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(t, &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 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 392 /*
 393  * Account a single tick of cpu time.
 394  * @p: the process that the cpu time gets accounted to
 395  * @user_tick: indicates if the tick is a user or a system tick
 396  */
 397 void account_process_tick(struct task_struct *p, int user_tick)
 398 {
 399         cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
 400         struct rq *rq = this_rq();
 401
 402         if (vtime_accounting_enabled())
 403                 return;
 404
 405         if (sched_clock_irqtime) {
 406                 irqtime_account_process_tick(p, user_tick, rq);
 407                 return;
 408         }
 409
 410         if (steal_account_process_tick())
 411                 return;
 412
 413         if (user_tick)
 414                 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
 415         else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
 416                 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
 417                                     one_jiffy_scaled);
 418         else
 419                 account_idle_time(cputime_one_jiffy);
 420 }
 421
 422 /*
 423  * Account multiple ticks of steal time.
 424  * @p: the process from which the cpu time has been stolen
 425  * @ticks: number of stolen ticks
 426  */
 427 void account_steal_ticks(unsigned long ticks)
 428 {
 429         account_steal_time(jiffies_to_cputime(ticks));
 430 }
 431
 432 /*
 433  * Account multiple ticks of idle time.
 434  * @ticks: number of stolen ticks
 435  */
 436 void account_idle_ticks(unsigned long ticks)
 437 {
 438
 439         if (sched_clock_irqtime) {
 440                 irqtime_account_idle_ticks(ticks);
 441                 return;
 442         }
 443
 444         account_idle_time(jiffies_to_cputime(ticks));
 445 }
 446 #endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 447
 448 /*
 449  * Use precise platform statistics if available:
 450  */
 451 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
 452 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 453 {
 454         *ut = p->utime;
 455         *st = p->stime;
 456 }
 457
 458 void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 459 {
 460         struct task_cputime cputime;
 461
 462         thread_group_cputime(p, &cputime);
 463
 464         *ut = cputime.utime;
 465         *st = cputime.stime;
 466 }
 467
 468 #ifndef __ARCH_HAS_VTIME_TASK_SWITCH
 469 void vtime_task_switch(struct task_struct *prev)
 470 {
 471         if (!vtime_accounting_enabled())
 472                 return;
 473
 474         if (is_idle_task(prev))
 475                 vtime_account_idle(prev);
 476         else
 477                 vtime_account_system(prev);
 478
 479 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 480         vtime_account_user(prev);
 481 #endif
 482         arch_vtime_task_switch(prev);
 483 }
 484 #endif
 485
 486 /*
 487  * Archs that account the whole time spent in the idle task
 488  * (outside irq) as idle time can rely on this and just implement
 489  * vtime_account_system() and vtime_account_idle(). Archs that
 490  * have other meaning of the idle time (s390 only includes the
 491  * time spent by the CPU when it's in low power mode) must override
 492  * vtime_account().
 493  */
 494 #ifndef __ARCH_HAS_VTIME_ACCOUNT
 495 void vtime_account_irq_enter(struct task_struct *tsk)
 496 {
 497         if (!vtime_accounting_enabled())
 498                 return;
 499
 500         if (!in_interrupt()) {
 501                 /*
 502                  * If we interrupted user, context_tracking_in_user()
 503                  * is 1 because the context tracking don't hook
 504                  * on irq entry/exit. This way we know if
 505                  * we need to flush user time on kernel entry.
 506                  */
 507                 if (context_tracking_in_user()) {
 508                         vtime_account_user(tsk);
 509                         return;
 510                 }
 511
 512                 if (is_idle_task(tsk)) {
 513                         vtime_account_idle(tsk);
 514                         return;
 515                 }
 516         }
 517         vtime_account_system(tsk);
 518 }
 519 EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
 520 #endif /* __ARCH_HAS_VTIME_ACCOUNT */
 521
 522 #else /* !CONFIG_VIRT_CPU_ACCOUNTING */
 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         stime = curr->stime;
 567         total = stime + curr->utime;
 568
 569         /*
 570          * Tick based cputime accounting depend on random scheduling
 571          * timeslices of a task to be interrupted or not by the timer.
 572          * Depending on these circumstances, the number of these interrupts
 573          * may be over or under-optimistic, matching the real user and system
 574          * cputime with a variable precision.
 575          *
 576          * Fix this by scaling these tick based values against the total
 577          * runtime accounted by the CFS scheduler.
 578          */
 579         rtime = nsecs_to_cputime(curr->sum_exec_runtime);
 580
 581         if (!rtime) {
 582                 stime = 0;
 583         } else if (!total) {
 584                 stime = rtime;
 585         } else {
 586                 stime = scale_stime((__force u64)stime,
 587                                     (__force u64)rtime, (__force u64)total);
 588         }
 589
 590         /*
 591          * If the tick based count grows faster than the scheduler one,
 592          * the result of the scaling may go backward.
 593          * Let's enforce monotonicity.
 594          */
 595         prev->stime = max(prev->stime, stime);
 596         prev->utime = max(prev->utime, rtime - prev->stime);
 597
 598         *ut = prev->utime;
 599         *st = prev->stime;
 600 }
 601
 602 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 603 {
 604         struct task_cputime cputime = {
 605                 .sum_exec_runtime = p->se.sum_exec_runtime,
 606         };
 607
 608         task_cputime(p, &cputime.utime, &cputime.stime);
 609         cputime_adjust(&cputime, &p->prev_cputime, ut, st);
 610 }
 611
 612 /*
 613  * Must be called with siglock held.
 614  */
 615 void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 616 {
 617         struct task_cputime cputime;
 618
 619         thread_group_cputime(p, &cputime);
 620         cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
 621 }
 622 #endif /* !CONFIG_VIRT_CPU_ACCOUNTING */
 623
 624 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
 625 static unsigned long long vtime_delta(struct task_struct *tsk)
 626 {
 627         unsigned long long clock;
 628
 629         clock = local_clock();
 630         if (clock < tsk->vtime_snap)
 631                 return 0;
 632
 633         return clock - tsk->vtime_snap;
 634 }
 635
 636 static cputime_t get_vtime_delta(struct task_struct *tsk)
 637 {
 638         unsigned long long delta = vtime_delta(tsk);
 639
 640         WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_SLEEPING);
 641         tsk->vtime_snap += delta;
 642
 643         /* CHECKME: always safe to convert nsecs to cputime? */
 644         return nsecs_to_cputime(delta);
 645 }
 646
 647 static void __vtime_account_system(struct task_struct *tsk)
 648 {
 649         cputime_t delta_cpu = get_vtime_delta(tsk);
 650
 651         account_system_time(tsk, irq_count(), delta_cpu, cputime_to_scaled(delta_cpu));
 652 }
 653
 654 void vtime_account_system(struct task_struct *tsk)
 655 {
 656         if (!vtime_accounting_enabled())
 657                 return;
 658
 659         write_seqlock(&tsk->vtime_seqlock);
 660         __vtime_account_system(tsk);
 661         write_sequnlock(&tsk->vtime_seqlock);
 662 }
 663
 664 void vtime_account_irq_exit(struct task_struct *tsk)
 665 {
 666         if (!vtime_accounting_enabled())
 667                 return;
 668
 669         write_seqlock(&tsk->vtime_seqlock);
 670         if (context_tracking_in_user())
 671                 tsk->vtime_snap_whence = VTIME_USER;
 672         __vtime_account_system(tsk);
 673         write_sequnlock(&tsk->vtime_seqlock);
 674 }
 675
 676 void vtime_account_user(struct task_struct *tsk)
 677 {
 678         cputime_t delta_cpu;
 679
 680         if (!vtime_accounting_enabled())
 681                 return;
 682
 683         delta_cpu = get_vtime_delta(tsk);
 684
 685         write_seqlock(&tsk->vtime_seqlock);
 686         tsk->vtime_snap_whence = VTIME_SYS;
 687         account_user_time(tsk, delta_cpu, cputime_to_scaled(delta_cpu));
 688         write_sequnlock(&tsk->vtime_seqlock);
 689 }
 690
 691 void vtime_user_enter(struct task_struct *tsk)
 692 {
 693         if (!vtime_accounting_enabled())
 694                 return;
 695
 696         write_seqlock(&tsk->vtime_seqlock);
 697         tsk->vtime_snap_whence = VTIME_USER;
 698         __vtime_account_system(tsk);
 699         write_sequnlock(&tsk->vtime_seqlock);
 700 }
 701
 702 void vtime_guest_enter(struct task_struct *tsk)
 703 {
 704         write_seqlock(&tsk->vtime_seqlock);
 705         __vtime_account_system(tsk);
 706         current->flags |= PF_VCPU;
 707         write_sequnlock(&tsk->vtime_seqlock);
 708 }
 709
 710 void vtime_guest_exit(struct task_struct *tsk)
 711 {
 712         write_seqlock(&tsk->vtime_seqlock);
 713         __vtime_account_system(tsk);
 714         current->flags &= ~PF_VCPU;
 715         write_sequnlock(&tsk->vtime_seqlock);
 716 }
 717
 718 void vtime_account_idle(struct task_struct *tsk)
 719 {
 720         cputime_t delta_cpu = get_vtime_delta(tsk);
 721
 722         account_idle_time(delta_cpu);
 723 }
 724
 725 bool vtime_accounting_enabled(void)
 726 {
 727         return context_tracking_active();
 728 }
 729
 730 void arch_vtime_task_switch(struct task_struct *prev)
 731 {
 732         write_seqlock(&prev->vtime_seqlock);
 733         prev->vtime_snap_whence = VTIME_SLEEPING;
 734         write_sequnlock(&prev->vtime_seqlock);
 735
 736         write_seqlock(&current->vtime_seqlock);
 737         current->vtime_snap_whence = VTIME_SYS;
 738         current->vtime_snap = sched_clock();
 739         write_sequnlock(&current->vtime_seqlock);
 740 }
 741
 742 void vtime_init_idle(struct task_struct *t)
 743 {
 744         unsigned long flags;
 745
 746         write_seqlock_irqsave(&t->vtime_seqlock, flags);
 747         t->vtime_snap_whence = VTIME_SYS;
 748         t->vtime_snap = sched_clock();
 749         write_sequnlock_irqrestore(&t->vtime_seqlock, flags);
 750 }
 751
 752 cputime_t task_gtime(struct task_struct *t)
 753 {
 754         unsigned int seq;
 755         cputime_t gtime;
 756
 757         do {
 758                 seq = read_seqbegin(&t->vtime_seqlock);
 759
 760                 gtime = t->gtime;
 761                 if (t->flags & PF_VCPU)
 762                         gtime += vtime_delta(t);
 763
 764         } while (read_seqretry(&t->vtime_seqlock, seq));
 765
 766         return gtime;
 767 }
 768
 769 /*
 770  * Fetch cputime raw values from fields of task_struct and
 771  * add up the pending nohz execution time since the last
 772  * cputime snapshot.
 773  */
 774 static void
 775 fetch_task_cputime(struct task_struct *t,
 776                    cputime_t *u_dst, cputime_t *s_dst,
 777                    cputime_t *u_src, cputime_t *s_src,
 778                    cputime_t *udelta, cputime_t *sdelta)
 779 {
 780         unsigned int seq;
 781         unsigned long long delta;
 782
 783         do {
 784                 *udelta = 0;
 785                 *sdelta = 0;
 786
 787                 seq = read_seqbegin(&t->vtime_seqlock);
 788
 789                 if (u_dst)
 790                         *u_dst = *u_src;
 791                 if (s_dst)
 792                         *s_dst = *s_src;
 793
 794                 /* Task is sleeping, nothing to add */
 795                 if (t->vtime_snap_whence == VTIME_SLEEPING ||
 796                     is_idle_task(t))
 797                         continue;
 798
 799                 delta = vtime_delta(t);
 800
 801                 /*
 802                  * Task runs either in user or kernel space, add pending nohz time to
 803                  * the right place.
 804                  */
 805                 if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU) {
 806                         *udelta = delta;
 807                 } else {
 808                         if (t->vtime_snap_whence == VTIME_SYS)
 809                                 *sdelta = delta;
 810                 }
 811         } while (read_seqretry(&t->vtime_seqlock, seq));
 812 }
 813
 814
 815 void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
 816 {
 817         cputime_t udelta, sdelta;
 818
 819         fetch_task_cputime(t, utime, stime, &t->utime,
 820                            &t->stime, &udelta, &sdelta);
 821         if (utime)
 822                 *utime += udelta;
 823         if (stime)
 824                 *stime += sdelta;
 825 }
 826
 827 void task_cputime_scaled(struct task_struct *t,
 828                          cputime_t *utimescaled, cputime_t *stimescaled)
 829 {
 830         cputime_t udelta, sdelta;
 831
 832         fetch_task_cputime(t, utimescaled, stimescaled,
 833                            &t->utimescaled, &t->stimescaled, &udelta, &sdelta);
 834         if (utimescaled)
 835                 *utimescaled += cputime_to_scaled(udelta);
 836         if (stimescaled)
 837                 *stimescaled += cputime_to_scaled(sdelta);
 838 }
 839 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */