blob: a95097cb4591b5bfa2466adb5600895e782fc661 [file] [log] [blame]
Frederic Weisbecker73fbec62012-06-16 15:57:37 +02001#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>
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +02006#include <linux/context_tracking.h>
Frederic Weisbecker73fbec62012-06-16 15:57:37 +02007#include "sched.h"
8
9
10#ifdef CONFIG_IRQ_TIME_ACCOUNTING
11
12/*
13 * There are no locks covering percpu hardirq/softirq time.
Frederic Weisbeckerbf9fae92012-09-08 15:23:11 +020014 * They are only modified in vtime_account, on corresponding CPU
Frederic Weisbecker73fbec62012-06-16 15:57:37 +020015 * 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
Frederic Weisbeckerbf9fae92012-09-08 15:23:11 +020018 * race with irq/vtime_account on this CPU. We would either get old
Frederic Weisbecker73fbec62012-06-16 15:57:37 +020019 * 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 */
23DEFINE_PER_CPU(u64, cpu_hardirq_time);
24DEFINE_PER_CPU(u64, cpu_softirq_time);
25
26static DEFINE_PER_CPU(u64, irq_start_time);
27static int sched_clock_irqtime;
28
29void enable_sched_clock_irqtime(void)
30{
31 sched_clock_irqtime = 1;
32}
33
34void disable_sched_clock_irqtime(void)
35{
36 sched_clock_irqtime = 0;
37}
38
39#ifndef CONFIG_64BIT
40DEFINE_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 */
Frederic Weisbecker3e1df4f52012-10-06 05:23:22 +020047void irqtime_account_irq(struct task_struct *curr)
Frederic Weisbecker73fbec62012-06-16 15:57:37 +020048{
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}
Frederic Weisbecker3e1df4f52012-10-06 05:23:22 +020077EXPORT_SYMBOL_GPL(irqtime_account_irq);
Frederic Weisbecker73fbec62012-06-16 15:57:37 +020078
79static 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
94static 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
115static inline void task_group_account_field(struct task_struct *p, int index,
116 u64 tmp)
117{
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200118 /*
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 */
Christoph Lametera4f61cc2013-08-07 15:38:24 +0000124 __this_cpu_add(kernel_cpustat.cpustat[index], tmp);
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200125
Li Zefan1966aaf2013-03-29 14:37:06 +0800126 cpuacct_account_field(p, index, tmp);
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200127}
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 */
135void 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
Dongsheng Yangd0ea0262014-01-27 22:00:45 -0500145 index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200146
147 /* Add user time to cpustat. */
148 task_group_account_field(p, index, (__force u64) cputime);
149
150 /* Account for user time used */
Frederic Weisbecker6fac4822012-11-13 14:20:55 +0100151 acct_account_cputime(p);
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200152}
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 */
160static 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. */
Dongsheng Yangd0ea0262014-01-27 22:00:45 -0500172 if (task_nice(p) > 0) {
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200173 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 */
188static inline
189void __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 */
Frederic Weisbecker6fac4822012-11-13 14:20:55 +0100201 acct_account_cputime(p);
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200202}
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 */
211void 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 */
235void 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 */
246void 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
257static __always_inline bool steal_account_process_tick(void)
258{
259#ifdef CONFIG_PARAVIRT
260 if (static_key_false(&paravirt_steal_enabled)) {
Frederic Weisbeckerdee08a72014-03-05 17:02:22 +0100261 u64 steal;
262 cputime_t steal_ct;
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200263
264 steal = paravirt_steal_clock(smp_processor_id());
265 steal -= this_rq()->prev_steal_time;
266
Frederic Weisbeckerdee08a72014-03-05 17:02:22 +0100267 /*
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);
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200274
Frederic Weisbeckerdee08a72014-03-05 17:02:22 +0100275 account_steal_time(steal_ct);
276 return steal_ct;
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200277 }
278#endif
279 return false;
280}
281
Frederic Weisbeckera634f932012-11-21 15:55:59 +0100282/*
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 */
286void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
287{
288 struct signal_struct *sig = tsk->signal;
Frederic Weisbecker6fac4822012-11-13 14:20:55 +0100289 cputime_t utime, stime;
Frederic Weisbeckera634f932012-11-21 15:55:59 +0100290 struct task_struct *t;
291
292 times->utime = sig->utime;
293 times->stime = sig->stime;
294 times->sum_exec_runtime = sig->sum_sched_runtime;
295
296 rcu_read_lock();
297 /* make sure we can trust tsk->thread_group list */
298 if (!likely(pid_alive(tsk)))
299 goto out;
300
301 t = tsk;
302 do {
Stanislaw Gruszkae614b332013-04-04 10:57:48 +0200303 task_cputime(t, &utime, &stime);
Frederic Weisbecker6fac4822012-11-13 14:20:55 +0100304 times->utime += utime;
305 times->stime += stime;
Frederic Weisbeckera634f932012-11-21 15:55:59 +0100306 times->sum_exec_runtime += task_sched_runtime(t);
307 } while_each_thread(tsk, t);
308out:
309 rcu_read_unlock();
310}
311
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200312#ifdef CONFIG_IRQ_TIME_ACCOUNTING
313/*
314 * Account a tick to a process and cpustat
315 * @p: the process that the cpu time gets accounted to
316 * @user_tick: is the tick from userspace
317 * @rq: the pointer to rq
318 *
319 * Tick demultiplexing follows the order
320 * - pending hardirq update
321 * - pending softirq update
322 * - user_time
323 * - idle_time
324 * - system time
325 * - check for guest_time
326 * - else account as system_time
327 *
328 * Check for hardirq is done both for system and user time as there is
329 * no timer going off while we are on hardirq and hence we may never get an
330 * opportunity to update it solely in system time.
331 * p->stime and friends are only updated on system time and not on irq
332 * softirq as those do not count in task exec_runtime any more.
333 */
334static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
335 struct rq *rq)
336{
337 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
338 u64 *cpustat = kcpustat_this_cpu->cpustat;
339
340 if (steal_account_process_tick())
341 return;
342
343 if (irqtime_account_hi_update()) {
344 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
345 } else if (irqtime_account_si_update()) {
346 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
347 } else if (this_cpu_ksoftirqd() == p) {
348 /*
349 * ksoftirqd time do not get accounted in cpu_softirq_time.
350 * So, we have to handle it separately here.
351 * Also, p->stime needs to be updated for ksoftirqd.
352 */
353 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
354 CPUTIME_SOFTIRQ);
355 } else if (user_tick) {
356 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
357 } else if (p == rq->idle) {
358 account_idle_time(cputime_one_jiffy);
359 } else if (p->flags & PF_VCPU) { /* System time or guest time */
360 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
361 } else {
362 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
363 CPUTIME_SYSTEM);
364 }
365}
366
367static void irqtime_account_idle_ticks(int ticks)
368{
369 int i;
370 struct rq *rq = this_rq();
371
372 for (i = 0; i < ticks; i++)
373 irqtime_account_process_tick(current, 0, rq);
374}
375#else /* CONFIG_IRQ_TIME_ACCOUNTING */
Frederic Weisbecker3f4724e2012-07-16 18:00:34 +0200376static inline void irqtime_account_idle_ticks(int ticks) {}
377static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200378 struct rq *rq) {}
379#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
380
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200381/*
382 * Use precise platform statistics if available:
383 */
384#ifdef CONFIG_VIRT_CPU_ACCOUNTING
Frederic Weisbeckera7e1a9e2012-09-08 16:14:02 +0200385
Frederic Weisbeckere3942ba2012-11-14 00:24:25 +0100386#ifndef __ARCH_HAS_VTIME_TASK_SWITCH
Frederic Weisbeckerb0493402013-07-12 03:10:15 +0200387void vtime_common_task_switch(struct task_struct *prev)
Frederic Weisbeckere3942ba2012-11-14 00:24:25 +0100388{
389 if (is_idle_task(prev))
390 vtime_account_idle(prev);
391 else
392 vtime_account_system(prev);
393
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200394#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
Frederic Weisbeckere3942ba2012-11-14 00:24:25 +0100395 vtime_account_user(prev);
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200396#endif
Frederic Weisbeckere3942ba2012-11-14 00:24:25 +0100397 arch_vtime_task_switch(prev);
398}
399#endif
Frederic Weisbecker11113332012-10-24 18:05:51 +0200400
Frederic Weisbeckera7e1a9e2012-09-08 16:14:02 +0200401/*
402 * Archs that account the whole time spent in the idle task
403 * (outside irq) as idle time can rely on this and just implement
Frederic Weisbeckerfd25b4c2012-11-13 18:21:22 +0100404 * vtime_account_system() and vtime_account_idle(). Archs that
Frederic Weisbeckera7e1a9e2012-09-08 16:14:02 +0200405 * have other meaning of the idle time (s390 only includes the
406 * time spent by the CPU when it's in low power mode) must override
407 * vtime_account().
408 */
409#ifndef __ARCH_HAS_VTIME_ACCOUNT
Frederic Weisbeckerb0493402013-07-12 03:10:15 +0200410void vtime_common_account_irq_enter(struct task_struct *tsk)
Frederic Weisbeckera7e1a9e2012-09-08 16:14:02 +0200411{
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200412 if (!in_interrupt()) {
413 /*
414 * If we interrupted user, context_tracking_in_user()
415 * is 1 because the context tracking don't hook
416 * on irq entry/exit. This way we know if
417 * we need to flush user time on kernel entry.
418 */
419 if (context_tracking_in_user()) {
420 vtime_account_user(tsk);
421 return;
422 }
423
424 if (is_idle_task(tsk)) {
425 vtime_account_idle(tsk);
426 return;
427 }
428 }
429 vtime_account_system(tsk);
Frederic Weisbeckera7e1a9e2012-09-08 16:14:02 +0200430}
Frederic Weisbeckerb0493402013-07-12 03:10:15 +0200431EXPORT_SYMBOL_GPL(vtime_common_account_irq_enter);
Frederic Weisbeckera7e1a9e2012-09-08 16:14:02 +0200432#endif /* __ARCH_HAS_VTIME_ACCOUNT */
Frederic Weisbecker9fbc42e2013-02-25 17:25:39 +0100433#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
Frederic Weisbeckera7e1a9e2012-09-08 16:14:02 +0200434
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200435
Frederic Weisbecker9fbc42e2013-02-25 17:25:39 +0100436#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
437void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200438{
Frederic Weisbecker9fbc42e2013-02-25 17:25:39 +0100439 *ut = p->utime;
440 *st = p->stime;
441}
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200442
Frederic Weisbecker9fbc42e2013-02-25 17:25:39 +0100443void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
444{
445 struct task_cputime cputime;
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200446
Frederic Weisbecker9fbc42e2013-02-25 17:25:39 +0100447 thread_group_cputime(p, &cputime);
448
449 *ut = cputime.utime;
450 *st = cputime.stime;
451}
452#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
453/*
454 * Account a single tick of cpu time.
455 * @p: the process that the cpu time gets accounted to
456 * @user_tick: indicates if the tick is a user or a system tick
457 */
458void account_process_tick(struct task_struct *p, int user_tick)
459{
460 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
461 struct rq *rq = this_rq();
462
463 if (vtime_accounting_enabled())
464 return;
465
466 if (sched_clock_irqtime) {
467 irqtime_account_process_tick(p, user_tick, rq);
468 return;
469 }
470
471 if (steal_account_process_tick())
472 return;
473
474 if (user_tick)
475 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
476 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
477 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
478 one_jiffy_scaled);
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200479 else
Frederic Weisbecker9fbc42e2013-02-25 17:25:39 +0100480 account_idle_time(cputime_one_jiffy);
481}
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200482
Frederic Weisbecker9fbc42e2013-02-25 17:25:39 +0100483/*
484 * Account multiple ticks of steal time.
485 * @p: the process from which the cpu time has been stolen
486 * @ticks: number of stolen ticks
487 */
488void account_steal_ticks(unsigned long ticks)
489{
490 account_steal_time(jiffies_to_cputime(ticks));
491}
492
493/*
494 * Account multiple ticks of idle time.
495 * @ticks: number of stolen ticks
496 */
497void account_idle_ticks(unsigned long ticks)
498{
499
500 if (sched_clock_irqtime) {
501 irqtime_account_idle_ticks(ticks);
502 return;
503 }
504
505 account_idle_time(jiffies_to_cputime(ticks));
506}
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200507
Frederic Weisbeckerd9a3c982013-02-20 18:54:55 +0100508/*
Stanislaw Gruszka55eaa7c2013-04-30 17:14:42 +0200509 * Perform (stime * rtime) / total, but avoid multiplication overflow by
510 * loosing precision when the numbers are big.
Frederic Weisbeckerd9a3c982013-02-20 18:54:55 +0100511 */
512static cputime_t scale_stime(u64 stime, u64 rtime, u64 total)
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200513{
Stanislaw Gruszka55eaa7c2013-04-30 17:14:42 +0200514 u64 scaled;
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200515
Stanislaw Gruszka55eaa7c2013-04-30 17:14:42 +0200516 for (;;) {
517 /* Make sure "rtime" is the bigger of stime/rtime */
Stanislaw Gruszka84f9f3a2013-05-02 15:34:33 +0200518 if (stime > rtime)
519 swap(rtime, stime);
Stanislaw Gruszka55eaa7c2013-04-30 17:14:42 +0200520
521 /* Make sure 'total' fits in 32 bits */
522 if (total >> 32)
523 goto drop_precision;
524
525 /* Does rtime (and thus stime) fit in 32 bits? */
526 if (!(rtime >> 32))
527 break;
528
529 /* Can we just balance rtime/stime rather than dropping bits? */
530 if (stime >> 31)
531 goto drop_precision;
532
533 /* We can grow stime and shrink rtime and try to make them both fit */
534 stime <<= 1;
535 rtime >>= 1;
536 continue;
537
538drop_precision:
539 /* We drop from rtime, it has more bits than stime */
540 rtime >>= 1;
541 total >>= 1;
Frederic Weisbeckerd9a3c982013-02-20 18:54:55 +0100542 }
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200543
Stanislaw Gruszka55eaa7c2013-04-30 17:14:42 +0200544 /*
545 * Make sure gcc understands that this is a 32x32->64 multiply,
546 * followed by a 64/32->64 divide.
547 */
548 scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
Frederic Weisbeckerd9a3c982013-02-20 18:54:55 +0100549 return (__force cputime_t) scaled;
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200550}
551
Frederic Weisbeckerfa092052012-11-28 17:00:57 +0100552/*
553 * Adjust tick based cputime random precision against scheduler
554 * runtime accounting.
555 */
Frederic Weisbeckerd37f761d2012-11-22 00:58:35 +0100556static void cputime_adjust(struct task_cputime *curr,
557 struct cputime *prev,
558 cputime_t *ut, cputime_t *st)
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200559{
Stanislaw Gruszka5a8e01f2013-09-04 15:16:03 +0200560 cputime_t rtime, stime, utime;
Frederic Weisbeckerfa092052012-11-28 17:00:57 +0100561
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200562 /*
Frederic Weisbeckerfa092052012-11-28 17:00:57 +0100563 * Tick based cputime accounting depend on random scheduling
564 * timeslices of a task to be interrupted or not by the timer.
565 * Depending on these circumstances, the number of these interrupts
566 * may be over or under-optimistic, matching the real user and system
567 * cputime with a variable precision.
568 *
569 * Fix this by scaling these tick based values against the total
570 * runtime accounted by the CFS scheduler.
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200571 */
Frederic Weisbeckerd37f761d2012-11-22 00:58:35 +0100572 rtime = nsecs_to_cputime(curr->sum_exec_runtime);
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200573
Stanislaw Gruszka772c8082013-04-30 11:35:05 +0200574 /*
575 * Update userspace visible utime/stime values only if actual execution
576 * time is bigger than already exported. Note that can happen, that we
577 * provided bigger values due to scaling inaccuracy on big numbers.
578 */
579 if (prev->stime + prev->utime >= rtime)
580 goto out;
581
Stanislaw Gruszka5a8e01f2013-09-04 15:16:03 +0200582 stime = curr->stime;
583 utime = curr->utime;
584
585 if (utime == 0) {
586 stime = rtime;
587 } else if (stime == 0) {
588 utime = rtime;
589 } else {
590 cputime_t total = stime + utime;
591
Frederic Weisbeckerd9a3c982013-02-20 18:54:55 +0100592 stime = scale_stime((__force u64)stime,
593 (__force u64)rtime, (__force u64)total);
Stanislaw Gruszka68aa8ef2013-04-30 11:35:06 +0200594 utime = rtime - stime;
Frederic Weisbeckerd9a3c982013-02-20 18:54:55 +0100595 }
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200596
597 /*
Frederic Weisbeckerfa092052012-11-28 17:00:57 +0100598 * If the tick based count grows faster than the scheduler one,
599 * the result of the scaling may go backward.
600 * Let's enforce monotonicity.
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200601 */
Frederic Weisbecker62188452013-01-26 17:19:42 +0100602 prev->stime = max(prev->stime, stime);
Stanislaw Gruszka68aa8ef2013-04-30 11:35:06 +0200603 prev->utime = max(prev->utime, utime);
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200604
Stanislaw Gruszka772c8082013-04-30 11:35:05 +0200605out:
Frederic Weisbeckerd37f761d2012-11-22 00:58:35 +0100606 *ut = prev->utime;
607 *st = prev->stime;
608}
609
610void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
611{
612 struct task_cputime cputime = {
Frederic Weisbeckerd37f761d2012-11-22 00:58:35 +0100613 .sum_exec_runtime = p->se.sum_exec_runtime,
614 };
615
Frederic Weisbecker6fac4822012-11-13 14:20:55 +0100616 task_cputime(p, &cputime.utime, &cputime.stime);
Frederic Weisbeckerd37f761d2012-11-22 00:58:35 +0100617 cputime_adjust(&cputime, &p->prev_cputime, ut, st);
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200618}
619
620/*
621 * Must be called with siglock held.
622 */
Frederic Weisbeckere80d0a1a2012-11-21 16:26:44 +0100623void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200624{
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200625 struct task_cputime cputime;
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200626
627 thread_group_cputime(p, &cputime);
Frederic Weisbeckerd37f761d2012-11-22 00:58:35 +0100628 cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
Frederic Weisbecker73fbec62012-06-16 15:57:37 +0200629}
Frederic Weisbecker9fbc42e2013-02-25 17:25:39 +0100630#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200631
632#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100633static unsigned long long vtime_delta(struct task_struct *tsk)
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200634{
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100635 unsigned long long clock;
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200636
Frederic Weisbecker7f6575f2013-02-23 17:28:45 +0100637 clock = local_clock();
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100638 if (clock < tsk->vtime_snap)
639 return 0;
640
641 return clock - tsk->vtime_snap;
642}
643
644static cputime_t get_vtime_delta(struct task_struct *tsk)
645{
646 unsigned long long delta = vtime_delta(tsk);
647
648 WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_SLEEPING);
649 tsk->vtime_snap += delta;
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200650
651 /* CHECKME: always safe to convert nsecs to cputime? */
652 return nsecs_to_cputime(delta);
653}
654
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100655static void __vtime_account_system(struct task_struct *tsk)
656{
657 cputime_t delta_cpu = get_vtime_delta(tsk);
658
659 account_system_time(tsk, irq_count(), delta_cpu, cputime_to_scaled(delta_cpu));
660}
661
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200662void vtime_account_system(struct task_struct *tsk)
663{
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100664 write_seqlock(&tsk->vtime_seqlock);
665 __vtime_account_system(tsk);
666 write_sequnlock(&tsk->vtime_seqlock);
667}
668
Frederic Weisbeckerb0493402013-07-12 03:10:15 +0200669void vtime_gen_account_irq_exit(struct task_struct *tsk)
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100670{
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100671 write_seqlock(&tsk->vtime_seqlock);
Frederic Weisbeckeraf2350b2013-07-15 16:35:55 +0200672 __vtime_account_system(tsk);
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100673 if (context_tracking_in_user())
674 tsk->vtime_snap_whence = VTIME_USER;
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100675 write_sequnlock(&tsk->vtime_seqlock);
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200676}
677
678void vtime_account_user(struct task_struct *tsk)
679{
Frederic Weisbecker3f4724e2012-07-16 18:00:34 +0200680 cputime_t delta_cpu;
681
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100682 write_seqlock(&tsk->vtime_seqlock);
Frederic Weisbecker54461562013-07-13 17:10:18 +0200683 delta_cpu = get_vtime_delta(tsk);
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100684 tsk->vtime_snap_whence = VTIME_SYS;
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200685 account_user_time(tsk, delta_cpu, cputime_to_scaled(delta_cpu));
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100686 write_sequnlock(&tsk->vtime_seqlock);
687}
688
689void vtime_user_enter(struct task_struct *tsk)
690{
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100691 write_seqlock(&tsk->vtime_seqlock);
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100692 __vtime_account_system(tsk);
Frederic Weisbeckeraf2350b2013-07-15 16:35:55 +0200693 tsk->vtime_snap_whence = VTIME_USER;
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100694 write_sequnlock(&tsk->vtime_seqlock);
695}
696
697void vtime_guest_enter(struct task_struct *tsk)
698{
Frederic Weisbecker5b206d42013-07-12 19:05:14 +0200699 /*
700 * The flags must be updated under the lock with
701 * the vtime_snap flush and update.
702 * That enforces a right ordering and update sequence
703 * synchronization against the reader (task_gtime())
704 * that can thus safely catch up with a tickless delta.
705 */
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100706 write_seqlock(&tsk->vtime_seqlock);
707 __vtime_account_system(tsk);
708 current->flags |= PF_VCPU;
709 write_sequnlock(&tsk->vtime_seqlock);
710}
Frederic Weisbecker48d6a812013-07-10 02:44:35 +0200711EXPORT_SYMBOL_GPL(vtime_guest_enter);
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100712
713void vtime_guest_exit(struct task_struct *tsk)
714{
715 write_seqlock(&tsk->vtime_seqlock);
716 __vtime_account_system(tsk);
717 current->flags &= ~PF_VCPU;
718 write_sequnlock(&tsk->vtime_seqlock);
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200719}
Frederic Weisbecker48d6a812013-07-10 02:44:35 +0200720EXPORT_SYMBOL_GPL(vtime_guest_exit);
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200721
722void vtime_account_idle(struct task_struct *tsk)
723{
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100724 cputime_t delta_cpu = get_vtime_delta(tsk);
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200725
726 account_idle_time(delta_cpu);
727}
Frederic Weisbecker3f4724e2012-07-16 18:00:34 +0200728
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100729void arch_vtime_task_switch(struct task_struct *prev)
730{
731 write_seqlock(&prev->vtime_seqlock);
732 prev->vtime_snap_whence = VTIME_SLEEPING;
733 write_sequnlock(&prev->vtime_seqlock);
734
735 write_seqlock(&current->vtime_seqlock);
736 current->vtime_snap_whence = VTIME_SYS;
Frederic Weisbecker45eacc62013-05-15 22:16:32 +0200737 current->vtime_snap = sched_clock_cpu(smp_processor_id());
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100738 write_sequnlock(&current->vtime_seqlock);
739}
740
Frederic Weisbecker45eacc62013-05-15 22:16:32 +0200741void vtime_init_idle(struct task_struct *t, int cpu)
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100742{
743 unsigned long flags;
744
745 write_seqlock_irqsave(&t->vtime_seqlock, flags);
746 t->vtime_snap_whence = VTIME_SYS;
Frederic Weisbecker45eacc62013-05-15 22:16:32 +0200747 t->vtime_snap = sched_clock_cpu(cpu);
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100748 write_sequnlock_irqrestore(&t->vtime_seqlock, flags);
749}
750
751cputime_t task_gtime(struct task_struct *t)
752{
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100753 unsigned int seq;
754 cputime_t gtime;
755
756 do {
Thomas Gleixnercdc4e862013-02-15 23:47:07 +0100757 seq = read_seqbegin(&t->vtime_seqlock);
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100758
759 gtime = t->gtime;
760 if (t->flags & PF_VCPU)
761 gtime += vtime_delta(t);
762
Thomas Gleixnercdc4e862013-02-15 23:47:07 +0100763 } while (read_seqretry(&t->vtime_seqlock, seq));
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100764
765 return gtime;
766}
767
768/*
769 * Fetch cputime raw values from fields of task_struct and
770 * add up the pending nohz execution time since the last
771 * cputime snapshot.
772 */
773static void
774fetch_task_cputime(struct task_struct *t,
775 cputime_t *u_dst, cputime_t *s_dst,
776 cputime_t *u_src, cputime_t *s_src,
777 cputime_t *udelta, cputime_t *sdelta)
778{
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100779 unsigned int seq;
780 unsigned long long delta;
781
782 do {
783 *udelta = 0;
784 *sdelta = 0;
785
Thomas Gleixnercdc4e862013-02-15 23:47:07 +0100786 seq = read_seqbegin(&t->vtime_seqlock);
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100787
788 if (u_dst)
789 *u_dst = *u_src;
790 if (s_dst)
791 *s_dst = *s_src;
792
793 /* Task is sleeping, nothing to add */
794 if (t->vtime_snap_whence == VTIME_SLEEPING ||
795 is_idle_task(t))
796 continue;
797
798 delta = vtime_delta(t);
799
800 /*
801 * Task runs either in user or kernel space, add pending nohz time to
802 * the right place.
803 */
804 if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU) {
805 *udelta = delta;
806 } else {
807 if (t->vtime_snap_whence == VTIME_SYS)
808 *sdelta = delta;
809 }
Thomas Gleixnercdc4e862013-02-15 23:47:07 +0100810 } while (read_seqretry(&t->vtime_seqlock, seq));
Frederic Weisbecker6a616712012-12-16 20:00:34 +0100811}
812
813
814void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
815{
816 cputime_t udelta, sdelta;
817
818 fetch_task_cputime(t, utime, stime, &t->utime,
819 &t->stime, &udelta, &sdelta);
820 if (utime)
821 *utime += udelta;
822 if (stime)
823 *stime += sdelta;
824}
825
826void task_cputime_scaled(struct task_struct *t,
827 cputime_t *utimescaled, cputime_t *stimescaled)
828{
829 cputime_t udelta, sdelta;
830
831 fetch_task_cputime(t, utimescaled, stimescaled,
832 &t->utimescaled, &t->stimescaled, &udelta, &sdelta);
833 if (utimescaled)
834 *utimescaled += cputime_to_scaled(udelta);
835 if (stimescaled)
836 *stimescaled += cputime_to_scaled(sdelta);
837}
Frederic Weisbeckerabf917c2012-07-25 07:56:04 +0200838#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */