4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/pid_namespace.h>
48 #include <linux/smp.h>
49 #include <linux/threads.h>
50 #include <linux/timer.h>
51 #include <linux/rcupdate.h>
52 #include <linux/cpu.h>
53 #include <linux/cpuset.h>
54 #include <linux/percpu.h>
55 #include <linux/kthread.h>
56 #include <linux/seq_file.h>
57 #include <linux/sysctl.h>
58 #include <linux/syscalls.h>
59 #include <linux/times.h>
60 #include <linux/tsacct_kern.h>
61 #include <linux/kprobes.h>
62 #include <linux/delayacct.h>
63 #include <linux/reciprocal_div.h>
64 #include <linux/unistd.h>
65 #include <linux/pagemap.h>
68 #include <asm/irq_regs.h>
71 * Scheduler clock - returns current time in nanosec units.
72 * This is default implementation.
73 * Architectures and sub-architectures can override this.
75 unsigned long long __attribute__((weak
)) sched_clock(void)
77 return (unsigned long long)jiffies
* (NSEC_PER_SEC
/ HZ
);
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Some helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / HZ))
104 #define NICE_0_LOAD SCHED_LOAD_SCALE
105 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
108 * These are the 'tuning knobs' of the scheduler:
110 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
111 * Timeslices get refilled after they expire.
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
122 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
131 sg
->__cpu_power
+= val
;
132 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
136 static inline int rt_policy(int policy
)
138 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
143 static inline int task_has_rt_policy(struct task_struct
*p
)
145 return rt_policy(p
->policy
);
149 * This is the priority-queue data structure of the RT scheduling class:
151 struct rt_prio_array
{
152 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
153 struct list_head queue
[MAX_RT_PRIO
];
156 #ifdef CONFIG_FAIR_GROUP_SCHED
158 #include <linux/cgroup.h>
162 /* task group related information */
164 #ifdef CONFIG_FAIR_CGROUP_SCHED
165 struct cgroup_subsys_state css
;
167 /* schedulable entities of this group on each cpu */
168 struct sched_entity
**se
;
169 /* runqueue "owned" by this group on each cpu */
170 struct cfs_rq
**cfs_rq
;
173 * shares assigned to a task group governs how much of cpu bandwidth
174 * is allocated to the group. The more shares a group has, the more is
175 * the cpu bandwidth allocated to it.
177 * For ex, lets say that there are three task groups, A, B and C which
178 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
179 * cpu bandwidth allocated by the scheduler to task groups A, B and C
182 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
183 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
184 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
186 * The weight assigned to a task group's schedulable entities on every
187 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
188 * group's shares. For ex: lets say that task group A has been
189 * assigned shares of 1000 and there are two CPUs in a system. Then,
191 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
193 * Note: It's not necessary that each of a task's group schedulable
194 * entity have the same weight on all CPUs. If the group
195 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
196 * better distribution of weight could be:
198 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
199 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
201 * rebalance_shares() is responsible for distributing the shares of a
202 * task groups like this among the group's schedulable entities across
206 unsigned long shares
;
211 /* Default task group's sched entity on each cpu */
212 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
213 /* Default task group's cfs_rq on each cpu */
214 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
216 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
217 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
219 /* task_group_mutex serializes add/remove of task groups and also changes to
220 * a task group's cpu shares.
222 static DEFINE_MUTEX(task_group_mutex
);
224 /* doms_cur_mutex serializes access to doms_cur[] array */
225 static DEFINE_MUTEX(doms_cur_mutex
);
228 /* kernel thread that runs rebalance_shares() periodically */
229 static struct task_struct
*lb_monitor_task
;
230 static int load_balance_monitor(void *unused
);
233 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
);
235 /* Default task group.
236 * Every task in system belong to this group at bootup.
238 struct task_group init_task_group
= {
239 .se
= init_sched_entity_p
,
240 .cfs_rq
= init_cfs_rq_p
,
243 #ifdef CONFIG_FAIR_USER_SCHED
244 # define INIT_TASK_GROUP_LOAD 2*NICE_0_LOAD
246 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
249 #define MIN_GROUP_SHARES 2
251 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
253 /* return group to which a task belongs */
254 static inline struct task_group
*task_group(struct task_struct
*p
)
256 struct task_group
*tg
;
258 #ifdef CONFIG_FAIR_USER_SCHED
260 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
261 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
262 struct task_group
, css
);
264 tg
= &init_task_group
;
269 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
270 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
)
272 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
273 p
->se
.parent
= task_group(p
)->se
[cpu
];
276 static inline void lock_task_group_list(void)
278 mutex_lock(&task_group_mutex
);
281 static inline void unlock_task_group_list(void)
283 mutex_unlock(&task_group_mutex
);
286 static inline void lock_doms_cur(void)
288 mutex_lock(&doms_cur_mutex
);
291 static inline void unlock_doms_cur(void)
293 mutex_unlock(&doms_cur_mutex
);
298 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
) { }
299 static inline void lock_task_group_list(void) { }
300 static inline void unlock_task_group_list(void) { }
301 static inline void lock_doms_cur(void) { }
302 static inline void unlock_doms_cur(void) { }
304 #endif /* CONFIG_FAIR_GROUP_SCHED */
306 /* CFS-related fields in a runqueue */
308 struct load_weight load
;
309 unsigned long nr_running
;
314 struct rb_root tasks_timeline
;
315 struct rb_node
*rb_leftmost
;
316 struct rb_node
*rb_load_balance_curr
;
317 /* 'curr' points to currently running entity on this cfs_rq.
318 * It is set to NULL otherwise (i.e when none are currently running).
320 struct sched_entity
*curr
;
322 unsigned long nr_spread_over
;
324 #ifdef CONFIG_FAIR_GROUP_SCHED
325 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
328 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
329 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
330 * (like users, containers etc.)
332 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
333 * list is used during load balance.
335 struct list_head leaf_cfs_rq_list
;
336 struct task_group
*tg
; /* group that "owns" this runqueue */
340 /* Real-Time classes' related field in a runqueue: */
342 struct rt_prio_array active
;
343 int rt_load_balance_idx
;
344 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
345 unsigned long rt_nr_running
;
346 unsigned long rt_nr_migratory
;
347 /* highest queued rt task prio */
353 * This is the main, per-CPU runqueue data structure.
355 * Locking rule: those places that want to lock multiple runqueues
356 * (such as the load balancing or the thread migration code), lock
357 * acquire operations must be ordered by ascending &runqueue.
364 * nr_running and cpu_load should be in the same cacheline because
365 * remote CPUs use both these fields when doing load calculation.
367 unsigned long nr_running
;
368 #define CPU_LOAD_IDX_MAX 5
369 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
370 unsigned char idle_at_tick
;
372 unsigned char in_nohz_recently
;
374 /* capture load from *all* tasks on this cpu: */
375 struct load_weight load
;
376 unsigned long nr_load_updates
;
380 #ifdef CONFIG_FAIR_GROUP_SCHED
381 /* list of leaf cfs_rq on this cpu: */
382 struct list_head leaf_cfs_rq_list
;
387 * This is part of a global counter where only the total sum
388 * over all CPUs matters. A task can increase this counter on
389 * one CPU and if it got migrated afterwards it may decrease
390 * it on another CPU. Always updated under the runqueue lock:
392 unsigned long nr_uninterruptible
;
394 struct task_struct
*curr
, *idle
;
395 unsigned long next_balance
;
396 struct mm_struct
*prev_mm
;
398 u64 clock
, prev_clock_raw
;
401 unsigned int clock_warps
, clock_overflows
;
403 unsigned int clock_deep_idle_events
;
409 struct sched_domain
*sd
;
411 /* For active balancing */
414 /* cpu of this runqueue: */
417 struct task_struct
*migration_thread
;
418 struct list_head migration_queue
;
421 #ifdef CONFIG_SCHEDSTATS
423 struct sched_info rq_sched_info
;
425 /* sys_sched_yield() stats */
426 unsigned int yld_exp_empty
;
427 unsigned int yld_act_empty
;
428 unsigned int yld_both_empty
;
429 unsigned int yld_count
;
431 /* schedule() stats */
432 unsigned int sched_switch
;
433 unsigned int sched_count
;
434 unsigned int sched_goidle
;
436 /* try_to_wake_up() stats */
437 unsigned int ttwu_count
;
438 unsigned int ttwu_local
;
441 unsigned int bkl_count
;
443 struct lock_class_key rq_lock_key
;
446 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
448 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
450 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
453 static inline int cpu_of(struct rq
*rq
)
463 * Update the per-runqueue clock, as finegrained as the platform can give
464 * us, but without assuming monotonicity, etc.:
466 static void __update_rq_clock(struct rq
*rq
)
468 u64 prev_raw
= rq
->prev_clock_raw
;
469 u64 now
= sched_clock();
470 s64 delta
= now
- prev_raw
;
471 u64 clock
= rq
->clock
;
473 #ifdef CONFIG_SCHED_DEBUG
474 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
477 * Protect against sched_clock() occasionally going backwards:
479 if (unlikely(delta
< 0)) {
484 * Catch too large forward jumps too:
486 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
487 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
488 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
491 rq
->clock_overflows
++;
493 if (unlikely(delta
> rq
->clock_max_delta
))
494 rq
->clock_max_delta
= delta
;
499 rq
->prev_clock_raw
= now
;
503 static void update_rq_clock(struct rq
*rq
)
505 if (likely(smp_processor_id() == cpu_of(rq
)))
506 __update_rq_clock(rq
);
510 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
511 * See detach_destroy_domains: synchronize_sched for details.
513 * The domain tree of any CPU may only be accessed from within
514 * preempt-disabled sections.
516 #define for_each_domain(cpu, __sd) \
517 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
519 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
520 #define this_rq() (&__get_cpu_var(runqueues))
521 #define task_rq(p) cpu_rq(task_cpu(p))
522 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
525 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
527 #ifdef CONFIG_SCHED_DEBUG
528 # define const_debug __read_mostly
530 # define const_debug static const
534 * Debugging: various feature bits
537 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
538 SCHED_FEAT_WAKEUP_PREEMPT
= 2,
539 SCHED_FEAT_START_DEBIT
= 4,
540 SCHED_FEAT_TREE_AVG
= 8,
541 SCHED_FEAT_APPROX_AVG
= 16,
544 const_debug
unsigned int sysctl_sched_features
=
545 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
546 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
547 SCHED_FEAT_START_DEBIT
* 1 |
548 SCHED_FEAT_TREE_AVG
* 0 |
549 SCHED_FEAT_APPROX_AVG
* 0;
551 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
554 * Number of tasks to iterate in a single balance run.
555 * Limited because this is done with IRQs disabled.
557 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
560 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
561 * clock constructed from sched_clock():
563 unsigned long long cpu_clock(int cpu
)
565 unsigned long long now
;
569 local_irq_save(flags
);
572 * Only call sched_clock() if the scheduler has already been
573 * initialized (some code might call cpu_clock() very early):
578 local_irq_restore(flags
);
582 EXPORT_SYMBOL_GPL(cpu_clock
);
584 #ifndef prepare_arch_switch
585 # define prepare_arch_switch(next) do { } while (0)
587 #ifndef finish_arch_switch
588 # define finish_arch_switch(prev) do { } while (0)
591 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
593 return rq
->curr
== p
;
596 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
597 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
599 return task_current(rq
, p
);
602 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
606 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
608 #ifdef CONFIG_DEBUG_SPINLOCK
609 /* this is a valid case when another task releases the spinlock */
610 rq
->lock
.owner
= current
;
613 * If we are tracking spinlock dependencies then we have to
614 * fix up the runqueue lock - which gets 'carried over' from
617 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
619 spin_unlock_irq(&rq
->lock
);
622 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
623 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
628 return task_current(rq
, p
);
632 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
636 * We can optimise this out completely for !SMP, because the
637 * SMP rebalancing from interrupt is the only thing that cares
642 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
643 spin_unlock_irq(&rq
->lock
);
645 spin_unlock(&rq
->lock
);
649 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
653 * After ->oncpu is cleared, the task can be moved to a different CPU.
654 * We must ensure this doesn't happen until the switch is completely
660 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
664 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
667 * __task_rq_lock - lock the runqueue a given task resides on.
668 * Must be called interrupts disabled.
670 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
674 struct rq
*rq
= task_rq(p
);
675 spin_lock(&rq
->lock
);
676 if (likely(rq
== task_rq(p
)))
678 spin_unlock(&rq
->lock
);
683 * task_rq_lock - lock the runqueue a given task resides on and disable
684 * interrupts. Note the ordering: we can safely lookup the task_rq without
685 * explicitly disabling preemption.
687 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
693 local_irq_save(*flags
);
695 spin_lock(&rq
->lock
);
696 if (likely(rq
== task_rq(p
)))
698 spin_unlock_irqrestore(&rq
->lock
, *flags
);
702 static void __task_rq_unlock(struct rq
*rq
)
705 spin_unlock(&rq
->lock
);
708 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
711 spin_unlock_irqrestore(&rq
->lock
, *flags
);
715 * this_rq_lock - lock this runqueue and disable interrupts.
717 static struct rq
*this_rq_lock(void)
724 spin_lock(&rq
->lock
);
730 * We are going deep-idle (irqs are disabled):
732 void sched_clock_idle_sleep_event(void)
734 struct rq
*rq
= cpu_rq(smp_processor_id());
736 spin_lock(&rq
->lock
);
737 __update_rq_clock(rq
);
738 spin_unlock(&rq
->lock
);
739 rq
->clock_deep_idle_events
++;
741 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
744 * We just idled delta nanoseconds (called with irqs disabled):
746 void sched_clock_idle_wakeup_event(u64 delta_ns
)
748 struct rq
*rq
= cpu_rq(smp_processor_id());
749 u64 now
= sched_clock();
751 touch_softlockup_watchdog();
752 rq
->idle_clock
+= delta_ns
;
754 * Override the previous timestamp and ignore all
755 * sched_clock() deltas that occured while we idled,
756 * and use the PM-provided delta_ns to advance the
759 spin_lock(&rq
->lock
);
760 rq
->prev_clock_raw
= now
;
761 rq
->clock
+= delta_ns
;
762 spin_unlock(&rq
->lock
);
764 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
767 * resched_task - mark a task 'to be rescheduled now'.
769 * On UP this means the setting of the need_resched flag, on SMP it
770 * might also involve a cross-CPU call to trigger the scheduler on
775 #ifndef tsk_is_polling
776 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
779 static void resched_task(struct task_struct
*p
)
783 assert_spin_locked(&task_rq(p
)->lock
);
785 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
788 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
791 if (cpu
== smp_processor_id())
794 /* NEED_RESCHED must be visible before we test polling */
796 if (!tsk_is_polling(p
))
797 smp_send_reschedule(cpu
);
800 static void resched_cpu(int cpu
)
802 struct rq
*rq
= cpu_rq(cpu
);
805 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
807 resched_task(cpu_curr(cpu
));
808 spin_unlock_irqrestore(&rq
->lock
, flags
);
811 static inline void resched_task(struct task_struct
*p
)
813 assert_spin_locked(&task_rq(p
)->lock
);
814 set_tsk_need_resched(p
);
818 #if BITS_PER_LONG == 32
819 # define WMULT_CONST (~0UL)
821 # define WMULT_CONST (1UL << 32)
824 #define WMULT_SHIFT 32
827 * Shift right and round:
829 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
832 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
833 struct load_weight
*lw
)
837 if (unlikely(!lw
->inv_weight
))
838 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
840 tmp
= (u64
)delta_exec
* weight
;
842 * Check whether we'd overflow the 64-bit multiplication:
844 if (unlikely(tmp
> WMULT_CONST
))
845 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
848 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
850 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
853 static inline unsigned long
854 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
856 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
859 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
864 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
870 * To aid in avoiding the subversion of "niceness" due to uneven distribution
871 * of tasks with abnormal "nice" values across CPUs the contribution that
872 * each task makes to its run queue's load is weighted according to its
873 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
874 * scaled version of the new time slice allocation that they receive on time
878 #define WEIGHT_IDLEPRIO 2
879 #define WMULT_IDLEPRIO (1 << 31)
882 * Nice levels are multiplicative, with a gentle 10% change for every
883 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
884 * nice 1, it will get ~10% less CPU time than another CPU-bound task
885 * that remained on nice 0.
887 * The "10% effect" is relative and cumulative: from _any_ nice level,
888 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
889 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
890 * If a task goes up by ~10% and another task goes down by ~10% then
891 * the relative distance between them is ~25%.)
893 static const int prio_to_weight
[40] = {
894 /* -20 */ 88761, 71755, 56483, 46273, 36291,
895 /* -15 */ 29154, 23254, 18705, 14949, 11916,
896 /* -10 */ 9548, 7620, 6100, 4904, 3906,
897 /* -5 */ 3121, 2501, 1991, 1586, 1277,
898 /* 0 */ 1024, 820, 655, 526, 423,
899 /* 5 */ 335, 272, 215, 172, 137,
900 /* 10 */ 110, 87, 70, 56, 45,
901 /* 15 */ 36, 29, 23, 18, 15,
905 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
907 * In cases where the weight does not change often, we can use the
908 * precalculated inverse to speed up arithmetics by turning divisions
909 * into multiplications:
911 static const u32 prio_to_wmult
[40] = {
912 /* -20 */ 48388, 59856, 76040, 92818, 118348,
913 /* -15 */ 147320, 184698, 229616, 287308, 360437,
914 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
915 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
916 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
917 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
918 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
919 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
922 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
925 * runqueue iterator, to support SMP load-balancing between different
926 * scheduling classes, without having to expose their internal data
927 * structures to the load-balancing proper:
931 struct task_struct
*(*start
)(void *);
932 struct task_struct
*(*next
)(void *);
937 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
938 unsigned long max_load_move
, struct sched_domain
*sd
,
939 enum cpu_idle_type idle
, int *all_pinned
,
940 int *this_best_prio
, struct rq_iterator
*iterator
);
943 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
944 struct sched_domain
*sd
, enum cpu_idle_type idle
,
945 struct rq_iterator
*iterator
);
948 #ifdef CONFIG_CGROUP_CPUACCT
949 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
951 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
954 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
956 update_load_add(&rq
->load
, load
);
959 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
961 update_load_sub(&rq
->load
, load
);
965 static unsigned long source_load(int cpu
, int type
);
966 static unsigned long target_load(int cpu
, int type
);
967 static unsigned long cpu_avg_load_per_task(int cpu
);
968 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
969 #endif /* CONFIG_SMP */
971 #include "sched_stats.h"
972 #include "sched_idletask.c"
973 #include "sched_fair.c"
974 #include "sched_rt.c"
975 #ifdef CONFIG_SCHED_DEBUG
976 # include "sched_debug.c"
979 #define sched_class_highest (&rt_sched_class)
981 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
986 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
991 static void set_load_weight(struct task_struct
*p
)
993 if (task_has_rt_policy(p
)) {
994 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
995 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1000 * SCHED_IDLE tasks get minimal weight:
1002 if (p
->policy
== SCHED_IDLE
) {
1003 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1004 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1008 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1009 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1012 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1014 sched_info_queued(p
);
1015 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1019 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1021 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1026 * __normal_prio - return the priority that is based on the static prio
1028 static inline int __normal_prio(struct task_struct
*p
)
1030 return p
->static_prio
;
1034 * Calculate the expected normal priority: i.e. priority
1035 * without taking RT-inheritance into account. Might be
1036 * boosted by interactivity modifiers. Changes upon fork,
1037 * setprio syscalls, and whenever the interactivity
1038 * estimator recalculates.
1040 static inline int normal_prio(struct task_struct
*p
)
1044 if (task_has_rt_policy(p
))
1045 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1047 prio
= __normal_prio(p
);
1052 * Calculate the current priority, i.e. the priority
1053 * taken into account by the scheduler. This value might
1054 * be boosted by RT tasks, or might be boosted by
1055 * interactivity modifiers. Will be RT if the task got
1056 * RT-boosted. If not then it returns p->normal_prio.
1058 static int effective_prio(struct task_struct
*p
)
1060 p
->normal_prio
= normal_prio(p
);
1062 * If we are RT tasks or we were boosted to RT priority,
1063 * keep the priority unchanged. Otherwise, update priority
1064 * to the normal priority:
1066 if (!rt_prio(p
->prio
))
1067 return p
->normal_prio
;
1072 * activate_task - move a task to the runqueue.
1074 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1076 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1077 rq
->nr_uninterruptible
--;
1079 enqueue_task(rq
, p
, wakeup
);
1080 inc_nr_running(p
, rq
);
1084 * deactivate_task - remove a task from the runqueue.
1086 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1088 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1089 rq
->nr_uninterruptible
++;
1091 dequeue_task(rq
, p
, sleep
);
1092 dec_nr_running(p
, rq
);
1096 * task_curr - is this task currently executing on a CPU?
1097 * @p: the task in question.
1099 inline int task_curr(const struct task_struct
*p
)
1101 return cpu_curr(task_cpu(p
)) == p
;
1104 /* Used instead of source_load when we know the type == 0 */
1105 unsigned long weighted_cpuload(const int cpu
)
1107 return cpu_rq(cpu
)->load
.weight
;
1110 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1112 set_task_cfs_rq(p
, cpu
);
1115 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1116 * successfuly executed on another CPU. We must ensure that updates of
1117 * per-task data have been completed by this moment.
1120 task_thread_info(p
)->cpu
= cpu
;
1127 * Is this task likely cache-hot:
1130 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1134 if (p
->sched_class
!= &fair_sched_class
)
1137 if (sysctl_sched_migration_cost
== -1)
1139 if (sysctl_sched_migration_cost
== 0)
1142 delta
= now
- p
->se
.exec_start
;
1144 return delta
< (s64
)sysctl_sched_migration_cost
;
1148 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1150 int old_cpu
= task_cpu(p
);
1151 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1152 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1153 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1156 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1158 #ifdef CONFIG_SCHEDSTATS
1159 if (p
->se
.wait_start
)
1160 p
->se
.wait_start
-= clock_offset
;
1161 if (p
->se
.sleep_start
)
1162 p
->se
.sleep_start
-= clock_offset
;
1163 if (p
->se
.block_start
)
1164 p
->se
.block_start
-= clock_offset
;
1165 if (old_cpu
!= new_cpu
) {
1166 schedstat_inc(p
, se
.nr_migrations
);
1167 if (task_hot(p
, old_rq
->clock
, NULL
))
1168 schedstat_inc(p
, se
.nr_forced2_migrations
);
1171 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1172 new_cfsrq
->min_vruntime
;
1174 __set_task_cpu(p
, new_cpu
);
1177 struct migration_req
{
1178 struct list_head list
;
1180 struct task_struct
*task
;
1183 struct completion done
;
1187 * The task's runqueue lock must be held.
1188 * Returns true if you have to wait for migration thread.
1191 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1193 struct rq
*rq
= task_rq(p
);
1196 * If the task is not on a runqueue (and not running), then
1197 * it is sufficient to simply update the task's cpu field.
1199 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1200 set_task_cpu(p
, dest_cpu
);
1204 init_completion(&req
->done
);
1206 req
->dest_cpu
= dest_cpu
;
1207 list_add(&req
->list
, &rq
->migration_queue
);
1213 * wait_task_inactive - wait for a thread to unschedule.
1215 * The caller must ensure that the task *will* unschedule sometime soon,
1216 * else this function might spin for a *long* time. This function can't
1217 * be called with interrupts off, or it may introduce deadlock with
1218 * smp_call_function() if an IPI is sent by the same process we are
1219 * waiting to become inactive.
1221 void wait_task_inactive(struct task_struct
*p
)
1223 unsigned long flags
;
1229 * We do the initial early heuristics without holding
1230 * any task-queue locks at all. We'll only try to get
1231 * the runqueue lock when things look like they will
1237 * If the task is actively running on another CPU
1238 * still, just relax and busy-wait without holding
1241 * NOTE! Since we don't hold any locks, it's not
1242 * even sure that "rq" stays as the right runqueue!
1243 * But we don't care, since "task_running()" will
1244 * return false if the runqueue has changed and p
1245 * is actually now running somewhere else!
1247 while (task_running(rq
, p
))
1251 * Ok, time to look more closely! We need the rq
1252 * lock now, to be *sure*. If we're wrong, we'll
1253 * just go back and repeat.
1255 rq
= task_rq_lock(p
, &flags
);
1256 running
= task_running(rq
, p
);
1257 on_rq
= p
->se
.on_rq
;
1258 task_rq_unlock(rq
, &flags
);
1261 * Was it really running after all now that we
1262 * checked with the proper locks actually held?
1264 * Oops. Go back and try again..
1266 if (unlikely(running
)) {
1272 * It's not enough that it's not actively running,
1273 * it must be off the runqueue _entirely_, and not
1276 * So if it wa still runnable (but just not actively
1277 * running right now), it's preempted, and we should
1278 * yield - it could be a while.
1280 if (unlikely(on_rq
)) {
1281 schedule_timeout_uninterruptible(1);
1286 * Ahh, all good. It wasn't running, and it wasn't
1287 * runnable, which means that it will never become
1288 * running in the future either. We're all done!
1295 * kick_process - kick a running thread to enter/exit the kernel
1296 * @p: the to-be-kicked thread
1298 * Cause a process which is running on another CPU to enter
1299 * kernel-mode, without any delay. (to get signals handled.)
1301 * NOTE: this function doesnt have to take the runqueue lock,
1302 * because all it wants to ensure is that the remote task enters
1303 * the kernel. If the IPI races and the task has been migrated
1304 * to another CPU then no harm is done and the purpose has been
1307 void kick_process(struct task_struct
*p
)
1313 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1314 smp_send_reschedule(cpu
);
1319 * Return a low guess at the load of a migration-source cpu weighted
1320 * according to the scheduling class and "nice" value.
1322 * We want to under-estimate the load of migration sources, to
1323 * balance conservatively.
1325 static unsigned long source_load(int cpu
, int type
)
1327 struct rq
*rq
= cpu_rq(cpu
);
1328 unsigned long total
= weighted_cpuload(cpu
);
1333 return min(rq
->cpu_load
[type
-1], total
);
1337 * Return a high guess at the load of a migration-target cpu weighted
1338 * according to the scheduling class and "nice" value.
1340 static unsigned long target_load(int cpu
, int type
)
1342 struct rq
*rq
= cpu_rq(cpu
);
1343 unsigned long total
= weighted_cpuload(cpu
);
1348 return max(rq
->cpu_load
[type
-1], total
);
1352 * Return the average load per task on the cpu's run queue
1354 static unsigned long cpu_avg_load_per_task(int cpu
)
1356 struct rq
*rq
= cpu_rq(cpu
);
1357 unsigned long total
= weighted_cpuload(cpu
);
1358 unsigned long n
= rq
->nr_running
;
1360 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1364 * find_idlest_group finds and returns the least busy CPU group within the
1367 static struct sched_group
*
1368 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1370 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1371 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1372 int load_idx
= sd
->forkexec_idx
;
1373 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1376 unsigned long load
, avg_load
;
1380 /* Skip over this group if it has no CPUs allowed */
1381 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1384 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1386 /* Tally up the load of all CPUs in the group */
1389 for_each_cpu_mask(i
, group
->cpumask
) {
1390 /* Bias balancing toward cpus of our domain */
1392 load
= source_load(i
, load_idx
);
1394 load
= target_load(i
, load_idx
);
1399 /* Adjust by relative CPU power of the group */
1400 avg_load
= sg_div_cpu_power(group
,
1401 avg_load
* SCHED_LOAD_SCALE
);
1404 this_load
= avg_load
;
1406 } else if (avg_load
< min_load
) {
1407 min_load
= avg_load
;
1410 } while (group
= group
->next
, group
!= sd
->groups
);
1412 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1418 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1421 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1424 unsigned long load
, min_load
= ULONG_MAX
;
1428 /* Traverse only the allowed CPUs */
1429 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1431 for_each_cpu_mask(i
, tmp
) {
1432 load
= weighted_cpuload(i
);
1434 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1444 * sched_balance_self: balance the current task (running on cpu) in domains
1445 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1448 * Balance, ie. select the least loaded group.
1450 * Returns the target CPU number, or the same CPU if no balancing is needed.
1452 * preempt must be disabled.
1454 static int sched_balance_self(int cpu
, int flag
)
1456 struct task_struct
*t
= current
;
1457 struct sched_domain
*tmp
, *sd
= NULL
;
1459 for_each_domain(cpu
, tmp
) {
1461 * If power savings logic is enabled for a domain, stop there.
1463 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1465 if (tmp
->flags
& flag
)
1471 struct sched_group
*group
;
1472 int new_cpu
, weight
;
1474 if (!(sd
->flags
& flag
)) {
1480 group
= find_idlest_group(sd
, t
, cpu
);
1486 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1487 if (new_cpu
== -1 || new_cpu
== cpu
) {
1488 /* Now try balancing at a lower domain level of cpu */
1493 /* Now try balancing at a lower domain level of new_cpu */
1496 weight
= cpus_weight(span
);
1497 for_each_domain(cpu
, tmp
) {
1498 if (weight
<= cpus_weight(tmp
->span
))
1500 if (tmp
->flags
& flag
)
1503 /* while loop will break here if sd == NULL */
1509 #endif /* CONFIG_SMP */
1512 * try_to_wake_up - wake up a thread
1513 * @p: the to-be-woken-up thread
1514 * @state: the mask of task states that can be woken
1515 * @sync: do a synchronous wakeup?
1517 * Put it on the run-queue if it's not already there. The "current"
1518 * thread is always on the run-queue (except when the actual
1519 * re-schedule is in progress), and as such you're allowed to do
1520 * the simpler "current->state = TASK_RUNNING" to mark yourself
1521 * runnable without the overhead of this.
1523 * returns failure only if the task is already active.
1525 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1527 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1528 unsigned long flags
;
1535 rq
= task_rq_lock(p
, &flags
);
1536 old_state
= p
->state
;
1537 if (!(old_state
& state
))
1545 this_cpu
= smp_processor_id();
1548 if (unlikely(task_running(rq
, p
)))
1551 new_cpu
= p
->sched_class
->select_task_rq(p
, sync
);
1552 if (new_cpu
!= cpu
) {
1553 set_task_cpu(p
, new_cpu
);
1554 task_rq_unlock(rq
, &flags
);
1555 /* might preempt at this point */
1556 rq
= task_rq_lock(p
, &flags
);
1557 old_state
= p
->state
;
1558 if (!(old_state
& state
))
1563 this_cpu
= smp_processor_id();
1567 #ifdef CONFIG_SCHEDSTATS
1568 schedstat_inc(rq
, ttwu_count
);
1569 if (cpu
== this_cpu
)
1570 schedstat_inc(rq
, ttwu_local
);
1572 struct sched_domain
*sd
;
1573 for_each_domain(this_cpu
, sd
) {
1574 if (cpu_isset(cpu
, sd
->span
)) {
1575 schedstat_inc(sd
, ttwu_wake_remote
);
1585 #endif /* CONFIG_SMP */
1586 schedstat_inc(p
, se
.nr_wakeups
);
1588 schedstat_inc(p
, se
.nr_wakeups_sync
);
1589 if (orig_cpu
!= cpu
)
1590 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1591 if (cpu
== this_cpu
)
1592 schedstat_inc(p
, se
.nr_wakeups_local
);
1594 schedstat_inc(p
, se
.nr_wakeups_remote
);
1595 update_rq_clock(rq
);
1596 activate_task(rq
, p
, 1);
1597 check_preempt_curr(rq
, p
);
1601 p
->state
= TASK_RUNNING
;
1602 wakeup_balance_rt(rq
, p
);
1604 task_rq_unlock(rq
, &flags
);
1609 int fastcall
wake_up_process(struct task_struct
*p
)
1611 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1612 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1614 EXPORT_SYMBOL(wake_up_process
);
1616 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1618 return try_to_wake_up(p
, state
, 0);
1622 * Perform scheduler related setup for a newly forked process p.
1623 * p is forked by current.
1625 * __sched_fork() is basic setup used by init_idle() too:
1627 static void __sched_fork(struct task_struct
*p
)
1629 p
->se
.exec_start
= 0;
1630 p
->se
.sum_exec_runtime
= 0;
1631 p
->se
.prev_sum_exec_runtime
= 0;
1633 #ifdef CONFIG_SCHEDSTATS
1634 p
->se
.wait_start
= 0;
1635 p
->se
.sum_sleep_runtime
= 0;
1636 p
->se
.sleep_start
= 0;
1637 p
->se
.block_start
= 0;
1638 p
->se
.sleep_max
= 0;
1639 p
->se
.block_max
= 0;
1641 p
->se
.slice_max
= 0;
1645 INIT_LIST_HEAD(&p
->run_list
);
1648 #ifdef CONFIG_PREEMPT_NOTIFIERS
1649 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1653 * We mark the process as running here, but have not actually
1654 * inserted it onto the runqueue yet. This guarantees that
1655 * nobody will actually run it, and a signal or other external
1656 * event cannot wake it up and insert it on the runqueue either.
1658 p
->state
= TASK_RUNNING
;
1662 * fork()/clone()-time setup:
1664 void sched_fork(struct task_struct
*p
, int clone_flags
)
1666 int cpu
= get_cpu();
1671 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1673 set_task_cpu(p
, cpu
);
1676 * Make sure we do not leak PI boosting priority to the child:
1678 p
->prio
= current
->normal_prio
;
1679 if (!rt_prio(p
->prio
))
1680 p
->sched_class
= &fair_sched_class
;
1682 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1683 if (likely(sched_info_on()))
1684 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1686 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1689 #ifdef CONFIG_PREEMPT
1690 /* Want to start with kernel preemption disabled. */
1691 task_thread_info(p
)->preempt_count
= 1;
1697 * wake_up_new_task - wake up a newly created task for the first time.
1699 * This function will do some initial scheduler statistics housekeeping
1700 * that must be done for every newly created context, then puts the task
1701 * on the runqueue and wakes it.
1703 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1705 unsigned long flags
;
1708 rq
= task_rq_lock(p
, &flags
);
1709 BUG_ON(p
->state
!= TASK_RUNNING
);
1710 update_rq_clock(rq
);
1712 p
->prio
= effective_prio(p
);
1714 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
1715 activate_task(rq
, p
, 0);
1718 * Let the scheduling class do new task startup
1719 * management (if any):
1721 p
->sched_class
->task_new(rq
, p
);
1722 inc_nr_running(p
, rq
);
1724 check_preempt_curr(rq
, p
);
1725 task_rq_unlock(rq
, &flags
);
1728 #ifdef CONFIG_PREEMPT_NOTIFIERS
1731 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1732 * @notifier: notifier struct to register
1734 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1736 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1738 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1741 * preempt_notifier_unregister - no longer interested in preemption notifications
1742 * @notifier: notifier struct to unregister
1744 * This is safe to call from within a preemption notifier.
1746 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1748 hlist_del(¬ifier
->link
);
1750 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1752 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1754 struct preempt_notifier
*notifier
;
1755 struct hlist_node
*node
;
1757 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1758 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1762 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1763 struct task_struct
*next
)
1765 struct preempt_notifier
*notifier
;
1766 struct hlist_node
*node
;
1768 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1769 notifier
->ops
->sched_out(notifier
, next
);
1774 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1779 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1780 struct task_struct
*next
)
1787 * prepare_task_switch - prepare to switch tasks
1788 * @rq: the runqueue preparing to switch
1789 * @prev: the current task that is being switched out
1790 * @next: the task we are going to switch to.
1792 * This is called with the rq lock held and interrupts off. It must
1793 * be paired with a subsequent finish_task_switch after the context
1796 * prepare_task_switch sets up locking and calls architecture specific
1800 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1801 struct task_struct
*next
)
1803 fire_sched_out_preempt_notifiers(prev
, next
);
1804 prepare_lock_switch(rq
, next
);
1805 prepare_arch_switch(next
);
1809 * finish_task_switch - clean up after a task-switch
1810 * @rq: runqueue associated with task-switch
1811 * @prev: the thread we just switched away from.
1813 * finish_task_switch must be called after the context switch, paired
1814 * with a prepare_task_switch call before the context switch.
1815 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1816 * and do any other architecture-specific cleanup actions.
1818 * Note that we may have delayed dropping an mm in context_switch(). If
1819 * so, we finish that here outside of the runqueue lock. (Doing it
1820 * with the lock held can cause deadlocks; see schedule() for
1823 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1824 __releases(rq
->lock
)
1826 struct mm_struct
*mm
= rq
->prev_mm
;
1832 * A task struct has one reference for the use as "current".
1833 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1834 * schedule one last time. The schedule call will never return, and
1835 * the scheduled task must drop that reference.
1836 * The test for TASK_DEAD must occur while the runqueue locks are
1837 * still held, otherwise prev could be scheduled on another cpu, die
1838 * there before we look at prev->state, and then the reference would
1840 * Manfred Spraul <manfred@colorfullife.com>
1842 prev_state
= prev
->state
;
1843 finish_arch_switch(prev
);
1844 finish_lock_switch(rq
, prev
);
1845 schedule_tail_balance_rt(rq
);
1847 fire_sched_in_preempt_notifiers(current
);
1850 if (unlikely(prev_state
== TASK_DEAD
)) {
1852 * Remove function-return probe instances associated with this
1853 * task and put them back on the free list.
1855 kprobe_flush_task(prev
);
1856 put_task_struct(prev
);
1861 * schedule_tail - first thing a freshly forked thread must call.
1862 * @prev: the thread we just switched away from.
1864 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1865 __releases(rq
->lock
)
1867 struct rq
*rq
= this_rq();
1869 finish_task_switch(rq
, prev
);
1870 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1871 /* In this case, finish_task_switch does not reenable preemption */
1874 if (current
->set_child_tid
)
1875 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1879 * context_switch - switch to the new MM and the new
1880 * thread's register state.
1883 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1884 struct task_struct
*next
)
1886 struct mm_struct
*mm
, *oldmm
;
1888 prepare_task_switch(rq
, prev
, next
);
1890 oldmm
= prev
->active_mm
;
1892 * For paravirt, this is coupled with an exit in switch_to to
1893 * combine the page table reload and the switch backend into
1896 arch_enter_lazy_cpu_mode();
1898 if (unlikely(!mm
)) {
1899 next
->active_mm
= oldmm
;
1900 atomic_inc(&oldmm
->mm_count
);
1901 enter_lazy_tlb(oldmm
, next
);
1903 switch_mm(oldmm
, mm
, next
);
1905 if (unlikely(!prev
->mm
)) {
1906 prev
->active_mm
= NULL
;
1907 rq
->prev_mm
= oldmm
;
1910 * Since the runqueue lock will be released by the next
1911 * task (which is an invalid locking op but in the case
1912 * of the scheduler it's an obvious special-case), so we
1913 * do an early lockdep release here:
1915 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1916 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1919 /* Here we just switch the register state and the stack. */
1920 switch_to(prev
, next
, prev
);
1924 * this_rq must be evaluated again because prev may have moved
1925 * CPUs since it called schedule(), thus the 'rq' on its stack
1926 * frame will be invalid.
1928 finish_task_switch(this_rq(), prev
);
1932 * nr_running, nr_uninterruptible and nr_context_switches:
1934 * externally visible scheduler statistics: current number of runnable
1935 * threads, current number of uninterruptible-sleeping threads, total
1936 * number of context switches performed since bootup.
1938 unsigned long nr_running(void)
1940 unsigned long i
, sum
= 0;
1942 for_each_online_cpu(i
)
1943 sum
+= cpu_rq(i
)->nr_running
;
1948 unsigned long nr_uninterruptible(void)
1950 unsigned long i
, sum
= 0;
1952 for_each_possible_cpu(i
)
1953 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1956 * Since we read the counters lockless, it might be slightly
1957 * inaccurate. Do not allow it to go below zero though:
1959 if (unlikely((long)sum
< 0))
1965 unsigned long long nr_context_switches(void)
1968 unsigned long long sum
= 0;
1970 for_each_possible_cpu(i
)
1971 sum
+= cpu_rq(i
)->nr_switches
;
1976 unsigned long nr_iowait(void)
1978 unsigned long i
, sum
= 0;
1980 for_each_possible_cpu(i
)
1981 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1986 unsigned long nr_active(void)
1988 unsigned long i
, running
= 0, uninterruptible
= 0;
1990 for_each_online_cpu(i
) {
1991 running
+= cpu_rq(i
)->nr_running
;
1992 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1995 if (unlikely((long)uninterruptible
< 0))
1996 uninterruptible
= 0;
1998 return running
+ uninterruptible
;
2002 * Update rq->cpu_load[] statistics. This function is usually called every
2003 * scheduler tick (TICK_NSEC).
2005 static void update_cpu_load(struct rq
*this_rq
)
2007 unsigned long this_load
= this_rq
->load
.weight
;
2010 this_rq
->nr_load_updates
++;
2012 /* Update our load: */
2013 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2014 unsigned long old_load
, new_load
;
2016 /* scale is effectively 1 << i now, and >> i divides by scale */
2018 old_load
= this_rq
->cpu_load
[i
];
2019 new_load
= this_load
;
2021 * Round up the averaging division if load is increasing. This
2022 * prevents us from getting stuck on 9 if the load is 10, for
2025 if (new_load
> old_load
)
2026 new_load
+= scale
-1;
2027 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2034 * double_rq_lock - safely lock two runqueues
2036 * Note this does not disable interrupts like task_rq_lock,
2037 * you need to do so manually before calling.
2039 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2040 __acquires(rq1
->lock
)
2041 __acquires(rq2
->lock
)
2043 BUG_ON(!irqs_disabled());
2045 spin_lock(&rq1
->lock
);
2046 __acquire(rq2
->lock
); /* Fake it out ;) */
2049 spin_lock(&rq1
->lock
);
2050 spin_lock(&rq2
->lock
);
2052 spin_lock(&rq2
->lock
);
2053 spin_lock(&rq1
->lock
);
2056 update_rq_clock(rq1
);
2057 update_rq_clock(rq2
);
2061 * double_rq_unlock - safely unlock two runqueues
2063 * Note this does not restore interrupts like task_rq_unlock,
2064 * you need to do so manually after calling.
2066 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2067 __releases(rq1
->lock
)
2068 __releases(rq2
->lock
)
2070 spin_unlock(&rq1
->lock
);
2072 spin_unlock(&rq2
->lock
);
2074 __release(rq2
->lock
);
2078 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2080 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2081 __releases(this_rq
->lock
)
2082 __acquires(busiest
->lock
)
2083 __acquires(this_rq
->lock
)
2087 if (unlikely(!irqs_disabled())) {
2088 /* printk() doesn't work good under rq->lock */
2089 spin_unlock(&this_rq
->lock
);
2092 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2093 if (busiest
< this_rq
) {
2094 spin_unlock(&this_rq
->lock
);
2095 spin_lock(&busiest
->lock
);
2096 spin_lock(&this_rq
->lock
);
2099 spin_lock(&busiest
->lock
);
2105 * If dest_cpu is allowed for this process, migrate the task to it.
2106 * This is accomplished by forcing the cpu_allowed mask to only
2107 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2108 * the cpu_allowed mask is restored.
2110 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2112 struct migration_req req
;
2113 unsigned long flags
;
2116 rq
= task_rq_lock(p
, &flags
);
2117 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2118 || unlikely(cpu_is_offline(dest_cpu
)))
2121 /* force the process onto the specified CPU */
2122 if (migrate_task(p
, dest_cpu
, &req
)) {
2123 /* Need to wait for migration thread (might exit: take ref). */
2124 struct task_struct
*mt
= rq
->migration_thread
;
2126 get_task_struct(mt
);
2127 task_rq_unlock(rq
, &flags
);
2128 wake_up_process(mt
);
2129 put_task_struct(mt
);
2130 wait_for_completion(&req
.done
);
2135 task_rq_unlock(rq
, &flags
);
2139 * sched_exec - execve() is a valuable balancing opportunity, because at
2140 * this point the task has the smallest effective memory and cache footprint.
2142 void sched_exec(void)
2144 int new_cpu
, this_cpu
= get_cpu();
2145 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2147 if (new_cpu
!= this_cpu
)
2148 sched_migrate_task(current
, new_cpu
);
2152 * pull_task - move a task from a remote runqueue to the local runqueue.
2153 * Both runqueues must be locked.
2155 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2156 struct rq
*this_rq
, int this_cpu
)
2158 deactivate_task(src_rq
, p
, 0);
2159 set_task_cpu(p
, this_cpu
);
2160 activate_task(this_rq
, p
, 0);
2162 * Note that idle threads have a prio of MAX_PRIO, for this test
2163 * to be always true for them.
2165 check_preempt_curr(this_rq
, p
);
2169 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2172 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2173 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2177 * We do not migrate tasks that are:
2178 * 1) running (obviously), or
2179 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2180 * 3) are cache-hot on their current CPU.
2182 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2183 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2188 if (task_running(rq
, p
)) {
2189 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2194 * Aggressive migration if:
2195 * 1) task is cache cold, or
2196 * 2) too many balance attempts have failed.
2199 if (!task_hot(p
, rq
->clock
, sd
) ||
2200 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2201 #ifdef CONFIG_SCHEDSTATS
2202 if (task_hot(p
, rq
->clock
, sd
)) {
2203 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2204 schedstat_inc(p
, se
.nr_forced_migrations
);
2210 if (task_hot(p
, rq
->clock
, sd
)) {
2211 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2217 static unsigned long
2218 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2219 unsigned long max_load_move
, struct sched_domain
*sd
,
2220 enum cpu_idle_type idle
, int *all_pinned
,
2221 int *this_best_prio
, struct rq_iterator
*iterator
)
2223 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2224 struct task_struct
*p
;
2225 long rem_load_move
= max_load_move
;
2227 if (max_load_move
== 0)
2233 * Start the load-balancing iterator:
2235 p
= iterator
->start(iterator
->arg
);
2237 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2240 * To help distribute high priority tasks across CPUs we don't
2241 * skip a task if it will be the highest priority task (i.e. smallest
2242 * prio value) on its new queue regardless of its load weight
2244 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2245 SCHED_LOAD_SCALE_FUZZ
;
2246 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2247 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2248 p
= iterator
->next(iterator
->arg
);
2252 pull_task(busiest
, p
, this_rq
, this_cpu
);
2254 rem_load_move
-= p
->se
.load
.weight
;
2257 * We only want to steal up to the prescribed amount of weighted load.
2259 if (rem_load_move
> 0) {
2260 if (p
->prio
< *this_best_prio
)
2261 *this_best_prio
= p
->prio
;
2262 p
= iterator
->next(iterator
->arg
);
2267 * Right now, this is one of only two places pull_task() is called,
2268 * so we can safely collect pull_task() stats here rather than
2269 * inside pull_task().
2271 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2274 *all_pinned
= pinned
;
2276 return max_load_move
- rem_load_move
;
2280 * move_tasks tries to move up to max_load_move weighted load from busiest to
2281 * this_rq, as part of a balancing operation within domain "sd".
2282 * Returns 1 if successful and 0 otherwise.
2284 * Called with both runqueues locked.
2286 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2287 unsigned long max_load_move
,
2288 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2291 const struct sched_class
*class = sched_class_highest
;
2292 unsigned long total_load_moved
= 0;
2293 int this_best_prio
= this_rq
->curr
->prio
;
2297 class->load_balance(this_rq
, this_cpu
, busiest
,
2298 max_load_move
- total_load_moved
,
2299 sd
, idle
, all_pinned
, &this_best_prio
);
2300 class = class->next
;
2301 } while (class && max_load_move
> total_load_moved
);
2303 return total_load_moved
> 0;
2307 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2308 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2309 struct rq_iterator
*iterator
)
2311 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2315 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2316 pull_task(busiest
, p
, this_rq
, this_cpu
);
2318 * Right now, this is only the second place pull_task()
2319 * is called, so we can safely collect pull_task()
2320 * stats here rather than inside pull_task().
2322 schedstat_inc(sd
, lb_gained
[idle
]);
2326 p
= iterator
->next(iterator
->arg
);
2333 * move_one_task tries to move exactly one task from busiest to this_rq, as
2334 * part of active balancing operations within "domain".
2335 * Returns 1 if successful and 0 otherwise.
2337 * Called with both runqueues locked.
2339 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2340 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2342 const struct sched_class
*class;
2344 for (class = sched_class_highest
; class; class = class->next
)
2345 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2352 * find_busiest_group finds and returns the busiest CPU group within the
2353 * domain. It calculates and returns the amount of weighted load which
2354 * should be moved to restore balance via the imbalance parameter.
2356 static struct sched_group
*
2357 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2358 unsigned long *imbalance
, enum cpu_idle_type idle
,
2359 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2361 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2362 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2363 unsigned long max_pull
;
2364 unsigned long busiest_load_per_task
, busiest_nr_running
;
2365 unsigned long this_load_per_task
, this_nr_running
;
2366 int load_idx
, group_imb
= 0;
2367 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2368 int power_savings_balance
= 1;
2369 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2370 unsigned long min_nr_running
= ULONG_MAX
;
2371 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2374 max_load
= this_load
= total_load
= total_pwr
= 0;
2375 busiest_load_per_task
= busiest_nr_running
= 0;
2376 this_load_per_task
= this_nr_running
= 0;
2377 if (idle
== CPU_NOT_IDLE
)
2378 load_idx
= sd
->busy_idx
;
2379 else if (idle
== CPU_NEWLY_IDLE
)
2380 load_idx
= sd
->newidle_idx
;
2382 load_idx
= sd
->idle_idx
;
2385 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2388 int __group_imb
= 0;
2389 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2390 unsigned long sum_nr_running
, sum_weighted_load
;
2392 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2395 balance_cpu
= first_cpu(group
->cpumask
);
2397 /* Tally up the load of all CPUs in the group */
2398 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2400 min_cpu_load
= ~0UL;
2402 for_each_cpu_mask(i
, group
->cpumask
) {
2405 if (!cpu_isset(i
, *cpus
))
2410 if (*sd_idle
&& rq
->nr_running
)
2413 /* Bias balancing toward cpus of our domain */
2415 if (idle_cpu(i
) && !first_idle_cpu
) {
2420 load
= target_load(i
, load_idx
);
2422 load
= source_load(i
, load_idx
);
2423 if (load
> max_cpu_load
)
2424 max_cpu_load
= load
;
2425 if (min_cpu_load
> load
)
2426 min_cpu_load
= load
;
2430 sum_nr_running
+= rq
->nr_running
;
2431 sum_weighted_load
+= weighted_cpuload(i
);
2435 * First idle cpu or the first cpu(busiest) in this sched group
2436 * is eligible for doing load balancing at this and above
2437 * domains. In the newly idle case, we will allow all the cpu's
2438 * to do the newly idle load balance.
2440 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2441 balance_cpu
!= this_cpu
&& balance
) {
2446 total_load
+= avg_load
;
2447 total_pwr
+= group
->__cpu_power
;
2449 /* Adjust by relative CPU power of the group */
2450 avg_load
= sg_div_cpu_power(group
,
2451 avg_load
* SCHED_LOAD_SCALE
);
2453 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2456 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2459 this_load
= avg_load
;
2461 this_nr_running
= sum_nr_running
;
2462 this_load_per_task
= sum_weighted_load
;
2463 } else if (avg_load
> max_load
&&
2464 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2465 max_load
= avg_load
;
2467 busiest_nr_running
= sum_nr_running
;
2468 busiest_load_per_task
= sum_weighted_load
;
2469 group_imb
= __group_imb
;
2472 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2474 * Busy processors will not participate in power savings
2477 if (idle
== CPU_NOT_IDLE
||
2478 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2482 * If the local group is idle or completely loaded
2483 * no need to do power savings balance at this domain
2485 if (local_group
&& (this_nr_running
>= group_capacity
||
2487 power_savings_balance
= 0;
2490 * If a group is already running at full capacity or idle,
2491 * don't include that group in power savings calculations
2493 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2498 * Calculate the group which has the least non-idle load.
2499 * This is the group from where we need to pick up the load
2502 if ((sum_nr_running
< min_nr_running
) ||
2503 (sum_nr_running
== min_nr_running
&&
2504 first_cpu(group
->cpumask
) <
2505 first_cpu(group_min
->cpumask
))) {
2507 min_nr_running
= sum_nr_running
;
2508 min_load_per_task
= sum_weighted_load
/
2513 * Calculate the group which is almost near its
2514 * capacity but still has some space to pick up some load
2515 * from other group and save more power
2517 if (sum_nr_running
<= group_capacity
- 1) {
2518 if (sum_nr_running
> leader_nr_running
||
2519 (sum_nr_running
== leader_nr_running
&&
2520 first_cpu(group
->cpumask
) >
2521 first_cpu(group_leader
->cpumask
))) {
2522 group_leader
= group
;
2523 leader_nr_running
= sum_nr_running
;
2528 group
= group
->next
;
2529 } while (group
!= sd
->groups
);
2531 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2534 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2536 if (this_load
>= avg_load
||
2537 100*max_load
<= sd
->imbalance_pct
*this_load
)
2540 busiest_load_per_task
/= busiest_nr_running
;
2542 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2545 * We're trying to get all the cpus to the average_load, so we don't
2546 * want to push ourselves above the average load, nor do we wish to
2547 * reduce the max loaded cpu below the average load, as either of these
2548 * actions would just result in more rebalancing later, and ping-pong
2549 * tasks around. Thus we look for the minimum possible imbalance.
2550 * Negative imbalances (*we* are more loaded than anyone else) will
2551 * be counted as no imbalance for these purposes -- we can't fix that
2552 * by pulling tasks to us. Be careful of negative numbers as they'll
2553 * appear as very large values with unsigned longs.
2555 if (max_load
<= busiest_load_per_task
)
2559 * In the presence of smp nice balancing, certain scenarios can have
2560 * max load less than avg load(as we skip the groups at or below
2561 * its cpu_power, while calculating max_load..)
2563 if (max_load
< avg_load
) {
2565 goto small_imbalance
;
2568 /* Don't want to pull so many tasks that a group would go idle */
2569 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2571 /* How much load to actually move to equalise the imbalance */
2572 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2573 (avg_load
- this_load
) * this->__cpu_power
)
2577 * if *imbalance is less than the average load per runnable task
2578 * there is no gaurantee that any tasks will be moved so we'll have
2579 * a think about bumping its value to force at least one task to be
2582 if (*imbalance
< busiest_load_per_task
) {
2583 unsigned long tmp
, pwr_now
, pwr_move
;
2587 pwr_move
= pwr_now
= 0;
2589 if (this_nr_running
) {
2590 this_load_per_task
/= this_nr_running
;
2591 if (busiest_load_per_task
> this_load_per_task
)
2594 this_load_per_task
= SCHED_LOAD_SCALE
;
2596 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2597 busiest_load_per_task
* imbn
) {
2598 *imbalance
= busiest_load_per_task
;
2603 * OK, we don't have enough imbalance to justify moving tasks,
2604 * however we may be able to increase total CPU power used by
2608 pwr_now
+= busiest
->__cpu_power
*
2609 min(busiest_load_per_task
, max_load
);
2610 pwr_now
+= this->__cpu_power
*
2611 min(this_load_per_task
, this_load
);
2612 pwr_now
/= SCHED_LOAD_SCALE
;
2614 /* Amount of load we'd subtract */
2615 tmp
= sg_div_cpu_power(busiest
,
2616 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2618 pwr_move
+= busiest
->__cpu_power
*
2619 min(busiest_load_per_task
, max_load
- tmp
);
2621 /* Amount of load we'd add */
2622 if (max_load
* busiest
->__cpu_power
<
2623 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2624 tmp
= sg_div_cpu_power(this,
2625 max_load
* busiest
->__cpu_power
);
2627 tmp
= sg_div_cpu_power(this,
2628 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2629 pwr_move
+= this->__cpu_power
*
2630 min(this_load_per_task
, this_load
+ tmp
);
2631 pwr_move
/= SCHED_LOAD_SCALE
;
2633 /* Move if we gain throughput */
2634 if (pwr_move
> pwr_now
)
2635 *imbalance
= busiest_load_per_task
;
2641 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2642 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2645 if (this == group_leader
&& group_leader
!= group_min
) {
2646 *imbalance
= min_load_per_task
;
2656 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2659 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2660 unsigned long imbalance
, cpumask_t
*cpus
)
2662 struct rq
*busiest
= NULL
, *rq
;
2663 unsigned long max_load
= 0;
2666 for_each_cpu_mask(i
, group
->cpumask
) {
2669 if (!cpu_isset(i
, *cpus
))
2673 wl
= weighted_cpuload(i
);
2675 if (rq
->nr_running
== 1 && wl
> imbalance
)
2678 if (wl
> max_load
) {
2688 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2689 * so long as it is large enough.
2691 #define MAX_PINNED_INTERVAL 512
2694 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2695 * tasks if there is an imbalance.
2697 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2698 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2701 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2702 struct sched_group
*group
;
2703 unsigned long imbalance
;
2705 cpumask_t cpus
= CPU_MASK_ALL
;
2706 unsigned long flags
;
2709 * When power savings policy is enabled for the parent domain, idle
2710 * sibling can pick up load irrespective of busy siblings. In this case,
2711 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2712 * portraying it as CPU_NOT_IDLE.
2714 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2715 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2718 schedstat_inc(sd
, lb_count
[idle
]);
2721 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2728 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2732 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2734 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2738 BUG_ON(busiest
== this_rq
);
2740 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2743 if (busiest
->nr_running
> 1) {
2745 * Attempt to move tasks. If find_busiest_group has found
2746 * an imbalance but busiest->nr_running <= 1, the group is
2747 * still unbalanced. ld_moved simply stays zero, so it is
2748 * correctly treated as an imbalance.
2750 local_irq_save(flags
);
2751 double_rq_lock(this_rq
, busiest
);
2752 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2753 imbalance
, sd
, idle
, &all_pinned
);
2754 double_rq_unlock(this_rq
, busiest
);
2755 local_irq_restore(flags
);
2758 * some other cpu did the load balance for us.
2760 if (ld_moved
&& this_cpu
!= smp_processor_id())
2761 resched_cpu(this_cpu
);
2763 /* All tasks on this runqueue were pinned by CPU affinity */
2764 if (unlikely(all_pinned
)) {
2765 cpu_clear(cpu_of(busiest
), cpus
);
2766 if (!cpus_empty(cpus
))
2773 schedstat_inc(sd
, lb_failed
[idle
]);
2774 sd
->nr_balance_failed
++;
2776 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2778 spin_lock_irqsave(&busiest
->lock
, flags
);
2780 /* don't kick the migration_thread, if the curr
2781 * task on busiest cpu can't be moved to this_cpu
2783 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2784 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2786 goto out_one_pinned
;
2789 if (!busiest
->active_balance
) {
2790 busiest
->active_balance
= 1;
2791 busiest
->push_cpu
= this_cpu
;
2794 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2796 wake_up_process(busiest
->migration_thread
);
2799 * We've kicked active balancing, reset the failure
2802 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2805 sd
->nr_balance_failed
= 0;
2807 if (likely(!active_balance
)) {
2808 /* We were unbalanced, so reset the balancing interval */
2809 sd
->balance_interval
= sd
->min_interval
;
2812 * If we've begun active balancing, start to back off. This
2813 * case may not be covered by the all_pinned logic if there
2814 * is only 1 task on the busy runqueue (because we don't call
2817 if (sd
->balance_interval
< sd
->max_interval
)
2818 sd
->balance_interval
*= 2;
2821 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2822 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2827 schedstat_inc(sd
, lb_balanced
[idle
]);
2829 sd
->nr_balance_failed
= 0;
2832 /* tune up the balancing interval */
2833 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2834 (sd
->balance_interval
< sd
->max_interval
))
2835 sd
->balance_interval
*= 2;
2837 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2838 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2844 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2845 * tasks if there is an imbalance.
2847 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2848 * this_rq is locked.
2851 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2853 struct sched_group
*group
;
2854 struct rq
*busiest
= NULL
;
2855 unsigned long imbalance
;
2859 cpumask_t cpus
= CPU_MASK_ALL
;
2862 * When power savings policy is enabled for the parent domain, idle
2863 * sibling can pick up load irrespective of busy siblings. In this case,
2864 * let the state of idle sibling percolate up as IDLE, instead of
2865 * portraying it as CPU_NOT_IDLE.
2867 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2868 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2871 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
2873 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2874 &sd_idle
, &cpus
, NULL
);
2876 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2880 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2883 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2887 BUG_ON(busiest
== this_rq
);
2889 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2892 if (busiest
->nr_running
> 1) {
2893 /* Attempt to move tasks */
2894 double_lock_balance(this_rq
, busiest
);
2895 /* this_rq->clock is already updated */
2896 update_rq_clock(busiest
);
2897 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2898 imbalance
, sd
, CPU_NEWLY_IDLE
,
2900 spin_unlock(&busiest
->lock
);
2902 if (unlikely(all_pinned
)) {
2903 cpu_clear(cpu_of(busiest
), cpus
);
2904 if (!cpus_empty(cpus
))
2910 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2911 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2912 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2915 sd
->nr_balance_failed
= 0;
2920 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2921 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2922 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2924 sd
->nr_balance_failed
= 0;
2930 * idle_balance is called by schedule() if this_cpu is about to become
2931 * idle. Attempts to pull tasks from other CPUs.
2933 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2935 struct sched_domain
*sd
;
2936 int pulled_task
= -1;
2937 unsigned long next_balance
= jiffies
+ HZ
;
2939 for_each_domain(this_cpu
, sd
) {
2940 unsigned long interval
;
2942 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2945 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2946 /* If we've pulled tasks over stop searching: */
2947 pulled_task
= load_balance_newidle(this_cpu
,
2950 interval
= msecs_to_jiffies(sd
->balance_interval
);
2951 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2952 next_balance
= sd
->last_balance
+ interval
;
2956 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2958 * We are going idle. next_balance may be set based on
2959 * a busy processor. So reset next_balance.
2961 this_rq
->next_balance
= next_balance
;
2966 * active_load_balance is run by migration threads. It pushes running tasks
2967 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2968 * running on each physical CPU where possible, and avoids physical /
2969 * logical imbalances.
2971 * Called with busiest_rq locked.
2973 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2975 int target_cpu
= busiest_rq
->push_cpu
;
2976 struct sched_domain
*sd
;
2977 struct rq
*target_rq
;
2979 /* Is there any task to move? */
2980 if (busiest_rq
->nr_running
<= 1)
2983 target_rq
= cpu_rq(target_cpu
);
2986 * This condition is "impossible", if it occurs
2987 * we need to fix it. Originally reported by
2988 * Bjorn Helgaas on a 128-cpu setup.
2990 BUG_ON(busiest_rq
== target_rq
);
2992 /* move a task from busiest_rq to target_rq */
2993 double_lock_balance(busiest_rq
, target_rq
);
2994 update_rq_clock(busiest_rq
);
2995 update_rq_clock(target_rq
);
2997 /* Search for an sd spanning us and the target CPU. */
2998 for_each_domain(target_cpu
, sd
) {
2999 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3000 cpu_isset(busiest_cpu
, sd
->span
))
3005 schedstat_inc(sd
, alb_count
);
3007 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3009 schedstat_inc(sd
, alb_pushed
);
3011 schedstat_inc(sd
, alb_failed
);
3013 spin_unlock(&target_rq
->lock
);
3018 atomic_t load_balancer
;
3020 } nohz ____cacheline_aligned
= {
3021 .load_balancer
= ATOMIC_INIT(-1),
3022 .cpu_mask
= CPU_MASK_NONE
,
3026 * This routine will try to nominate the ilb (idle load balancing)
3027 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3028 * load balancing on behalf of all those cpus. If all the cpus in the system
3029 * go into this tickless mode, then there will be no ilb owner (as there is
3030 * no need for one) and all the cpus will sleep till the next wakeup event
3033 * For the ilb owner, tick is not stopped. And this tick will be used
3034 * for idle load balancing. ilb owner will still be part of
3037 * While stopping the tick, this cpu will become the ilb owner if there
3038 * is no other owner. And will be the owner till that cpu becomes busy
3039 * or if all cpus in the system stop their ticks at which point
3040 * there is no need for ilb owner.
3042 * When the ilb owner becomes busy, it nominates another owner, during the
3043 * next busy scheduler_tick()
3045 int select_nohz_load_balancer(int stop_tick
)
3047 int cpu
= smp_processor_id();
3050 cpu_set(cpu
, nohz
.cpu_mask
);
3051 cpu_rq(cpu
)->in_nohz_recently
= 1;
3054 * If we are going offline and still the leader, give up!
3056 if (cpu_is_offline(cpu
) &&
3057 atomic_read(&nohz
.load_balancer
) == cpu
) {
3058 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3063 /* time for ilb owner also to sleep */
3064 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3065 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3066 atomic_set(&nohz
.load_balancer
, -1);
3070 if (atomic_read(&nohz
.load_balancer
) == -1) {
3071 /* make me the ilb owner */
3072 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3074 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3077 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3080 cpu_clear(cpu
, nohz
.cpu_mask
);
3082 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3083 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3090 static DEFINE_SPINLOCK(balancing
);
3093 * It checks each scheduling domain to see if it is due to be balanced,
3094 * and initiates a balancing operation if so.
3096 * Balancing parameters are set up in arch_init_sched_domains.
3098 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3101 struct rq
*rq
= cpu_rq(cpu
);
3102 unsigned long interval
;
3103 struct sched_domain
*sd
;
3104 /* Earliest time when we have to do rebalance again */
3105 unsigned long next_balance
= jiffies
+ 60*HZ
;
3106 int update_next_balance
= 0;
3108 for_each_domain(cpu
, sd
) {
3109 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3112 interval
= sd
->balance_interval
;
3113 if (idle
!= CPU_IDLE
)
3114 interval
*= sd
->busy_factor
;
3116 /* scale ms to jiffies */
3117 interval
= msecs_to_jiffies(interval
);
3118 if (unlikely(!interval
))
3120 if (interval
> HZ
*NR_CPUS
/10)
3121 interval
= HZ
*NR_CPUS
/10;
3124 if (sd
->flags
& SD_SERIALIZE
) {
3125 if (!spin_trylock(&balancing
))
3129 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3130 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3132 * We've pulled tasks over so either we're no
3133 * longer idle, or one of our SMT siblings is
3136 idle
= CPU_NOT_IDLE
;
3138 sd
->last_balance
= jiffies
;
3140 if (sd
->flags
& SD_SERIALIZE
)
3141 spin_unlock(&balancing
);
3143 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3144 next_balance
= sd
->last_balance
+ interval
;
3145 update_next_balance
= 1;
3149 * Stop the load balance at this level. There is another
3150 * CPU in our sched group which is doing load balancing more
3158 * next_balance will be updated only when there is a need.
3159 * When the cpu is attached to null domain for ex, it will not be
3162 if (likely(update_next_balance
))
3163 rq
->next_balance
= next_balance
;
3167 * run_rebalance_domains is triggered when needed from the scheduler tick.
3168 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3169 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3171 static void run_rebalance_domains(struct softirq_action
*h
)
3173 int this_cpu
= smp_processor_id();
3174 struct rq
*this_rq
= cpu_rq(this_cpu
);
3175 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3176 CPU_IDLE
: CPU_NOT_IDLE
;
3178 rebalance_domains(this_cpu
, idle
);
3182 * If this cpu is the owner for idle load balancing, then do the
3183 * balancing on behalf of the other idle cpus whose ticks are
3186 if (this_rq
->idle_at_tick
&&
3187 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3188 cpumask_t cpus
= nohz
.cpu_mask
;
3192 cpu_clear(this_cpu
, cpus
);
3193 for_each_cpu_mask(balance_cpu
, cpus
) {
3195 * If this cpu gets work to do, stop the load balancing
3196 * work being done for other cpus. Next load
3197 * balancing owner will pick it up.
3202 rebalance_domains(balance_cpu
, CPU_IDLE
);
3204 rq
= cpu_rq(balance_cpu
);
3205 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3206 this_rq
->next_balance
= rq
->next_balance
;
3213 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3215 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3216 * idle load balancing owner or decide to stop the periodic load balancing,
3217 * if the whole system is idle.
3219 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3223 * If we were in the nohz mode recently and busy at the current
3224 * scheduler tick, then check if we need to nominate new idle
3227 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3228 rq
->in_nohz_recently
= 0;
3230 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3231 cpu_clear(cpu
, nohz
.cpu_mask
);
3232 atomic_set(&nohz
.load_balancer
, -1);
3235 if (atomic_read(&nohz
.load_balancer
) == -1) {
3237 * simple selection for now: Nominate the
3238 * first cpu in the nohz list to be the next
3241 * TBD: Traverse the sched domains and nominate
3242 * the nearest cpu in the nohz.cpu_mask.
3244 int ilb
= first_cpu(nohz
.cpu_mask
);
3252 * If this cpu is idle and doing idle load balancing for all the
3253 * cpus with ticks stopped, is it time for that to stop?
3255 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3256 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3262 * If this cpu is idle and the idle load balancing is done by
3263 * someone else, then no need raise the SCHED_SOFTIRQ
3265 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3266 cpu_isset(cpu
, nohz
.cpu_mask
))
3269 if (time_after_eq(jiffies
, rq
->next_balance
))
3270 raise_softirq(SCHED_SOFTIRQ
);
3273 #else /* CONFIG_SMP */
3276 * on UP we do not need to balance between CPUs:
3278 static inline void idle_balance(int cpu
, struct rq
*rq
)
3284 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3286 EXPORT_PER_CPU_SYMBOL(kstat
);
3289 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3290 * that have not yet been banked in case the task is currently running.
3292 unsigned long long task_sched_runtime(struct task_struct
*p
)
3294 unsigned long flags
;
3298 rq
= task_rq_lock(p
, &flags
);
3299 ns
= p
->se
.sum_exec_runtime
;
3300 if (task_current(rq
, p
)) {
3301 update_rq_clock(rq
);
3302 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3303 if ((s64
)delta_exec
> 0)
3306 task_rq_unlock(rq
, &flags
);
3312 * Account user cpu time to a process.
3313 * @p: the process that the cpu time gets accounted to
3314 * @cputime: the cpu time spent in user space since the last update
3316 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3318 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3321 p
->utime
= cputime_add(p
->utime
, cputime
);
3323 /* Add user time to cpustat. */
3324 tmp
= cputime_to_cputime64(cputime
);
3325 if (TASK_NICE(p
) > 0)
3326 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3328 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3332 * Account guest cpu time to a process.
3333 * @p: the process that the cpu time gets accounted to
3334 * @cputime: the cpu time spent in virtual machine since the last update
3336 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3339 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3341 tmp
= cputime_to_cputime64(cputime
);
3343 p
->utime
= cputime_add(p
->utime
, cputime
);
3344 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3346 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3347 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3351 * Account scaled user cpu time to a process.
3352 * @p: the process that the cpu time gets accounted to
3353 * @cputime: the cpu time spent in user space since the last update
3355 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3357 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3361 * Account system cpu time to a process.
3362 * @p: the process that the cpu time gets accounted to
3363 * @hardirq_offset: the offset to subtract from hardirq_count()
3364 * @cputime: the cpu time spent in kernel space since the last update
3366 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3369 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3370 struct rq
*rq
= this_rq();
3373 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3374 return account_guest_time(p
, cputime
);
3376 p
->stime
= cputime_add(p
->stime
, cputime
);
3378 /* Add system time to cpustat. */
3379 tmp
= cputime_to_cputime64(cputime
);
3380 if (hardirq_count() - hardirq_offset
)
3381 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3382 else if (softirq_count())
3383 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3384 else if (p
!= rq
->idle
)
3385 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3386 else if (atomic_read(&rq
->nr_iowait
) > 0)
3387 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3389 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3390 /* Account for system time used */
3391 acct_update_integrals(p
);
3395 * Account scaled system cpu time to a process.
3396 * @p: the process that the cpu time gets accounted to
3397 * @hardirq_offset: the offset to subtract from hardirq_count()
3398 * @cputime: the cpu time spent in kernel space since the last update
3400 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3402 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3406 * Account for involuntary wait time.
3407 * @p: the process from which the cpu time has been stolen
3408 * @steal: the cpu time spent in involuntary wait
3410 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3412 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3413 cputime64_t tmp
= cputime_to_cputime64(steal
);
3414 struct rq
*rq
= this_rq();
3416 if (p
== rq
->idle
) {
3417 p
->stime
= cputime_add(p
->stime
, steal
);
3418 if (atomic_read(&rq
->nr_iowait
) > 0)
3419 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3421 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3423 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3427 * This function gets called by the timer code, with HZ frequency.
3428 * We call it with interrupts disabled.
3430 * It also gets called by the fork code, when changing the parent's
3433 void scheduler_tick(void)
3435 int cpu
= smp_processor_id();
3436 struct rq
*rq
= cpu_rq(cpu
);
3437 struct task_struct
*curr
= rq
->curr
;
3438 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3440 spin_lock(&rq
->lock
);
3441 __update_rq_clock(rq
);
3443 * Let rq->clock advance by at least TICK_NSEC:
3445 if (unlikely(rq
->clock
< next_tick
))
3446 rq
->clock
= next_tick
;
3447 rq
->tick_timestamp
= rq
->clock
;
3448 update_cpu_load(rq
);
3449 if (curr
!= rq
->idle
) /* FIXME: needed? */
3450 curr
->sched_class
->task_tick(rq
, curr
);
3451 spin_unlock(&rq
->lock
);
3454 rq
->idle_at_tick
= idle_cpu(cpu
);
3455 trigger_load_balance(rq
, cpu
);
3459 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3461 void fastcall
add_preempt_count(int val
)
3466 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3468 preempt_count() += val
;
3470 * Spinlock count overflowing soon?
3472 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3475 EXPORT_SYMBOL(add_preempt_count
);
3477 void fastcall
sub_preempt_count(int val
)
3482 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3485 * Is the spinlock portion underflowing?
3487 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3488 !(preempt_count() & PREEMPT_MASK
)))
3491 preempt_count() -= val
;
3493 EXPORT_SYMBOL(sub_preempt_count
);
3498 * Print scheduling while atomic bug:
3500 static noinline
void __schedule_bug(struct task_struct
*prev
)
3502 struct pt_regs
*regs
= get_irq_regs();
3504 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3505 prev
->comm
, prev
->pid
, preempt_count());
3507 debug_show_held_locks(prev
);
3508 if (irqs_disabled())
3509 print_irqtrace_events(prev
);
3518 * Various schedule()-time debugging checks and statistics:
3520 static inline void schedule_debug(struct task_struct
*prev
)
3523 * Test if we are atomic. Since do_exit() needs to call into
3524 * schedule() atomically, we ignore that path for now.
3525 * Otherwise, whine if we are scheduling when we should not be.
3527 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3528 __schedule_bug(prev
);
3530 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3532 schedstat_inc(this_rq(), sched_count
);
3533 #ifdef CONFIG_SCHEDSTATS
3534 if (unlikely(prev
->lock_depth
>= 0)) {
3535 schedstat_inc(this_rq(), bkl_count
);
3536 schedstat_inc(prev
, sched_info
.bkl_count
);
3542 * Pick up the highest-prio task:
3544 static inline struct task_struct
*
3545 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3547 const struct sched_class
*class;
3548 struct task_struct
*p
;
3551 * Optimization: we know that if all tasks are in
3552 * the fair class we can call that function directly:
3554 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3555 p
= fair_sched_class
.pick_next_task(rq
);
3560 class = sched_class_highest
;
3562 p
= class->pick_next_task(rq
);
3566 * Will never be NULL as the idle class always
3567 * returns a non-NULL p:
3569 class = class->next
;
3574 * schedule() is the main scheduler function.
3576 asmlinkage
void __sched
schedule(void)
3578 struct task_struct
*prev
, *next
;
3585 cpu
= smp_processor_id();
3589 switch_count
= &prev
->nivcsw
;
3591 release_kernel_lock(prev
);
3592 need_resched_nonpreemptible
:
3594 schedule_debug(prev
);
3597 * Do the rq-clock update outside the rq lock:
3599 local_irq_disable();
3600 __update_rq_clock(rq
);
3601 spin_lock(&rq
->lock
);
3602 clear_tsk_need_resched(prev
);
3604 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3605 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3606 unlikely(signal_pending(prev
)))) {
3607 prev
->state
= TASK_RUNNING
;
3609 deactivate_task(rq
, prev
, 1);
3611 switch_count
= &prev
->nvcsw
;
3614 schedule_balance_rt(rq
, prev
);
3616 if (unlikely(!rq
->nr_running
))
3617 idle_balance(cpu
, rq
);
3619 prev
->sched_class
->put_prev_task(rq
, prev
);
3620 next
= pick_next_task(rq
, prev
);
3622 sched_info_switch(prev
, next
);
3624 if (likely(prev
!= next
)) {
3629 context_switch(rq
, prev
, next
); /* unlocks the rq */
3631 spin_unlock_irq(&rq
->lock
);
3633 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3634 cpu
= smp_processor_id();
3636 goto need_resched_nonpreemptible
;
3638 preempt_enable_no_resched();
3639 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3642 EXPORT_SYMBOL(schedule
);
3644 #ifdef CONFIG_PREEMPT
3646 * this is the entry point to schedule() from in-kernel preemption
3647 * off of preempt_enable. Kernel preemptions off return from interrupt
3648 * occur there and call schedule directly.
3650 asmlinkage
void __sched
preempt_schedule(void)
3652 struct thread_info
*ti
= current_thread_info();
3653 #ifdef CONFIG_PREEMPT_BKL
3654 struct task_struct
*task
= current
;
3655 int saved_lock_depth
;
3658 * If there is a non-zero preempt_count or interrupts are disabled,
3659 * we do not want to preempt the current task. Just return..
3661 if (likely(ti
->preempt_count
|| irqs_disabled()))
3665 add_preempt_count(PREEMPT_ACTIVE
);
3668 * We keep the big kernel semaphore locked, but we
3669 * clear ->lock_depth so that schedule() doesnt
3670 * auto-release the semaphore:
3672 #ifdef CONFIG_PREEMPT_BKL
3673 saved_lock_depth
= task
->lock_depth
;
3674 task
->lock_depth
= -1;
3677 #ifdef CONFIG_PREEMPT_BKL
3678 task
->lock_depth
= saved_lock_depth
;
3680 sub_preempt_count(PREEMPT_ACTIVE
);
3683 * Check again in case we missed a preemption opportunity
3684 * between schedule and now.
3687 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3689 EXPORT_SYMBOL(preempt_schedule
);
3692 * this is the entry point to schedule() from kernel preemption
3693 * off of irq context.
3694 * Note, that this is called and return with irqs disabled. This will
3695 * protect us against recursive calling from irq.
3697 asmlinkage
void __sched
preempt_schedule_irq(void)
3699 struct thread_info
*ti
= current_thread_info();
3700 #ifdef CONFIG_PREEMPT_BKL
3701 struct task_struct
*task
= current
;
3702 int saved_lock_depth
;
3704 /* Catch callers which need to be fixed */
3705 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3708 add_preempt_count(PREEMPT_ACTIVE
);
3711 * We keep the big kernel semaphore locked, but we
3712 * clear ->lock_depth so that schedule() doesnt
3713 * auto-release the semaphore:
3715 #ifdef CONFIG_PREEMPT_BKL
3716 saved_lock_depth
= task
->lock_depth
;
3717 task
->lock_depth
= -1;
3721 local_irq_disable();
3722 #ifdef CONFIG_PREEMPT_BKL
3723 task
->lock_depth
= saved_lock_depth
;
3725 sub_preempt_count(PREEMPT_ACTIVE
);
3728 * Check again in case we missed a preemption opportunity
3729 * between schedule and now.
3732 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3735 #endif /* CONFIG_PREEMPT */
3737 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3740 return try_to_wake_up(curr
->private, mode
, sync
);
3742 EXPORT_SYMBOL(default_wake_function
);
3745 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3746 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3747 * number) then we wake all the non-exclusive tasks and one exclusive task.
3749 * There are circumstances in which we can try to wake a task which has already
3750 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3751 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3753 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3754 int nr_exclusive
, int sync
, void *key
)
3756 wait_queue_t
*curr
, *next
;
3758 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3759 unsigned flags
= curr
->flags
;
3761 if (curr
->func(curr
, mode
, sync
, key
) &&
3762 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3768 * __wake_up - wake up threads blocked on a waitqueue.
3770 * @mode: which threads
3771 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3772 * @key: is directly passed to the wakeup function
3774 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3775 int nr_exclusive
, void *key
)
3777 unsigned long flags
;
3779 spin_lock_irqsave(&q
->lock
, flags
);
3780 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3781 spin_unlock_irqrestore(&q
->lock
, flags
);
3783 EXPORT_SYMBOL(__wake_up
);
3786 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3788 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3790 __wake_up_common(q
, mode
, 1, 0, NULL
);
3794 * __wake_up_sync - wake up threads blocked on a waitqueue.
3796 * @mode: which threads
3797 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3799 * The sync wakeup differs that the waker knows that it will schedule
3800 * away soon, so while the target thread will be woken up, it will not
3801 * be migrated to another CPU - ie. the two threads are 'synchronized'
3802 * with each other. This can prevent needless bouncing between CPUs.
3804 * On UP it can prevent extra preemption.
3807 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3809 unsigned long flags
;
3815 if (unlikely(!nr_exclusive
))
3818 spin_lock_irqsave(&q
->lock
, flags
);
3819 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3820 spin_unlock_irqrestore(&q
->lock
, flags
);
3822 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3824 void complete(struct completion
*x
)
3826 unsigned long flags
;
3828 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3830 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3832 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3834 EXPORT_SYMBOL(complete
);
3836 void complete_all(struct completion
*x
)
3838 unsigned long flags
;
3840 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3841 x
->done
+= UINT_MAX
/2;
3842 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3844 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3846 EXPORT_SYMBOL(complete_all
);
3848 static inline long __sched
3849 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3852 DECLARE_WAITQUEUE(wait
, current
);
3854 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3855 __add_wait_queue_tail(&x
->wait
, &wait
);
3857 if (state
== TASK_INTERRUPTIBLE
&&
3858 signal_pending(current
)) {
3859 __remove_wait_queue(&x
->wait
, &wait
);
3860 return -ERESTARTSYS
;
3862 __set_current_state(state
);
3863 spin_unlock_irq(&x
->wait
.lock
);
3864 timeout
= schedule_timeout(timeout
);
3865 spin_lock_irq(&x
->wait
.lock
);
3867 __remove_wait_queue(&x
->wait
, &wait
);
3871 __remove_wait_queue(&x
->wait
, &wait
);
3878 wait_for_common(struct completion
*x
, long timeout
, int state
)
3882 spin_lock_irq(&x
->wait
.lock
);
3883 timeout
= do_wait_for_common(x
, timeout
, state
);
3884 spin_unlock_irq(&x
->wait
.lock
);
3888 void __sched
wait_for_completion(struct completion
*x
)
3890 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3892 EXPORT_SYMBOL(wait_for_completion
);
3894 unsigned long __sched
3895 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3897 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3899 EXPORT_SYMBOL(wait_for_completion_timeout
);
3901 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3903 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3904 if (t
== -ERESTARTSYS
)
3908 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3910 unsigned long __sched
3911 wait_for_completion_interruptible_timeout(struct completion
*x
,
3912 unsigned long timeout
)
3914 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3916 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3919 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3921 unsigned long flags
;
3924 init_waitqueue_entry(&wait
, current
);
3926 __set_current_state(state
);
3928 spin_lock_irqsave(&q
->lock
, flags
);
3929 __add_wait_queue(q
, &wait
);
3930 spin_unlock(&q
->lock
);
3931 timeout
= schedule_timeout(timeout
);
3932 spin_lock_irq(&q
->lock
);
3933 __remove_wait_queue(q
, &wait
);
3934 spin_unlock_irqrestore(&q
->lock
, flags
);
3939 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3941 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3943 EXPORT_SYMBOL(interruptible_sleep_on
);
3946 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3948 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3950 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3952 void __sched
sleep_on(wait_queue_head_t
*q
)
3954 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3956 EXPORT_SYMBOL(sleep_on
);
3958 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3960 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3962 EXPORT_SYMBOL(sleep_on_timeout
);
3964 #ifdef CONFIG_RT_MUTEXES
3967 * rt_mutex_setprio - set the current priority of a task
3969 * @prio: prio value (kernel-internal form)
3971 * This function changes the 'effective' priority of a task. It does
3972 * not touch ->normal_prio like __setscheduler().
3974 * Used by the rt_mutex code to implement priority inheritance logic.
3976 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3978 unsigned long flags
;
3979 int oldprio
, on_rq
, running
;
3982 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3984 rq
= task_rq_lock(p
, &flags
);
3985 update_rq_clock(rq
);
3988 on_rq
= p
->se
.on_rq
;
3989 running
= task_current(rq
, p
);
3991 dequeue_task(rq
, p
, 0);
3993 p
->sched_class
->put_prev_task(rq
, p
);
3997 p
->sched_class
= &rt_sched_class
;
3999 p
->sched_class
= &fair_sched_class
;
4005 p
->sched_class
->set_curr_task(rq
);
4006 enqueue_task(rq
, p
, 0);
4008 * Reschedule if we are currently running on this runqueue and
4009 * our priority decreased, or if we are not currently running on
4010 * this runqueue and our priority is higher than the current's
4013 if (p
->prio
> oldprio
)
4014 resched_task(rq
->curr
);
4016 check_preempt_curr(rq
, p
);
4019 task_rq_unlock(rq
, &flags
);
4024 void set_user_nice(struct task_struct
*p
, long nice
)
4026 int old_prio
, delta
, on_rq
;
4027 unsigned long flags
;
4030 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4033 * We have to be careful, if called from sys_setpriority(),
4034 * the task might be in the middle of scheduling on another CPU.
4036 rq
= task_rq_lock(p
, &flags
);
4037 update_rq_clock(rq
);
4039 * The RT priorities are set via sched_setscheduler(), but we still
4040 * allow the 'normal' nice value to be set - but as expected
4041 * it wont have any effect on scheduling until the task is
4042 * SCHED_FIFO/SCHED_RR:
4044 if (task_has_rt_policy(p
)) {
4045 p
->static_prio
= NICE_TO_PRIO(nice
);
4048 on_rq
= p
->se
.on_rq
;
4050 dequeue_task(rq
, p
, 0);
4052 p
->static_prio
= NICE_TO_PRIO(nice
);
4055 p
->prio
= effective_prio(p
);
4056 delta
= p
->prio
- old_prio
;
4059 enqueue_task(rq
, p
, 0);
4061 * If the task increased its priority or is running and
4062 * lowered its priority, then reschedule its CPU:
4064 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4065 resched_task(rq
->curr
);
4068 task_rq_unlock(rq
, &flags
);
4070 EXPORT_SYMBOL(set_user_nice
);
4073 * can_nice - check if a task can reduce its nice value
4077 int can_nice(const struct task_struct
*p
, const int nice
)
4079 /* convert nice value [19,-20] to rlimit style value [1,40] */
4080 int nice_rlim
= 20 - nice
;
4082 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4083 capable(CAP_SYS_NICE
));
4086 #ifdef __ARCH_WANT_SYS_NICE
4089 * sys_nice - change the priority of the current process.
4090 * @increment: priority increment
4092 * sys_setpriority is a more generic, but much slower function that
4093 * does similar things.
4095 asmlinkage
long sys_nice(int increment
)
4100 * Setpriority might change our priority at the same moment.
4101 * We don't have to worry. Conceptually one call occurs first
4102 * and we have a single winner.
4104 if (increment
< -40)
4109 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4115 if (increment
< 0 && !can_nice(current
, nice
))
4118 retval
= security_task_setnice(current
, nice
);
4122 set_user_nice(current
, nice
);
4129 * task_prio - return the priority value of a given task.
4130 * @p: the task in question.
4132 * This is the priority value as seen by users in /proc.
4133 * RT tasks are offset by -200. Normal tasks are centered
4134 * around 0, value goes from -16 to +15.
4136 int task_prio(const struct task_struct
*p
)
4138 return p
->prio
- MAX_RT_PRIO
;
4142 * task_nice - return the nice value of a given task.
4143 * @p: the task in question.
4145 int task_nice(const struct task_struct
*p
)
4147 return TASK_NICE(p
);
4149 EXPORT_SYMBOL_GPL(task_nice
);
4152 * idle_cpu - is a given cpu idle currently?
4153 * @cpu: the processor in question.
4155 int idle_cpu(int cpu
)
4157 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4161 * idle_task - return the idle task for a given cpu.
4162 * @cpu: the processor in question.
4164 struct task_struct
*idle_task(int cpu
)
4166 return cpu_rq(cpu
)->idle
;
4170 * find_process_by_pid - find a process with a matching PID value.
4171 * @pid: the pid in question.
4173 static struct task_struct
*find_process_by_pid(pid_t pid
)
4175 return pid
? find_task_by_vpid(pid
) : current
;
4178 /* Actually do priority change: must hold rq lock. */
4180 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4182 BUG_ON(p
->se
.on_rq
);
4185 switch (p
->policy
) {
4189 p
->sched_class
= &fair_sched_class
;
4193 p
->sched_class
= &rt_sched_class
;
4197 p
->rt_priority
= prio
;
4198 p
->normal_prio
= normal_prio(p
);
4199 /* we are holding p->pi_lock already */
4200 p
->prio
= rt_mutex_getprio(p
);
4205 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4206 * @p: the task in question.
4207 * @policy: new policy.
4208 * @param: structure containing the new RT priority.
4210 * NOTE that the task may be already dead.
4212 int sched_setscheduler(struct task_struct
*p
, int policy
,
4213 struct sched_param
*param
)
4215 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4216 unsigned long flags
;
4219 /* may grab non-irq protected spin_locks */
4220 BUG_ON(in_interrupt());
4222 /* double check policy once rq lock held */
4224 policy
= oldpolicy
= p
->policy
;
4225 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4226 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4227 policy
!= SCHED_IDLE
)
4230 * Valid priorities for SCHED_FIFO and SCHED_RR are
4231 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4232 * SCHED_BATCH and SCHED_IDLE is 0.
4234 if (param
->sched_priority
< 0 ||
4235 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4236 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4238 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4242 * Allow unprivileged RT tasks to decrease priority:
4244 if (!capable(CAP_SYS_NICE
)) {
4245 if (rt_policy(policy
)) {
4246 unsigned long rlim_rtprio
;
4248 if (!lock_task_sighand(p
, &flags
))
4250 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4251 unlock_task_sighand(p
, &flags
);
4253 /* can't set/change the rt policy */
4254 if (policy
!= p
->policy
&& !rlim_rtprio
)
4257 /* can't increase priority */
4258 if (param
->sched_priority
> p
->rt_priority
&&
4259 param
->sched_priority
> rlim_rtprio
)
4263 * Like positive nice levels, dont allow tasks to
4264 * move out of SCHED_IDLE either:
4266 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4269 /* can't change other user's priorities */
4270 if ((current
->euid
!= p
->euid
) &&
4271 (current
->euid
!= p
->uid
))
4275 retval
= security_task_setscheduler(p
, policy
, param
);
4279 * make sure no PI-waiters arrive (or leave) while we are
4280 * changing the priority of the task:
4282 spin_lock_irqsave(&p
->pi_lock
, flags
);
4284 * To be able to change p->policy safely, the apropriate
4285 * runqueue lock must be held.
4287 rq
= __task_rq_lock(p
);
4288 /* recheck policy now with rq lock held */
4289 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4290 policy
= oldpolicy
= -1;
4291 __task_rq_unlock(rq
);
4292 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4295 update_rq_clock(rq
);
4296 on_rq
= p
->se
.on_rq
;
4297 running
= task_current(rq
, p
);
4299 deactivate_task(rq
, p
, 0);
4301 p
->sched_class
->put_prev_task(rq
, p
);
4305 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4309 p
->sched_class
->set_curr_task(rq
);
4310 activate_task(rq
, p
, 0);
4312 * Reschedule if we are currently running on this runqueue and
4313 * our priority decreased, or if we are not currently running on
4314 * this runqueue and our priority is higher than the current's
4317 if (p
->prio
> oldprio
)
4318 resched_task(rq
->curr
);
4320 check_preempt_curr(rq
, p
);
4323 __task_rq_unlock(rq
);
4324 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4326 rt_mutex_adjust_pi(p
);
4330 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4333 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4335 struct sched_param lparam
;
4336 struct task_struct
*p
;
4339 if (!param
|| pid
< 0)
4341 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4346 p
= find_process_by_pid(pid
);
4348 retval
= sched_setscheduler(p
, policy
, &lparam
);
4355 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4356 * @pid: the pid in question.
4357 * @policy: new policy.
4358 * @param: structure containing the new RT priority.
4361 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4363 /* negative values for policy are not valid */
4367 return do_sched_setscheduler(pid
, policy
, param
);
4371 * sys_sched_setparam - set/change the RT priority of a thread
4372 * @pid: the pid in question.
4373 * @param: structure containing the new RT priority.
4375 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4377 return do_sched_setscheduler(pid
, -1, param
);
4381 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4382 * @pid: the pid in question.
4384 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4386 struct task_struct
*p
;
4393 read_lock(&tasklist_lock
);
4394 p
= find_process_by_pid(pid
);
4396 retval
= security_task_getscheduler(p
);
4400 read_unlock(&tasklist_lock
);
4405 * sys_sched_getscheduler - get the RT priority of a thread
4406 * @pid: the pid in question.
4407 * @param: structure containing the RT priority.
4409 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4411 struct sched_param lp
;
4412 struct task_struct
*p
;
4415 if (!param
|| pid
< 0)
4418 read_lock(&tasklist_lock
);
4419 p
= find_process_by_pid(pid
);
4424 retval
= security_task_getscheduler(p
);
4428 lp
.sched_priority
= p
->rt_priority
;
4429 read_unlock(&tasklist_lock
);
4432 * This one might sleep, we cannot do it with a spinlock held ...
4434 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4439 read_unlock(&tasklist_lock
);
4443 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4445 cpumask_t cpus_allowed
;
4446 struct task_struct
*p
;
4450 read_lock(&tasklist_lock
);
4452 p
= find_process_by_pid(pid
);
4454 read_unlock(&tasklist_lock
);
4460 * It is not safe to call set_cpus_allowed with the
4461 * tasklist_lock held. We will bump the task_struct's
4462 * usage count and then drop tasklist_lock.
4465 read_unlock(&tasklist_lock
);
4468 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4469 !capable(CAP_SYS_NICE
))
4472 retval
= security_task_setscheduler(p
, 0, NULL
);
4476 cpus_allowed
= cpuset_cpus_allowed(p
);
4477 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4479 retval
= set_cpus_allowed(p
, new_mask
);
4482 cpus_allowed
= cpuset_cpus_allowed(p
);
4483 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4485 * We must have raced with a concurrent cpuset
4486 * update. Just reset the cpus_allowed to the
4487 * cpuset's cpus_allowed
4489 new_mask
= cpus_allowed
;
4499 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4500 cpumask_t
*new_mask
)
4502 if (len
< sizeof(cpumask_t
)) {
4503 memset(new_mask
, 0, sizeof(cpumask_t
));
4504 } else if (len
> sizeof(cpumask_t
)) {
4505 len
= sizeof(cpumask_t
);
4507 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4511 * sys_sched_setaffinity - set the cpu affinity of a process
4512 * @pid: pid of the process
4513 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4514 * @user_mask_ptr: user-space pointer to the new cpu mask
4516 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4517 unsigned long __user
*user_mask_ptr
)
4522 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4526 return sched_setaffinity(pid
, new_mask
);
4530 * Represents all cpu's present in the system
4531 * In systems capable of hotplug, this map could dynamically grow
4532 * as new cpu's are detected in the system via any platform specific
4533 * method, such as ACPI for e.g.
4536 cpumask_t cpu_present_map __read_mostly
;
4537 EXPORT_SYMBOL(cpu_present_map
);
4540 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4541 EXPORT_SYMBOL(cpu_online_map
);
4543 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4544 EXPORT_SYMBOL(cpu_possible_map
);
4547 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4549 struct task_struct
*p
;
4553 read_lock(&tasklist_lock
);
4556 p
= find_process_by_pid(pid
);
4560 retval
= security_task_getscheduler(p
);
4564 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4567 read_unlock(&tasklist_lock
);
4574 * sys_sched_getaffinity - get the cpu affinity of a process
4575 * @pid: pid of the process
4576 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4577 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4579 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4580 unsigned long __user
*user_mask_ptr
)
4585 if (len
< sizeof(cpumask_t
))
4588 ret
= sched_getaffinity(pid
, &mask
);
4592 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4595 return sizeof(cpumask_t
);
4599 * sys_sched_yield - yield the current processor to other threads.
4601 * This function yields the current CPU to other tasks. If there are no
4602 * other threads running on this CPU then this function will return.
4604 asmlinkage
long sys_sched_yield(void)
4606 struct rq
*rq
= this_rq_lock();
4608 schedstat_inc(rq
, yld_count
);
4609 current
->sched_class
->yield_task(rq
);
4612 * Since we are going to call schedule() anyway, there's
4613 * no need to preempt or enable interrupts:
4615 __release(rq
->lock
);
4616 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4617 _raw_spin_unlock(&rq
->lock
);
4618 preempt_enable_no_resched();
4625 static void __cond_resched(void)
4627 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4628 __might_sleep(__FILE__
, __LINE__
);
4631 * The BKS might be reacquired before we have dropped
4632 * PREEMPT_ACTIVE, which could trigger a second
4633 * cond_resched() call.
4636 add_preempt_count(PREEMPT_ACTIVE
);
4638 sub_preempt_count(PREEMPT_ACTIVE
);
4639 } while (need_resched());
4642 int __sched
cond_resched(void)
4644 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4645 system_state
== SYSTEM_RUNNING
) {
4651 EXPORT_SYMBOL(cond_resched
);
4654 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4655 * call schedule, and on return reacquire the lock.
4657 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4658 * operations here to prevent schedule() from being called twice (once via
4659 * spin_unlock(), once by hand).
4661 int cond_resched_lock(spinlock_t
*lock
)
4665 if (need_lockbreak(lock
)) {
4671 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4672 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4673 _raw_spin_unlock(lock
);
4674 preempt_enable_no_resched();
4681 EXPORT_SYMBOL(cond_resched_lock
);
4683 int __sched
cond_resched_softirq(void)
4685 BUG_ON(!in_softirq());
4687 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4695 EXPORT_SYMBOL(cond_resched_softirq
);
4698 * yield - yield the current processor to other threads.
4700 * This is a shortcut for kernel-space yielding - it marks the
4701 * thread runnable and calls sys_sched_yield().
4703 void __sched
yield(void)
4705 set_current_state(TASK_RUNNING
);
4708 EXPORT_SYMBOL(yield
);
4711 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4712 * that process accounting knows that this is a task in IO wait state.
4714 * But don't do that if it is a deliberate, throttling IO wait (this task
4715 * has set its backing_dev_info: the queue against which it should throttle)
4717 void __sched
io_schedule(void)
4719 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4721 delayacct_blkio_start();
4722 atomic_inc(&rq
->nr_iowait
);
4724 atomic_dec(&rq
->nr_iowait
);
4725 delayacct_blkio_end();
4727 EXPORT_SYMBOL(io_schedule
);
4729 long __sched
io_schedule_timeout(long timeout
)
4731 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4734 delayacct_blkio_start();
4735 atomic_inc(&rq
->nr_iowait
);
4736 ret
= schedule_timeout(timeout
);
4737 atomic_dec(&rq
->nr_iowait
);
4738 delayacct_blkio_end();
4743 * sys_sched_get_priority_max - return maximum RT priority.
4744 * @policy: scheduling class.
4746 * this syscall returns the maximum rt_priority that can be used
4747 * by a given scheduling class.
4749 asmlinkage
long sys_sched_get_priority_max(int policy
)
4756 ret
= MAX_USER_RT_PRIO
-1;
4768 * sys_sched_get_priority_min - return minimum RT priority.
4769 * @policy: scheduling class.
4771 * this syscall returns the minimum rt_priority that can be used
4772 * by a given scheduling class.
4774 asmlinkage
long sys_sched_get_priority_min(int policy
)
4792 * sys_sched_rr_get_interval - return the default timeslice of a process.
4793 * @pid: pid of the process.
4794 * @interval: userspace pointer to the timeslice value.
4796 * this syscall writes the default timeslice value of a given process
4797 * into the user-space timespec buffer. A value of '0' means infinity.
4800 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4802 struct task_struct
*p
;
4803 unsigned int time_slice
;
4811 read_lock(&tasklist_lock
);
4812 p
= find_process_by_pid(pid
);
4816 retval
= security_task_getscheduler(p
);
4821 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4822 * tasks that are on an otherwise idle runqueue:
4825 if (p
->policy
== SCHED_RR
) {
4826 time_slice
= DEF_TIMESLICE
;
4828 struct sched_entity
*se
= &p
->se
;
4829 unsigned long flags
;
4832 rq
= task_rq_lock(p
, &flags
);
4833 if (rq
->cfs
.load
.weight
)
4834 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
4835 task_rq_unlock(rq
, &flags
);
4837 read_unlock(&tasklist_lock
);
4838 jiffies_to_timespec(time_slice
, &t
);
4839 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4843 read_unlock(&tasklist_lock
);
4847 static const char stat_nam
[] = "RSDTtZX";
4849 void sched_show_task(struct task_struct
*p
)
4851 unsigned long free
= 0;
4854 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4855 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
4856 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4857 #if BITS_PER_LONG == 32
4858 if (state
== TASK_RUNNING
)
4859 printk(KERN_CONT
" running ");
4861 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4863 if (state
== TASK_RUNNING
)
4864 printk(KERN_CONT
" running task ");
4866 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4868 #ifdef CONFIG_DEBUG_STACK_USAGE
4870 unsigned long *n
= end_of_stack(p
);
4873 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4876 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
4877 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
4879 if (state
!= TASK_RUNNING
)
4880 show_stack(p
, NULL
);
4883 void show_state_filter(unsigned long state_filter
)
4885 struct task_struct
*g
, *p
;
4887 #if BITS_PER_LONG == 32
4889 " task PC stack pid father\n");
4892 " task PC stack pid father\n");
4894 read_lock(&tasklist_lock
);
4895 do_each_thread(g
, p
) {
4897 * reset the NMI-timeout, listing all files on a slow
4898 * console might take alot of time:
4900 touch_nmi_watchdog();
4901 if (!state_filter
|| (p
->state
& state_filter
))
4903 } while_each_thread(g
, p
);
4905 touch_all_softlockup_watchdogs();
4907 #ifdef CONFIG_SCHED_DEBUG
4908 sysrq_sched_debug_show();
4910 read_unlock(&tasklist_lock
);
4912 * Only show locks if all tasks are dumped:
4914 if (state_filter
== -1)
4915 debug_show_all_locks();
4918 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4920 idle
->sched_class
= &idle_sched_class
;
4924 * init_idle - set up an idle thread for a given CPU
4925 * @idle: task in question
4926 * @cpu: cpu the idle task belongs to
4928 * NOTE: this function does not set the idle thread's NEED_RESCHED
4929 * flag, to make booting more robust.
4931 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4933 struct rq
*rq
= cpu_rq(cpu
);
4934 unsigned long flags
;
4937 idle
->se
.exec_start
= sched_clock();
4939 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4940 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4941 __set_task_cpu(idle
, cpu
);
4943 spin_lock_irqsave(&rq
->lock
, flags
);
4944 rq
->curr
= rq
->idle
= idle
;
4945 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4948 spin_unlock_irqrestore(&rq
->lock
, flags
);
4950 /* Set the preempt count _outside_ the spinlocks! */
4951 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4952 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4954 task_thread_info(idle
)->preempt_count
= 0;
4957 * The idle tasks have their own, simple scheduling class:
4959 idle
->sched_class
= &idle_sched_class
;
4963 * In a system that switches off the HZ timer nohz_cpu_mask
4964 * indicates which cpus entered this state. This is used
4965 * in the rcu update to wait only for active cpus. For system
4966 * which do not switch off the HZ timer nohz_cpu_mask should
4967 * always be CPU_MASK_NONE.
4969 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4972 * Increase the granularity value when there are more CPUs,
4973 * because with more CPUs the 'effective latency' as visible
4974 * to users decreases. But the relationship is not linear,
4975 * so pick a second-best guess by going with the log2 of the
4978 * This idea comes from the SD scheduler of Con Kolivas:
4980 static inline void sched_init_granularity(void)
4982 unsigned int factor
= 1 + ilog2(num_online_cpus());
4983 const unsigned long limit
= 200000000;
4985 sysctl_sched_min_granularity
*= factor
;
4986 if (sysctl_sched_min_granularity
> limit
)
4987 sysctl_sched_min_granularity
= limit
;
4989 sysctl_sched_latency
*= factor
;
4990 if (sysctl_sched_latency
> limit
)
4991 sysctl_sched_latency
= limit
;
4993 sysctl_sched_wakeup_granularity
*= factor
;
4994 sysctl_sched_batch_wakeup_granularity
*= factor
;
4999 * This is how migration works:
5001 * 1) we queue a struct migration_req structure in the source CPU's
5002 * runqueue and wake up that CPU's migration thread.
5003 * 2) we down() the locked semaphore => thread blocks.
5004 * 3) migration thread wakes up (implicitly it forces the migrated
5005 * thread off the CPU)
5006 * 4) it gets the migration request and checks whether the migrated
5007 * task is still in the wrong runqueue.
5008 * 5) if it's in the wrong runqueue then the migration thread removes
5009 * it and puts it into the right queue.
5010 * 6) migration thread up()s the semaphore.
5011 * 7) we wake up and the migration is done.
5015 * Change a given task's CPU affinity. Migrate the thread to a
5016 * proper CPU and schedule it away if the CPU it's executing on
5017 * is removed from the allowed bitmask.
5019 * NOTE: the caller must have a valid reference to the task, the
5020 * task must not exit() & deallocate itself prematurely. The
5021 * call is not atomic; no spinlocks may be held.
5023 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5025 struct migration_req req
;
5026 unsigned long flags
;
5030 rq
= task_rq_lock(p
, &flags
);
5031 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5036 if (p
->sched_class
->set_cpus_allowed
)
5037 p
->sched_class
->set_cpus_allowed(p
, &new_mask
);
5039 p
->cpus_allowed
= new_mask
;
5040 p
->nr_cpus_allowed
= cpus_weight(new_mask
);
5043 /* Can the task run on the task's current CPU? If so, we're done */
5044 if (cpu_isset(task_cpu(p
), new_mask
))
5047 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5048 /* Need help from migration thread: drop lock and wait. */
5049 task_rq_unlock(rq
, &flags
);
5050 wake_up_process(rq
->migration_thread
);
5051 wait_for_completion(&req
.done
);
5052 tlb_migrate_finish(p
->mm
);
5056 task_rq_unlock(rq
, &flags
);
5060 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5063 * Move (not current) task off this cpu, onto dest cpu. We're doing
5064 * this because either it can't run here any more (set_cpus_allowed()
5065 * away from this CPU, or CPU going down), or because we're
5066 * attempting to rebalance this task on exec (sched_exec).
5068 * So we race with normal scheduler movements, but that's OK, as long
5069 * as the task is no longer on this CPU.
5071 * Returns non-zero if task was successfully migrated.
5073 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5075 struct rq
*rq_dest
, *rq_src
;
5078 if (unlikely(cpu_is_offline(dest_cpu
)))
5081 rq_src
= cpu_rq(src_cpu
);
5082 rq_dest
= cpu_rq(dest_cpu
);
5084 double_rq_lock(rq_src
, rq_dest
);
5085 /* Already moved. */
5086 if (task_cpu(p
) != src_cpu
)
5088 /* Affinity changed (again). */
5089 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5092 on_rq
= p
->se
.on_rq
;
5094 deactivate_task(rq_src
, p
, 0);
5096 set_task_cpu(p
, dest_cpu
);
5098 activate_task(rq_dest
, p
, 0);
5099 check_preempt_curr(rq_dest
, p
);
5103 double_rq_unlock(rq_src
, rq_dest
);
5108 * migration_thread - this is a highprio system thread that performs
5109 * thread migration by bumping thread off CPU then 'pushing' onto
5112 static int migration_thread(void *data
)
5114 int cpu
= (long)data
;
5118 BUG_ON(rq
->migration_thread
!= current
);
5120 set_current_state(TASK_INTERRUPTIBLE
);
5121 while (!kthread_should_stop()) {
5122 struct migration_req
*req
;
5123 struct list_head
*head
;
5125 spin_lock_irq(&rq
->lock
);
5127 if (cpu_is_offline(cpu
)) {
5128 spin_unlock_irq(&rq
->lock
);
5132 if (rq
->active_balance
) {
5133 active_load_balance(rq
, cpu
);
5134 rq
->active_balance
= 0;
5137 head
= &rq
->migration_queue
;
5139 if (list_empty(head
)) {
5140 spin_unlock_irq(&rq
->lock
);
5142 set_current_state(TASK_INTERRUPTIBLE
);
5145 req
= list_entry(head
->next
, struct migration_req
, list
);
5146 list_del_init(head
->next
);
5148 spin_unlock(&rq
->lock
);
5149 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5152 complete(&req
->done
);
5154 __set_current_state(TASK_RUNNING
);
5158 /* Wait for kthread_stop */
5159 set_current_state(TASK_INTERRUPTIBLE
);
5160 while (!kthread_should_stop()) {
5162 set_current_state(TASK_INTERRUPTIBLE
);
5164 __set_current_state(TASK_RUNNING
);
5168 #ifdef CONFIG_HOTPLUG_CPU
5170 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5174 local_irq_disable();
5175 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5181 * Figure out where task on dead CPU should go, use force if necessary.
5182 * NOTE: interrupts should be disabled by the caller
5184 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5186 unsigned long flags
;
5193 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5194 cpus_and(mask
, mask
, p
->cpus_allowed
);
5195 dest_cpu
= any_online_cpu(mask
);
5197 /* On any allowed CPU? */
5198 if (dest_cpu
== NR_CPUS
)
5199 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5201 /* No more Mr. Nice Guy. */
5202 if (dest_cpu
== NR_CPUS
) {
5203 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5205 * Try to stay on the same cpuset, where the
5206 * current cpuset may be a subset of all cpus.
5207 * The cpuset_cpus_allowed_locked() variant of
5208 * cpuset_cpus_allowed() will not block. It must be
5209 * called within calls to cpuset_lock/cpuset_unlock.
5211 rq
= task_rq_lock(p
, &flags
);
5212 p
->cpus_allowed
= cpus_allowed
;
5213 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5214 task_rq_unlock(rq
, &flags
);
5217 * Don't tell them about moving exiting tasks or
5218 * kernel threads (both mm NULL), since they never
5221 if (p
->mm
&& printk_ratelimit()) {
5222 printk(KERN_INFO
"process %d (%s) no "
5223 "longer affine to cpu%d\n",
5224 task_pid_nr(p
), p
->comm
, dead_cpu
);
5227 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5231 * While a dead CPU has no uninterruptible tasks queued at this point,
5232 * it might still have a nonzero ->nr_uninterruptible counter, because
5233 * for performance reasons the counter is not stricly tracking tasks to
5234 * their home CPUs. So we just add the counter to another CPU's counter,
5235 * to keep the global sum constant after CPU-down:
5237 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5239 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5240 unsigned long flags
;
5242 local_irq_save(flags
);
5243 double_rq_lock(rq_src
, rq_dest
);
5244 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5245 rq_src
->nr_uninterruptible
= 0;
5246 double_rq_unlock(rq_src
, rq_dest
);
5247 local_irq_restore(flags
);
5250 /* Run through task list and migrate tasks from the dead cpu. */
5251 static void migrate_live_tasks(int src_cpu
)
5253 struct task_struct
*p
, *t
;
5255 read_lock(&tasklist_lock
);
5257 do_each_thread(t
, p
) {
5261 if (task_cpu(p
) == src_cpu
)
5262 move_task_off_dead_cpu(src_cpu
, p
);
5263 } while_each_thread(t
, p
);
5265 read_unlock(&tasklist_lock
);
5269 * Schedules idle task to be the next runnable task on current CPU.
5270 * It does so by boosting its priority to highest possible.
5271 * Used by CPU offline code.
5273 void sched_idle_next(void)
5275 int this_cpu
= smp_processor_id();
5276 struct rq
*rq
= cpu_rq(this_cpu
);
5277 struct task_struct
*p
= rq
->idle
;
5278 unsigned long flags
;
5280 /* cpu has to be offline */
5281 BUG_ON(cpu_online(this_cpu
));
5284 * Strictly not necessary since rest of the CPUs are stopped by now
5285 * and interrupts disabled on the current cpu.
5287 spin_lock_irqsave(&rq
->lock
, flags
);
5289 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5291 update_rq_clock(rq
);
5292 activate_task(rq
, p
, 0);
5294 spin_unlock_irqrestore(&rq
->lock
, flags
);
5298 * Ensures that the idle task is using init_mm right before its cpu goes
5301 void idle_task_exit(void)
5303 struct mm_struct
*mm
= current
->active_mm
;
5305 BUG_ON(cpu_online(smp_processor_id()));
5308 switch_mm(mm
, &init_mm
, current
);
5312 /* called under rq->lock with disabled interrupts */
5313 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5315 struct rq
*rq
= cpu_rq(dead_cpu
);
5317 /* Must be exiting, otherwise would be on tasklist. */
5318 BUG_ON(!p
->exit_state
);
5320 /* Cannot have done final schedule yet: would have vanished. */
5321 BUG_ON(p
->state
== TASK_DEAD
);
5326 * Drop lock around migration; if someone else moves it,
5327 * that's OK. No task can be added to this CPU, so iteration is
5330 spin_unlock_irq(&rq
->lock
);
5331 move_task_off_dead_cpu(dead_cpu
, p
);
5332 spin_lock_irq(&rq
->lock
);
5337 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5338 static void migrate_dead_tasks(unsigned int dead_cpu
)
5340 struct rq
*rq
= cpu_rq(dead_cpu
);
5341 struct task_struct
*next
;
5344 if (!rq
->nr_running
)
5346 update_rq_clock(rq
);
5347 next
= pick_next_task(rq
, rq
->curr
);
5350 migrate_dead(dead_cpu
, next
);
5354 #endif /* CONFIG_HOTPLUG_CPU */
5356 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5358 static struct ctl_table sd_ctl_dir
[] = {
5360 .procname
= "sched_domain",
5366 static struct ctl_table sd_ctl_root
[] = {
5368 .ctl_name
= CTL_KERN
,
5369 .procname
= "kernel",
5371 .child
= sd_ctl_dir
,
5376 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5378 struct ctl_table
*entry
=
5379 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5384 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5386 struct ctl_table
*entry
;
5389 * In the intermediate directories, both the child directory and
5390 * procname are dynamically allocated and could fail but the mode
5391 * will always be set. In the lowest directory the names are
5392 * static strings and all have proc handlers.
5394 for (entry
= *tablep
; entry
->mode
; entry
++) {
5396 sd_free_ctl_entry(&entry
->child
);
5397 if (entry
->proc_handler
== NULL
)
5398 kfree(entry
->procname
);
5406 set_table_entry(struct ctl_table
*entry
,
5407 const char *procname
, void *data
, int maxlen
,
5408 mode_t mode
, proc_handler
*proc_handler
)
5410 entry
->procname
= procname
;
5412 entry
->maxlen
= maxlen
;
5414 entry
->proc_handler
= proc_handler
;
5417 static struct ctl_table
*
5418 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5420 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5425 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5426 sizeof(long), 0644, proc_doulongvec_minmax
);
5427 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5428 sizeof(long), 0644, proc_doulongvec_minmax
);
5429 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5430 sizeof(int), 0644, proc_dointvec_minmax
);
5431 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5432 sizeof(int), 0644, proc_dointvec_minmax
);
5433 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5434 sizeof(int), 0644, proc_dointvec_minmax
);
5435 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5436 sizeof(int), 0644, proc_dointvec_minmax
);
5437 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5438 sizeof(int), 0644, proc_dointvec_minmax
);
5439 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5440 sizeof(int), 0644, proc_dointvec_minmax
);
5441 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5442 sizeof(int), 0644, proc_dointvec_minmax
);
5443 set_table_entry(&table
[9], "cache_nice_tries",
5444 &sd
->cache_nice_tries
,
5445 sizeof(int), 0644, proc_dointvec_minmax
);
5446 set_table_entry(&table
[10], "flags", &sd
->flags
,
5447 sizeof(int), 0644, proc_dointvec_minmax
);
5448 /* &table[11] is terminator */
5453 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5455 struct ctl_table
*entry
, *table
;
5456 struct sched_domain
*sd
;
5457 int domain_num
= 0, i
;
5460 for_each_domain(cpu
, sd
)
5462 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5467 for_each_domain(cpu
, sd
) {
5468 snprintf(buf
, 32, "domain%d", i
);
5469 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5471 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5478 static struct ctl_table_header
*sd_sysctl_header
;
5479 static void register_sched_domain_sysctl(void)
5481 int i
, cpu_num
= num_online_cpus();
5482 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5485 WARN_ON(sd_ctl_dir
[0].child
);
5486 sd_ctl_dir
[0].child
= entry
;
5491 for_each_online_cpu(i
) {
5492 snprintf(buf
, 32, "cpu%d", i
);
5493 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5495 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5499 WARN_ON(sd_sysctl_header
);
5500 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5503 /* may be called multiple times per register */
5504 static void unregister_sched_domain_sysctl(void)
5506 if (sd_sysctl_header
)
5507 unregister_sysctl_table(sd_sysctl_header
);
5508 sd_sysctl_header
= NULL
;
5509 if (sd_ctl_dir
[0].child
)
5510 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5513 static void register_sched_domain_sysctl(void)
5516 static void unregister_sched_domain_sysctl(void)
5522 * migration_call - callback that gets triggered when a CPU is added.
5523 * Here we can start up the necessary migration thread for the new CPU.
5525 static int __cpuinit
5526 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5528 struct task_struct
*p
;
5529 int cpu
= (long)hcpu
;
5530 unsigned long flags
;
5535 case CPU_UP_PREPARE
:
5536 case CPU_UP_PREPARE_FROZEN
:
5537 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5540 kthread_bind(p
, cpu
);
5541 /* Must be high prio: stop_machine expects to yield to it. */
5542 rq
= task_rq_lock(p
, &flags
);
5543 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5544 task_rq_unlock(rq
, &flags
);
5545 cpu_rq(cpu
)->migration_thread
= p
;
5549 case CPU_ONLINE_FROZEN
:
5550 /* Strictly unnecessary, as first user will wake it. */
5551 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5554 #ifdef CONFIG_HOTPLUG_CPU
5555 case CPU_UP_CANCELED
:
5556 case CPU_UP_CANCELED_FROZEN
:
5557 if (!cpu_rq(cpu
)->migration_thread
)
5559 /* Unbind it from offline cpu so it can run. Fall thru. */
5560 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5561 any_online_cpu(cpu_online_map
));
5562 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5563 cpu_rq(cpu
)->migration_thread
= NULL
;
5567 case CPU_DEAD_FROZEN
:
5568 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5569 migrate_live_tasks(cpu
);
5571 kthread_stop(rq
->migration_thread
);
5572 rq
->migration_thread
= NULL
;
5573 /* Idle task back to normal (off runqueue, low prio) */
5574 spin_lock_irq(&rq
->lock
);
5575 update_rq_clock(rq
);
5576 deactivate_task(rq
, rq
->idle
, 0);
5577 rq
->idle
->static_prio
= MAX_PRIO
;
5578 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5579 rq
->idle
->sched_class
= &idle_sched_class
;
5580 migrate_dead_tasks(cpu
);
5581 spin_unlock_irq(&rq
->lock
);
5583 migrate_nr_uninterruptible(rq
);
5584 BUG_ON(rq
->nr_running
!= 0);
5587 * No need to migrate the tasks: it was best-effort if
5588 * they didn't take sched_hotcpu_mutex. Just wake up
5591 spin_lock_irq(&rq
->lock
);
5592 while (!list_empty(&rq
->migration_queue
)) {
5593 struct migration_req
*req
;
5595 req
= list_entry(rq
->migration_queue
.next
,
5596 struct migration_req
, list
);
5597 list_del_init(&req
->list
);
5598 complete(&req
->done
);
5600 spin_unlock_irq(&rq
->lock
);
5607 /* Register at highest priority so that task migration (migrate_all_tasks)
5608 * happens before everything else.
5610 static struct notifier_block __cpuinitdata migration_notifier
= {
5611 .notifier_call
= migration_call
,
5615 void __init
migration_init(void)
5617 void *cpu
= (void *)(long)smp_processor_id();
5620 /* Start one for the boot CPU: */
5621 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5622 BUG_ON(err
== NOTIFY_BAD
);
5623 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5624 register_cpu_notifier(&migration_notifier
);
5630 /* Number of possible processor ids */
5631 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5632 EXPORT_SYMBOL(nr_cpu_ids
);
5634 #ifdef CONFIG_SCHED_DEBUG
5636 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
5638 struct sched_group
*group
= sd
->groups
;
5639 cpumask_t groupmask
;
5642 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5643 cpus_clear(groupmask
);
5645 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5647 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5648 printk("does not load-balance\n");
5650 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5655 printk(KERN_CONT
"span %s\n", str
);
5657 if (!cpu_isset(cpu
, sd
->span
)) {
5658 printk(KERN_ERR
"ERROR: domain->span does not contain "
5661 if (!cpu_isset(cpu
, group
->cpumask
)) {
5662 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5666 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5670 printk(KERN_ERR
"ERROR: group is NULL\n");
5674 if (!group
->__cpu_power
) {
5675 printk(KERN_CONT
"\n");
5676 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5681 if (!cpus_weight(group
->cpumask
)) {
5682 printk(KERN_CONT
"\n");
5683 printk(KERN_ERR
"ERROR: empty group\n");
5687 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5688 printk(KERN_CONT
"\n");
5689 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5693 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5695 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5696 printk(KERN_CONT
" %s", str
);
5698 group
= group
->next
;
5699 } while (group
!= sd
->groups
);
5700 printk(KERN_CONT
"\n");
5702 if (!cpus_equal(sd
->span
, groupmask
))
5703 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5705 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
5706 printk(KERN_ERR
"ERROR: parent span is not a superset "
5707 "of domain->span\n");
5711 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5716 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5720 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5723 if (sched_domain_debug_one(sd
, cpu
, level
))
5732 # define sched_domain_debug(sd, cpu) do { } while (0)
5735 static int sd_degenerate(struct sched_domain
*sd
)
5737 if (cpus_weight(sd
->span
) == 1)
5740 /* Following flags need at least 2 groups */
5741 if (sd
->flags
& (SD_LOAD_BALANCE
|
5742 SD_BALANCE_NEWIDLE
|
5746 SD_SHARE_PKG_RESOURCES
)) {
5747 if (sd
->groups
!= sd
->groups
->next
)
5751 /* Following flags don't use groups */
5752 if (sd
->flags
& (SD_WAKE_IDLE
|
5761 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5763 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5765 if (sd_degenerate(parent
))
5768 if (!cpus_equal(sd
->span
, parent
->span
))
5771 /* Does parent contain flags not in child? */
5772 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5773 if (cflags
& SD_WAKE_AFFINE
)
5774 pflags
&= ~SD_WAKE_BALANCE
;
5775 /* Flags needing groups don't count if only 1 group in parent */
5776 if (parent
->groups
== parent
->groups
->next
) {
5777 pflags
&= ~(SD_LOAD_BALANCE
|
5778 SD_BALANCE_NEWIDLE
|
5782 SD_SHARE_PKG_RESOURCES
);
5784 if (~cflags
& pflags
)
5791 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5792 * hold the hotplug lock.
5794 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5796 struct rq
*rq
= cpu_rq(cpu
);
5797 struct sched_domain
*tmp
;
5799 /* Remove the sched domains which do not contribute to scheduling. */
5800 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5801 struct sched_domain
*parent
= tmp
->parent
;
5804 if (sd_parent_degenerate(tmp
, parent
)) {
5805 tmp
->parent
= parent
->parent
;
5807 parent
->parent
->child
= tmp
;
5811 if (sd
&& sd_degenerate(sd
)) {
5817 sched_domain_debug(sd
, cpu
);
5819 rcu_assign_pointer(rq
->sd
, sd
);
5822 /* cpus with isolated domains */
5823 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5825 /* Setup the mask of cpus configured for isolated domains */
5826 static int __init
isolated_cpu_setup(char *str
)
5828 int ints
[NR_CPUS
], i
;
5830 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5831 cpus_clear(cpu_isolated_map
);
5832 for (i
= 1; i
<= ints
[0]; i
++)
5833 if (ints
[i
] < NR_CPUS
)
5834 cpu_set(ints
[i
], cpu_isolated_map
);
5838 __setup("isolcpus=", isolated_cpu_setup
);
5841 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5842 * to a function which identifies what group(along with sched group) a CPU
5843 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5844 * (due to the fact that we keep track of groups covered with a cpumask_t).
5846 * init_sched_build_groups will build a circular linked list of the groups
5847 * covered by the given span, and will set each group's ->cpumask correctly,
5848 * and ->cpu_power to 0.
5851 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5852 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5853 struct sched_group
**sg
))
5855 struct sched_group
*first
= NULL
, *last
= NULL
;
5856 cpumask_t covered
= CPU_MASK_NONE
;
5859 for_each_cpu_mask(i
, span
) {
5860 struct sched_group
*sg
;
5861 int group
= group_fn(i
, cpu_map
, &sg
);
5864 if (cpu_isset(i
, covered
))
5867 sg
->cpumask
= CPU_MASK_NONE
;
5868 sg
->__cpu_power
= 0;
5870 for_each_cpu_mask(j
, span
) {
5871 if (group_fn(j
, cpu_map
, NULL
) != group
)
5874 cpu_set(j
, covered
);
5875 cpu_set(j
, sg
->cpumask
);
5886 #define SD_NODES_PER_DOMAIN 16
5891 * find_next_best_node - find the next node to include in a sched_domain
5892 * @node: node whose sched_domain we're building
5893 * @used_nodes: nodes already in the sched_domain
5895 * Find the next node to include in a given scheduling domain. Simply
5896 * finds the closest node not already in the @used_nodes map.
5898 * Should use nodemask_t.
5900 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5902 int i
, n
, val
, min_val
, best_node
= 0;
5906 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5907 /* Start at @node */
5908 n
= (node
+ i
) % MAX_NUMNODES
;
5910 if (!nr_cpus_node(n
))
5913 /* Skip already used nodes */
5914 if (test_bit(n
, used_nodes
))
5917 /* Simple min distance search */
5918 val
= node_distance(node
, n
);
5920 if (val
< min_val
) {
5926 set_bit(best_node
, used_nodes
);
5931 * sched_domain_node_span - get a cpumask for a node's sched_domain
5932 * @node: node whose cpumask we're constructing
5933 * @size: number of nodes to include in this span
5935 * Given a node, construct a good cpumask for its sched_domain to span. It
5936 * should be one that prevents unnecessary balancing, but also spreads tasks
5939 static cpumask_t
sched_domain_node_span(int node
)
5941 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5942 cpumask_t span
, nodemask
;
5946 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5948 nodemask
= node_to_cpumask(node
);
5949 cpus_or(span
, span
, nodemask
);
5950 set_bit(node
, used_nodes
);
5952 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5953 int next_node
= find_next_best_node(node
, used_nodes
);
5955 nodemask
= node_to_cpumask(next_node
);
5956 cpus_or(span
, span
, nodemask
);
5963 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5966 * SMT sched-domains:
5968 #ifdef CONFIG_SCHED_SMT
5969 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5970 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5973 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
5976 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5982 * multi-core sched-domains:
5984 #ifdef CONFIG_SCHED_MC
5985 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5986 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5989 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5991 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
5994 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
5995 cpus_and(mask
, mask
, *cpu_map
);
5996 group
= first_cpu(mask
);
5998 *sg
= &per_cpu(sched_group_core
, group
);
6001 #elif defined(CONFIG_SCHED_MC)
6003 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6006 *sg
= &per_cpu(sched_group_core
, cpu
);
6011 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6012 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6015 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6018 #ifdef CONFIG_SCHED_MC
6019 cpumask_t mask
= cpu_coregroup_map(cpu
);
6020 cpus_and(mask
, mask
, *cpu_map
);
6021 group
= first_cpu(mask
);
6022 #elif defined(CONFIG_SCHED_SMT)
6023 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6024 cpus_and(mask
, mask
, *cpu_map
);
6025 group
= first_cpu(mask
);
6030 *sg
= &per_cpu(sched_group_phys
, group
);
6036 * The init_sched_build_groups can't handle what we want to do with node
6037 * groups, so roll our own. Now each node has its own list of groups which
6038 * gets dynamically allocated.
6040 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6041 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6043 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6044 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6046 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6047 struct sched_group
**sg
)
6049 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6052 cpus_and(nodemask
, nodemask
, *cpu_map
);
6053 group
= first_cpu(nodemask
);
6056 *sg
= &per_cpu(sched_group_allnodes
, group
);
6060 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6062 struct sched_group
*sg
= group_head
;
6068 for_each_cpu_mask(j
, sg
->cpumask
) {
6069 struct sched_domain
*sd
;
6071 sd
= &per_cpu(phys_domains
, j
);
6072 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6074 * Only add "power" once for each
6080 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6083 } while (sg
!= group_head
);
6088 /* Free memory allocated for various sched_group structures */
6089 static void free_sched_groups(const cpumask_t
*cpu_map
)
6093 for_each_cpu_mask(cpu
, *cpu_map
) {
6094 struct sched_group
**sched_group_nodes
6095 = sched_group_nodes_bycpu
[cpu
];
6097 if (!sched_group_nodes
)
6100 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6101 cpumask_t nodemask
= node_to_cpumask(i
);
6102 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6104 cpus_and(nodemask
, nodemask
, *cpu_map
);
6105 if (cpus_empty(nodemask
))
6115 if (oldsg
!= sched_group_nodes
[i
])
6118 kfree(sched_group_nodes
);
6119 sched_group_nodes_bycpu
[cpu
] = NULL
;
6123 static void free_sched_groups(const cpumask_t
*cpu_map
)
6129 * Initialize sched groups cpu_power.
6131 * cpu_power indicates the capacity of sched group, which is used while
6132 * distributing the load between different sched groups in a sched domain.
6133 * Typically cpu_power for all the groups in a sched domain will be same unless
6134 * there are asymmetries in the topology. If there are asymmetries, group
6135 * having more cpu_power will pickup more load compared to the group having
6138 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6139 * the maximum number of tasks a group can handle in the presence of other idle
6140 * or lightly loaded groups in the same sched domain.
6142 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6144 struct sched_domain
*child
;
6145 struct sched_group
*group
;
6147 WARN_ON(!sd
|| !sd
->groups
);
6149 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6154 sd
->groups
->__cpu_power
= 0;
6157 * For perf policy, if the groups in child domain share resources
6158 * (for example cores sharing some portions of the cache hierarchy
6159 * or SMT), then set this domain groups cpu_power such that each group
6160 * can handle only one task, when there are other idle groups in the
6161 * same sched domain.
6163 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6165 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6166 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6171 * add cpu_power of each child group to this groups cpu_power
6173 group
= child
->groups
;
6175 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6176 group
= group
->next
;
6177 } while (group
!= child
->groups
);
6181 * Build sched domains for a given set of cpus and attach the sched domains
6182 * to the individual cpus
6184 static int build_sched_domains(const cpumask_t
*cpu_map
)
6188 struct sched_group
**sched_group_nodes
= NULL
;
6189 int sd_allnodes
= 0;
6192 * Allocate the per-node list of sched groups
6194 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6196 if (!sched_group_nodes
) {
6197 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6200 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6204 * Set up domains for cpus specified by the cpu_map.
6206 for_each_cpu_mask(i
, *cpu_map
) {
6207 struct sched_domain
*sd
= NULL
, *p
;
6208 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6210 cpus_and(nodemask
, nodemask
, *cpu_map
);
6213 if (cpus_weight(*cpu_map
) >
6214 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6215 sd
= &per_cpu(allnodes_domains
, i
);
6216 *sd
= SD_ALLNODES_INIT
;
6217 sd
->span
= *cpu_map
;
6218 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6224 sd
= &per_cpu(node_domains
, i
);
6226 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6230 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6234 sd
= &per_cpu(phys_domains
, i
);
6236 sd
->span
= nodemask
;
6240 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6242 #ifdef CONFIG_SCHED_MC
6244 sd
= &per_cpu(core_domains
, i
);
6246 sd
->span
= cpu_coregroup_map(i
);
6247 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6250 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6253 #ifdef CONFIG_SCHED_SMT
6255 sd
= &per_cpu(cpu_domains
, i
);
6256 *sd
= SD_SIBLING_INIT
;
6257 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6258 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6261 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6265 #ifdef CONFIG_SCHED_SMT
6266 /* Set up CPU (sibling) groups */
6267 for_each_cpu_mask(i
, *cpu_map
) {
6268 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6269 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6270 if (i
!= first_cpu(this_sibling_map
))
6273 init_sched_build_groups(this_sibling_map
, cpu_map
,
6278 #ifdef CONFIG_SCHED_MC
6279 /* Set up multi-core groups */
6280 for_each_cpu_mask(i
, *cpu_map
) {
6281 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6282 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6283 if (i
!= first_cpu(this_core_map
))
6285 init_sched_build_groups(this_core_map
, cpu_map
,
6286 &cpu_to_core_group
);
6290 /* Set up physical groups */
6291 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6292 cpumask_t nodemask
= node_to_cpumask(i
);
6294 cpus_and(nodemask
, nodemask
, *cpu_map
);
6295 if (cpus_empty(nodemask
))
6298 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6302 /* Set up node groups */
6304 init_sched_build_groups(*cpu_map
, cpu_map
,
6305 &cpu_to_allnodes_group
);
6307 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6308 /* Set up node groups */
6309 struct sched_group
*sg
, *prev
;
6310 cpumask_t nodemask
= node_to_cpumask(i
);
6311 cpumask_t domainspan
;
6312 cpumask_t covered
= CPU_MASK_NONE
;
6315 cpus_and(nodemask
, nodemask
, *cpu_map
);
6316 if (cpus_empty(nodemask
)) {
6317 sched_group_nodes
[i
] = NULL
;
6321 domainspan
= sched_domain_node_span(i
);
6322 cpus_and(domainspan
, domainspan
, *cpu_map
);
6324 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6326 printk(KERN_WARNING
"Can not alloc domain group for "
6330 sched_group_nodes
[i
] = sg
;
6331 for_each_cpu_mask(j
, nodemask
) {
6332 struct sched_domain
*sd
;
6334 sd
= &per_cpu(node_domains
, j
);
6337 sg
->__cpu_power
= 0;
6338 sg
->cpumask
= nodemask
;
6340 cpus_or(covered
, covered
, nodemask
);
6343 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6344 cpumask_t tmp
, notcovered
;
6345 int n
= (i
+ j
) % MAX_NUMNODES
;
6347 cpus_complement(notcovered
, covered
);
6348 cpus_and(tmp
, notcovered
, *cpu_map
);
6349 cpus_and(tmp
, tmp
, domainspan
);
6350 if (cpus_empty(tmp
))
6353 nodemask
= node_to_cpumask(n
);
6354 cpus_and(tmp
, tmp
, nodemask
);
6355 if (cpus_empty(tmp
))
6358 sg
= kmalloc_node(sizeof(struct sched_group
),
6362 "Can not alloc domain group for node %d\n", j
);
6365 sg
->__cpu_power
= 0;
6367 sg
->next
= prev
->next
;
6368 cpus_or(covered
, covered
, tmp
);
6375 /* Calculate CPU power for physical packages and nodes */
6376 #ifdef CONFIG_SCHED_SMT
6377 for_each_cpu_mask(i
, *cpu_map
) {
6378 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6380 init_sched_groups_power(i
, sd
);
6383 #ifdef CONFIG_SCHED_MC
6384 for_each_cpu_mask(i
, *cpu_map
) {
6385 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6387 init_sched_groups_power(i
, sd
);
6391 for_each_cpu_mask(i
, *cpu_map
) {
6392 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6394 init_sched_groups_power(i
, sd
);
6398 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6399 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6402 struct sched_group
*sg
;
6404 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6405 init_numa_sched_groups_power(sg
);
6409 /* Attach the domains */
6410 for_each_cpu_mask(i
, *cpu_map
) {
6411 struct sched_domain
*sd
;
6412 #ifdef CONFIG_SCHED_SMT
6413 sd
= &per_cpu(cpu_domains
, i
);
6414 #elif defined(CONFIG_SCHED_MC)
6415 sd
= &per_cpu(core_domains
, i
);
6417 sd
= &per_cpu(phys_domains
, i
);
6419 cpu_attach_domain(sd
, i
);
6426 free_sched_groups(cpu_map
);
6431 static cpumask_t
*doms_cur
; /* current sched domains */
6432 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6435 * Special case: If a kmalloc of a doms_cur partition (array of
6436 * cpumask_t) fails, then fallback to a single sched domain,
6437 * as determined by the single cpumask_t fallback_doms.
6439 static cpumask_t fallback_doms
;
6442 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6443 * For now this just excludes isolated cpus, but could be used to
6444 * exclude other special cases in the future.
6446 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6451 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6453 doms_cur
= &fallback_doms
;
6454 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6455 err
= build_sched_domains(doms_cur
);
6456 register_sched_domain_sysctl();
6461 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6463 free_sched_groups(cpu_map
);
6467 * Detach sched domains from a group of cpus specified in cpu_map
6468 * These cpus will now be attached to the NULL domain
6470 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6474 unregister_sched_domain_sysctl();
6476 for_each_cpu_mask(i
, *cpu_map
)
6477 cpu_attach_domain(NULL
, i
);
6478 synchronize_sched();
6479 arch_destroy_sched_domains(cpu_map
);
6483 * Partition sched domains as specified by the 'ndoms_new'
6484 * cpumasks in the array doms_new[] of cpumasks. This compares
6485 * doms_new[] to the current sched domain partitioning, doms_cur[].
6486 * It destroys each deleted domain and builds each new domain.
6488 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6489 * The masks don't intersect (don't overlap.) We should setup one
6490 * sched domain for each mask. CPUs not in any of the cpumasks will
6491 * not be load balanced. If the same cpumask appears both in the
6492 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6495 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6496 * ownership of it and will kfree it when done with it. If the caller
6497 * failed the kmalloc call, then it can pass in doms_new == NULL,
6498 * and partition_sched_domains() will fallback to the single partition
6501 * Call with hotplug lock held
6503 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6509 /* always unregister in case we don't destroy any domains */
6510 unregister_sched_domain_sysctl();
6512 if (doms_new
== NULL
) {
6514 doms_new
= &fallback_doms
;
6515 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6518 /* Destroy deleted domains */
6519 for (i
= 0; i
< ndoms_cur
; i
++) {
6520 for (j
= 0; j
< ndoms_new
; j
++) {
6521 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
6524 /* no match - a current sched domain not in new doms_new[] */
6525 detach_destroy_domains(doms_cur
+ i
);
6530 /* Build new domains */
6531 for (i
= 0; i
< ndoms_new
; i
++) {
6532 for (j
= 0; j
< ndoms_cur
; j
++) {
6533 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
6536 /* no match - add a new doms_new */
6537 build_sched_domains(doms_new
+ i
);
6542 /* Remember the new sched domains */
6543 if (doms_cur
!= &fallback_doms
)
6545 doms_cur
= doms_new
;
6546 ndoms_cur
= ndoms_new
;
6548 register_sched_domain_sysctl();
6553 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6554 static int arch_reinit_sched_domains(void)
6559 detach_destroy_domains(&cpu_online_map
);
6560 err
= arch_init_sched_domains(&cpu_online_map
);
6566 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6570 if (buf
[0] != '0' && buf
[0] != '1')
6574 sched_smt_power_savings
= (buf
[0] == '1');
6576 sched_mc_power_savings
= (buf
[0] == '1');
6578 ret
= arch_reinit_sched_domains();
6580 return ret
? ret
: count
;
6583 #ifdef CONFIG_SCHED_MC
6584 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6586 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6588 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6589 const char *buf
, size_t count
)
6591 return sched_power_savings_store(buf
, count
, 0);
6593 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6594 sched_mc_power_savings_store
);
6597 #ifdef CONFIG_SCHED_SMT
6598 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6600 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6602 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6603 const char *buf
, size_t count
)
6605 return sched_power_savings_store(buf
, count
, 1);
6607 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6608 sched_smt_power_savings_store
);
6611 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6615 #ifdef CONFIG_SCHED_SMT
6617 err
= sysfs_create_file(&cls
->kset
.kobj
,
6618 &attr_sched_smt_power_savings
.attr
);
6620 #ifdef CONFIG_SCHED_MC
6621 if (!err
&& mc_capable())
6622 err
= sysfs_create_file(&cls
->kset
.kobj
,
6623 &attr_sched_mc_power_savings
.attr
);
6630 * Force a reinitialization of the sched domains hierarchy. The domains
6631 * and groups cannot be updated in place without racing with the balancing
6632 * code, so we temporarily attach all running cpus to the NULL domain
6633 * which will prevent rebalancing while the sched domains are recalculated.
6635 static int update_sched_domains(struct notifier_block
*nfb
,
6636 unsigned long action
, void *hcpu
)
6639 case CPU_UP_PREPARE
:
6640 case CPU_UP_PREPARE_FROZEN
:
6641 case CPU_DOWN_PREPARE
:
6642 case CPU_DOWN_PREPARE_FROZEN
:
6643 detach_destroy_domains(&cpu_online_map
);
6646 case CPU_UP_CANCELED
:
6647 case CPU_UP_CANCELED_FROZEN
:
6648 case CPU_DOWN_FAILED
:
6649 case CPU_DOWN_FAILED_FROZEN
:
6651 case CPU_ONLINE_FROZEN
:
6653 case CPU_DEAD_FROZEN
:
6655 * Fall through and re-initialise the domains.
6662 /* The hotplug lock is already held by cpu_up/cpu_down */
6663 arch_init_sched_domains(&cpu_online_map
);
6668 void __init
sched_init_smp(void)
6670 cpumask_t non_isolated_cpus
;
6673 arch_init_sched_domains(&cpu_online_map
);
6674 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6675 if (cpus_empty(non_isolated_cpus
))
6676 cpu_set(smp_processor_id(), non_isolated_cpus
);
6678 /* XXX: Theoretical race here - CPU may be hotplugged now */
6679 hotcpu_notifier(update_sched_domains
, 0);
6681 /* Move init over to a non-isolated CPU */
6682 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6684 sched_init_granularity();
6686 #ifdef CONFIG_FAIR_GROUP_SCHED
6687 if (nr_cpu_ids
== 1)
6690 lb_monitor_task
= kthread_create(load_balance_monitor
, NULL
,
6692 if (!IS_ERR(lb_monitor_task
)) {
6693 lb_monitor_task
->flags
|= PF_NOFREEZE
;
6694 wake_up_process(lb_monitor_task
);
6696 printk(KERN_ERR
"Could not create load balance monitor thread"
6697 "(error = %ld) \n", PTR_ERR(lb_monitor_task
));
6702 void __init
sched_init_smp(void)
6704 sched_init_granularity();
6706 #endif /* CONFIG_SMP */
6708 int in_sched_functions(unsigned long addr
)
6710 return in_lock_functions(addr
) ||
6711 (addr
>= (unsigned long)__sched_text_start
6712 && addr
< (unsigned long)__sched_text_end
);
6715 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6717 cfs_rq
->tasks_timeline
= RB_ROOT
;
6718 #ifdef CONFIG_FAIR_GROUP_SCHED
6721 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6724 void __init
sched_init(void)
6726 int highest_cpu
= 0;
6729 for_each_possible_cpu(i
) {
6730 struct rt_prio_array
*array
;
6734 spin_lock_init(&rq
->lock
);
6735 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6738 init_cfs_rq(&rq
->cfs
, rq
);
6739 #ifdef CONFIG_FAIR_GROUP_SCHED
6740 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6742 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6743 struct sched_entity
*se
=
6744 &per_cpu(init_sched_entity
, i
);
6746 init_cfs_rq_p
[i
] = cfs_rq
;
6747 init_cfs_rq(cfs_rq
, rq
);
6748 cfs_rq
->tg
= &init_task_group
;
6749 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6750 &rq
->leaf_cfs_rq_list
);
6752 init_sched_entity_p
[i
] = se
;
6753 se
->cfs_rq
= &rq
->cfs
;
6755 se
->load
.weight
= init_task_group_load
;
6756 se
->load
.inv_weight
=
6757 div64_64(1ULL<<32, init_task_group_load
);
6760 init_task_group
.shares
= init_task_group_load
;
6763 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6764 rq
->cpu_load
[j
] = 0;
6767 rq
->active_balance
= 0;
6768 rq
->next_balance
= jiffies
;
6771 rq
->migration_thread
= NULL
;
6772 INIT_LIST_HEAD(&rq
->migration_queue
);
6773 rq
->rt
.highest_prio
= MAX_RT_PRIO
;
6774 rq
->rt
.overloaded
= 0;
6776 atomic_set(&rq
->nr_iowait
, 0);
6778 array
= &rq
->rt
.active
;
6779 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6780 INIT_LIST_HEAD(array
->queue
+ j
);
6781 __clear_bit(j
, array
->bitmap
);
6784 /* delimiter for bitsearch: */
6785 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6788 set_load_weight(&init_task
);
6790 #ifdef CONFIG_PREEMPT_NOTIFIERS
6791 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6795 nr_cpu_ids
= highest_cpu
+ 1;
6796 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6799 #ifdef CONFIG_RT_MUTEXES
6800 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6804 * The boot idle thread does lazy MMU switching as well:
6806 atomic_inc(&init_mm
.mm_count
);
6807 enter_lazy_tlb(&init_mm
, current
);
6810 * Make us the idle thread. Technically, schedule() should not be
6811 * called from this thread, however somewhere below it might be,
6812 * but because we are the idle thread, we just pick up running again
6813 * when this runqueue becomes "idle".
6815 init_idle(current
, smp_processor_id());
6817 * During early bootup we pretend to be a normal task:
6819 current
->sched_class
= &fair_sched_class
;
6822 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6823 void __might_sleep(char *file
, int line
)
6826 static unsigned long prev_jiffy
; /* ratelimiting */
6828 if ((in_atomic() || irqs_disabled()) &&
6829 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6830 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6832 prev_jiffy
= jiffies
;
6833 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6834 " context at %s:%d\n", file
, line
);
6835 printk("in_atomic():%d, irqs_disabled():%d\n",
6836 in_atomic(), irqs_disabled());
6837 debug_show_held_locks(current
);
6838 if (irqs_disabled())
6839 print_irqtrace_events(current
);
6844 EXPORT_SYMBOL(__might_sleep
);
6847 #ifdef CONFIG_MAGIC_SYSRQ
6848 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6851 update_rq_clock(rq
);
6852 on_rq
= p
->se
.on_rq
;
6854 deactivate_task(rq
, p
, 0);
6855 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6857 activate_task(rq
, p
, 0);
6858 resched_task(rq
->curr
);
6862 void normalize_rt_tasks(void)
6864 struct task_struct
*g
, *p
;
6865 unsigned long flags
;
6868 read_lock_irq(&tasklist_lock
);
6869 do_each_thread(g
, p
) {
6871 * Only normalize user tasks:
6876 p
->se
.exec_start
= 0;
6877 #ifdef CONFIG_SCHEDSTATS
6878 p
->se
.wait_start
= 0;
6879 p
->se
.sleep_start
= 0;
6880 p
->se
.block_start
= 0;
6882 task_rq(p
)->clock
= 0;
6886 * Renice negative nice level userspace
6889 if (TASK_NICE(p
) < 0 && p
->mm
)
6890 set_user_nice(p
, 0);
6894 spin_lock_irqsave(&p
->pi_lock
, flags
);
6895 rq
= __task_rq_lock(p
);
6897 normalize_task(rq
, p
);
6899 __task_rq_unlock(rq
);
6900 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6901 } while_each_thread(g
, p
);
6903 read_unlock_irq(&tasklist_lock
);
6906 #endif /* CONFIG_MAGIC_SYSRQ */
6910 * These functions are only useful for the IA64 MCA handling.
6912 * They can only be called when the whole system has been
6913 * stopped - every CPU needs to be quiescent, and no scheduling
6914 * activity can take place. Using them for anything else would
6915 * be a serious bug, and as a result, they aren't even visible
6916 * under any other configuration.
6920 * curr_task - return the current task for a given cpu.
6921 * @cpu: the processor in question.
6923 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6925 struct task_struct
*curr_task(int cpu
)
6927 return cpu_curr(cpu
);
6931 * set_curr_task - set the current task for a given cpu.
6932 * @cpu: the processor in question.
6933 * @p: the task pointer to set.
6935 * Description: This function must only be used when non-maskable interrupts
6936 * are serviced on a separate stack. It allows the architecture to switch the
6937 * notion of the current task on a cpu in a non-blocking manner. This function
6938 * must be called with all CPU's synchronized, and interrupts disabled, the
6939 * and caller must save the original value of the current task (see
6940 * curr_task() above) and restore that value before reenabling interrupts and
6941 * re-starting the system.
6943 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6945 void set_curr_task(int cpu
, struct task_struct
*p
)
6952 #ifdef CONFIG_FAIR_GROUP_SCHED
6956 * distribute shares of all task groups among their schedulable entities,
6957 * to reflect load distrbution across cpus.
6959 static int rebalance_shares(struct sched_domain
*sd
, int this_cpu
)
6961 struct cfs_rq
*cfs_rq
;
6962 struct rq
*rq
= cpu_rq(this_cpu
);
6963 cpumask_t sdspan
= sd
->span
;
6966 /* Walk thr' all the task groups that we have */
6967 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
6969 unsigned long total_load
= 0, total_shares
;
6970 struct task_group
*tg
= cfs_rq
->tg
;
6972 /* Gather total task load of this group across cpus */
6973 for_each_cpu_mask(i
, sdspan
)
6974 total_load
+= tg
->cfs_rq
[i
]->load
.weight
;
6976 /* Nothing to do if this group has no load */
6981 * tg->shares represents the number of cpu shares the task group
6982 * is eligible to hold on a single cpu. On N cpus, it is
6983 * eligible to hold (N * tg->shares) number of cpu shares.
6985 total_shares
= tg
->shares
* cpus_weight(sdspan
);
6988 * redistribute total_shares across cpus as per the task load
6991 for_each_cpu_mask(i
, sdspan
) {
6992 unsigned long local_load
, local_shares
;
6994 local_load
= tg
->cfs_rq
[i
]->load
.weight
;
6995 local_shares
= (local_load
* total_shares
) / total_load
;
6997 local_shares
= MIN_GROUP_SHARES
;
6998 if (local_shares
== tg
->se
[i
]->load
.weight
)
7001 spin_lock_irq(&cpu_rq(i
)->lock
);
7002 set_se_shares(tg
->se
[i
], local_shares
);
7003 spin_unlock_irq(&cpu_rq(i
)->lock
);
7012 * How frequently should we rebalance_shares() across cpus?
7014 * The more frequently we rebalance shares, the more accurate is the fairness
7015 * of cpu bandwidth distribution between task groups. However higher frequency
7016 * also implies increased scheduling overhead.
7018 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7019 * consecutive calls to rebalance_shares() in the same sched domain.
7021 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7022 * consecutive calls to rebalance_shares() in the same sched domain.
7024 * These settings allows for the appropriate tradeoff between accuracy of
7025 * fairness and the associated overhead.
7029 /* default: 8ms, units: milliseconds */
7030 const_debug
unsigned int sysctl_sched_min_bal_int_shares
= 8;
7032 /* default: 128ms, units: milliseconds */
7033 const_debug
unsigned int sysctl_sched_max_bal_int_shares
= 128;
7035 /* kernel thread that runs rebalance_shares() periodically */
7036 static int load_balance_monitor(void *unused
)
7038 unsigned int timeout
= sysctl_sched_min_bal_int_shares
;
7039 struct sched_param schedparm
;
7043 * We don't want this thread's execution to be limited by the shares
7044 * assigned to default group (init_task_group). Hence make it run
7045 * as a SCHED_RR RT task at the lowest priority.
7047 schedparm
.sched_priority
= 1;
7048 ret
= sched_setscheduler(current
, SCHED_RR
, &schedparm
);
7050 printk(KERN_ERR
"Couldn't set SCHED_RR policy for load balance"
7051 " monitor thread (error = %d) \n", ret
);
7053 while (!kthread_should_stop()) {
7054 int i
, cpu
, balanced
= 1;
7056 /* Prevent cpus going down or coming up */
7058 /* lockout changes to doms_cur[] array */
7061 * Enter a rcu read-side critical section to safely walk rq->sd
7062 * chain on various cpus and to walk task group list
7063 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7067 for (i
= 0; i
< ndoms_cur
; i
++) {
7068 cpumask_t cpumap
= doms_cur
[i
];
7069 struct sched_domain
*sd
= NULL
, *sd_prev
= NULL
;
7071 cpu
= first_cpu(cpumap
);
7073 /* Find the highest domain at which to balance shares */
7074 for_each_domain(cpu
, sd
) {
7075 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7081 /* sd == NULL? No load balance reqd in this domain */
7085 balanced
&= rebalance_shares(sd
, cpu
);
7094 timeout
= sysctl_sched_min_bal_int_shares
;
7095 else if (timeout
< sysctl_sched_max_bal_int_shares
)
7098 msleep_interruptible(timeout
);
7103 #endif /* CONFIG_SMP */
7105 /* allocate runqueue etc for a new task group */
7106 struct task_group
*sched_create_group(void)
7108 struct task_group
*tg
;
7109 struct cfs_rq
*cfs_rq
;
7110 struct sched_entity
*se
;
7114 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7116 return ERR_PTR(-ENOMEM
);
7118 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
7121 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
7125 for_each_possible_cpu(i
) {
7128 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
7133 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
7138 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
7139 memset(se
, 0, sizeof(struct sched_entity
));
7141 tg
->cfs_rq
[i
] = cfs_rq
;
7142 init_cfs_rq(cfs_rq
, rq
);
7146 se
->cfs_rq
= &rq
->cfs
;
7148 se
->load
.weight
= NICE_0_LOAD
;
7149 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
7153 tg
->shares
= NICE_0_LOAD
;
7155 lock_task_group_list();
7156 for_each_possible_cpu(i
) {
7158 cfs_rq
= tg
->cfs_rq
[i
];
7159 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7161 unlock_task_group_list();
7166 for_each_possible_cpu(i
) {
7168 kfree(tg
->cfs_rq
[i
]);
7176 return ERR_PTR(-ENOMEM
);
7179 /* rcu callback to free various structures associated with a task group */
7180 static void free_sched_group(struct rcu_head
*rhp
)
7182 struct task_group
*tg
= container_of(rhp
, struct task_group
, rcu
);
7183 struct cfs_rq
*cfs_rq
;
7184 struct sched_entity
*se
;
7187 /* now it should be safe to free those cfs_rqs */
7188 for_each_possible_cpu(i
) {
7189 cfs_rq
= tg
->cfs_rq
[i
];
7201 /* Destroy runqueue etc associated with a task group */
7202 void sched_destroy_group(struct task_group
*tg
)
7204 struct cfs_rq
*cfs_rq
= NULL
;
7207 lock_task_group_list();
7208 for_each_possible_cpu(i
) {
7209 cfs_rq
= tg
->cfs_rq
[i
];
7210 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7212 unlock_task_group_list();
7216 /* wait for possible concurrent references to cfs_rqs complete */
7217 call_rcu(&tg
->rcu
, free_sched_group
);
7220 /* change task's runqueue when it moves between groups.
7221 * The caller of this function should have put the task in its new group
7222 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7223 * reflect its new group.
7225 void sched_move_task(struct task_struct
*tsk
)
7228 unsigned long flags
;
7231 rq
= task_rq_lock(tsk
, &flags
);
7233 if (tsk
->sched_class
!= &fair_sched_class
) {
7234 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7238 update_rq_clock(rq
);
7240 running
= task_current(rq
, tsk
);
7241 on_rq
= tsk
->se
.on_rq
;
7244 dequeue_task(rq
, tsk
, 0);
7245 if (unlikely(running
))
7246 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7249 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7252 if (unlikely(running
))
7253 tsk
->sched_class
->set_curr_task(rq
);
7254 enqueue_task(rq
, tsk
, 0);
7258 task_rq_unlock(rq
, &flags
);
7261 /* rq->lock to be locked by caller */
7262 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7264 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7265 struct rq
*rq
= cfs_rq
->rq
;
7269 shares
= MIN_GROUP_SHARES
;
7273 dequeue_entity(cfs_rq
, se
, 0);
7274 dec_cpu_load(rq
, se
->load
.weight
);
7277 se
->load
.weight
= shares
;
7278 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7281 enqueue_entity(cfs_rq
, se
, 0);
7282 inc_cpu_load(rq
, se
->load
.weight
);
7286 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7289 struct cfs_rq
*cfs_rq
;
7292 lock_task_group_list();
7293 if (tg
->shares
== shares
)
7296 if (shares
< MIN_GROUP_SHARES
)
7297 shares
= MIN_GROUP_SHARES
;
7300 * Prevent any load balance activity (rebalance_shares,
7301 * load_balance_fair) from referring to this group first,
7302 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7304 for_each_possible_cpu(i
) {
7305 cfs_rq
= tg
->cfs_rq
[i
];
7306 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7309 /* wait for any ongoing reference to this group to finish */
7310 synchronize_sched();
7313 * Now we are free to modify the group's share on each cpu
7314 * w/o tripping rebalance_share or load_balance_fair.
7316 tg
->shares
= shares
;
7317 for_each_possible_cpu(i
) {
7318 spin_lock_irq(&cpu_rq(i
)->lock
);
7319 set_se_shares(tg
->se
[i
], shares
);
7320 spin_unlock_irq(&cpu_rq(i
)->lock
);
7324 * Enable load balance activity on this group, by inserting it back on
7325 * each cpu's rq->leaf_cfs_rq_list.
7327 for_each_possible_cpu(i
) {
7329 cfs_rq
= tg
->cfs_rq
[i
];
7330 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7333 unlock_task_group_list();
7337 unsigned long sched_group_shares(struct task_group
*tg
)
7342 #endif /* CONFIG_FAIR_GROUP_SCHED */
7344 #ifdef CONFIG_FAIR_CGROUP_SCHED
7346 /* return corresponding task_group object of a cgroup */
7347 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7349 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7350 struct task_group
, css
);
7353 static struct cgroup_subsys_state
*
7354 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7356 struct task_group
*tg
;
7358 if (!cgrp
->parent
) {
7359 /* This is early initialization for the top cgroup */
7360 init_task_group
.css
.cgroup
= cgrp
;
7361 return &init_task_group
.css
;
7364 /* we support only 1-level deep hierarchical scheduler atm */
7365 if (cgrp
->parent
->parent
)
7366 return ERR_PTR(-EINVAL
);
7368 tg
= sched_create_group();
7370 return ERR_PTR(-ENOMEM
);
7372 /* Bind the cgroup to task_group object we just created */
7373 tg
->css
.cgroup
= cgrp
;
7379 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7381 struct task_group
*tg
= cgroup_tg(cgrp
);
7383 sched_destroy_group(tg
);
7387 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7388 struct task_struct
*tsk
)
7390 /* We don't support RT-tasks being in separate groups */
7391 if (tsk
->sched_class
!= &fair_sched_class
)
7398 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7399 struct cgroup
*old_cont
, struct task_struct
*tsk
)
7401 sched_move_task(tsk
);
7404 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7407 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7410 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7412 struct task_group
*tg
= cgroup_tg(cgrp
);
7414 return (u64
) tg
->shares
;
7417 static struct cftype cpu_files
[] = {
7420 .read_uint
= cpu_shares_read_uint
,
7421 .write_uint
= cpu_shares_write_uint
,
7425 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7427 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7430 struct cgroup_subsys cpu_cgroup_subsys
= {
7432 .create
= cpu_cgroup_create
,
7433 .destroy
= cpu_cgroup_destroy
,
7434 .can_attach
= cpu_cgroup_can_attach
,
7435 .attach
= cpu_cgroup_attach
,
7436 .populate
= cpu_cgroup_populate
,
7437 .subsys_id
= cpu_cgroup_subsys_id
,
7441 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7443 #ifdef CONFIG_CGROUP_CPUACCT
7446 * CPU accounting code for task groups.
7448 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7449 * (balbir@in.ibm.com).
7452 /* track cpu usage of a group of tasks */
7454 struct cgroup_subsys_state css
;
7455 /* cpuusage holds pointer to a u64-type object on every cpu */
7459 struct cgroup_subsys cpuacct_subsys
;
7461 /* return cpu accounting group corresponding to this container */
7462 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
7464 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
7465 struct cpuacct
, css
);
7468 /* return cpu accounting group to which this task belongs */
7469 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
7471 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
7472 struct cpuacct
, css
);
7475 /* create a new cpu accounting group */
7476 static struct cgroup_subsys_state
*cpuacct_create(
7477 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7479 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7482 return ERR_PTR(-ENOMEM
);
7484 ca
->cpuusage
= alloc_percpu(u64
);
7485 if (!ca
->cpuusage
) {
7487 return ERR_PTR(-ENOMEM
);
7493 /* destroy an existing cpu accounting group */
7495 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7497 struct cpuacct
*ca
= cgroup_ca(cont
);
7499 free_percpu(ca
->cpuusage
);
7503 /* return total cpu usage (in nanoseconds) of a group */
7504 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
7506 struct cpuacct
*ca
= cgroup_ca(cont
);
7507 u64 totalcpuusage
= 0;
7510 for_each_possible_cpu(i
) {
7511 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
7514 * Take rq->lock to make 64-bit addition safe on 32-bit
7517 spin_lock_irq(&cpu_rq(i
)->lock
);
7518 totalcpuusage
+= *cpuusage
;
7519 spin_unlock_irq(&cpu_rq(i
)->lock
);
7522 return totalcpuusage
;
7525 static struct cftype files
[] = {
7528 .read_uint
= cpuusage_read
,
7532 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7534 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
7538 * charge this task's execution time to its accounting group.
7540 * called with rq->lock held.
7542 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
7546 if (!cpuacct_subsys
.active
)
7551 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
7553 *cpuusage
+= cputime
;
7557 struct cgroup_subsys cpuacct_subsys
= {
7559 .create
= cpuacct_create
,
7560 .destroy
= cpuacct_destroy
,
7561 .populate
= cpuacct_populate
,
7562 .subsys_id
= cpuacct_subsys_id
,
7564 #endif /* CONFIG_CGROUP_CPUACCT */