process_tick()) 406 return; 407 408 if (user_tick) 409 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 410 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET)) 411 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy, 412 one_jiffy_scaled); 413 else 414 account_idle_time(cputime_one_jiffy); 415 } 416 417 /* 418 * Account multiple ticks of steal time. 419 * @p: the process from which the cpu time has been stolen 420 * @ticks: number of stolen ticks 421 */ 422 void account_steal_ticks(unsigned long ticks) 423 { 424 account_steal_time(jiffies_to_cputime(ticks)); 425 } 426 427 /* 428 * Account multiple ticks of idle time. 429 * @ticks: number of stolen ticks 430 */ 431 void account_idle_ticks(unsigned long ticks) 432 { 433 434 if (sched_clock_irqtime) { 435 irqtime_account_idle_ticks(ticks); 436 return; 437 } 438 439 account_idle_time(jiffies_to_cputime(ticks)); 440 } 441 442 #endif 443 444 /* 445 * Use precise platform statistics if available: 446 */ 447 #ifdef CONFIG_VIRT_CPU_ACCOUNTING 448 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st) 449 { 450 *ut = p->utime; 451 *st = p->stime; 452 } 453 454 void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st) 455 { 456 struct task_cputime cputime; 457 458 thread_group_cputime(p, &cputime); 459 460 *ut = cputime.utime; 461 *st = cputime.stime; 462 } 463 464 void vtime_account_system_irqsafe(struct task_struct *tsk) 465 { 466 unsigned long flags; 467 468 local_irq_save(flags); 469 vtime_account_system(tsk); 470 local_irq_restore(flags); 471 } 472 EXPORT_SYMBOL_GPL(vtime_account_system_irqsafe); 473 474 #ifndef __ARCH_HAS_VTIME_TASK_SWITCH 475 void vtime_task_switch(struct task_struct *prev) 476 { 477 if (is_idle_task(prev)) 478 vtime_account_idle(prev); 479 else 480 vtime_account_system(prev); 481 482 vtime_account_user(prev); 483 arch_vtime_task_switch(prev); 484 } 485 #endif 486 487 /* 488 * Archs that account the whole time spent in the idle task 489 * (outside irq) as idle time can rely on this and just implement 490 * vtime_account_system() and vtime_account_idle(). Archs that 491 * have other meaning of the idle time (s390 only includes the 492 * time spent by the CPU when it's in low power mode) must override 493 * vtime_account(). 494 */ 495 #ifndef __ARCH_HAS_VTIME_ACCOUNT 496 void vtime_account(struct task_struct *tsk) 497 { 498 if (in_interrupt() || !is_idle_task(tsk)) 499 vtime_account_system(tsk); 500 else 501 vtime_account_idle(tsk); 502 } 503 EXPORT_SYMBOL_GPL(vtime_account); 504 #endif /* __ARCH_HAS_VTIME_ACCOUNT */ 505 506 #else 507 508 #ifndef nsecs_to_cputime 509 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs) 510 #endif 511 512 static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total) 513 { 514 u64 temp = (__force u64) rtime; 515 516 temp *= (__force u64) utime; 517 518 if (sizeof(cputime_t) == 4) 519 temp = div_u64(temp, (__force u32) total); 520 else 521 temp = div64_u64(temp, (__force u64) total); 522 523 return (__force cputime_t) temp; 524 } 525 526 /* 527 * Adjust tick based cputime random precision against scheduler 528 * runtime accounting. 529 */ 530 static void cputime_adjust(struct task_cputime *curr, 531 struct cputime *prev, 532 cputime_t *ut, cputime_t *st) 533 { 534 cputime_t rtime, utime, total; 535 536 utime = curr->utime; 537 total = utime + curr->stime; 538 539 /* 540 * Tick based cputime accounting depend on random scheduling 541 * timeslices of a task to be interrupted or not by the timer. 542 * Depending on these circumstances, the number of these interrupts 543 * may be over or under-optimistic, matching the real user and system 544 * cputime with a variable precision. 545 * 546 * Fix this by scaling these tick based values against the total 547 * runtime accounted by the CFS scheduler. 548 */ 549 rtime = nsecs_to_cputime(curr->sum_exec_runtime); 550 551 if (total) 552 utime = scale_utime(utime, rtime, total); 553 else 554 utime = rtime; 555 556 /* 557 * If the tick based count grows faster than the scheduler one, 558 * the result of the scaling may go backward. 559 * Let's enforce monotonicity. 560 */ 561 prev->utime = max(prev->utime, utime); 562 prev->stime = max(prev->stime, rtime - prev->utime); 563 564 *ut = prev->utime; 565 *st = prev->stime; 566 } 567 568 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st) 569 { 570 struct task_cputime cputime = { 571 .utime = p->utime, 572 .stime = p->stime, 573 .sum_exec_runtime = p->se.sum_exec_runtime, 574 }; 575 576 cputime_adjust(&cputime, &p->prev_cputime, ut, st); 577 } 578 579 /* 580 * Must be called with siglock held. 581 */ 582 void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st) 583 { 584 struct task_cputime cputime; 585 586 thread_group_cputime(p, &cputime); 587 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st); 588 } 589 #endif 590