kernel/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         /*
 119          * Since all updates are sure to touch the root cgroup, we
 120          * get ourselves ahead and touch it first. If the root cgroup
 121          * is the only cgroup, then nothing else should be necessary.
 122          *
 123          */
 124         __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
 125
 126         cpuacct_account_field(p, index, tmp);
 127 }
 128
 129 /*
 130  * Account user cpu time to a process.
 131  * @p: the process that the cpu time gets accounted to
 132  * @cputime: the cpu time spent in user space since the last update
 133  * @cputime_scaled: cputime scaled by cpu frequency
 134  */
 135 void account_user_time(struct task_struct *p, cputime_t cputime,
 136                        cputime_t cputime_scaled)
 137 {
 138         int index;
 139
 140         /* Add user time to process. */
 141         p->utime += cputime;
 142         p->utimescaled += cputime_scaled;
 143         account_group_user_time(p, cputime);
 144
 145         index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
 146
 147         /* Add user time to cpustat. */
 148         task_group_account_field(p, index, (__force u64) cputime);
 149
 150         /* Account for user time used */
 151         acct_account_cputime(p);
 152 }
 153
 154 /*
 155  * Account guest cpu time to a process.
 156  * @p: the process that the cpu time gets accounted to
 157  * @cputime: the cpu time spent in virtual machine since the last update
 158  * @cputime_scaled: cputime scaled by cpu frequency
 159  */
 160 static void account_guest_time(struct task_struct *p, cputime_t cputime,
 161                                cputime_t cputime_scaled)
 162 {
 163         u64 *cpustat = kcpustat_this_cpu->cpustat;
 164
 165         /* Add guest time to process. */
 166         p->utime += cputime;
 167         p->utimescaled += cputime_scaled;
 168         account_group_user_time(p, cputime);
 169         p->gtime += cputime;
 170
 171         /* Add guest time to cpustat. */
 172         if (task_nice(p) > 0) {
 173                 cpustat[CPUTIME_NICE] += (__force u64) cputime;
 174                 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
 175         } else {
 176                 cpustat[CPUTIME_USER] += (__force u64) cputime;
 177                 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
 178         }
 179 }
 180
 181 /*
 182  * Account system cpu time to a process and desired cpustat field
 183  * @p: the process that the cpu time gets accounted to
 184  * @cputime: the cpu time spent in kernel space since the last update
 185  * @cputime_scaled: cputime scaled by cpu frequency
 186  * @target_cputime64: pointer to cpustat field that has to be updated
 187  */
 188 static inline
 189 void __account_system_time(struct task_struct *p, cputime_t cputime,
 190                         cputime_t cputime_scaled, int index)
 191 {
 192         /* Add system time to process. */
 193         p->stime += cputime;
 194         p->stimescaled += cputime_scaled;
 195         account_group_system_time(p, cputime);
 196
 197         /* Add system time to cpustat. */
 198         task_group_account_field(p, index, (__force u64) cputime);
 199
 200         /* Account for system time used */
 201         acct_account_cputime(p);
 202 }
 203
 204 /*
 205  * Account system cpu time to a process.
 206  * @p: the process that the cpu time gets accounted to
 207  * @hardirq_offset: the offset to subtract from hardirq_count()
 208  * @cputime: the cpu time spent in kernel space since the last update
 209  * @cputime_scaled: cputime scaled by cpu frequency
 210  */
 211 void account_system_time(struct task_struct *p, int hardirq_offset,
 212                          cputime_t cputime, cputime_t cputime_scaled)
 213 {
 214         int index;
 215
 216         if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
 217                 account_guest_time(p, cputime, cputime_scaled);
 218                 return;
 219         }
 220
 221         if (hardirq_count() - hardirq_offset)
 222                 index = CPUTIME_IRQ;
 223         else if (in_serving_softirq())
 224                 index = CPUTIME_SOFTIRQ;
 225         else
 226                 index = CPUTIME_SYSTEM;
 227
 228         __account_system_time(p, cputime, cputime_scaled, index);
 229 }
 230
 231 /*
 232  * Account for involuntary wait time.
 233  * @cputime: the cpu time spent in involuntary wait
 234  */
 235 void account_steal_time(cputime_t cputime)
 236 {
 237         u64 *cpustat = kcpustat_this_cpu->cpustat;
 238
 239         cpustat[CPUTIME_STEAL] += (__force u64) cputime;
 240 }
 241
 242 /*
 243  * Account for idle time.
 244  * @cputime: the cpu time spent in idle wait
 245  */
 246 void account_idle_time(cputime_t cputime)
 247 {
 248         u64 *cpustat = kcpustat_this_cpu->cpustat;
 249         struct rq *rq = this_rq();
 250
 251         if (atomic_read(&rq->nr_iowait) > 0)
 252                 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
 253         else
 254                 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
 255 }
 256
 257 static __always_inline bool steal_account_process_tick(void)
 258 {
 259 #ifdef CONFIG_PARAVIRT
 260         if (static_key_false(&paravirt_steal_enabled)) {
 261                 u64 steal;
 262                 cputime_t steal_ct;
 263
 264                 steal = paravirt_steal_clock(smp_processor_id());
 265                 steal -= this_rq()->prev_steal_time;
 266
 267                 /*
 268                  * cputime_t may be less precise than nsecs (eg: if it's
 269                  * based on jiffies). Lets cast the result to cputime
 270                  * granularity and account the rest on the next rounds.
 271                  */
 272                 steal_ct = nsecs_to_cputime(steal);
 273                 this_rq()->prev_steal_time += cputime_to_nsecs(steal_ct);
 274
 275                 account_steal_time(steal_ct);
 276                 return steal_ct;
 277         }
 278 #endif
 279         return false;
 280 }
 281
 282 /*
 283  * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
 284  * tasks (sum on group iteration) belonging to @tsk's group.
 285  */
 286 void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
 287 {
 288         struct signal_struct *sig = tsk->signal;
 289         cputime_t utime, stime;
 290         struct task_struct *t;
 291         unsigned int seq, nextseq;
 292         unsigned long flags;
 293
 294         rcu_read_lock();
 295         /* Attempt a lockless read on the first round. */
 296         nextseq = 0;
 297         do {
 298                 seq = nextseq;
 299                 flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
 300                 times->utime = sig->utime;
 301                 times->stime = sig->stime;
 302                 times->sum_exec_runtime = sig->sum_sched_runtime;
 303
 304                 for_each_thread(tsk, t) {
 305                         task_cputime(t, &utime, &stime);
 306                         times->utime += utime;
 307                         times->stime += stime;
 308                         times->sum_exec_runtime += task_sched_runtime(t);
 309                 }
 310                 /* If lockless access failed, take the lock. */
 311                 nextseq = 1;
 312         } while (need_seqretry(&sig->stats_lock, seq));
 313         done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
 314         rcu_read_unlock();
 315 }
 316
 317 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
 318 /*
 319  * Account a tick to a process and cpustat
 320  * @p: the process that the cpu time gets accounted to
 321  * @user_tick: is the tick from userspace
 322  * @rq: the pointer to rq
 323  *
 324  * Tick demultiplexing follows the order
 325  * - pending hardirq update
 326  * - pending softirq update
 327  * - user_time
 328  * - idle_time
 329  * - system time
 330  *   - check for guest_time
 331  *   - else account as system_time
 332  *
 333  * Check for hardirq is done both for system and user time as there is
 334  * no timer going off while we are on hardirq and hence we may never get an
 335  * opportunity to update it solely in system time.
 336  * p->stime and friends are only updated on system time and not on irq
 337  * softirq as those do not count in task exec_runtime any more.
 338  */
 339 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
 340                                          struct rq *rq, int ticks)
 341 {
 342         cputime_t scaled = cputime_to_scaled(cputime_one_jiffy);
 343         u64 cputime = (__force u64) cputime_one_jiffy;
 344         u64 *cpustat = kcpustat_this_cpu->cpustat;
 345
 346         if (steal_account_process_tick())
 347                 return;
 348
 349         cputime *= ticks;
 350         scaled *= ticks;
 351
 352         if (irqtime_account_hi_update()) {
 353                 cpustat[CPUTIME_IRQ] += cputime;
 354         } else if (irqtime_account_si_update()) {
 355                 cpustat[CPUTIME_SOFTIRQ] += cputime;
 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, scaled, CPUTIME_SOFTIRQ);
 363         } else if (user_tick) {
 364                 account_user_time(p, cputime, scaled);
 365         } else if (p == rq->idle) {
 366                 account_idle_time(cputime);
 367         } else if (p->flags & PF_VCPU) { /* System time or guest time */
 368                 account_guest_time(p, cputime, scaled);
 369         } else {
 370                 __account_system_time(p, cputime, scaled,       CPUTIME_SYSTEM);
 371         }
 372 }
 373
 374 static void irqtime_account_idle_ticks(int ticks)
 375 {
 376         struct rq *rq = this_rq();
 377
 378         irqtime_account_process_tick(current, 0, rq, ticks);
 379 }
 380 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
 381 static inline void irqtime_account_idle_ticks(int ticks) {}
 382 static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
 383                                                 struct rq *rq, int nr_ticks) {}
 384 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
 385
 386 /*
 387  * Use precise platform statistics if available:
 388  */
 389 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
 390
 391 #ifndef __ARCH_HAS_VTIME_TASK_SWITCH
 392 void vtime_common_task_switch(struct task_struct *prev)
 393 {
 394         if (is_idle_task(prev))
 395                 vtime_account_idle(prev);
 396         else
 397                 vtime_account_system(prev);
 398
 399 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 400         vtime_account_user(prev);
 401 #endif
 402         arch_vtime_task_switch(prev);
 403 }
 404 #endif
 405
 406 /*
 407  * Archs that account the whole time spent in the idle task
 408  * (outside irq) as idle time can rely on this and just implement
 409  * vtime_account_system() and vtime_account_idle(). Archs that
 410  * have other meaning of the idle time (s390 only includes the
 411  * time spent by the CPU when it's in low power mode) must override
 412  * vtime_account().
 413  */
 414 #ifndef __ARCH_HAS_VTIME_ACCOUNT
 415 void vtime_common_account_irq_enter(struct task_struct *tsk)
 416 {
 417         if (!in_interrupt()) {
 418                 /*
 419                  * If we interrupted user, context_tracking_in_user()
 420                  * is 1 because the context tracking don't hook
 421                  * on irq entry/exit. This way we know if
 422                  * we need to flush user time on kernel entry.
 423                  */
 424                 if (context_tracking_in_user()) {
 425                         vtime_account_user(tsk);
 426                         return;
 427                 }
 428
 429                 if (is_idle_task(tsk)) {
 430                         vtime_account_idle(tsk);
 431                         return;
 432                 }
 433         }
 434         vtime_account_system(tsk);
 435 }
 436 EXPORT_SYMBOL_GPL(vtime_common_account_irq_enter);
 437 #endif /* __ARCH_HAS_VTIME_ACCOUNT */
 438 #endif /* CONFIG_VIRT_CPU_ACCOUNTING */
 439
 440
 441 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 442 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 443 {
 444         *ut = p->utime;
 445         *st = p->stime;
 446 }
 447
 448 void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 449 {
 450         struct task_cputime cputime;
 451
 452         thread_group_cputime(p, &cputime);
 453
 454         *ut = cputime.utime;
 455         *st = cputime.stime;
 456 }
 457 #else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 458 /*
 459  * Account a single tick of cpu time.
 460  * @p: the process that the cpu time gets accounted to
 461  * @user_tick: indicates if the tick is a user or a system tick
 462  */
 463 void account_process_tick(struct task_struct *p, int user_tick)
 464 {
 465         cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
 466         struct rq *rq = this_rq();
 467
 468         if (vtime_accounting_enabled())
 469                 return;
 470
 471         if (sched_clock_irqtime) {
 472                 irqtime_account_process_tick(p, user_tick, rq, 1);
 473                 return;
 474         }
 475
 476         if (steal_account_process_tick())
 477                 return;
 478
 479         if (user_tick)
 480                 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
 481         else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
 482                 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
 483                                     one_jiffy_scaled);
 484         else
 485                 account_idle_time(cputime_one_jiffy);
 486 }
 487
 488 /*
 489  * Account multiple ticks of steal time.
 490  * @p: the process from which the cpu time has been stolen
 491  * @ticks: number of stolen ticks
 492  */
 493 void account_steal_ticks(unsigned long ticks)
 494 {
 495         account_steal_time(jiffies_to_cputime(ticks));
 496 }
 497
 498 /*
 499  * Account multiple ticks of idle time.
 500  * @ticks: number of stolen ticks
 501  */
 502 void account_idle_ticks(unsigned long ticks)
 503 {
 504
 505         if (sched_clock_irqtime) {
 506                 irqtime_account_idle_ticks(ticks);
 507                 return;
 508         }
 509
 510         account_idle_time(jiffies_to_cputime(ticks));
 511 }
 512
 513 /*
 514  * Perform (stime * rtime) / total, but avoid multiplication overflow by
 515  * loosing precision when the numbers are big.
 516  */
 517 static cputime_t scale_stime(u64 stime, u64 rtime, u64 total)
 518 {
 519         u64 scaled;
 520
 521         for (;;) {
 522                 /* Make sure "rtime" is the bigger of stime/rtime */
 523                 if (stime > rtime)
 524                         swap(rtime, stime);
 525
 526                 /* Make sure 'total' fits in 32 bits */
 527                 if (total >> 32)
 528                         goto drop_precision;
 529
 530                 /* Does rtime (and thus stime) fit in 32 bits? */
 531                 if (!(rtime >> 32))
 532                         break;
 533
 534                 /* Can we just balance rtime/stime rather than dropping bits? */
 535                 if (stime >> 31)
 536                         goto drop_precision;
 537
 538                 /* We can grow stime and shrink rtime and try to make them both fit */
 539                 stime <<= 1;
 540                 rtime >>= 1;
 541                 continue;
 542
 543 drop_precision:
 544                 /* We drop from rtime, it has more bits than stime */
 545                 rtime >>= 1;
 546                 total >>= 1;
 547         }
 548
 549         /*
 550          * Make sure gcc understands that this is a 32x32->64 multiply,
 551          * followed by a 64/32->64 divide.
 552          */
 553         scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
 554         return (__force cputime_t) scaled;
 555 }
 556
 557 /*
 558  * Atomically advance counter to the new value. Interrupts, vcpu
 559  * scheduling, and scaling inaccuracies can cause cputime_advance
 560  * to be occasionally called with a new value smaller than counter.
 561  * Let's enforce atomicity.
 562  *
 563  * Normally a caller will only go through this loop once, or not
 564  * at all in case a previous caller updated counter the same jiffy.
 565  */
 566 static void cputime_advance(cputime_t *counter, cputime_t new)
 567 {
 568         cputime_t old;
 569
 570         while (new > (old = ACCESS_ONCE(*counter)))
 571                 cmpxchg_cputime(counter, old, new);
 572 }
 573
 574 /*
 575  * Adjust tick based cputime random precision against scheduler
 576  * runtime accounting.
 577  */
 578 static void cputime_adjust(struct task_cputime *curr,
 579                            struct cputime *prev,
 580                            cputime_t *ut, cputime_t *st)
 581 {
 582         cputime_t rtime, stime, utime;
 583
 584         /*
 585          * Tick based cputime accounting depend on random scheduling
 586          * timeslices of a task to be interrupted or not by the timer.
 587          * Depending on these circumstances, the number of these interrupts
 588          * may be over or under-optimistic, matching the real user and system
 589          * cputime with a variable precision.
 590          *
 591          * Fix this by scaling these tick based values against the total
 592          * runtime accounted by the CFS scheduler.
 593          */
 594         rtime = nsecs_to_cputime(curr->sum_exec_runtime);
 595
 596         /*
 597          * Update userspace visible utime/stime values only if actual execution
 598          * time is bigger than already exported. Note that can happen, that we
 599          * provided bigger values due to scaling inaccuracy on big numbers.
 600          */
 601         if (prev->stime + prev->utime >= rtime)
 602                 goto out;
 603
 604         stime = curr->stime;
 605         utime = curr->utime;
 606
 607         if (utime == 0) {
 608                 stime = rtime;
 609         } else if (stime == 0) {
 610                 utime = rtime;
 611         } else {
 612                 cputime_t total = stime + utime;
 613
 614                 stime = scale_stime((__force u64)stime,
 615                                     (__force u64)rtime, (__force u64)total);
 616                 utime = rtime - stime;
 617         }
 618
 619         cputime_advance(&prev->stime, stime);
 620         cputime_advance(&prev->utime, utime);
 621
 622 out:
 623         *ut = prev->utime;
 624         *st = prev->stime;
 625 }
 626
 627 void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 628 {
 629         struct task_cputime cputime = {
 630                 .sum_exec_runtime = p->se.sum_exec_runtime,
 631         };
 632
 633         task_cputime(p, &cputime.utime, &cputime.stime);
 634         cputime_adjust(&cputime, &p->prev_cputime, ut, st);
 635 }
 636
 637 void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
 638 {
 639         struct task_cputime cputime;
 640
 641         thread_group_cputime(p, &cputime);
 642         cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
 643 }
 644 #endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 645
 646 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
 647 static unsigned long long vtime_delta(struct task_struct *tsk)
 648 {
 649         unsigned long long clock;
 650
 651         clock = local_clock();
 652         if (clock < tsk->vtime_snap)
 653                 return 0;
 654
 655         return clock - tsk->vtime_snap;
 656 }
 657
 658 static cputime_t get_vtime_delta(struct task_struct *tsk)
 659 {
 660         unsigned long long delta = vtime_delta(tsk);
 661
 662         WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_SLEEPING);
 663         tsk->vtime_snap += delta;
 664
 665         /* CHECKME: always safe to convert nsecs to cputime? */
 666         return nsecs_to_cputime(delta);
 667 }
 668
 669 static void __vtime_account_system(struct task_struct *tsk)
 670 {
 671         cputime_t delta_cpu = get_vtime_delta(tsk);
 672
 673         account_system_time(tsk, irq_count(), delta_cpu, cputime_to_scaled(delta_cpu));
 674 }
 675
 676 void vtime_account_system(struct task_struct *tsk)
 677 {
 678         raw_spin_lock(&tsk->vtime_lock);
 679         write_seqcount_begin(&tsk->vtime_seq);
 680         __vtime_account_system(tsk);
 681         write_seqcount_end(&tsk->vtime_seq);
 682         raw_spin_unlock(&tsk->vtime_lock);
 683 }
 684
 685 void vtime_gen_account_irq_exit(struct task_struct *tsk)
 686 {
 687         raw_spin_lock(&tsk->vtime_lock);
 688         write_seqcount_begin(&tsk->vtime_seq);
 689         __vtime_account_system(tsk);
 690         if (context_tracking_in_user())
 691                 tsk->vtime_snap_whence = VTIME_USER;
 692         write_seqcount_end(&tsk->vtime_seq);
 693         raw_spin_unlock(&tsk->vtime_lock);
 694 }
 695
 696 void vtime_account_user(struct task_struct *tsk)
 697 {
 698         cputime_t delta_cpu;
 699
 700         raw_spin_lock(&tsk->vtime_lock);
 701         write_seqcount_begin(&tsk->vtime_seq);
 702         delta_cpu = get_vtime_delta(tsk);
 703         tsk->vtime_snap_whence = VTIME_SYS;
 704         account_user_time(tsk, delta_cpu, cputime_to_scaled(delta_cpu));
 705         write_seqcount_end(&tsk->vtime_seq);
 706         raw_spin_unlock(&tsk->vtime_lock);
 707 }
 708
 709 void vtime_user_enter(struct task_struct *tsk)
 710 {
 711         raw_spin_lock(&tsk->vtime_lock);
 712         write_seqcount_begin(&tsk->vtime_seq);
 713         __vtime_account_system(tsk);
 714         tsk->vtime_snap_whence = VTIME_USER;
 715         write_seqcount_end(&tsk->vtime_seq);
 716         raw_spin_unlock(&tsk->vtime_lock);
 717 }
 718
 719 void vtime_guest_enter(struct task_struct *tsk)
 720 {
 721         /*
 722          * The flags must be updated under the lock with
 723          * the vtime_snap flush and update.
 724          * That enforces a right ordering and update sequence
 725          * synchronization against the reader (task_gtime())
 726          * that can thus safely catch up with a tickless delta.
 727          */
 728         raw_spin_lock(&tsk->vtime_lock);
 729         write_seqcount_begin(&tsk->vtime_seq);
 730         __vtime_account_system(tsk);
 731         current->flags |= PF_VCPU;
 732         write_seqcount_end(&tsk->vtime_seq);
 733         raw_spin_unlock(&tsk->vtime_lock);
 734 }
 735 EXPORT_SYMBOL_GPL(vtime_guest_enter);
 736
 737 void vtime_guest_exit(struct task_struct *tsk)
 738 {
 739         raw_spin_lock(&tsk->vtime_lock);
 740         write_seqcount_begin(&tsk->vtime_seq);
 741         __vtime_account_system(tsk);
 742         current->flags &= ~PF_VCPU;
 743         write_seqcount_end(&tsk->vtime_seq);
 744         raw_spin_unlock(&tsk->vtime_lock);
 745 }
 746 EXPORT_SYMBOL_GPL(vtime_guest_exit);
 747
 748 void vtime_account_idle(struct task_struct *tsk)
 749 {
 750         cputime_t delta_cpu = get_vtime_delta(tsk);
 751
 752         account_idle_time(delta_cpu);
 753 }
 754
 755 void arch_vtime_task_switch(struct task_struct *prev)
 756 {
 757         raw_spin_lock(&prev->vtime_lock);
 758         write_seqcount_begin(&prev->vtime_seq);
 759         prev->vtime_snap_whence = VTIME_SLEEPING;
 760         write_seqcount_end(&prev->vtime_seq);
 761         raw_spin_unlock(&prev->vtime_lock);
 762
 763         raw_spin_lock(&current->vtime_lock);
 764         write_seqcount_begin(&current->vtime_seq);
 765         current->vtime_snap_whence = VTIME_SYS;
 766         current->vtime_snap = sched_clock_cpu(smp_processor_id());
 767         write_seqcount_end(&current->vtime_seq);
 768         raw_spin_unlock(&current->vtime_lock);
 769 }
 770
 771 void vtime_init_idle(struct task_struct *t, int cpu)
 772 {
 773         unsigned long flags;
 774
 775         raw_spin_lock_irqsave(&t->vtime_lock, flags);
 776         write_seqcount_begin(&t->vtime_seq);
 777         t->vtime_snap_whence = VTIME_SYS;
 778         t->vtime_snap = sched_clock_cpu(cpu);
 779         write_seqcount_end(&t->vtime_seq);
 780         raw_spin_unlock_irqrestore(&t->vtime_lock, flags);
 781 }
 782
 783 cputime_t task_gtime(struct task_struct *t)
 784 {
 785         unsigned int seq;
 786         cputime_t gtime;
 787
 788         do {
 789                 seq = read_seqcount_begin(&t->vtime_seq);
 790
 791                 gtime = t->gtime;
 792                 if (t->flags & PF_VCPU)
 793                         gtime += vtime_delta(t);
 794
 795         } while (read_seqcount_retry(&t->vtime_seq, seq));
 796
 797         return gtime;
 798 }
 799
 800 /*
 801  * Fetch cputime raw values from fields of task_struct and
 802  * add up the pending nohz execution time since the last
 803  * cputime snapshot.
 804  */
 805 static void
 806 fetch_task_cputime(struct task_struct *t,
 807                    cputime_t *u_dst, cputime_t *s_dst,
 808                    cputime_t *u_src, cputime_t *s_src,
 809                    cputime_t *udelta, cputime_t *sdelta)
 810 {
 811         unsigned int seq;
 812         unsigned long long delta;
 813
 814         do {
 815                 *udelta = 0;
 816                 *sdelta = 0;
 817
 818                 seq = read_seqcount_begin(&t->vtime_seq);
 819
 820                 if (u_dst)
 821                         *u_dst = *u_src;
 822                 if (s_dst)
 823                         *s_dst = *s_src;
 824
 825                 /* Task is sleeping, nothing to add */
 826                 if (t->vtime_snap_whence == VTIME_SLEEPING ||
 827                     is_idle_task(t))
 828                         continue;
 829
 830                 delta = vtime_delta(t);
 831
 832                 /*
 833                  * Task runs either in user or kernel space, add pending nohz time to
 834                  * the right place.
 835                  */
 836                 if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU) {
 837                         *udelta = delta;
 838                 } else {
 839                         if (t->vtime_snap_whence == VTIME_SYS)
 840                                 *sdelta = delta;
 841                 }
 842         } while (read_seqcount_retry(&t->vtime_seq, seq));
 843 }
 844
 845
 846 void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
 847 {
 848         cputime_t udelta, sdelta;
 849
 850         fetch_task_cputime(t, utime, stime, &t->utime,
 851                            &t->stime, &udelta, &sdelta);
 852         if (utime)
 853                 *utime += udelta;
 854         if (stime)
 855                 *stime += sdelta;
 856 }
 857
 858 void task_cputime_scaled(struct task_struct *t,
 859                          cputime_t *utimescaled, cputime_t *stimescaled)
 860 {
 861         cputime_t udelta, sdelta;
 862
 863         fetch_task_cputime(t, utimescaled, stimescaled,
 864                            &t->utimescaled, &t->stimescaled, &udelta, &sdelta);
 865         if (utimescaled)
 866                 *utimescaled += cputime_to_scaled(udelta);
 867         if (stimescaled)
 868                 *stimescaled += cputime_to_scaled(sdelta);
 869 }
 870 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */