kernel/sched/cputime.c

   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 <linux/context_tracking.h>
   7 #include "sched.h"
   8
   9
  10 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
  11
  12 /*
  13  * There are no locks covering percpu hardirq/softirq time.
  14  * They are only modified in vtime_account, on corresponding CPU
  15  * with interrupts disabled. So, writes are safe.
  16  * They are read and saved off onto struct rq in update_rq_clock().
  17  * This may result in other CPU reading this CPU's irq time and can
  18  * race with irq/vtime_account on this CPU. We would either get old
  19  * or new value with a side effect of accounting a slice of irq time to wrong
  20  * task when irq is in progress while we read rq->clock. That is a worthy
  21  * compromise in place of having locks on each irq in account_system_time.
  22  */
  23 DEFINE_PER_CPU(u64, cpu_hardirq_time);
  24 DEFINE_PER_CPU(u64, cpu_softirq_time);
  25
  26 static DEFINE_PER_CPU(u64, irq_start_time);
  27 static int sched_clock_irqtime;
  28
  29 void enable_sched_clock_irqtime(void)
  30 {
  31         sched_clock_irqtime = 1;
  32 }
  33
  34 void disable_sched_clock_irqtime(void)
  35 {
  36         sched_clock_irqtime = 0;
  37 }
  38
  39 #ifndef CONFIG_64BIT
  40 DEFINE_PER_CPU(seqcount_t, irq_time_seq);
  41 #endif /* CONFIG_64BIT */
  42
  43 /*
  44  * Called before incrementing preempt_count on {soft,}irq_enter
  45  * and before decrementing preempt_count on {soft,}irq_exit.
  46  */
  47 void irqtime_account_irq(struct task_struct *curr)
  48 {
  49         unsigned long flags;
  50         s64 delta;
  51         int cpu;
  52
  53         if (!sched_clock_irqtime)
  54                 return;
  55
  56         local_irq_save(flags);
  57
  58         cpu = smp_processor_id();
  59         delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
  60         __this_cpu_add(irq_start_time, delta);
  61
  62         irq_time_write_begin();
  63         /*
  64          * We do not account for softirq time from ksoftirqd here.
  65          * We want to continue accounting softirq time to ksoftirqd thread
  66          * in that case, so as not to confuse scheduler with a special task
  67          * that do not consume any time, but still wants to run.
  68          */
  69         if (hardirq_count())
  70                 __this_cpu_add(cpu_hardirq_time, delta);
  71         else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
  72                 __this_cpu_add(cpu_softirq_time, delta);
  73
  74         irq_time_write_end();
  75         local_irq_restore(flags);
  76 }
  77 EXPORT_SYMBOL_GPL(irqtime_account_irq);
  78
  79 static int irqtime_account_hi_update(void)
  80 {
  81         u64 *cpustat = kcpustat_this_cpu->cpustat;
  82         unsigned long flags;
  83         u64 latest_ns;
  84         int ret = 0;
  85
  86         local_irq_save(flags);
  87         latest_ns = this_cpu_read(cpu_hardirq_time);
  88         if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
  89                 ret = 1;
  90         local_irq_restore(flags);
  91         return ret;
  92 }
  93
  94 static int irqtime_account_si_update(void)
  95 {
  96         u64 *cpustat = kcpustat_this_cpu->cpustat;
  97         unsigned long flags;
  98         u64 latest_ns;
  99         int ret = 0;
 100
 101         local_irq_save(flags);
 102         latest_ns = this_cpu_read(cpu_softirq_time);
 103         if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
 104                 ret = 1;
 105         local_irq_restore(flags);
 106         return ret;
 107 }
 108
 109 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
 110
 111 #define sched_clock_irqtime     (0)
 112
 113 #endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
 114
 115 static inline void task_group_account_field(struct task_struct *p, int index,
 116                                             u64 tmp)
 117 {
 118 #ifdef CONFIG_CGROUP_CPUACCT
 119         struct kernel_cpustat *kcpustat;
 120         struct cpuacct *ca;
 121 #endif
 122         /*
 123          * Since all updates are sure to touch the root cgroup, we
 124          * get ourselves ahead and touch it first. If the root cgroup
 125          * is the only cgroup, then nothing else should be necessary.
 126          *
 127          */
 128         __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
 129
 130 #ifdef CONFIG_CGROUP_CPUACCT
 131         if (unlikely(!cpuacct_subsys.active))
 132                 return;
 133
 134         rcu_read_lock();
 135         ca = task_ca(p);
 136         while (ca && (ca != &root_cpuacct)) {
 137                 kcpustat = this_cpu_ptr(ca->cpustat);
 138                 kcpustat->cpustat[index] += tmp;
 139                 ca = parent_ca(ca);
 140         }
 141         rcu_read_unlock();
 142 #endif
 143 }
 144
 145 /*
 146  * Account user cpu time to a process.
 147  * @p: the process that the cpu time gets accounted to
 148  * @cputime: the cpu time spent in user space since the last update
 149  * @cputime_scaled: cputime scaled by cpu frequency
 150  */
 151 void account_user_time(struct task_struct *p, cputime_t cputime,
 152                        cputime_t cputime_scaled)
 153 {
 154         int index;
 155
 156         /* Add user time to process. */
 157         p->utime += cputime;
 158         p->utimescaled += cputime_scaled;
 159         account_group_user_time(p, cputime);
 160
 161         index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
 162
 163         /* Add user time to cpustat. */
 164         task_group_account_field(p, index, (__force u64) cputime);
 165
 166         /* Account for user time used */
 167         acct_account_cputime(p);
 168 }
 169
 170 /*
 171  * Account guest cpu time to a process.
 172  * @p: the process that the cpu time gets accounted to
 173  * @cputime: the cpu time spent in virtual machine since the last update
 174  * @cputime_scaled: cputime scaled by cpu frequency
 175  */
 176 static void account_guest_time(struct task_struct *p, cputime_t cputime,
 177                                cputime_t cputime_scaled)
 178 {
 179         u64 *cpustat = kcpustat_this_cpu->cpustat;
 180
 181         /* Add guest time to process. */
 182         p->utime += cputime;
 183         p->utimescaled += cputime_scaled;
 184         account_group_user_time(p, cputime);
 185         p->gtime += cputime;
 186
 187         /* Add guest time to cpustat. */
 188         if (TASK_NICE(p) > 0) {
 189                 cpustat[CPUTIME_NICE] += (__force u64) cputime;
 190                 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
 191         } else {
 192                 cpustat[CPUTIME_USER] += (__force u64) cputime;
 193                 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
 194         }
 195 }
 196
 197 /*
 198  * Account system cpu time to a process and desired cpustat field
 199  * @p: the process that the cpu time gets accounted to
 200  * @cputime: the cpu time spent in kernel space since the last update
 201  * @cputime_scaled: cputime scaled by cpu frequency
 202  * @target_cputime64: pointer to cpustat field that has to be updated
 203  */
 204 static inline
 205 void __account_system_time(struct task_struct *p, cputime_t cputime,
 206                         cputime_t cputime_scaled, int index)
 207 {
 208         /* Add system time to process. */
 209         p->stime += cputime;
 210         p->stimescaled += cputime_scaled;
 211         account_group_system_time(p, cputime);
 212
 213         /* Add system time to cpustat. */
 214         task_group_account_field(p, index, (__force u64) cputime);
 215
 216         /* Account for system time used */
 217         acct_account_cputime(p);
 218 }
 219
 220 /*
 221  * Account system cpu time to a process.
 222  * @p: the process that the cpu time gets accounted to
 223  * @hardirq_offset: the offset to subtract from hardirq_count()
 224  * @cputime: the cpu time spent in kernel space since the last update
 225  * @cputime_scaled: cputime scaled by cpu frequency
 226  */
 227 void account_system_time(struct task_struct *p, int hardirq_offset,
 228                          cputime_t cputime, cputime_t cputime_scaled)
 229 {
 230         int index;
 231
 232         if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
 233                 account_guest_time(p, cputime, cputime_scaled);
 234                 return;
 235         }
 236
 237         if (hardirq_count() - hardirq_offset)
 238                 index = CPUTIME_IRQ;
 239         else if (in_serving_softirq())
 240                 index = CPUTIME_SOFTIRQ;
 241         else
 242                 index = CPUTIME_SYSTEM;
 243
 244         __account_system_time(p, cputime, cputime_scaled, index);
 245 }
 246
 247 /*
 248  * Account for involuntary wait time.
 249  * @cputime: the cpu time spent in involuntary wait
 250  */
 251 void account_steal_time(cputime_t cputime)
 252 {
 253         u64 *cpustat = kcpustat_this_cpu->cpustat;
 254
 255         cpustat[CPUTIME_STEAL] += (__force u64) cputime;
 256 }
 257
 258 /*
 259  * Account for idle time.
 260  * @cputime: the cpu time spent in idle wait
 261  */
 262 void account_idle_time(cputime_t cputime)
 263 {
 264         u64 *cpustat = kcpustat_this_cpu->cpustat;
 265         struct rq *rq = this_rq();
 266
 267         if (atomic_read(&rq->nr_iowait) > 0)
 268                 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
 269         else
 270                 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
 271 }
 272
 273 static __always_inline bool steal_account_process_tick(void)
 274 {
 275 #ifdef CONFIG_PARAVIRT
 276         if (static_key_false(&paravirt_steal_enabled)) {
 277                 u64 steal, st = 0;
 278
 279                 steal = paravirt_steal_clock(smp_processor_id());
 280                 steal -= this_rq()->prev_steal_time;
 281
 282                 st = steal_ticks(steal);
 283                 this_rq()->prev_steal_time += st * TICK_NSEC;
 284
 285                 account_steal_time(st);
 286                 return st;
 287         }
 288 #endif
 289         return false;
 290 }
 291
 292 /*
 293  * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
 294  * tasks (sum on group iteration) belonging to @tsk's group.
 295  */
 296 void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
 297 {
 298         struct signal_struct *sig = tsk->signal;
 299         cputime_t utime, stime;
 300         struct task_struct *t;
 301
 302         times->utime = sig->utime;
 303         times->stime = sig->stime;
 304         times->sum_exec_runtime = sig->sum_sched_runtime;
 305
 306         rcu_read_lock();
 307         /* make sure we can trust tsk->thread_group list */
 308         if (!likely(pid_alive(tsk)))
 309                 goto out;
 310
 311         t = tsk;
 312         do {
 313                 task_cputime(tsk, &utime, &stime);
 314                 times->utime += utime;
 315                 times->stime += stime;
 316                 times->sum_exec_runtime += task_sched_runtime(t);
 317         } while_each_thread(tsk, t);
 318 out:
 319         rcu_read_unlock();
 320 }
 321
 322 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
 323 /*
 324  * Account a tick to a process and cpustat
 325  * @p: the process that the cpu time gets accounted to
 326  * @user_tick: is the tick from userspace
 327  * @rq: the pointer to rq
 328  *
 329  * Tick demultiplexing follows the order
 330  * - pending hardirq update
 331  * - pending softirq update
 332  * - user_time
 333  * - idle_time
 334  * - system time
 335  *   - check for guest_time
 336  *   - else account as system_time
 337  *
 338  * Check for hardirq is done both for system and user time as there is
 339  * no timer going off while we are on hardirq and hence we may never get an
 340  * opportunity to update it solely in system time.
 341  * p->stime and friends are only updated on system time and not on irq
 342  * softirq as those do not count in task exec_runtime any more.
 343  */
 344 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
 345                                                 struct rq *rq)
 346 {
 347         cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
 348         u64 *cpustat = kcpustat_this_cpu->cpustat;
 349
 350         if (steal_account_process_tick())
 351                 return;
 352
 353         if (irqtime_account_hi_update()) {
 354                 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
 355         } else if (irqtime_account_si_update()) {
 356                 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
 357         } else if (this_cpu_ksoftirqd() == p) {
 358                 /*
 359                  * ksoftirqd time do not get accounted in cpu_softirq_time.
 360                  * So, we have to handle it separately here.
 361                  * Also, p->stime needs to be updated for ksoftirqd.
 362                  */
 363                 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
 364                                         CPUTIME_SOFTIRQ);
 365         } else if (user_tick) {
 366                 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
 367         } else if (p == rq->idle) {
 368                 account_idle_time(cputime_one_jiffy);
 369         } else if (p->flags & PF_VCPU) { /* System time or guest time */
 370                 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
 371         } else {
 372                 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
 373                                         CPUTIME_SYSTEM);
 374         }
 375 }
 376
 377 static void irqtime_account_idle_ticks(int ticks)
 378 {
 379         int i;
 380         struct rq *rq = this_rq();
 381
 382         for (i = 0; i < ticks; i++)
 383                 irqtime_account_process_tick(current, 0, rq);
 384 }
 385 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
 386 static inline void irqtime_account_idle_ticks(int ticks) {}
 387 static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
 388                                                 struct rq *rq) {}
 389 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
 390
 391 /*
 392  * Use precise platform statistics if available:
 393  */
 394 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
 395
 396 #ifndef __ARCH_HAS_VTIME_TASK_SWITCH
 397 void vtime_task_switch(struct task_struct *prev)
 398 {
 399         if (!vtime_accounting_enabled())
 400                 return;
 401
 402         if (is_idle_task(prev))
 403                 vtime_account_idle(prev);
 404         else
 405                 vtime_account_system(prev);
 406
 407 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 408         vtime_account_user(prev);
 409 #endif
 410         arch_vtime_task_switch(prev);
 411 }
 412 #endif
 413
 414 /*
 415  * Archs that account the whole time spent in the idle task
 416  * (outside irq) as idle time can rely on this and just implement
 417  * vtime_account_system() and vtime_account_idle(). Archs that
 418  * have other meaning of the idle time (s390 only includes the
 419  * time spent by the CPU when it's in low power mode) must override
 420  * vtime_account().
 421  */
 422 #ifndef __ARCH_HAS_VTIME_ACCOUNT
 423 void vtime_account_irq_enter(struct task_struct *tsk)
 424 {
 425         if (!vtime_accounting_enabled())
 426                 return;
 427
 428         if (!in_interrupt()) {
 429                 /*
 430                  * If we interrupted user, context_tracking_in_user()
 431                  * is 1 because the context tracking don't hook
 432                  * on irq entry/exit. This way we know if
 433                  * we need to flush user time on kernel entry.
 434                  */
 435                 if (context_tracking_in_user()) {
 436                         vtime_account_user(tsk);
 437                         return;
 438                 }
 439
 440                 if (is_idle_task(tsk)) {
 441                         vtime_account_idle(tsk);
 442                         return;
 443                 }
 444         }
 445         vtime_account_system(tsk);
 446 }
 447 EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
 448 #endif /* __ARCH_HAS_VTIME_ACCOUNT */
 449 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
 450
 451
 452 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 453 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 454 {
 455         *ut = p->utime;
 456         *st = p->stime;
 457 }
 458
 459 void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 460 {
 461         struct task_cputime cputime;
 462
 463         thread_group_cputime(p, &cputime);
 464
 465         *ut = cputime.utime;
 466         *st = cputime.stime;
 467 }
 468 #else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 469 /*
 470  * Account a single tick of cpu time.
 471  * @p: the process that the cpu time gets accounted to
 472  * @user_tick: indicates if the tick is a user or a system tick
 473  */
 474 void account_process_tick(struct task_struct *p, int user_tick)
 475 {
 476         cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
 477         struct rq *rq = this_rq();
 478
 479         if (vtime_accounting_enabled())
 480                 return;
 481
 482         if (sched_clock_irqtime) {
 483                 irqtime_account_process_tick(p, user_tick, rq);
 484                 return;
 485         }
 486
 487         if (steal_account_process_tick())
 488                 return;
 489
 490         if (user_tick)
 491                 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
 492         else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
 493                 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
 494                                     one_jiffy_scaled);
 495         else
 496                 account_idle_time(cputime_one_jiffy);
 497 }
 498
 499 /*
 500  * Account multiple ticks of steal time.
 501  * @p: the process from which the cpu time has been stolen
 502  * @ticks: number of stolen ticks
 503  */
 504 void account_steal_ticks(unsigned long ticks)
 505 {
 506         account_steal_time(jiffies_to_cputime(ticks));
 507 }
 508
 509 /*
 510  * Account multiple ticks of idle time.
 511  * @ticks: number of stolen ticks
 512  */
 513 void account_idle_ticks(unsigned long ticks)
 514 {
 515
 516         if (sched_clock_irqtime) {
 517                 irqtime_account_idle_ticks(ticks);
 518                 return;
 519         }
 520
 521         account_idle_time(jiffies_to_cputime(ticks));
 522 }
 523
 524 /*
 525  * Perform (stime * rtime) / total with reduced chances
 526  * of multiplication overflows by using smaller factors
 527  * like quotient and remainders of divisions between
 528  * rtime and total.
 529  */
 530 static cputime_t scale_stime(u64 stime, u64 rtime, u64 total)
 531 {
 532         u64 rem, res, scaled;
 533
 534         if (rtime >= total) {
 535                 /*
 536                  * Scale up to rtime / total then add
 537                  * the remainder scaled to stime / total.
 538                  */
 539                 res = div64_u64_rem(rtime, total, &rem);
 540                 scaled = stime * res;
 541                 scaled += div64_u64(stime * rem, total);
 542         } else {
 543                 /*
 544                  * Same in reverse: scale down to total / rtime
 545                  * then substract that result scaled to
 546                  * to the remaining part.
 547                  */
 548                 res = div64_u64_rem(total, rtime, &rem);
 549                 scaled = div64_u64(stime, res);
 550                 scaled -= div64_u64(scaled * rem, total);
 551         }
 552
 553         return (__force cputime_t) scaled;
 554 }
 555
 556 /*
 557  * Adjust tick based cputime random precision against scheduler
 558  * runtime accounting.
 559  */
 560 static void cputime_adjust(struct task_cputime *curr,
 561                            struct cputime *prev,
 562                            cputime_t *ut, cputime_t *st)
 563 {
 564         cputime_t rtime, stime, total;
 565
 566         if (vtime_accounting_enabled()) {
 567                 *ut = curr->utime;
 568                 *st = curr->stime;
 569                 return;
 570         }
 571
 572         stime = curr->stime;
 573         total = stime + curr->utime;
 574
 575         /*
 576          * Tick based cputime accounting depend on random scheduling
 577          * timeslices of a task to be interrupted or not by the timer.
 578          * Depending on these circumstances, the number of these interrupts
 579          * may be over or under-optimistic, matching the real user and system
 580          * cputime with a variable precision.
 581          *
 582          * Fix this by scaling these tick based values against the total
 583          * runtime accounted by the CFS scheduler.
 584          */
 585         rtime = nsecs_to_cputime(curr->sum_exec_runtime);
 586
 587         if (!rtime) {
 588                 stime = 0;
 589         } else if (!total) {
 590                 stime = rtime;
 591         } else {
 592                 stime = scale_stime((__force u64)stime,
 593                                     (__force u64)rtime, (__force u64)total);
 594         }
 595
 596         /*
 597          * If the tick based count grows faster than the scheduler one,
 598          * the result of the scaling may go backward.
 599          * Let's enforce monotonicity.
 600          */
 601         prev->stime = max(prev->stime, stime);
 602         prev->utime = max(prev->utime, rtime - prev->stime);
 603
 604         *ut = prev->utime;
 605         *st = prev->stime;
 606 }
 607
 608 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 609 {
 610         struct task_cputime cputime = {
 611                 .sum_exec_runtime = p->se.sum_exec_runtime,
 612         };
 613
 614         task_cputime(p, &cputime.utime, &cputime.stime);
 615         cputime_adjust(&cputime, &p->prev_cputime, ut, st);
 616 }
 617
 618 /*
 619  * Must be called with siglock held.
 620  */
 621 void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 622 {
 623         struct task_cputime cputime;
 624
 625         thread_group_cputime(p, &cputime);
 626         cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
 627 }
 628 #endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 629
 630 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
 631 static unsigned long long vtime_delta(struct task_struct *tsk)
 632 {
 633         unsigned long long clock;
 634
 635         clock = local_clock();
 636         if (clock < tsk->vtime_snap)
 637                 return 0;
 638
 639         return clock - tsk->vtime_snap;
 640 }
 641
 642 static cputime_t get_vtime_delta(struct task_struct *tsk)
 643 {
 644         unsigned long long delta = vtime_delta(tsk);
 645
 646         WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_SLEEPING);
 647         tsk->vtime_snap += delta;
 648
 649         /* CHECKME: always safe to convert nsecs to cputime? */
 650         return nsecs_to_cputime(delta);
 651 }
 652
 653 static void __vtime_account_system(struct task_struct *tsk)
 654 {
 655         cputime_t delta_cpu = get_vtime_delta(tsk);
 656
 657         account_system_time(tsk, irq_count(), delta_cpu, cputime_to_scaled(delta_cpu));
 658 }
 659
 660 void vtime_account_system(struct task_struct *tsk)
 661 {
 662         if (!vtime_accounting_enabled())
 663                 return;
 664
 665         write_seqlock(&tsk->vtime_seqlock);
 666         __vtime_account_system(tsk);
 667         write_sequnlock(&tsk->vtime_seqlock);
 668 }
 669
 670 void vtime_account_irq_exit(struct task_struct *tsk)
 671 {
 672         if (!vtime_accounting_enabled())
 673                 return;
 674
 675         write_seqlock(&tsk->vtime_seqlock);
 676         if (context_tracking_in_user())
 677                 tsk->vtime_snap_whence = VTIME_USER;
 678         __vtime_account_system(tsk);
 679         write_sequnlock(&tsk->vtime_seqlock);
 680 }
 681
 682 void vtime_account_user(struct task_struct *tsk)
 683 {
 684         cputime_t delta_cpu;
 685
 686         if (!vtime_accounting_enabled())
 687                 return;
 688
 689         delta_cpu = get_vtime_delta(tsk);
 690
 691         write_seqlock(&tsk->vtime_seqlock);
 692         tsk->vtime_snap_whence = VTIME_SYS;
 693         account_user_time(tsk, delta_cpu, cputime_to_scaled(delta_cpu));
 694         write_sequnlock(&tsk->vtime_seqlock);
 695 }
 696
 697 void vtime_user_enter(struct task_struct *tsk)
 698 {
 699         if (!vtime_accounting_enabled())
 700                 return;
 701
 702         write_seqlock(&tsk->vtime_seqlock);
 703         tsk->vtime_snap_whence = VTIME_USER;
 704         __vtime_account_system(tsk);
 705         write_sequnlock(&tsk->vtime_seqlock);
 706 }
 707
 708 void vtime_guest_enter(struct task_struct *tsk)
 709 {
 710         write_seqlock(&tsk->vtime_seqlock);
 711         __vtime_account_system(tsk);
 712         current->flags |= PF_VCPU;
 713         write_sequnlock(&tsk->vtime_seqlock);
 714 }
 715
 716 void vtime_guest_exit(struct task_struct *tsk)
 717 {
 718         write_seqlock(&tsk->vtime_seqlock);
 719         __vtime_account_system(tsk);
 720         current->flags &= ~PF_VCPU;
 721         write_sequnlock(&tsk->vtime_seqlock);
 722 }
 723
 724 void vtime_account_idle(struct task_struct *tsk)
 725 {
 726         cputime_t delta_cpu = get_vtime_delta(tsk);
 727
 728         account_idle_time(delta_cpu);
 729 }
 730
 731 bool vtime_accounting_enabled(void)
 732 {
 733         return context_tracking_active();
 734 }
 735
 736 void arch_vtime_task_switch(struct task_struct *prev)
 737 {
 738         write_seqlock(&prev->vtime_seqlock);
 739         prev->vtime_snap_whence = VTIME_SLEEPING;
 740         write_sequnlock(&prev->vtime_seqlock);
 741
 742         write_seqlock(&current->vtime_seqlock);
 743         current->vtime_snap_whence = VTIME_SYS;
 744         current->vtime_snap = sched_clock();
 745         write_sequnlock(&current->vtime_seqlock);
 746 }
 747
 748 void vtime_init_idle(struct task_struct *t)
 749 {
 750         unsigned long flags;
 751
 752         write_seqlock_irqsave(&t->vtime_seqlock, flags);
 753         t->vtime_snap_whence = VTIME_SYS;
 754         t->vtime_snap = sched_clock();
 755         write_sequnlock_irqrestore(&t->vtime_seqlock, flags);
 756 }
 757
 758 cputime_t task_gtime(struct task_struct *t)
 759 {
 760         unsigned int seq;
 761         cputime_t gtime;
 762
 763         do {
 764                 seq = read_seqbegin(&t->vtime_seqlock);
 765
 766                 gtime = t->gtime;
 767                 if (t->flags & PF_VCPU)
 768                         gtime += vtime_delta(t);
 769
 770         } while (read_seqretry(&t->vtime_seqlock, seq));
 771
 772         return gtime;
 773 }
 774
 775 /*
 776  * Fetch cputime raw values from fields of task_struct and
 777  * add up the pending nohz execution time since the last
 778  * cputime snapshot.
 779  */
 780 static void
 781 fetch_task_cputime(struct task_struct *t,
 782                    cputime_t *u_dst, cputime_t *s_dst,
 783                    cputime_t *u_src, cputime_t *s_src,
 784                    cputime_t *udelta, cputime_t *sdelta)
 785 {
 786         unsigned int seq;
 787         unsigned long long delta;
 788
 789         do {
 790                 *udelta = 0;
 791                 *sdelta = 0;
 792
 793                 seq = read_seqbegin(&t->vtime_seqlock);
 794
 795                 if (u_dst)
 796                         *u_dst = *u_src;
 797                 if (s_dst)
 798                         *s_dst = *s_src;
 799
 800                 /* Task is sleeping, nothing to add */
 801                 if (t->vtime_snap_whence == VTIME_SLEEPING ||
 802                     is_idle_task(t))
 803                         continue;
 804
 805                 delta = vtime_delta(t);
 806
 807                 /*
 808                  * Task runs either in user or kernel space, add pending nohz time to
 809                  * the right place.
 810                  */
 811                 if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU) {
 812                         *udelta = delta;
 813                 } else {
 814                         if (t->vtime_snap_whence == VTIME_SYS)
 815                                 *sdelta = delta;
 816                 }
 817         } while (read_seqretry(&t->vtime_seqlock, seq));
 818 }
 819
 820
 821 void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
 822 {
 823         cputime_t udelta, sdelta;
 824
 825         fetch_task_cputime(t, utime, stime, &t->utime,
 826                            &t->stime, &udelta, &sdelta);
 827         if (utime)
 828                 *utime += udelta;
 829         if (stime)
 830                 *stime += sdelta;
 831 }
 832
 833 void task_cputime_scaled(struct task_struct *t,
 834                          cputime_t *utimescaled, cputime_t *stimescaled)
 835 {
 836         cputime_t udelta, sdelta;
 837
 838         fetch_task_cputime(t, utimescaled, stimescaled,
 839                            &t->utimescaled, &t->stimescaled, &udelta, &sdelta);
 840         if (utimescaled)
 841                 *utimescaled += cputime_to_scaled(udelta);
 842         if (stimescaled)
 843                 *stimescaled += cputime_to_scaled(sdelta);
 844 }
 845 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */