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 * Helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
103 #define NICE_0_LOAD SCHED_LOAD_SCALE
104 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
107 * These are the 'tuning knobs' of the scheduler:
109 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
110 * Timeslices get refilled after they expire.
112 #define DEF_TIMESLICE (100 * HZ / 1000)
116 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
117 * Since cpu_power is a 'constant', we can use a reciprocal divide.
119 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
121 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
125 * Each time a sched group cpu_power is changed,
126 * we must compute its reciprocal value
128 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
130 sg
->__cpu_power
+= val
;
131 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
135 static inline int rt_policy(int policy
)
137 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
142 static inline int task_has_rt_policy(struct task_struct
*p
)
144 return rt_policy(p
->policy
);
148 * This is the priority-queue data structure of the RT scheduling class:
150 struct rt_prio_array
{
151 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
152 struct list_head queue
[MAX_RT_PRIO
];
155 #ifdef CONFIG_FAIR_GROUP_SCHED
157 #include <linux/cgroup.h>
161 /* task group related information */
163 #ifdef CONFIG_FAIR_CGROUP_SCHED
164 struct cgroup_subsys_state css
;
166 /* schedulable entities of this group on each cpu */
167 struct sched_entity
**se
;
168 /* runqueue "owned" by this group on each cpu */
169 struct cfs_rq
**cfs_rq
;
172 * shares assigned to a task group governs how much of cpu bandwidth
173 * is allocated to the group. The more shares a group has, the more is
174 * the cpu bandwidth allocated to it.
176 * For ex, lets say that there are three task groups, A, B and C which
177 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
178 * cpu bandwidth allocated by the scheduler to task groups A, B and C
181 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
182 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
183 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
185 * The weight assigned to a task group's schedulable entities on every
186 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
187 * group's shares. For ex: lets say that task group A has been
188 * assigned shares of 1000 and there are two CPUs in a system. Then,
190 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
192 * Note: It's not necessary that each of a task's group schedulable
193 * entity have the same weight on all CPUs. If the group
194 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
195 * better distribution of weight could be:
197 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
198 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
200 * rebalance_shares() is responsible for distributing the shares of a
201 * task groups like this among the group's schedulable entities across
205 unsigned long shares
;
210 /* Default task group's sched entity on each cpu */
211 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
212 /* Default task group's cfs_rq on each cpu */
213 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
215 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
216 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
218 /* task_group_mutex serializes add/remove of task groups and also changes to
219 * a task group's cpu shares.
221 static DEFINE_MUTEX(task_group_mutex
);
223 /* doms_cur_mutex serializes access to doms_cur[] array */
224 static DEFINE_MUTEX(doms_cur_mutex
);
227 /* kernel thread that runs rebalance_shares() periodically */
228 static struct task_struct
*lb_monitor_task
;
229 static int load_balance_monitor(void *unused
);
232 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
);
234 /* Default task group.
235 * Every task in system belong to this group at bootup.
237 struct task_group init_task_group
= {
238 .se
= init_sched_entity_p
,
239 .cfs_rq
= init_cfs_rq_p
,
242 #ifdef CONFIG_FAIR_USER_SCHED
243 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
245 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
248 #define MIN_GROUP_SHARES 2
250 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
252 /* return group to which a task belongs */
253 static inline struct task_group
*task_group(struct task_struct
*p
)
255 struct task_group
*tg
;
257 #ifdef CONFIG_FAIR_USER_SCHED
259 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
260 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
261 struct task_group
, css
);
263 tg
= &init_task_group
;
268 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
269 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
)
271 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
272 p
->se
.parent
= task_group(p
)->se
[cpu
];
275 static inline void lock_task_group_list(void)
277 mutex_lock(&task_group_mutex
);
280 static inline void unlock_task_group_list(void)
282 mutex_unlock(&task_group_mutex
);
285 static inline void lock_doms_cur(void)
287 mutex_lock(&doms_cur_mutex
);
290 static inline void unlock_doms_cur(void)
292 mutex_unlock(&doms_cur_mutex
);
297 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
) { }
298 static inline void lock_task_group_list(void) { }
299 static inline void unlock_task_group_list(void) { }
300 static inline void lock_doms_cur(void) { }
301 static inline void unlock_doms_cur(void) { }
303 #endif /* CONFIG_FAIR_GROUP_SCHED */
305 /* CFS-related fields in a runqueue */
307 struct load_weight load
;
308 unsigned long nr_running
;
313 struct rb_root tasks_timeline
;
314 struct rb_node
*rb_leftmost
;
315 struct rb_node
*rb_load_balance_curr
;
316 /* 'curr' points to currently running entity on this cfs_rq.
317 * It is set to NULL otherwise (i.e when none are currently running).
319 struct sched_entity
*curr
;
321 unsigned long nr_spread_over
;
323 #ifdef CONFIG_FAIR_GROUP_SCHED
324 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
327 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
328 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
329 * (like users, containers etc.)
331 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
332 * list is used during load balance.
334 struct list_head leaf_cfs_rq_list
;
335 struct task_group
*tg
; /* group that "owns" this runqueue */
339 /* Real-Time classes' related field in a runqueue: */
341 struct rt_prio_array active
;
342 int rt_load_balance_idx
;
343 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
344 unsigned long rt_nr_running
;
345 unsigned long rt_nr_migratory
;
346 /* highest queued rt task prio */
354 * We add the notion of a root-domain which will be used to define per-domain
355 * variables. Each exclusive cpuset essentially defines an island domain by
356 * fully partitioning the member cpus from any other cpuset. Whenever a new
357 * exclusive cpuset is created, we also create and attach a new root-domain
360 * By default the system creates a single root-domain with all cpus as
361 * members (mimicking the global state we have today).
369 * The "RT overload" flag: it gets set if a CPU has more than
370 * one runnable RT task.
376 static struct root_domain def_root_domain
;
381 * This is the main, per-CPU runqueue data structure.
383 * Locking rule: those places that want to lock multiple runqueues
384 * (such as the load balancing or the thread migration code), lock
385 * acquire operations must be ordered by ascending &runqueue.
392 * nr_running and cpu_load should be in the same cacheline because
393 * remote CPUs use both these fields when doing load calculation.
395 unsigned long nr_running
;
396 #define CPU_LOAD_IDX_MAX 5
397 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
398 unsigned char idle_at_tick
;
400 unsigned char in_nohz_recently
;
402 /* capture load from *all* tasks on this cpu: */
403 struct load_weight load
;
404 unsigned long nr_load_updates
;
408 #ifdef CONFIG_FAIR_GROUP_SCHED
409 /* list of leaf cfs_rq on this cpu: */
410 struct list_head leaf_cfs_rq_list
;
415 * This is part of a global counter where only the total sum
416 * over all CPUs matters. A task can increase this counter on
417 * one CPU and if it got migrated afterwards it may decrease
418 * it on another CPU. Always updated under the runqueue lock:
420 unsigned long nr_uninterruptible
;
422 struct task_struct
*curr
, *idle
;
423 unsigned long next_balance
;
424 struct mm_struct
*prev_mm
;
426 u64 clock
, prev_clock_raw
;
429 unsigned int clock_warps
, clock_overflows
;
431 unsigned int clock_deep_idle_events
;
437 struct root_domain
*rd
;
438 struct sched_domain
*sd
;
440 /* For active balancing */
443 /* cpu of this runqueue: */
446 struct task_struct
*migration_thread
;
447 struct list_head migration_queue
;
450 #ifdef CONFIG_SCHEDSTATS
452 struct sched_info rq_sched_info
;
454 /* sys_sched_yield() stats */
455 unsigned int yld_exp_empty
;
456 unsigned int yld_act_empty
;
457 unsigned int yld_both_empty
;
458 unsigned int yld_count
;
460 /* schedule() stats */
461 unsigned int sched_switch
;
462 unsigned int sched_count
;
463 unsigned int sched_goidle
;
465 /* try_to_wake_up() stats */
466 unsigned int ttwu_count
;
467 unsigned int ttwu_local
;
470 unsigned int bkl_count
;
472 struct lock_class_key rq_lock_key
;
475 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
477 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
479 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
482 static inline int cpu_of(struct rq
*rq
)
492 * Update the per-runqueue clock, as finegrained as the platform can give
493 * us, but without assuming monotonicity, etc.:
495 static void __update_rq_clock(struct rq
*rq
)
497 u64 prev_raw
= rq
->prev_clock_raw
;
498 u64 now
= sched_clock();
499 s64 delta
= now
- prev_raw
;
500 u64 clock
= rq
->clock
;
502 #ifdef CONFIG_SCHED_DEBUG
503 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
506 * Protect against sched_clock() occasionally going backwards:
508 if (unlikely(delta
< 0)) {
513 * Catch too large forward jumps too:
515 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
516 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
517 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
520 rq
->clock_overflows
++;
522 if (unlikely(delta
> rq
->clock_max_delta
))
523 rq
->clock_max_delta
= delta
;
528 rq
->prev_clock_raw
= now
;
532 static void update_rq_clock(struct rq
*rq
)
534 if (likely(smp_processor_id() == cpu_of(rq
)))
535 __update_rq_clock(rq
);
539 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
540 * See detach_destroy_domains: synchronize_sched for details.
542 * The domain tree of any CPU may only be accessed from within
543 * preempt-disabled sections.
545 #define for_each_domain(cpu, __sd) \
546 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
548 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
549 #define this_rq() (&__get_cpu_var(runqueues))
550 #define task_rq(p) cpu_rq(task_cpu(p))
551 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
554 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
556 #ifdef CONFIG_SCHED_DEBUG
557 # define const_debug __read_mostly
559 # define const_debug static const
563 * Debugging: various feature bits
566 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
567 SCHED_FEAT_WAKEUP_PREEMPT
= 2,
568 SCHED_FEAT_START_DEBIT
= 4,
569 SCHED_FEAT_TREE_AVG
= 8,
570 SCHED_FEAT_APPROX_AVG
= 16,
573 const_debug
unsigned int sysctl_sched_features
=
574 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
575 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
576 SCHED_FEAT_START_DEBIT
* 1 |
577 SCHED_FEAT_TREE_AVG
* 0 |
578 SCHED_FEAT_APPROX_AVG
* 0;
580 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
583 * Number of tasks to iterate in a single balance run.
584 * Limited because this is done with IRQs disabled.
586 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
589 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
590 * clock constructed from sched_clock():
592 unsigned long long cpu_clock(int cpu
)
594 unsigned long long now
;
598 local_irq_save(flags
);
601 * Only call sched_clock() if the scheduler has already been
602 * initialized (some code might call cpu_clock() very early):
607 local_irq_restore(flags
);
611 EXPORT_SYMBOL_GPL(cpu_clock
);
613 #ifndef prepare_arch_switch
614 # define prepare_arch_switch(next) do { } while (0)
616 #ifndef finish_arch_switch
617 # define finish_arch_switch(prev) do { } while (0)
620 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
622 return rq
->curr
== p
;
625 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
626 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
628 return task_current(rq
, p
);
631 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
635 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
637 #ifdef CONFIG_DEBUG_SPINLOCK
638 /* this is a valid case when another task releases the spinlock */
639 rq
->lock
.owner
= current
;
642 * If we are tracking spinlock dependencies then we have to
643 * fix up the runqueue lock - which gets 'carried over' from
646 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
648 spin_unlock_irq(&rq
->lock
);
651 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
652 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
657 return task_current(rq
, p
);
661 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
665 * We can optimise this out completely for !SMP, because the
666 * SMP rebalancing from interrupt is the only thing that cares
671 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
672 spin_unlock_irq(&rq
->lock
);
674 spin_unlock(&rq
->lock
);
678 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
682 * After ->oncpu is cleared, the task can be moved to a different CPU.
683 * We must ensure this doesn't happen until the switch is completely
689 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
693 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
696 * __task_rq_lock - lock the runqueue a given task resides on.
697 * Must be called interrupts disabled.
699 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
703 struct rq
*rq
= task_rq(p
);
704 spin_lock(&rq
->lock
);
705 if (likely(rq
== task_rq(p
)))
707 spin_unlock(&rq
->lock
);
712 * task_rq_lock - lock the runqueue a given task resides on and disable
713 * interrupts. Note the ordering: we can safely lookup the task_rq without
714 * explicitly disabling preemption.
716 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
722 local_irq_save(*flags
);
724 spin_lock(&rq
->lock
);
725 if (likely(rq
== task_rq(p
)))
727 spin_unlock_irqrestore(&rq
->lock
, *flags
);
731 static void __task_rq_unlock(struct rq
*rq
)
734 spin_unlock(&rq
->lock
);
737 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
740 spin_unlock_irqrestore(&rq
->lock
, *flags
);
744 * this_rq_lock - lock this runqueue and disable interrupts.
746 static struct rq
*this_rq_lock(void)
753 spin_lock(&rq
->lock
);
759 * We are going deep-idle (irqs are disabled):
761 void sched_clock_idle_sleep_event(void)
763 struct rq
*rq
= cpu_rq(smp_processor_id());
765 spin_lock(&rq
->lock
);
766 __update_rq_clock(rq
);
767 spin_unlock(&rq
->lock
);
768 rq
->clock_deep_idle_events
++;
770 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
773 * We just idled delta nanoseconds (called with irqs disabled):
775 void sched_clock_idle_wakeup_event(u64 delta_ns
)
777 struct rq
*rq
= cpu_rq(smp_processor_id());
778 u64 now
= sched_clock();
780 touch_softlockup_watchdog();
781 rq
->idle_clock
+= delta_ns
;
783 * Override the previous timestamp and ignore all
784 * sched_clock() deltas that occured while we idled,
785 * and use the PM-provided delta_ns to advance the
788 spin_lock(&rq
->lock
);
789 rq
->prev_clock_raw
= now
;
790 rq
->clock
+= delta_ns
;
791 spin_unlock(&rq
->lock
);
793 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
796 * resched_task - mark a task 'to be rescheduled now'.
798 * On UP this means the setting of the need_resched flag, on SMP it
799 * might also involve a cross-CPU call to trigger the scheduler on
804 #ifndef tsk_is_polling
805 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
808 static void resched_task(struct task_struct
*p
)
812 assert_spin_locked(&task_rq(p
)->lock
);
814 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
817 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
820 if (cpu
== smp_processor_id())
823 /* NEED_RESCHED must be visible before we test polling */
825 if (!tsk_is_polling(p
))
826 smp_send_reschedule(cpu
);
829 static void resched_cpu(int cpu
)
831 struct rq
*rq
= cpu_rq(cpu
);
834 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
836 resched_task(cpu_curr(cpu
));
837 spin_unlock_irqrestore(&rq
->lock
, flags
);
840 static inline void resched_task(struct task_struct
*p
)
842 assert_spin_locked(&task_rq(p
)->lock
);
843 set_tsk_need_resched(p
);
847 #if BITS_PER_LONG == 32
848 # define WMULT_CONST (~0UL)
850 # define WMULT_CONST (1UL << 32)
853 #define WMULT_SHIFT 32
856 * Shift right and round:
858 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
861 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
862 struct load_weight
*lw
)
866 if (unlikely(!lw
->inv_weight
))
867 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
869 tmp
= (u64
)delta_exec
* weight
;
871 * Check whether we'd overflow the 64-bit multiplication:
873 if (unlikely(tmp
> WMULT_CONST
))
874 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
877 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
879 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
882 static inline unsigned long
883 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
885 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
888 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
893 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
899 * To aid in avoiding the subversion of "niceness" due to uneven distribution
900 * of tasks with abnormal "nice" values across CPUs the contribution that
901 * each task makes to its run queue's load is weighted according to its
902 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
903 * scaled version of the new time slice allocation that they receive on time
907 #define WEIGHT_IDLEPRIO 2
908 #define WMULT_IDLEPRIO (1 << 31)
911 * Nice levels are multiplicative, with a gentle 10% change for every
912 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
913 * nice 1, it will get ~10% less CPU time than another CPU-bound task
914 * that remained on nice 0.
916 * The "10% effect" is relative and cumulative: from _any_ nice level,
917 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
918 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
919 * If a task goes up by ~10% and another task goes down by ~10% then
920 * the relative distance between them is ~25%.)
922 static const int prio_to_weight
[40] = {
923 /* -20 */ 88761, 71755, 56483, 46273, 36291,
924 /* -15 */ 29154, 23254, 18705, 14949, 11916,
925 /* -10 */ 9548, 7620, 6100, 4904, 3906,
926 /* -5 */ 3121, 2501, 1991, 1586, 1277,
927 /* 0 */ 1024, 820, 655, 526, 423,
928 /* 5 */ 335, 272, 215, 172, 137,
929 /* 10 */ 110, 87, 70, 56, 45,
930 /* 15 */ 36, 29, 23, 18, 15,
934 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
936 * In cases where the weight does not change often, we can use the
937 * precalculated inverse to speed up arithmetics by turning divisions
938 * into multiplications:
940 static const u32 prio_to_wmult
[40] = {
941 /* -20 */ 48388, 59856, 76040, 92818, 118348,
942 /* -15 */ 147320, 184698, 229616, 287308, 360437,
943 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
944 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
945 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
946 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
947 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
948 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
951 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
954 * runqueue iterator, to support SMP load-balancing between different
955 * scheduling classes, without having to expose their internal data
956 * structures to the load-balancing proper:
960 struct task_struct
*(*start
)(void *);
961 struct task_struct
*(*next
)(void *);
966 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
967 unsigned long max_load_move
, struct sched_domain
*sd
,
968 enum cpu_idle_type idle
, int *all_pinned
,
969 int *this_best_prio
, struct rq_iterator
*iterator
);
972 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
973 struct sched_domain
*sd
, enum cpu_idle_type idle
,
974 struct rq_iterator
*iterator
);
977 #ifdef CONFIG_CGROUP_CPUACCT
978 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
980 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
983 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
985 update_load_add(&rq
->load
, load
);
988 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
990 update_load_sub(&rq
->load
, load
);
994 static unsigned long source_load(int cpu
, int type
);
995 static unsigned long target_load(int cpu
, int type
);
996 static unsigned long cpu_avg_load_per_task(int cpu
);
997 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
998 #endif /* CONFIG_SMP */
1000 #include "sched_stats.h"
1001 #include "sched_idletask.c"
1002 #include "sched_fair.c"
1003 #include "sched_rt.c"
1004 #ifdef CONFIG_SCHED_DEBUG
1005 # include "sched_debug.c"
1008 #define sched_class_highest (&rt_sched_class)
1010 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
1015 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
1020 static void set_load_weight(struct task_struct
*p
)
1022 if (task_has_rt_policy(p
)) {
1023 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1024 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1029 * SCHED_IDLE tasks get minimal weight:
1031 if (p
->policy
== SCHED_IDLE
) {
1032 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1033 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1037 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1038 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1041 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1043 sched_info_queued(p
);
1044 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1048 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1050 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1055 * __normal_prio - return the priority that is based on the static prio
1057 static inline int __normal_prio(struct task_struct
*p
)
1059 return p
->static_prio
;
1063 * Calculate the expected normal priority: i.e. priority
1064 * without taking RT-inheritance into account. Might be
1065 * boosted by interactivity modifiers. Changes upon fork,
1066 * setprio syscalls, and whenever the interactivity
1067 * estimator recalculates.
1069 static inline int normal_prio(struct task_struct
*p
)
1073 if (task_has_rt_policy(p
))
1074 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1076 prio
= __normal_prio(p
);
1081 * Calculate the current priority, i.e. the priority
1082 * taken into account by the scheduler. This value might
1083 * be boosted by RT tasks, or might be boosted by
1084 * interactivity modifiers. Will be RT if the task got
1085 * RT-boosted. If not then it returns p->normal_prio.
1087 static int effective_prio(struct task_struct
*p
)
1089 p
->normal_prio
= normal_prio(p
);
1091 * If we are RT tasks or we were boosted to RT priority,
1092 * keep the priority unchanged. Otherwise, update priority
1093 * to the normal priority:
1095 if (!rt_prio(p
->prio
))
1096 return p
->normal_prio
;
1101 * activate_task - move a task to the runqueue.
1103 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1105 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1106 rq
->nr_uninterruptible
--;
1108 enqueue_task(rq
, p
, wakeup
);
1109 inc_nr_running(p
, rq
);
1113 * deactivate_task - remove a task from the runqueue.
1115 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1117 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1118 rq
->nr_uninterruptible
++;
1120 dequeue_task(rq
, p
, sleep
);
1121 dec_nr_running(p
, rq
);
1125 * task_curr - is this task currently executing on a CPU?
1126 * @p: the task in question.
1128 inline int task_curr(const struct task_struct
*p
)
1130 return cpu_curr(task_cpu(p
)) == p
;
1133 /* Used instead of source_load when we know the type == 0 */
1134 unsigned long weighted_cpuload(const int cpu
)
1136 return cpu_rq(cpu
)->load
.weight
;
1139 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1141 set_task_cfs_rq(p
, cpu
);
1144 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1145 * successfuly executed on another CPU. We must ensure that updates of
1146 * per-task data have been completed by this moment.
1149 task_thread_info(p
)->cpu
= cpu
;
1156 * Is this task likely cache-hot:
1159 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1163 if (p
->sched_class
!= &fair_sched_class
)
1166 if (sysctl_sched_migration_cost
== -1)
1168 if (sysctl_sched_migration_cost
== 0)
1171 delta
= now
- p
->se
.exec_start
;
1173 return delta
< (s64
)sysctl_sched_migration_cost
;
1177 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1179 int old_cpu
= task_cpu(p
);
1180 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1181 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1182 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1185 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1187 #ifdef CONFIG_SCHEDSTATS
1188 if (p
->se
.wait_start
)
1189 p
->se
.wait_start
-= clock_offset
;
1190 if (p
->se
.sleep_start
)
1191 p
->se
.sleep_start
-= clock_offset
;
1192 if (p
->se
.block_start
)
1193 p
->se
.block_start
-= clock_offset
;
1194 if (old_cpu
!= new_cpu
) {
1195 schedstat_inc(p
, se
.nr_migrations
);
1196 if (task_hot(p
, old_rq
->clock
, NULL
))
1197 schedstat_inc(p
, se
.nr_forced2_migrations
);
1200 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1201 new_cfsrq
->min_vruntime
;
1203 __set_task_cpu(p
, new_cpu
);
1206 struct migration_req
{
1207 struct list_head list
;
1209 struct task_struct
*task
;
1212 struct completion done
;
1216 * The task's runqueue lock must be held.
1217 * Returns true if you have to wait for migration thread.
1220 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1222 struct rq
*rq
= task_rq(p
);
1225 * If the task is not on a runqueue (and not running), then
1226 * it is sufficient to simply update the task's cpu field.
1228 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1229 set_task_cpu(p
, dest_cpu
);
1233 init_completion(&req
->done
);
1235 req
->dest_cpu
= dest_cpu
;
1236 list_add(&req
->list
, &rq
->migration_queue
);
1242 * wait_task_inactive - wait for a thread to unschedule.
1244 * The caller must ensure that the task *will* unschedule sometime soon,
1245 * else this function might spin for a *long* time. This function can't
1246 * be called with interrupts off, or it may introduce deadlock with
1247 * smp_call_function() if an IPI is sent by the same process we are
1248 * waiting to become inactive.
1250 void wait_task_inactive(struct task_struct
*p
)
1252 unsigned long flags
;
1258 * We do the initial early heuristics without holding
1259 * any task-queue locks at all. We'll only try to get
1260 * the runqueue lock when things look like they will
1266 * If the task is actively running on another CPU
1267 * still, just relax and busy-wait without holding
1270 * NOTE! Since we don't hold any locks, it's not
1271 * even sure that "rq" stays as the right runqueue!
1272 * But we don't care, since "task_running()" will
1273 * return false if the runqueue has changed and p
1274 * is actually now running somewhere else!
1276 while (task_running(rq
, p
))
1280 * Ok, time to look more closely! We need the rq
1281 * lock now, to be *sure*. If we're wrong, we'll
1282 * just go back and repeat.
1284 rq
= task_rq_lock(p
, &flags
);
1285 running
= task_running(rq
, p
);
1286 on_rq
= p
->se
.on_rq
;
1287 task_rq_unlock(rq
, &flags
);
1290 * Was it really running after all now that we
1291 * checked with the proper locks actually held?
1293 * Oops. Go back and try again..
1295 if (unlikely(running
)) {
1301 * It's not enough that it's not actively running,
1302 * it must be off the runqueue _entirely_, and not
1305 * So if it wa still runnable (but just not actively
1306 * running right now), it's preempted, and we should
1307 * yield - it could be a while.
1309 if (unlikely(on_rq
)) {
1310 schedule_timeout_uninterruptible(1);
1315 * Ahh, all good. It wasn't running, and it wasn't
1316 * runnable, which means that it will never become
1317 * running in the future either. We're all done!
1324 * kick_process - kick a running thread to enter/exit the kernel
1325 * @p: the to-be-kicked thread
1327 * Cause a process which is running on another CPU to enter
1328 * kernel-mode, without any delay. (to get signals handled.)
1330 * NOTE: this function doesnt have to take the runqueue lock,
1331 * because all it wants to ensure is that the remote task enters
1332 * the kernel. If the IPI races and the task has been migrated
1333 * to another CPU then no harm is done and the purpose has been
1336 void kick_process(struct task_struct
*p
)
1342 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1343 smp_send_reschedule(cpu
);
1348 * Return a low guess at the load of a migration-source cpu weighted
1349 * according to the scheduling class and "nice" value.
1351 * We want to under-estimate the load of migration sources, to
1352 * balance conservatively.
1354 static unsigned long source_load(int cpu
, int type
)
1356 struct rq
*rq
= cpu_rq(cpu
);
1357 unsigned long total
= weighted_cpuload(cpu
);
1362 return min(rq
->cpu_load
[type
-1], total
);
1366 * Return a high guess at the load of a migration-target cpu weighted
1367 * according to the scheduling class and "nice" value.
1369 static unsigned long target_load(int cpu
, int type
)
1371 struct rq
*rq
= cpu_rq(cpu
);
1372 unsigned long total
= weighted_cpuload(cpu
);
1377 return max(rq
->cpu_load
[type
-1], total
);
1381 * Return the average load per task on the cpu's run queue
1383 static unsigned long cpu_avg_load_per_task(int cpu
)
1385 struct rq
*rq
= cpu_rq(cpu
);
1386 unsigned long total
= weighted_cpuload(cpu
);
1387 unsigned long n
= rq
->nr_running
;
1389 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1393 * find_idlest_group finds and returns the least busy CPU group within the
1396 static struct sched_group
*
1397 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1399 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1400 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1401 int load_idx
= sd
->forkexec_idx
;
1402 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1405 unsigned long load
, avg_load
;
1409 /* Skip over this group if it has no CPUs allowed */
1410 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1413 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1415 /* Tally up the load of all CPUs in the group */
1418 for_each_cpu_mask(i
, group
->cpumask
) {
1419 /* Bias balancing toward cpus of our domain */
1421 load
= source_load(i
, load_idx
);
1423 load
= target_load(i
, load_idx
);
1428 /* Adjust by relative CPU power of the group */
1429 avg_load
= sg_div_cpu_power(group
,
1430 avg_load
* SCHED_LOAD_SCALE
);
1433 this_load
= avg_load
;
1435 } else if (avg_load
< min_load
) {
1436 min_load
= avg_load
;
1439 } while (group
= group
->next
, group
!= sd
->groups
);
1441 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1447 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1450 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1453 unsigned long load
, min_load
= ULONG_MAX
;
1457 /* Traverse only the allowed CPUs */
1458 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1460 for_each_cpu_mask(i
, tmp
) {
1461 load
= weighted_cpuload(i
);
1463 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1473 * sched_balance_self: balance the current task (running on cpu) in domains
1474 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1477 * Balance, ie. select the least loaded group.
1479 * Returns the target CPU number, or the same CPU if no balancing is needed.
1481 * preempt must be disabled.
1483 static int sched_balance_self(int cpu
, int flag
)
1485 struct task_struct
*t
= current
;
1486 struct sched_domain
*tmp
, *sd
= NULL
;
1488 for_each_domain(cpu
, tmp
) {
1490 * If power savings logic is enabled for a domain, stop there.
1492 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1494 if (tmp
->flags
& flag
)
1500 struct sched_group
*group
;
1501 int new_cpu
, weight
;
1503 if (!(sd
->flags
& flag
)) {
1509 group
= find_idlest_group(sd
, t
, cpu
);
1515 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1516 if (new_cpu
== -1 || new_cpu
== cpu
) {
1517 /* Now try balancing at a lower domain level of cpu */
1522 /* Now try balancing at a lower domain level of new_cpu */
1525 weight
= cpus_weight(span
);
1526 for_each_domain(cpu
, tmp
) {
1527 if (weight
<= cpus_weight(tmp
->span
))
1529 if (tmp
->flags
& flag
)
1532 /* while loop will break here if sd == NULL */
1538 #endif /* CONFIG_SMP */
1541 * try_to_wake_up - wake up a thread
1542 * @p: the to-be-woken-up thread
1543 * @state: the mask of task states that can be woken
1544 * @sync: do a synchronous wakeup?
1546 * Put it on the run-queue if it's not already there. The "current"
1547 * thread is always on the run-queue (except when the actual
1548 * re-schedule is in progress), and as such you're allowed to do
1549 * the simpler "current->state = TASK_RUNNING" to mark yourself
1550 * runnable without the overhead of this.
1552 * returns failure only if the task is already active.
1554 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1556 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1557 unsigned long flags
;
1564 rq
= task_rq_lock(p
, &flags
);
1565 old_state
= p
->state
;
1566 if (!(old_state
& state
))
1574 this_cpu
= smp_processor_id();
1577 if (unlikely(task_running(rq
, p
)))
1580 new_cpu
= p
->sched_class
->select_task_rq(p
, sync
);
1581 if (new_cpu
!= cpu
) {
1582 set_task_cpu(p
, new_cpu
);
1583 task_rq_unlock(rq
, &flags
);
1584 /* might preempt at this point */
1585 rq
= task_rq_lock(p
, &flags
);
1586 old_state
= p
->state
;
1587 if (!(old_state
& state
))
1592 this_cpu
= smp_processor_id();
1596 #ifdef CONFIG_SCHEDSTATS
1597 schedstat_inc(rq
, ttwu_count
);
1598 if (cpu
== this_cpu
)
1599 schedstat_inc(rq
, ttwu_local
);
1601 struct sched_domain
*sd
;
1602 for_each_domain(this_cpu
, sd
) {
1603 if (cpu_isset(cpu
, sd
->span
)) {
1604 schedstat_inc(sd
, ttwu_wake_remote
);
1614 #endif /* CONFIG_SMP */
1615 schedstat_inc(p
, se
.nr_wakeups
);
1617 schedstat_inc(p
, se
.nr_wakeups_sync
);
1618 if (orig_cpu
!= cpu
)
1619 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1620 if (cpu
== this_cpu
)
1621 schedstat_inc(p
, se
.nr_wakeups_local
);
1623 schedstat_inc(p
, se
.nr_wakeups_remote
);
1624 update_rq_clock(rq
);
1625 activate_task(rq
, p
, 1);
1626 check_preempt_curr(rq
, p
);
1630 p
->state
= TASK_RUNNING
;
1631 wakeup_balance_rt(rq
, p
);
1633 task_rq_unlock(rq
, &flags
);
1638 int fastcall
wake_up_process(struct task_struct
*p
)
1640 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1641 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1643 EXPORT_SYMBOL(wake_up_process
);
1645 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1647 return try_to_wake_up(p
, state
, 0);
1651 * Perform scheduler related setup for a newly forked process p.
1652 * p is forked by current.
1654 * __sched_fork() is basic setup used by init_idle() too:
1656 static void __sched_fork(struct task_struct
*p
)
1658 p
->se
.exec_start
= 0;
1659 p
->se
.sum_exec_runtime
= 0;
1660 p
->se
.prev_sum_exec_runtime
= 0;
1662 #ifdef CONFIG_SCHEDSTATS
1663 p
->se
.wait_start
= 0;
1664 p
->se
.sum_sleep_runtime
= 0;
1665 p
->se
.sleep_start
= 0;
1666 p
->se
.block_start
= 0;
1667 p
->se
.sleep_max
= 0;
1668 p
->se
.block_max
= 0;
1670 p
->se
.slice_max
= 0;
1674 INIT_LIST_HEAD(&p
->run_list
);
1677 #ifdef CONFIG_PREEMPT_NOTIFIERS
1678 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1682 * We mark the process as running here, but have not actually
1683 * inserted it onto the runqueue yet. This guarantees that
1684 * nobody will actually run it, and a signal or other external
1685 * event cannot wake it up and insert it on the runqueue either.
1687 p
->state
= TASK_RUNNING
;
1691 * fork()/clone()-time setup:
1693 void sched_fork(struct task_struct
*p
, int clone_flags
)
1695 int cpu
= get_cpu();
1700 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1702 set_task_cpu(p
, cpu
);
1705 * Make sure we do not leak PI boosting priority to the child:
1707 p
->prio
= current
->normal_prio
;
1708 if (!rt_prio(p
->prio
))
1709 p
->sched_class
= &fair_sched_class
;
1711 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1712 if (likely(sched_info_on()))
1713 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1715 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1718 #ifdef CONFIG_PREEMPT
1719 /* Want to start with kernel preemption disabled. */
1720 task_thread_info(p
)->preempt_count
= 1;
1726 * wake_up_new_task - wake up a newly created task for the first time.
1728 * This function will do some initial scheduler statistics housekeeping
1729 * that must be done for every newly created context, then puts the task
1730 * on the runqueue and wakes it.
1732 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1734 unsigned long flags
;
1737 rq
= task_rq_lock(p
, &flags
);
1738 BUG_ON(p
->state
!= TASK_RUNNING
);
1739 update_rq_clock(rq
);
1741 p
->prio
= effective_prio(p
);
1743 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
1744 activate_task(rq
, p
, 0);
1747 * Let the scheduling class do new task startup
1748 * management (if any):
1750 p
->sched_class
->task_new(rq
, p
);
1751 inc_nr_running(p
, rq
);
1753 check_preempt_curr(rq
, p
);
1754 wakeup_balance_rt(rq
, p
);
1755 task_rq_unlock(rq
, &flags
);
1758 #ifdef CONFIG_PREEMPT_NOTIFIERS
1761 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1762 * @notifier: notifier struct to register
1764 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1766 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1768 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1771 * preempt_notifier_unregister - no longer interested in preemption notifications
1772 * @notifier: notifier struct to unregister
1774 * This is safe to call from within a preemption notifier.
1776 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1778 hlist_del(¬ifier
->link
);
1780 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1782 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1784 struct preempt_notifier
*notifier
;
1785 struct hlist_node
*node
;
1787 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1788 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1792 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1793 struct task_struct
*next
)
1795 struct preempt_notifier
*notifier
;
1796 struct hlist_node
*node
;
1798 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1799 notifier
->ops
->sched_out(notifier
, next
);
1804 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1809 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1810 struct task_struct
*next
)
1817 * prepare_task_switch - prepare to switch tasks
1818 * @rq: the runqueue preparing to switch
1819 * @prev: the current task that is being switched out
1820 * @next: the task we are going to switch to.
1822 * This is called with the rq lock held and interrupts off. It must
1823 * be paired with a subsequent finish_task_switch after the context
1826 * prepare_task_switch sets up locking and calls architecture specific
1830 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1831 struct task_struct
*next
)
1833 fire_sched_out_preempt_notifiers(prev
, next
);
1834 prepare_lock_switch(rq
, next
);
1835 prepare_arch_switch(next
);
1839 * finish_task_switch - clean up after a task-switch
1840 * @rq: runqueue associated with task-switch
1841 * @prev: the thread we just switched away from.
1843 * finish_task_switch must be called after the context switch, paired
1844 * with a prepare_task_switch call before the context switch.
1845 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1846 * and do any other architecture-specific cleanup actions.
1848 * Note that we may have delayed dropping an mm in context_switch(). If
1849 * so, we finish that here outside of the runqueue lock. (Doing it
1850 * with the lock held can cause deadlocks; see schedule() for
1853 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1854 __releases(rq
->lock
)
1856 struct mm_struct
*mm
= rq
->prev_mm
;
1862 * A task struct has one reference for the use as "current".
1863 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1864 * schedule one last time. The schedule call will never return, and
1865 * the scheduled task must drop that reference.
1866 * The test for TASK_DEAD must occur while the runqueue locks are
1867 * still held, otherwise prev could be scheduled on another cpu, die
1868 * there before we look at prev->state, and then the reference would
1870 * Manfred Spraul <manfred@colorfullife.com>
1872 prev_state
= prev
->state
;
1873 finish_arch_switch(prev
);
1874 finish_lock_switch(rq
, prev
);
1875 schedule_tail_balance_rt(rq
);
1877 fire_sched_in_preempt_notifiers(current
);
1880 if (unlikely(prev_state
== TASK_DEAD
)) {
1882 * Remove function-return probe instances associated with this
1883 * task and put them back on the free list.
1885 kprobe_flush_task(prev
);
1886 put_task_struct(prev
);
1891 * schedule_tail - first thing a freshly forked thread must call.
1892 * @prev: the thread we just switched away from.
1894 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1895 __releases(rq
->lock
)
1897 struct rq
*rq
= this_rq();
1899 finish_task_switch(rq
, prev
);
1900 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1901 /* In this case, finish_task_switch does not reenable preemption */
1904 if (current
->set_child_tid
)
1905 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1909 * context_switch - switch to the new MM and the new
1910 * thread's register state.
1913 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1914 struct task_struct
*next
)
1916 struct mm_struct
*mm
, *oldmm
;
1918 prepare_task_switch(rq
, prev
, next
);
1920 oldmm
= prev
->active_mm
;
1922 * For paravirt, this is coupled with an exit in switch_to to
1923 * combine the page table reload and the switch backend into
1926 arch_enter_lazy_cpu_mode();
1928 if (unlikely(!mm
)) {
1929 next
->active_mm
= oldmm
;
1930 atomic_inc(&oldmm
->mm_count
);
1931 enter_lazy_tlb(oldmm
, next
);
1933 switch_mm(oldmm
, mm
, next
);
1935 if (unlikely(!prev
->mm
)) {
1936 prev
->active_mm
= NULL
;
1937 rq
->prev_mm
= oldmm
;
1940 * Since the runqueue lock will be released by the next
1941 * task (which is an invalid locking op but in the case
1942 * of the scheduler it's an obvious special-case), so we
1943 * do an early lockdep release here:
1945 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1946 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1949 /* Here we just switch the register state and the stack. */
1950 switch_to(prev
, next
, prev
);
1954 * this_rq must be evaluated again because prev may have moved
1955 * CPUs since it called schedule(), thus the 'rq' on its stack
1956 * frame will be invalid.
1958 finish_task_switch(this_rq(), prev
);
1962 * nr_running, nr_uninterruptible and nr_context_switches:
1964 * externally visible scheduler statistics: current number of runnable
1965 * threads, current number of uninterruptible-sleeping threads, total
1966 * number of context switches performed since bootup.
1968 unsigned long nr_running(void)
1970 unsigned long i
, sum
= 0;
1972 for_each_online_cpu(i
)
1973 sum
+= cpu_rq(i
)->nr_running
;
1978 unsigned long nr_uninterruptible(void)
1980 unsigned long i
, sum
= 0;
1982 for_each_possible_cpu(i
)
1983 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1986 * Since we read the counters lockless, it might be slightly
1987 * inaccurate. Do not allow it to go below zero though:
1989 if (unlikely((long)sum
< 0))
1995 unsigned long long nr_context_switches(void)
1998 unsigned long long sum
= 0;
2000 for_each_possible_cpu(i
)
2001 sum
+= cpu_rq(i
)->nr_switches
;
2006 unsigned long nr_iowait(void)
2008 unsigned long i
, sum
= 0;
2010 for_each_possible_cpu(i
)
2011 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2016 unsigned long nr_active(void)
2018 unsigned long i
, running
= 0, uninterruptible
= 0;
2020 for_each_online_cpu(i
) {
2021 running
+= cpu_rq(i
)->nr_running
;
2022 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2025 if (unlikely((long)uninterruptible
< 0))
2026 uninterruptible
= 0;
2028 return running
+ uninterruptible
;
2032 * Update rq->cpu_load[] statistics. This function is usually called every
2033 * scheduler tick (TICK_NSEC).
2035 static void update_cpu_load(struct rq
*this_rq
)
2037 unsigned long this_load
= this_rq
->load
.weight
;
2040 this_rq
->nr_load_updates
++;
2042 /* Update our load: */
2043 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2044 unsigned long old_load
, new_load
;
2046 /* scale is effectively 1 << i now, and >> i divides by scale */
2048 old_load
= this_rq
->cpu_load
[i
];
2049 new_load
= this_load
;
2051 * Round up the averaging division if load is increasing. This
2052 * prevents us from getting stuck on 9 if the load is 10, for
2055 if (new_load
> old_load
)
2056 new_load
+= scale
-1;
2057 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2064 * double_rq_lock - safely lock two runqueues
2066 * Note this does not disable interrupts like task_rq_lock,
2067 * you need to do so manually before calling.
2069 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2070 __acquires(rq1
->lock
)
2071 __acquires(rq2
->lock
)
2073 BUG_ON(!irqs_disabled());
2075 spin_lock(&rq1
->lock
);
2076 __acquire(rq2
->lock
); /* Fake it out ;) */
2079 spin_lock(&rq1
->lock
);
2080 spin_lock(&rq2
->lock
);
2082 spin_lock(&rq2
->lock
);
2083 spin_lock(&rq1
->lock
);
2086 update_rq_clock(rq1
);
2087 update_rq_clock(rq2
);
2091 * double_rq_unlock - safely unlock two runqueues
2093 * Note this does not restore interrupts like task_rq_unlock,
2094 * you need to do so manually after calling.
2096 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2097 __releases(rq1
->lock
)
2098 __releases(rq2
->lock
)
2100 spin_unlock(&rq1
->lock
);
2102 spin_unlock(&rq2
->lock
);
2104 __release(rq2
->lock
);
2108 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2110 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2111 __releases(this_rq
->lock
)
2112 __acquires(busiest
->lock
)
2113 __acquires(this_rq
->lock
)
2117 if (unlikely(!irqs_disabled())) {
2118 /* printk() doesn't work good under rq->lock */
2119 spin_unlock(&this_rq
->lock
);
2122 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2123 if (busiest
< this_rq
) {
2124 spin_unlock(&this_rq
->lock
);
2125 spin_lock(&busiest
->lock
);
2126 spin_lock(&this_rq
->lock
);
2129 spin_lock(&busiest
->lock
);
2135 * If dest_cpu is allowed for this process, migrate the task to it.
2136 * This is accomplished by forcing the cpu_allowed mask to only
2137 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2138 * the cpu_allowed mask is restored.
2140 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2142 struct migration_req req
;
2143 unsigned long flags
;
2146 rq
= task_rq_lock(p
, &flags
);
2147 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2148 || unlikely(cpu_is_offline(dest_cpu
)))
2151 /* force the process onto the specified CPU */
2152 if (migrate_task(p
, dest_cpu
, &req
)) {
2153 /* Need to wait for migration thread (might exit: take ref). */
2154 struct task_struct
*mt
= rq
->migration_thread
;
2156 get_task_struct(mt
);
2157 task_rq_unlock(rq
, &flags
);
2158 wake_up_process(mt
);
2159 put_task_struct(mt
);
2160 wait_for_completion(&req
.done
);
2165 task_rq_unlock(rq
, &flags
);
2169 * sched_exec - execve() is a valuable balancing opportunity, because at
2170 * this point the task has the smallest effective memory and cache footprint.
2172 void sched_exec(void)
2174 int new_cpu
, this_cpu
= get_cpu();
2175 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2177 if (new_cpu
!= this_cpu
)
2178 sched_migrate_task(current
, new_cpu
);
2182 * pull_task - move a task from a remote runqueue to the local runqueue.
2183 * Both runqueues must be locked.
2185 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2186 struct rq
*this_rq
, int this_cpu
)
2188 deactivate_task(src_rq
, p
, 0);
2189 set_task_cpu(p
, this_cpu
);
2190 activate_task(this_rq
, p
, 0);
2192 * Note that idle threads have a prio of MAX_PRIO, for this test
2193 * to be always true for them.
2195 check_preempt_curr(this_rq
, p
);
2199 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2202 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2203 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2207 * We do not migrate tasks that are:
2208 * 1) running (obviously), or
2209 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2210 * 3) are cache-hot on their current CPU.
2212 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2213 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2218 if (task_running(rq
, p
)) {
2219 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2224 * Aggressive migration if:
2225 * 1) task is cache cold, or
2226 * 2) too many balance attempts have failed.
2229 if (!task_hot(p
, rq
->clock
, sd
) ||
2230 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2231 #ifdef CONFIG_SCHEDSTATS
2232 if (task_hot(p
, rq
->clock
, sd
)) {
2233 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2234 schedstat_inc(p
, se
.nr_forced_migrations
);
2240 if (task_hot(p
, rq
->clock
, sd
)) {
2241 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2247 static unsigned long
2248 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2249 unsigned long max_load_move
, struct sched_domain
*sd
,
2250 enum cpu_idle_type idle
, int *all_pinned
,
2251 int *this_best_prio
, struct rq_iterator
*iterator
)
2253 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2254 struct task_struct
*p
;
2255 long rem_load_move
= max_load_move
;
2257 if (max_load_move
== 0)
2263 * Start the load-balancing iterator:
2265 p
= iterator
->start(iterator
->arg
);
2267 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2270 * To help distribute high priority tasks across CPUs we don't
2271 * skip a task if it will be the highest priority task (i.e. smallest
2272 * prio value) on its new queue regardless of its load weight
2274 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2275 SCHED_LOAD_SCALE_FUZZ
;
2276 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2277 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2278 p
= iterator
->next(iterator
->arg
);
2282 pull_task(busiest
, p
, this_rq
, this_cpu
);
2284 rem_load_move
-= p
->se
.load
.weight
;
2287 * We only want to steal up to the prescribed amount of weighted load.
2289 if (rem_load_move
> 0) {
2290 if (p
->prio
< *this_best_prio
)
2291 *this_best_prio
= p
->prio
;
2292 p
= iterator
->next(iterator
->arg
);
2297 * Right now, this is one of only two places pull_task() is called,
2298 * so we can safely collect pull_task() stats here rather than
2299 * inside pull_task().
2301 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2304 *all_pinned
= pinned
;
2306 return max_load_move
- rem_load_move
;
2310 * move_tasks tries to move up to max_load_move weighted load from busiest to
2311 * this_rq, as part of a balancing operation within domain "sd".
2312 * Returns 1 if successful and 0 otherwise.
2314 * Called with both runqueues locked.
2316 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2317 unsigned long max_load_move
,
2318 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2321 const struct sched_class
*class = sched_class_highest
;
2322 unsigned long total_load_moved
= 0;
2323 int this_best_prio
= this_rq
->curr
->prio
;
2327 class->load_balance(this_rq
, this_cpu
, busiest
,
2328 max_load_move
- total_load_moved
,
2329 sd
, idle
, all_pinned
, &this_best_prio
);
2330 class = class->next
;
2331 } while (class && max_load_move
> total_load_moved
);
2333 return total_load_moved
> 0;
2337 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2338 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2339 struct rq_iterator
*iterator
)
2341 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2345 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2346 pull_task(busiest
, p
, this_rq
, this_cpu
);
2348 * Right now, this is only the second place pull_task()
2349 * is called, so we can safely collect pull_task()
2350 * stats here rather than inside pull_task().
2352 schedstat_inc(sd
, lb_gained
[idle
]);
2356 p
= iterator
->next(iterator
->arg
);
2363 * move_one_task tries to move exactly one task from busiest to this_rq, as
2364 * part of active balancing operations within "domain".
2365 * Returns 1 if successful and 0 otherwise.
2367 * Called with both runqueues locked.
2369 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2370 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2372 const struct sched_class
*class;
2374 for (class = sched_class_highest
; class; class = class->next
)
2375 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2382 * find_busiest_group finds and returns the busiest CPU group within the
2383 * domain. It calculates and returns the amount of weighted load which
2384 * should be moved to restore balance via the imbalance parameter.
2386 static struct sched_group
*
2387 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2388 unsigned long *imbalance
, enum cpu_idle_type idle
,
2389 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2391 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2392 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2393 unsigned long max_pull
;
2394 unsigned long busiest_load_per_task
, busiest_nr_running
;
2395 unsigned long this_load_per_task
, this_nr_running
;
2396 int load_idx
, group_imb
= 0;
2397 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2398 int power_savings_balance
= 1;
2399 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2400 unsigned long min_nr_running
= ULONG_MAX
;
2401 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2404 max_load
= this_load
= total_load
= total_pwr
= 0;
2405 busiest_load_per_task
= busiest_nr_running
= 0;
2406 this_load_per_task
= this_nr_running
= 0;
2407 if (idle
== CPU_NOT_IDLE
)
2408 load_idx
= sd
->busy_idx
;
2409 else if (idle
== CPU_NEWLY_IDLE
)
2410 load_idx
= sd
->newidle_idx
;
2412 load_idx
= sd
->idle_idx
;
2415 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2418 int __group_imb
= 0;
2419 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2420 unsigned long sum_nr_running
, sum_weighted_load
;
2422 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2425 balance_cpu
= first_cpu(group
->cpumask
);
2427 /* Tally up the load of all CPUs in the group */
2428 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2430 min_cpu_load
= ~0UL;
2432 for_each_cpu_mask(i
, group
->cpumask
) {
2435 if (!cpu_isset(i
, *cpus
))
2440 if (*sd_idle
&& rq
->nr_running
)
2443 /* Bias balancing toward cpus of our domain */
2445 if (idle_cpu(i
) && !first_idle_cpu
) {
2450 load
= target_load(i
, load_idx
);
2452 load
= source_load(i
, load_idx
);
2453 if (load
> max_cpu_load
)
2454 max_cpu_load
= load
;
2455 if (min_cpu_load
> load
)
2456 min_cpu_load
= load
;
2460 sum_nr_running
+= rq
->nr_running
;
2461 sum_weighted_load
+= weighted_cpuload(i
);
2465 * First idle cpu or the first cpu(busiest) in this sched group
2466 * is eligible for doing load balancing at this and above
2467 * domains. In the newly idle case, we will allow all the cpu's
2468 * to do the newly idle load balance.
2470 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2471 balance_cpu
!= this_cpu
&& balance
) {
2476 total_load
+= avg_load
;
2477 total_pwr
+= group
->__cpu_power
;
2479 /* Adjust by relative CPU power of the group */
2480 avg_load
= sg_div_cpu_power(group
,
2481 avg_load
* SCHED_LOAD_SCALE
);
2483 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2486 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2489 this_load
= avg_load
;
2491 this_nr_running
= sum_nr_running
;
2492 this_load_per_task
= sum_weighted_load
;
2493 } else if (avg_load
> max_load
&&
2494 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2495 max_load
= avg_load
;
2497 busiest_nr_running
= sum_nr_running
;
2498 busiest_load_per_task
= sum_weighted_load
;
2499 group_imb
= __group_imb
;
2502 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2504 * Busy processors will not participate in power savings
2507 if (idle
== CPU_NOT_IDLE
||
2508 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2512 * If the local group is idle or completely loaded
2513 * no need to do power savings balance at this domain
2515 if (local_group
&& (this_nr_running
>= group_capacity
||
2517 power_savings_balance
= 0;
2520 * If a group is already running at full capacity or idle,
2521 * don't include that group in power savings calculations
2523 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2528 * Calculate the group which has the least non-idle load.
2529 * This is the group from where we need to pick up the load
2532 if ((sum_nr_running
< min_nr_running
) ||
2533 (sum_nr_running
== min_nr_running
&&
2534 first_cpu(group
->cpumask
) <
2535 first_cpu(group_min
->cpumask
))) {
2537 min_nr_running
= sum_nr_running
;
2538 min_load_per_task
= sum_weighted_load
/
2543 * Calculate the group which is almost near its
2544 * capacity but still has some space to pick up some load
2545 * from other group and save more power
2547 if (sum_nr_running
<= group_capacity
- 1) {
2548 if (sum_nr_running
> leader_nr_running
||
2549 (sum_nr_running
== leader_nr_running
&&
2550 first_cpu(group
->cpumask
) >
2551 first_cpu(group_leader
->cpumask
))) {
2552 group_leader
= group
;
2553 leader_nr_running
= sum_nr_running
;
2558 group
= group
->next
;
2559 } while (group
!= sd
->groups
);
2561 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2564 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2566 if (this_load
>= avg_load
||
2567 100*max_load
<= sd
->imbalance_pct
*this_load
)
2570 busiest_load_per_task
/= busiest_nr_running
;
2572 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2575 * We're trying to get all the cpus to the average_load, so we don't
2576 * want to push ourselves above the average load, nor do we wish to
2577 * reduce the max loaded cpu below the average load, as either of these
2578 * actions would just result in more rebalancing later, and ping-pong
2579 * tasks around. Thus we look for the minimum possible imbalance.
2580 * Negative imbalances (*we* are more loaded than anyone else) will
2581 * be counted as no imbalance for these purposes -- we can't fix that
2582 * by pulling tasks to us. Be careful of negative numbers as they'll
2583 * appear as very large values with unsigned longs.
2585 if (max_load
<= busiest_load_per_task
)
2589 * In the presence of smp nice balancing, certain scenarios can have
2590 * max load less than avg load(as we skip the groups at or below
2591 * its cpu_power, while calculating max_load..)
2593 if (max_load
< avg_load
) {
2595 goto small_imbalance
;
2598 /* Don't want to pull so many tasks that a group would go idle */
2599 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2601 /* How much load to actually move to equalise the imbalance */
2602 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2603 (avg_load
- this_load
) * this->__cpu_power
)
2607 * if *imbalance is less than the average load per runnable task
2608 * there is no gaurantee that any tasks will be moved so we'll have
2609 * a think about bumping its value to force at least one task to be
2612 if (*imbalance
< busiest_load_per_task
) {
2613 unsigned long tmp
, pwr_now
, pwr_move
;
2617 pwr_move
= pwr_now
= 0;
2619 if (this_nr_running
) {
2620 this_load_per_task
/= this_nr_running
;
2621 if (busiest_load_per_task
> this_load_per_task
)
2624 this_load_per_task
= SCHED_LOAD_SCALE
;
2626 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2627 busiest_load_per_task
* imbn
) {
2628 *imbalance
= busiest_load_per_task
;
2633 * OK, we don't have enough imbalance to justify moving tasks,
2634 * however we may be able to increase total CPU power used by
2638 pwr_now
+= busiest
->__cpu_power
*
2639 min(busiest_load_per_task
, max_load
);
2640 pwr_now
+= this->__cpu_power
*
2641 min(this_load_per_task
, this_load
);
2642 pwr_now
/= SCHED_LOAD_SCALE
;
2644 /* Amount of load we'd subtract */
2645 tmp
= sg_div_cpu_power(busiest
,
2646 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2648 pwr_move
+= busiest
->__cpu_power
*
2649 min(busiest_load_per_task
, max_load
- tmp
);
2651 /* Amount of load we'd add */
2652 if (max_load
* busiest
->__cpu_power
<
2653 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2654 tmp
= sg_div_cpu_power(this,
2655 max_load
* busiest
->__cpu_power
);
2657 tmp
= sg_div_cpu_power(this,
2658 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2659 pwr_move
+= this->__cpu_power
*
2660 min(this_load_per_task
, this_load
+ tmp
);
2661 pwr_move
/= SCHED_LOAD_SCALE
;
2663 /* Move if we gain throughput */
2664 if (pwr_move
> pwr_now
)
2665 *imbalance
= busiest_load_per_task
;
2671 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2672 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2675 if (this == group_leader
&& group_leader
!= group_min
) {
2676 *imbalance
= min_load_per_task
;
2686 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2689 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2690 unsigned long imbalance
, cpumask_t
*cpus
)
2692 struct rq
*busiest
= NULL
, *rq
;
2693 unsigned long max_load
= 0;
2696 for_each_cpu_mask(i
, group
->cpumask
) {
2699 if (!cpu_isset(i
, *cpus
))
2703 wl
= weighted_cpuload(i
);
2705 if (rq
->nr_running
== 1 && wl
> imbalance
)
2708 if (wl
> max_load
) {
2718 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2719 * so long as it is large enough.
2721 #define MAX_PINNED_INTERVAL 512
2724 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2725 * tasks if there is an imbalance.
2727 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2728 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2731 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2732 struct sched_group
*group
;
2733 unsigned long imbalance
;
2735 cpumask_t cpus
= CPU_MASK_ALL
;
2736 unsigned long flags
;
2739 * When power savings policy is enabled for the parent domain, idle
2740 * sibling can pick up load irrespective of busy siblings. In this case,
2741 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2742 * portraying it as CPU_NOT_IDLE.
2744 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2745 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2748 schedstat_inc(sd
, lb_count
[idle
]);
2751 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2758 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2762 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2764 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2768 BUG_ON(busiest
== this_rq
);
2770 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2773 if (busiest
->nr_running
> 1) {
2775 * Attempt to move tasks. If find_busiest_group has found
2776 * an imbalance but busiest->nr_running <= 1, the group is
2777 * still unbalanced. ld_moved simply stays zero, so it is
2778 * correctly treated as an imbalance.
2780 local_irq_save(flags
);
2781 double_rq_lock(this_rq
, busiest
);
2782 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2783 imbalance
, sd
, idle
, &all_pinned
);
2784 double_rq_unlock(this_rq
, busiest
);
2785 local_irq_restore(flags
);
2788 * some other cpu did the load balance for us.
2790 if (ld_moved
&& this_cpu
!= smp_processor_id())
2791 resched_cpu(this_cpu
);
2793 /* All tasks on this runqueue were pinned by CPU affinity */
2794 if (unlikely(all_pinned
)) {
2795 cpu_clear(cpu_of(busiest
), cpus
);
2796 if (!cpus_empty(cpus
))
2803 schedstat_inc(sd
, lb_failed
[idle
]);
2804 sd
->nr_balance_failed
++;
2806 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2808 spin_lock_irqsave(&busiest
->lock
, flags
);
2810 /* don't kick the migration_thread, if the curr
2811 * task on busiest cpu can't be moved to this_cpu
2813 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2814 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2816 goto out_one_pinned
;
2819 if (!busiest
->active_balance
) {
2820 busiest
->active_balance
= 1;
2821 busiest
->push_cpu
= this_cpu
;
2824 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2826 wake_up_process(busiest
->migration_thread
);
2829 * We've kicked active balancing, reset the failure
2832 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2835 sd
->nr_balance_failed
= 0;
2837 if (likely(!active_balance
)) {
2838 /* We were unbalanced, so reset the balancing interval */
2839 sd
->balance_interval
= sd
->min_interval
;
2842 * If we've begun active balancing, start to back off. This
2843 * case may not be covered by the all_pinned logic if there
2844 * is only 1 task on the busy runqueue (because we don't call
2847 if (sd
->balance_interval
< sd
->max_interval
)
2848 sd
->balance_interval
*= 2;
2851 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2852 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2857 schedstat_inc(sd
, lb_balanced
[idle
]);
2859 sd
->nr_balance_failed
= 0;
2862 /* tune up the balancing interval */
2863 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2864 (sd
->balance_interval
< sd
->max_interval
))
2865 sd
->balance_interval
*= 2;
2867 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2868 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2874 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2875 * tasks if there is an imbalance.
2877 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2878 * this_rq is locked.
2881 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2883 struct sched_group
*group
;
2884 struct rq
*busiest
= NULL
;
2885 unsigned long imbalance
;
2889 cpumask_t cpus
= CPU_MASK_ALL
;
2892 * When power savings policy is enabled for the parent domain, idle
2893 * sibling can pick up load irrespective of busy siblings. In this case,
2894 * let the state of idle sibling percolate up as IDLE, instead of
2895 * portraying it as CPU_NOT_IDLE.
2897 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2898 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2901 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
2903 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2904 &sd_idle
, &cpus
, NULL
);
2906 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2910 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2913 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2917 BUG_ON(busiest
== this_rq
);
2919 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2922 if (busiest
->nr_running
> 1) {
2923 /* Attempt to move tasks */
2924 double_lock_balance(this_rq
, busiest
);
2925 /* this_rq->clock is already updated */
2926 update_rq_clock(busiest
);
2927 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2928 imbalance
, sd
, CPU_NEWLY_IDLE
,
2930 spin_unlock(&busiest
->lock
);
2932 if (unlikely(all_pinned
)) {
2933 cpu_clear(cpu_of(busiest
), cpus
);
2934 if (!cpus_empty(cpus
))
2940 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2941 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2942 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2945 sd
->nr_balance_failed
= 0;
2950 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2951 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2952 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2954 sd
->nr_balance_failed
= 0;
2960 * idle_balance is called by schedule() if this_cpu is about to become
2961 * idle. Attempts to pull tasks from other CPUs.
2963 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2965 struct sched_domain
*sd
;
2966 int pulled_task
= -1;
2967 unsigned long next_balance
= jiffies
+ HZ
;
2969 for_each_domain(this_cpu
, sd
) {
2970 unsigned long interval
;
2972 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2975 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2976 /* If we've pulled tasks over stop searching: */
2977 pulled_task
= load_balance_newidle(this_cpu
,
2980 interval
= msecs_to_jiffies(sd
->balance_interval
);
2981 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2982 next_balance
= sd
->last_balance
+ interval
;
2986 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2988 * We are going idle. next_balance may be set based on
2989 * a busy processor. So reset next_balance.
2991 this_rq
->next_balance
= next_balance
;
2996 * active_load_balance is run by migration threads. It pushes running tasks
2997 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2998 * running on each physical CPU where possible, and avoids physical /
2999 * logical imbalances.
3001 * Called with busiest_rq locked.
3003 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3005 int target_cpu
= busiest_rq
->push_cpu
;
3006 struct sched_domain
*sd
;
3007 struct rq
*target_rq
;
3009 /* Is there any task to move? */
3010 if (busiest_rq
->nr_running
<= 1)
3013 target_rq
= cpu_rq(target_cpu
);
3016 * This condition is "impossible", if it occurs
3017 * we need to fix it. Originally reported by
3018 * Bjorn Helgaas on a 128-cpu setup.
3020 BUG_ON(busiest_rq
== target_rq
);
3022 /* move a task from busiest_rq to target_rq */
3023 double_lock_balance(busiest_rq
, target_rq
);
3024 update_rq_clock(busiest_rq
);
3025 update_rq_clock(target_rq
);
3027 /* Search for an sd spanning us and the target CPU. */
3028 for_each_domain(target_cpu
, sd
) {
3029 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3030 cpu_isset(busiest_cpu
, sd
->span
))
3035 schedstat_inc(sd
, alb_count
);
3037 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3039 schedstat_inc(sd
, alb_pushed
);
3041 schedstat_inc(sd
, alb_failed
);
3043 spin_unlock(&target_rq
->lock
);
3048 atomic_t load_balancer
;
3050 } nohz ____cacheline_aligned
= {
3051 .load_balancer
= ATOMIC_INIT(-1),
3052 .cpu_mask
= CPU_MASK_NONE
,
3056 * This routine will try to nominate the ilb (idle load balancing)
3057 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3058 * load balancing on behalf of all those cpus. If all the cpus in the system
3059 * go into this tickless mode, then there will be no ilb owner (as there is
3060 * no need for one) and all the cpus will sleep till the next wakeup event
3063 * For the ilb owner, tick is not stopped. And this tick will be used
3064 * for idle load balancing. ilb owner will still be part of
3067 * While stopping the tick, this cpu will become the ilb owner if there
3068 * is no other owner. And will be the owner till that cpu becomes busy
3069 * or if all cpus in the system stop their ticks at which point
3070 * there is no need for ilb owner.
3072 * When the ilb owner becomes busy, it nominates another owner, during the
3073 * next busy scheduler_tick()
3075 int select_nohz_load_balancer(int stop_tick
)
3077 int cpu
= smp_processor_id();
3080 cpu_set(cpu
, nohz
.cpu_mask
);
3081 cpu_rq(cpu
)->in_nohz_recently
= 1;
3084 * If we are going offline and still the leader, give up!
3086 if (cpu_is_offline(cpu
) &&
3087 atomic_read(&nohz
.load_balancer
) == cpu
) {
3088 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3093 /* time for ilb owner also to sleep */
3094 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3095 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3096 atomic_set(&nohz
.load_balancer
, -1);
3100 if (atomic_read(&nohz
.load_balancer
) == -1) {
3101 /* make me the ilb owner */
3102 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3104 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3107 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3110 cpu_clear(cpu
, nohz
.cpu_mask
);
3112 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3113 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3120 static DEFINE_SPINLOCK(balancing
);
3123 * It checks each scheduling domain to see if it is due to be balanced,
3124 * and initiates a balancing operation if so.
3126 * Balancing parameters are set up in arch_init_sched_domains.
3128 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3131 struct rq
*rq
= cpu_rq(cpu
);
3132 unsigned long interval
;
3133 struct sched_domain
*sd
;
3134 /* Earliest time when we have to do rebalance again */
3135 unsigned long next_balance
= jiffies
+ 60*HZ
;
3136 int update_next_balance
= 0;
3138 for_each_domain(cpu
, sd
) {
3139 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3142 interval
= sd
->balance_interval
;
3143 if (idle
!= CPU_IDLE
)
3144 interval
*= sd
->busy_factor
;
3146 /* scale ms to jiffies */
3147 interval
= msecs_to_jiffies(interval
);
3148 if (unlikely(!interval
))
3150 if (interval
> HZ
*NR_CPUS
/10)
3151 interval
= HZ
*NR_CPUS
/10;
3154 if (sd
->flags
& SD_SERIALIZE
) {
3155 if (!spin_trylock(&balancing
))
3159 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3160 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3162 * We've pulled tasks over so either we're no
3163 * longer idle, or one of our SMT siblings is
3166 idle
= CPU_NOT_IDLE
;
3168 sd
->last_balance
= jiffies
;
3170 if (sd
->flags
& SD_SERIALIZE
)
3171 spin_unlock(&balancing
);
3173 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3174 next_balance
= sd
->last_balance
+ interval
;
3175 update_next_balance
= 1;
3179 * Stop the load balance at this level. There is another
3180 * CPU in our sched group which is doing load balancing more
3188 * next_balance will be updated only when there is a need.
3189 * When the cpu is attached to null domain for ex, it will not be
3192 if (likely(update_next_balance
))
3193 rq
->next_balance
= next_balance
;
3197 * run_rebalance_domains is triggered when needed from the scheduler tick.
3198 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3199 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3201 static void run_rebalance_domains(struct softirq_action
*h
)
3203 int this_cpu
= smp_processor_id();
3204 struct rq
*this_rq
= cpu_rq(this_cpu
);
3205 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3206 CPU_IDLE
: CPU_NOT_IDLE
;
3208 rebalance_domains(this_cpu
, idle
);
3212 * If this cpu is the owner for idle load balancing, then do the
3213 * balancing on behalf of the other idle cpus whose ticks are
3216 if (this_rq
->idle_at_tick
&&
3217 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3218 cpumask_t cpus
= nohz
.cpu_mask
;
3222 cpu_clear(this_cpu
, cpus
);
3223 for_each_cpu_mask(balance_cpu
, cpus
) {
3225 * If this cpu gets work to do, stop the load balancing
3226 * work being done for other cpus. Next load
3227 * balancing owner will pick it up.
3232 rebalance_domains(balance_cpu
, CPU_IDLE
);
3234 rq
= cpu_rq(balance_cpu
);
3235 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3236 this_rq
->next_balance
= rq
->next_balance
;
3243 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3245 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3246 * idle load balancing owner or decide to stop the periodic load balancing,
3247 * if the whole system is idle.
3249 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3253 * If we were in the nohz mode recently and busy at the current
3254 * scheduler tick, then check if we need to nominate new idle
3257 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3258 rq
->in_nohz_recently
= 0;
3260 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3261 cpu_clear(cpu
, nohz
.cpu_mask
);
3262 atomic_set(&nohz
.load_balancer
, -1);
3265 if (atomic_read(&nohz
.load_balancer
) == -1) {
3267 * simple selection for now: Nominate the
3268 * first cpu in the nohz list to be the next
3271 * TBD: Traverse the sched domains and nominate
3272 * the nearest cpu in the nohz.cpu_mask.
3274 int ilb
= first_cpu(nohz
.cpu_mask
);
3282 * If this cpu is idle and doing idle load balancing for all the
3283 * cpus with ticks stopped, is it time for that to stop?
3285 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3286 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3292 * If this cpu is idle and the idle load balancing is done by
3293 * someone else, then no need raise the SCHED_SOFTIRQ
3295 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3296 cpu_isset(cpu
, nohz
.cpu_mask
))
3299 if (time_after_eq(jiffies
, rq
->next_balance
))
3300 raise_softirq(SCHED_SOFTIRQ
);
3303 #else /* CONFIG_SMP */
3306 * on UP we do not need to balance between CPUs:
3308 static inline void idle_balance(int cpu
, struct rq
*rq
)
3314 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3316 EXPORT_PER_CPU_SYMBOL(kstat
);
3319 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3320 * that have not yet been banked in case the task is currently running.
3322 unsigned long long task_sched_runtime(struct task_struct
*p
)
3324 unsigned long flags
;
3328 rq
= task_rq_lock(p
, &flags
);
3329 ns
= p
->se
.sum_exec_runtime
;
3330 if (task_current(rq
, p
)) {
3331 update_rq_clock(rq
);
3332 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3333 if ((s64
)delta_exec
> 0)
3336 task_rq_unlock(rq
, &flags
);
3342 * Account user cpu time to a process.
3343 * @p: the process that the cpu time gets accounted to
3344 * @cputime: the cpu time spent in user space since the last update
3346 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3348 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3351 p
->utime
= cputime_add(p
->utime
, cputime
);
3353 /* Add user time to cpustat. */
3354 tmp
= cputime_to_cputime64(cputime
);
3355 if (TASK_NICE(p
) > 0)
3356 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3358 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3362 * Account guest cpu time to a process.
3363 * @p: the process that the cpu time gets accounted to
3364 * @cputime: the cpu time spent in virtual machine since the last update
3366 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3369 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3371 tmp
= cputime_to_cputime64(cputime
);
3373 p
->utime
= cputime_add(p
->utime
, cputime
);
3374 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3376 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3377 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3381 * Account scaled user cpu time to a process.
3382 * @p: the process that the cpu time gets accounted to
3383 * @cputime: the cpu time spent in user space since the last update
3385 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3387 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3391 * Account system cpu time to a process.
3392 * @p: the process that the cpu time gets accounted to
3393 * @hardirq_offset: the offset to subtract from hardirq_count()
3394 * @cputime: the cpu time spent in kernel space since the last update
3396 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3399 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3400 struct rq
*rq
= this_rq();
3403 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3404 return account_guest_time(p
, cputime
);
3406 p
->stime
= cputime_add(p
->stime
, cputime
);
3408 /* Add system time to cpustat. */
3409 tmp
= cputime_to_cputime64(cputime
);
3410 if (hardirq_count() - hardirq_offset
)
3411 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3412 else if (softirq_count())
3413 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3414 else if (p
!= rq
->idle
)
3415 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3416 else if (atomic_read(&rq
->nr_iowait
) > 0)
3417 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3419 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3420 /* Account for system time used */
3421 acct_update_integrals(p
);
3425 * Account scaled system cpu time to a process.
3426 * @p: the process that the cpu time gets accounted to
3427 * @hardirq_offset: the offset to subtract from hardirq_count()
3428 * @cputime: the cpu time spent in kernel space since the last update
3430 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3432 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3436 * Account for involuntary wait time.
3437 * @p: the process from which the cpu time has been stolen
3438 * @steal: the cpu time spent in involuntary wait
3440 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3442 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3443 cputime64_t tmp
= cputime_to_cputime64(steal
);
3444 struct rq
*rq
= this_rq();
3446 if (p
== rq
->idle
) {
3447 p
->stime
= cputime_add(p
->stime
, steal
);
3448 if (atomic_read(&rq
->nr_iowait
) > 0)
3449 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3451 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3453 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3457 * This function gets called by the timer code, with HZ frequency.
3458 * We call it with interrupts disabled.
3460 * It also gets called by the fork code, when changing the parent's
3463 void scheduler_tick(void)
3465 int cpu
= smp_processor_id();
3466 struct rq
*rq
= cpu_rq(cpu
);
3467 struct task_struct
*curr
= rq
->curr
;
3468 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3470 spin_lock(&rq
->lock
);
3471 __update_rq_clock(rq
);
3473 * Let rq->clock advance by at least TICK_NSEC:
3475 if (unlikely(rq
->clock
< next_tick
))
3476 rq
->clock
= next_tick
;
3477 rq
->tick_timestamp
= rq
->clock
;
3478 update_cpu_load(rq
);
3479 if (curr
!= rq
->idle
) /* FIXME: needed? */
3480 curr
->sched_class
->task_tick(rq
, curr
);
3481 spin_unlock(&rq
->lock
);
3484 rq
->idle_at_tick
= idle_cpu(cpu
);
3485 trigger_load_balance(rq
, cpu
);
3489 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3491 void fastcall
add_preempt_count(int val
)
3496 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3498 preempt_count() += val
;
3500 * Spinlock count overflowing soon?
3502 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3505 EXPORT_SYMBOL(add_preempt_count
);
3507 void fastcall
sub_preempt_count(int val
)
3512 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3515 * Is the spinlock portion underflowing?
3517 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3518 !(preempt_count() & PREEMPT_MASK
)))
3521 preempt_count() -= val
;
3523 EXPORT_SYMBOL(sub_preempt_count
);
3528 * Print scheduling while atomic bug:
3530 static noinline
void __schedule_bug(struct task_struct
*prev
)
3532 struct pt_regs
*regs
= get_irq_regs();
3534 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3535 prev
->comm
, prev
->pid
, preempt_count());
3537 debug_show_held_locks(prev
);
3538 if (irqs_disabled())
3539 print_irqtrace_events(prev
);
3548 * Various schedule()-time debugging checks and statistics:
3550 static inline void schedule_debug(struct task_struct
*prev
)
3553 * Test if we are atomic. Since do_exit() needs to call into
3554 * schedule() atomically, we ignore that path for now.
3555 * Otherwise, whine if we are scheduling when we should not be.
3557 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3558 __schedule_bug(prev
);
3560 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3562 schedstat_inc(this_rq(), sched_count
);
3563 #ifdef CONFIG_SCHEDSTATS
3564 if (unlikely(prev
->lock_depth
>= 0)) {
3565 schedstat_inc(this_rq(), bkl_count
);
3566 schedstat_inc(prev
, sched_info
.bkl_count
);
3572 * Pick up the highest-prio task:
3574 static inline struct task_struct
*
3575 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3577 const struct sched_class
*class;
3578 struct task_struct
*p
;
3581 * Optimization: we know that if all tasks are in
3582 * the fair class we can call that function directly:
3584 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3585 p
= fair_sched_class
.pick_next_task(rq
);
3590 class = sched_class_highest
;
3592 p
= class->pick_next_task(rq
);
3596 * Will never be NULL as the idle class always
3597 * returns a non-NULL p:
3599 class = class->next
;
3604 * schedule() is the main scheduler function.
3606 asmlinkage
void __sched
schedule(void)
3608 struct task_struct
*prev
, *next
;
3615 cpu
= smp_processor_id();
3619 switch_count
= &prev
->nivcsw
;
3621 release_kernel_lock(prev
);
3622 need_resched_nonpreemptible
:
3624 schedule_debug(prev
);
3627 * Do the rq-clock update outside the rq lock:
3629 local_irq_disable();
3630 __update_rq_clock(rq
);
3631 spin_lock(&rq
->lock
);
3632 clear_tsk_need_resched(prev
);
3634 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3635 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3636 unlikely(signal_pending(prev
)))) {
3637 prev
->state
= TASK_RUNNING
;
3639 deactivate_task(rq
, prev
, 1);
3641 switch_count
= &prev
->nvcsw
;
3644 schedule_balance_rt(rq
, prev
);
3646 if (unlikely(!rq
->nr_running
))
3647 idle_balance(cpu
, rq
);
3649 prev
->sched_class
->put_prev_task(rq
, prev
);
3650 next
= pick_next_task(rq
, prev
);
3652 sched_info_switch(prev
, next
);
3654 if (likely(prev
!= next
)) {
3659 context_switch(rq
, prev
, next
); /* unlocks the rq */
3661 spin_unlock_irq(&rq
->lock
);
3663 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3664 cpu
= smp_processor_id();
3666 goto need_resched_nonpreemptible
;
3668 preempt_enable_no_resched();
3669 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3672 EXPORT_SYMBOL(schedule
);
3674 #ifdef CONFIG_PREEMPT
3676 * this is the entry point to schedule() from in-kernel preemption
3677 * off of preempt_enable. Kernel preemptions off return from interrupt
3678 * occur there and call schedule directly.
3680 asmlinkage
void __sched
preempt_schedule(void)
3682 struct thread_info
*ti
= current_thread_info();
3683 #ifdef CONFIG_PREEMPT_BKL
3684 struct task_struct
*task
= current
;
3685 int saved_lock_depth
;
3688 * If there is a non-zero preempt_count or interrupts are disabled,
3689 * we do not want to preempt the current task. Just return..
3691 if (likely(ti
->preempt_count
|| irqs_disabled()))
3695 add_preempt_count(PREEMPT_ACTIVE
);
3698 * We keep the big kernel semaphore locked, but we
3699 * clear ->lock_depth so that schedule() doesnt
3700 * auto-release the semaphore:
3702 #ifdef CONFIG_PREEMPT_BKL
3703 saved_lock_depth
= task
->lock_depth
;
3704 task
->lock_depth
= -1;
3707 #ifdef CONFIG_PREEMPT_BKL
3708 task
->lock_depth
= saved_lock_depth
;
3710 sub_preempt_count(PREEMPT_ACTIVE
);
3713 * Check again in case we missed a preemption opportunity
3714 * between schedule and now.
3717 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3719 EXPORT_SYMBOL(preempt_schedule
);
3722 * this is the entry point to schedule() from kernel preemption
3723 * off of irq context.
3724 * Note, that this is called and return with irqs disabled. This will
3725 * protect us against recursive calling from irq.
3727 asmlinkage
void __sched
preempt_schedule_irq(void)
3729 struct thread_info
*ti
= current_thread_info();
3730 #ifdef CONFIG_PREEMPT_BKL
3731 struct task_struct
*task
= current
;
3732 int saved_lock_depth
;
3734 /* Catch callers which need to be fixed */
3735 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3738 add_preempt_count(PREEMPT_ACTIVE
);
3741 * We keep the big kernel semaphore locked, but we
3742 * clear ->lock_depth so that schedule() doesnt
3743 * auto-release the semaphore:
3745 #ifdef CONFIG_PREEMPT_BKL
3746 saved_lock_depth
= task
->lock_depth
;
3747 task
->lock_depth
= -1;
3751 local_irq_disable();
3752 #ifdef CONFIG_PREEMPT_BKL
3753 task
->lock_depth
= saved_lock_depth
;
3755 sub_preempt_count(PREEMPT_ACTIVE
);
3758 * Check again in case we missed a preemption opportunity
3759 * between schedule and now.
3762 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3765 #endif /* CONFIG_PREEMPT */
3767 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3770 return try_to_wake_up(curr
->private, mode
, sync
);
3772 EXPORT_SYMBOL(default_wake_function
);
3775 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3776 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3777 * number) then we wake all the non-exclusive tasks and one exclusive task.
3779 * There are circumstances in which we can try to wake a task which has already
3780 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3781 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3783 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3784 int nr_exclusive
, int sync
, void *key
)
3786 wait_queue_t
*curr
, *next
;
3788 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3789 unsigned flags
= curr
->flags
;
3791 if (curr
->func(curr
, mode
, sync
, key
) &&
3792 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3798 * __wake_up - wake up threads blocked on a waitqueue.
3800 * @mode: which threads
3801 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3802 * @key: is directly passed to the wakeup function
3804 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3805 int nr_exclusive
, void *key
)
3807 unsigned long flags
;
3809 spin_lock_irqsave(&q
->lock
, flags
);
3810 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3811 spin_unlock_irqrestore(&q
->lock
, flags
);
3813 EXPORT_SYMBOL(__wake_up
);
3816 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3818 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3820 __wake_up_common(q
, mode
, 1, 0, NULL
);
3824 * __wake_up_sync - wake up threads blocked on a waitqueue.
3826 * @mode: which threads
3827 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3829 * The sync wakeup differs that the waker knows that it will schedule
3830 * away soon, so while the target thread will be woken up, it will not
3831 * be migrated to another CPU - ie. the two threads are 'synchronized'
3832 * with each other. This can prevent needless bouncing between CPUs.
3834 * On UP it can prevent extra preemption.
3837 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3839 unsigned long flags
;
3845 if (unlikely(!nr_exclusive
))
3848 spin_lock_irqsave(&q
->lock
, flags
);
3849 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3850 spin_unlock_irqrestore(&q
->lock
, flags
);
3852 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3854 void complete(struct completion
*x
)
3856 unsigned long flags
;
3858 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3860 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3862 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3864 EXPORT_SYMBOL(complete
);
3866 void complete_all(struct completion
*x
)
3868 unsigned long flags
;
3870 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3871 x
->done
+= UINT_MAX
/2;
3872 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3874 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3876 EXPORT_SYMBOL(complete_all
);
3878 static inline long __sched
3879 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3882 DECLARE_WAITQUEUE(wait
, current
);
3884 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3885 __add_wait_queue_tail(&x
->wait
, &wait
);
3887 if (state
== TASK_INTERRUPTIBLE
&&
3888 signal_pending(current
)) {
3889 __remove_wait_queue(&x
->wait
, &wait
);
3890 return -ERESTARTSYS
;
3892 __set_current_state(state
);
3893 spin_unlock_irq(&x
->wait
.lock
);
3894 timeout
= schedule_timeout(timeout
);
3895 spin_lock_irq(&x
->wait
.lock
);
3897 __remove_wait_queue(&x
->wait
, &wait
);
3901 __remove_wait_queue(&x
->wait
, &wait
);
3908 wait_for_common(struct completion
*x
, long timeout
, int state
)
3912 spin_lock_irq(&x
->wait
.lock
);
3913 timeout
= do_wait_for_common(x
, timeout
, state
);
3914 spin_unlock_irq(&x
->wait
.lock
);
3918 void __sched
wait_for_completion(struct completion
*x
)
3920 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3922 EXPORT_SYMBOL(wait_for_completion
);
3924 unsigned long __sched
3925 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3927 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3929 EXPORT_SYMBOL(wait_for_completion_timeout
);
3931 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3933 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3934 if (t
== -ERESTARTSYS
)
3938 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3940 unsigned long __sched
3941 wait_for_completion_interruptible_timeout(struct completion
*x
,
3942 unsigned long timeout
)
3944 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3946 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3949 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3951 unsigned long flags
;
3954 init_waitqueue_entry(&wait
, current
);
3956 __set_current_state(state
);
3958 spin_lock_irqsave(&q
->lock
, flags
);
3959 __add_wait_queue(q
, &wait
);
3960 spin_unlock(&q
->lock
);
3961 timeout
= schedule_timeout(timeout
);
3962 spin_lock_irq(&q
->lock
);
3963 __remove_wait_queue(q
, &wait
);
3964 spin_unlock_irqrestore(&q
->lock
, flags
);
3969 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3971 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3973 EXPORT_SYMBOL(interruptible_sleep_on
);
3976 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3978 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3980 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3982 void __sched
sleep_on(wait_queue_head_t
*q
)
3984 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3986 EXPORT_SYMBOL(sleep_on
);
3988 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3990 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3992 EXPORT_SYMBOL(sleep_on_timeout
);
3994 #ifdef CONFIG_RT_MUTEXES
3997 * rt_mutex_setprio - set the current priority of a task
3999 * @prio: prio value (kernel-internal form)
4001 * This function changes the 'effective' priority of a task. It does
4002 * not touch ->normal_prio like __setscheduler().
4004 * Used by the rt_mutex code to implement priority inheritance logic.
4006 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4008 unsigned long flags
;
4009 int oldprio
, on_rq
, running
;
4012 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4014 rq
= task_rq_lock(p
, &flags
);
4015 update_rq_clock(rq
);
4018 on_rq
= p
->se
.on_rq
;
4019 running
= task_current(rq
, p
);
4021 dequeue_task(rq
, p
, 0);
4023 p
->sched_class
->put_prev_task(rq
, p
);
4027 p
->sched_class
= &rt_sched_class
;
4029 p
->sched_class
= &fair_sched_class
;
4035 p
->sched_class
->set_curr_task(rq
);
4036 enqueue_task(rq
, p
, 0);
4038 * Reschedule if we are currently running on this runqueue and
4039 * our priority decreased, or if we are not currently running on
4040 * this runqueue and our priority is higher than the current's
4043 if (p
->prio
> oldprio
)
4044 resched_task(rq
->curr
);
4046 check_preempt_curr(rq
, p
);
4049 task_rq_unlock(rq
, &flags
);
4054 void set_user_nice(struct task_struct
*p
, long nice
)
4056 int old_prio
, delta
, on_rq
;
4057 unsigned long flags
;
4060 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4063 * We have to be careful, if called from sys_setpriority(),
4064 * the task might be in the middle of scheduling on another CPU.
4066 rq
= task_rq_lock(p
, &flags
);
4067 update_rq_clock(rq
);
4069 * The RT priorities are set via sched_setscheduler(), but we still
4070 * allow the 'normal' nice value to be set - but as expected
4071 * it wont have any effect on scheduling until the task is
4072 * SCHED_FIFO/SCHED_RR:
4074 if (task_has_rt_policy(p
)) {
4075 p
->static_prio
= NICE_TO_PRIO(nice
);
4078 on_rq
= p
->se
.on_rq
;
4080 dequeue_task(rq
, p
, 0);
4082 p
->static_prio
= NICE_TO_PRIO(nice
);
4085 p
->prio
= effective_prio(p
);
4086 delta
= p
->prio
- old_prio
;
4089 enqueue_task(rq
, p
, 0);
4091 * If the task increased its priority or is running and
4092 * lowered its priority, then reschedule its CPU:
4094 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4095 resched_task(rq
->curr
);
4098 task_rq_unlock(rq
, &flags
);
4100 EXPORT_SYMBOL(set_user_nice
);
4103 * can_nice - check if a task can reduce its nice value
4107 int can_nice(const struct task_struct
*p
, const int nice
)
4109 /* convert nice value [19,-20] to rlimit style value [1,40] */
4110 int nice_rlim
= 20 - nice
;
4112 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4113 capable(CAP_SYS_NICE
));
4116 #ifdef __ARCH_WANT_SYS_NICE
4119 * sys_nice - change the priority of the current process.
4120 * @increment: priority increment
4122 * sys_setpriority is a more generic, but much slower function that
4123 * does similar things.
4125 asmlinkage
long sys_nice(int increment
)
4130 * Setpriority might change our priority at the same moment.
4131 * We don't have to worry. Conceptually one call occurs first
4132 * and we have a single winner.
4134 if (increment
< -40)
4139 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4145 if (increment
< 0 && !can_nice(current
, nice
))
4148 retval
= security_task_setnice(current
, nice
);
4152 set_user_nice(current
, nice
);
4159 * task_prio - return the priority value of a given task.
4160 * @p: the task in question.
4162 * This is the priority value as seen by users in /proc.
4163 * RT tasks are offset by -200. Normal tasks are centered
4164 * around 0, value goes from -16 to +15.
4166 int task_prio(const struct task_struct
*p
)
4168 return p
->prio
- MAX_RT_PRIO
;
4172 * task_nice - return the nice value of a given task.
4173 * @p: the task in question.
4175 int task_nice(const struct task_struct
*p
)
4177 return TASK_NICE(p
);
4179 EXPORT_SYMBOL_GPL(task_nice
);
4182 * idle_cpu - is a given cpu idle currently?
4183 * @cpu: the processor in question.
4185 int idle_cpu(int cpu
)
4187 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4191 * idle_task - return the idle task for a given cpu.
4192 * @cpu: the processor in question.
4194 struct task_struct
*idle_task(int cpu
)
4196 return cpu_rq(cpu
)->idle
;
4200 * find_process_by_pid - find a process with a matching PID value.
4201 * @pid: the pid in question.
4203 static struct task_struct
*find_process_by_pid(pid_t pid
)
4205 return pid
? find_task_by_vpid(pid
) : current
;
4208 /* Actually do priority change: must hold rq lock. */
4210 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4212 BUG_ON(p
->se
.on_rq
);
4215 switch (p
->policy
) {
4219 p
->sched_class
= &fair_sched_class
;
4223 p
->sched_class
= &rt_sched_class
;
4227 p
->rt_priority
= prio
;
4228 p
->normal_prio
= normal_prio(p
);
4229 /* we are holding p->pi_lock already */
4230 p
->prio
= rt_mutex_getprio(p
);
4235 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4236 * @p: the task in question.
4237 * @policy: new policy.
4238 * @param: structure containing the new RT priority.
4240 * NOTE that the task may be already dead.
4242 int sched_setscheduler(struct task_struct
*p
, int policy
,
4243 struct sched_param
*param
)
4245 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4246 unsigned long flags
;
4249 /* may grab non-irq protected spin_locks */
4250 BUG_ON(in_interrupt());
4252 /* double check policy once rq lock held */
4254 policy
= oldpolicy
= p
->policy
;
4255 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4256 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4257 policy
!= SCHED_IDLE
)
4260 * Valid priorities for SCHED_FIFO and SCHED_RR are
4261 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4262 * SCHED_BATCH and SCHED_IDLE is 0.
4264 if (param
->sched_priority
< 0 ||
4265 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4266 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4268 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4272 * Allow unprivileged RT tasks to decrease priority:
4274 if (!capable(CAP_SYS_NICE
)) {
4275 if (rt_policy(policy
)) {
4276 unsigned long rlim_rtprio
;
4278 if (!lock_task_sighand(p
, &flags
))
4280 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4281 unlock_task_sighand(p
, &flags
);
4283 /* can't set/change the rt policy */
4284 if (policy
!= p
->policy
&& !rlim_rtprio
)
4287 /* can't increase priority */
4288 if (param
->sched_priority
> p
->rt_priority
&&
4289 param
->sched_priority
> rlim_rtprio
)
4293 * Like positive nice levels, dont allow tasks to
4294 * move out of SCHED_IDLE either:
4296 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4299 /* can't change other user's priorities */
4300 if ((current
->euid
!= p
->euid
) &&
4301 (current
->euid
!= p
->uid
))
4305 retval
= security_task_setscheduler(p
, policy
, param
);
4309 * make sure no PI-waiters arrive (or leave) while we are
4310 * changing the priority of the task:
4312 spin_lock_irqsave(&p
->pi_lock
, flags
);
4314 * To be able to change p->policy safely, the apropriate
4315 * runqueue lock must be held.
4317 rq
= __task_rq_lock(p
);
4318 /* recheck policy now with rq lock held */
4319 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4320 policy
= oldpolicy
= -1;
4321 __task_rq_unlock(rq
);
4322 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4325 update_rq_clock(rq
);
4326 on_rq
= p
->se
.on_rq
;
4327 running
= task_current(rq
, p
);
4329 deactivate_task(rq
, p
, 0);
4331 p
->sched_class
->put_prev_task(rq
, p
);
4335 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4339 p
->sched_class
->set_curr_task(rq
);
4340 activate_task(rq
, p
, 0);
4342 * Reschedule if we are currently running on this runqueue and
4343 * our priority decreased, or if we are not currently running on
4344 * this runqueue and our priority is higher than the current's
4347 if (p
->prio
> oldprio
)
4348 resched_task(rq
->curr
);
4350 check_preempt_curr(rq
, p
);
4353 __task_rq_unlock(rq
);
4354 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4356 rt_mutex_adjust_pi(p
);
4360 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4363 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4365 struct sched_param lparam
;
4366 struct task_struct
*p
;
4369 if (!param
|| pid
< 0)
4371 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4376 p
= find_process_by_pid(pid
);
4378 retval
= sched_setscheduler(p
, policy
, &lparam
);
4385 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4386 * @pid: the pid in question.
4387 * @policy: new policy.
4388 * @param: structure containing the new RT priority.
4391 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4393 /* negative values for policy are not valid */
4397 return do_sched_setscheduler(pid
, policy
, param
);
4401 * sys_sched_setparam - set/change the RT priority of a thread
4402 * @pid: the pid in question.
4403 * @param: structure containing the new RT priority.
4405 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4407 return do_sched_setscheduler(pid
, -1, param
);
4411 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4412 * @pid: the pid in question.
4414 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4416 struct task_struct
*p
;
4423 read_lock(&tasklist_lock
);
4424 p
= find_process_by_pid(pid
);
4426 retval
= security_task_getscheduler(p
);
4430 read_unlock(&tasklist_lock
);
4435 * sys_sched_getscheduler - get the RT priority of a thread
4436 * @pid: the pid in question.
4437 * @param: structure containing the RT priority.
4439 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4441 struct sched_param lp
;
4442 struct task_struct
*p
;
4445 if (!param
|| pid
< 0)
4448 read_lock(&tasklist_lock
);
4449 p
= find_process_by_pid(pid
);
4454 retval
= security_task_getscheduler(p
);
4458 lp
.sched_priority
= p
->rt_priority
;
4459 read_unlock(&tasklist_lock
);
4462 * This one might sleep, we cannot do it with a spinlock held ...
4464 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4469 read_unlock(&tasklist_lock
);
4473 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4475 cpumask_t cpus_allowed
;
4476 struct task_struct
*p
;
4480 read_lock(&tasklist_lock
);
4482 p
= find_process_by_pid(pid
);
4484 read_unlock(&tasklist_lock
);
4490 * It is not safe to call set_cpus_allowed with the
4491 * tasklist_lock held. We will bump the task_struct's
4492 * usage count and then drop tasklist_lock.
4495 read_unlock(&tasklist_lock
);
4498 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4499 !capable(CAP_SYS_NICE
))
4502 retval
= security_task_setscheduler(p
, 0, NULL
);
4506 cpus_allowed
= cpuset_cpus_allowed(p
);
4507 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4509 retval
= set_cpus_allowed(p
, new_mask
);
4512 cpus_allowed
= cpuset_cpus_allowed(p
);
4513 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4515 * We must have raced with a concurrent cpuset
4516 * update. Just reset the cpus_allowed to the
4517 * cpuset's cpus_allowed
4519 new_mask
= cpus_allowed
;
4529 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4530 cpumask_t
*new_mask
)
4532 if (len
< sizeof(cpumask_t
)) {
4533 memset(new_mask
, 0, sizeof(cpumask_t
));
4534 } else if (len
> sizeof(cpumask_t
)) {
4535 len
= sizeof(cpumask_t
);
4537 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4541 * sys_sched_setaffinity - set the cpu affinity of a process
4542 * @pid: pid of the process
4543 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4544 * @user_mask_ptr: user-space pointer to the new cpu mask
4546 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4547 unsigned long __user
*user_mask_ptr
)
4552 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4556 return sched_setaffinity(pid
, new_mask
);
4560 * Represents all cpu's present in the system
4561 * In systems capable of hotplug, this map could dynamically grow
4562 * as new cpu's are detected in the system via any platform specific
4563 * method, such as ACPI for e.g.
4566 cpumask_t cpu_present_map __read_mostly
;
4567 EXPORT_SYMBOL(cpu_present_map
);
4570 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4571 EXPORT_SYMBOL(cpu_online_map
);
4573 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4574 EXPORT_SYMBOL(cpu_possible_map
);
4577 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4579 struct task_struct
*p
;
4583 read_lock(&tasklist_lock
);
4586 p
= find_process_by_pid(pid
);
4590 retval
= security_task_getscheduler(p
);
4594 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4597 read_unlock(&tasklist_lock
);
4604 * sys_sched_getaffinity - get the cpu affinity of a process
4605 * @pid: pid of the process
4606 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4607 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4609 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4610 unsigned long __user
*user_mask_ptr
)
4615 if (len
< sizeof(cpumask_t
))
4618 ret
= sched_getaffinity(pid
, &mask
);
4622 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4625 return sizeof(cpumask_t
);
4629 * sys_sched_yield - yield the current processor to other threads.
4631 * This function yields the current CPU to other tasks. If there are no
4632 * other threads running on this CPU then this function will return.
4634 asmlinkage
long sys_sched_yield(void)
4636 struct rq
*rq
= this_rq_lock();
4638 schedstat_inc(rq
, yld_count
);
4639 current
->sched_class
->yield_task(rq
);
4642 * Since we are going to call schedule() anyway, there's
4643 * no need to preempt or enable interrupts:
4645 __release(rq
->lock
);
4646 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4647 _raw_spin_unlock(&rq
->lock
);
4648 preempt_enable_no_resched();
4655 static void __cond_resched(void)
4657 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4658 __might_sleep(__FILE__
, __LINE__
);
4661 * The BKS might be reacquired before we have dropped
4662 * PREEMPT_ACTIVE, which could trigger a second
4663 * cond_resched() call.
4666 add_preempt_count(PREEMPT_ACTIVE
);
4668 sub_preempt_count(PREEMPT_ACTIVE
);
4669 } while (need_resched());
4672 int __sched
cond_resched(void)
4674 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4675 system_state
== SYSTEM_RUNNING
) {
4681 EXPORT_SYMBOL(cond_resched
);
4684 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4685 * call schedule, and on return reacquire the lock.
4687 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4688 * operations here to prevent schedule() from being called twice (once via
4689 * spin_unlock(), once by hand).
4691 int cond_resched_lock(spinlock_t
*lock
)
4695 if (need_lockbreak(lock
)) {
4701 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4702 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4703 _raw_spin_unlock(lock
);
4704 preempt_enable_no_resched();
4711 EXPORT_SYMBOL(cond_resched_lock
);
4713 int __sched
cond_resched_softirq(void)
4715 BUG_ON(!in_softirq());
4717 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4725 EXPORT_SYMBOL(cond_resched_softirq
);
4728 * yield - yield the current processor to other threads.
4730 * This is a shortcut for kernel-space yielding - it marks the
4731 * thread runnable and calls sys_sched_yield().
4733 void __sched
yield(void)
4735 set_current_state(TASK_RUNNING
);
4738 EXPORT_SYMBOL(yield
);
4741 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4742 * that process accounting knows that this is a task in IO wait state.
4744 * But don't do that if it is a deliberate, throttling IO wait (this task
4745 * has set its backing_dev_info: the queue against which it should throttle)
4747 void __sched
io_schedule(void)
4749 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4751 delayacct_blkio_start();
4752 atomic_inc(&rq
->nr_iowait
);
4754 atomic_dec(&rq
->nr_iowait
);
4755 delayacct_blkio_end();
4757 EXPORT_SYMBOL(io_schedule
);
4759 long __sched
io_schedule_timeout(long timeout
)
4761 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4764 delayacct_blkio_start();
4765 atomic_inc(&rq
->nr_iowait
);
4766 ret
= schedule_timeout(timeout
);
4767 atomic_dec(&rq
->nr_iowait
);
4768 delayacct_blkio_end();
4773 * sys_sched_get_priority_max - return maximum RT priority.
4774 * @policy: scheduling class.
4776 * this syscall returns the maximum rt_priority that can be used
4777 * by a given scheduling class.
4779 asmlinkage
long sys_sched_get_priority_max(int policy
)
4786 ret
= MAX_USER_RT_PRIO
-1;
4798 * sys_sched_get_priority_min - return minimum RT priority.
4799 * @policy: scheduling class.
4801 * this syscall returns the minimum rt_priority that can be used
4802 * by a given scheduling class.
4804 asmlinkage
long sys_sched_get_priority_min(int policy
)
4822 * sys_sched_rr_get_interval - return the default timeslice of a process.
4823 * @pid: pid of the process.
4824 * @interval: userspace pointer to the timeslice value.
4826 * this syscall writes the default timeslice value of a given process
4827 * into the user-space timespec buffer. A value of '0' means infinity.
4830 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4832 struct task_struct
*p
;
4833 unsigned int time_slice
;
4841 read_lock(&tasklist_lock
);
4842 p
= find_process_by_pid(pid
);
4846 retval
= security_task_getscheduler(p
);
4851 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4852 * tasks that are on an otherwise idle runqueue:
4855 if (p
->policy
== SCHED_RR
) {
4856 time_slice
= DEF_TIMESLICE
;
4858 struct sched_entity
*se
= &p
->se
;
4859 unsigned long flags
;
4862 rq
= task_rq_lock(p
, &flags
);
4863 if (rq
->cfs
.load
.weight
)
4864 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
4865 task_rq_unlock(rq
, &flags
);
4867 read_unlock(&tasklist_lock
);
4868 jiffies_to_timespec(time_slice
, &t
);
4869 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4873 read_unlock(&tasklist_lock
);
4877 static const char stat_nam
[] = "RSDTtZX";
4879 void sched_show_task(struct task_struct
*p
)
4881 unsigned long free
= 0;
4884 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4885 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
4886 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4887 #if BITS_PER_LONG == 32
4888 if (state
== TASK_RUNNING
)
4889 printk(KERN_CONT
" running ");
4891 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4893 if (state
== TASK_RUNNING
)
4894 printk(KERN_CONT
" running task ");
4896 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4898 #ifdef CONFIG_DEBUG_STACK_USAGE
4900 unsigned long *n
= end_of_stack(p
);
4903 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4906 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
4907 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
4909 if (state
!= TASK_RUNNING
)
4910 show_stack(p
, NULL
);
4913 void show_state_filter(unsigned long state_filter
)
4915 struct task_struct
*g
, *p
;
4917 #if BITS_PER_LONG == 32
4919 " task PC stack pid father\n");
4922 " task PC stack pid father\n");
4924 read_lock(&tasklist_lock
);
4925 do_each_thread(g
, p
) {
4927 * reset the NMI-timeout, listing all files on a slow
4928 * console might take alot of time:
4930 touch_nmi_watchdog();
4931 if (!state_filter
|| (p
->state
& state_filter
))
4933 } while_each_thread(g
, p
);
4935 touch_all_softlockup_watchdogs();
4937 #ifdef CONFIG_SCHED_DEBUG
4938 sysrq_sched_debug_show();
4940 read_unlock(&tasklist_lock
);
4942 * Only show locks if all tasks are dumped:
4944 if (state_filter
== -1)
4945 debug_show_all_locks();
4948 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4950 idle
->sched_class
= &idle_sched_class
;
4954 * init_idle - set up an idle thread for a given CPU
4955 * @idle: task in question
4956 * @cpu: cpu the idle task belongs to
4958 * NOTE: this function does not set the idle thread's NEED_RESCHED
4959 * flag, to make booting more robust.
4961 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4963 struct rq
*rq
= cpu_rq(cpu
);
4964 unsigned long flags
;
4967 idle
->se
.exec_start
= sched_clock();
4969 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4970 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4971 __set_task_cpu(idle
, cpu
);
4973 spin_lock_irqsave(&rq
->lock
, flags
);
4974 rq
->curr
= rq
->idle
= idle
;
4975 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4978 spin_unlock_irqrestore(&rq
->lock
, flags
);
4980 /* Set the preempt count _outside_ the spinlocks! */
4981 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4982 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4984 task_thread_info(idle
)->preempt_count
= 0;
4987 * The idle tasks have their own, simple scheduling class:
4989 idle
->sched_class
= &idle_sched_class
;
4993 * In a system that switches off the HZ timer nohz_cpu_mask
4994 * indicates which cpus entered this state. This is used
4995 * in the rcu update to wait only for active cpus. For system
4996 * which do not switch off the HZ timer nohz_cpu_mask should
4997 * always be CPU_MASK_NONE.
4999 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5002 * Increase the granularity value when there are more CPUs,
5003 * because with more CPUs the 'effective latency' as visible
5004 * to users decreases. But the relationship is not linear,
5005 * so pick a second-best guess by going with the log2 of the
5008 * This idea comes from the SD scheduler of Con Kolivas:
5010 static inline void sched_init_granularity(void)
5012 unsigned int factor
= 1 + ilog2(num_online_cpus());
5013 const unsigned long limit
= 200000000;
5015 sysctl_sched_min_granularity
*= factor
;
5016 if (sysctl_sched_min_granularity
> limit
)
5017 sysctl_sched_min_granularity
= limit
;
5019 sysctl_sched_latency
*= factor
;
5020 if (sysctl_sched_latency
> limit
)
5021 sysctl_sched_latency
= limit
;
5023 sysctl_sched_wakeup_granularity
*= factor
;
5024 sysctl_sched_batch_wakeup_granularity
*= factor
;
5029 * This is how migration works:
5031 * 1) we queue a struct migration_req structure in the source CPU's
5032 * runqueue and wake up that CPU's migration thread.
5033 * 2) we down() the locked semaphore => thread blocks.
5034 * 3) migration thread wakes up (implicitly it forces the migrated
5035 * thread off the CPU)
5036 * 4) it gets the migration request and checks whether the migrated
5037 * task is still in the wrong runqueue.
5038 * 5) if it's in the wrong runqueue then the migration thread removes
5039 * it and puts it into the right queue.
5040 * 6) migration thread up()s the semaphore.
5041 * 7) we wake up and the migration is done.
5045 * Change a given task's CPU affinity. Migrate the thread to a
5046 * proper CPU and schedule it away if the CPU it's executing on
5047 * is removed from the allowed bitmask.
5049 * NOTE: the caller must have a valid reference to the task, the
5050 * task must not exit() & deallocate itself prematurely. The
5051 * call is not atomic; no spinlocks may be held.
5053 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5055 struct migration_req req
;
5056 unsigned long flags
;
5060 rq
= task_rq_lock(p
, &flags
);
5061 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5066 if (p
->sched_class
->set_cpus_allowed
)
5067 p
->sched_class
->set_cpus_allowed(p
, &new_mask
);
5069 p
->cpus_allowed
= new_mask
;
5070 p
->nr_cpus_allowed
= cpus_weight(new_mask
);
5073 /* Can the task run on the task's current CPU? If so, we're done */
5074 if (cpu_isset(task_cpu(p
), new_mask
))
5077 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5078 /* Need help from migration thread: drop lock and wait. */
5079 task_rq_unlock(rq
, &flags
);
5080 wake_up_process(rq
->migration_thread
);
5081 wait_for_completion(&req
.done
);
5082 tlb_migrate_finish(p
->mm
);
5086 task_rq_unlock(rq
, &flags
);
5090 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5093 * Move (not current) task off this cpu, onto dest cpu. We're doing
5094 * this because either it can't run here any more (set_cpus_allowed()
5095 * away from this CPU, or CPU going down), or because we're
5096 * attempting to rebalance this task on exec (sched_exec).
5098 * So we race with normal scheduler movements, but that's OK, as long
5099 * as the task is no longer on this CPU.
5101 * Returns non-zero if task was successfully migrated.
5103 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5105 struct rq
*rq_dest
, *rq_src
;
5108 if (unlikely(cpu_is_offline(dest_cpu
)))
5111 rq_src
= cpu_rq(src_cpu
);
5112 rq_dest
= cpu_rq(dest_cpu
);
5114 double_rq_lock(rq_src
, rq_dest
);
5115 /* Already moved. */
5116 if (task_cpu(p
) != src_cpu
)
5118 /* Affinity changed (again). */
5119 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5122 on_rq
= p
->se
.on_rq
;
5124 deactivate_task(rq_src
, p
, 0);
5126 set_task_cpu(p
, dest_cpu
);
5128 activate_task(rq_dest
, p
, 0);
5129 check_preempt_curr(rq_dest
, p
);
5133 double_rq_unlock(rq_src
, rq_dest
);
5138 * migration_thread - this is a highprio system thread that performs
5139 * thread migration by bumping thread off CPU then 'pushing' onto
5142 static int migration_thread(void *data
)
5144 int cpu
= (long)data
;
5148 BUG_ON(rq
->migration_thread
!= current
);
5150 set_current_state(TASK_INTERRUPTIBLE
);
5151 while (!kthread_should_stop()) {
5152 struct migration_req
*req
;
5153 struct list_head
*head
;
5155 spin_lock_irq(&rq
->lock
);
5157 if (cpu_is_offline(cpu
)) {
5158 spin_unlock_irq(&rq
->lock
);
5162 if (rq
->active_balance
) {
5163 active_load_balance(rq
, cpu
);
5164 rq
->active_balance
= 0;
5167 head
= &rq
->migration_queue
;
5169 if (list_empty(head
)) {
5170 spin_unlock_irq(&rq
->lock
);
5172 set_current_state(TASK_INTERRUPTIBLE
);
5175 req
= list_entry(head
->next
, struct migration_req
, list
);
5176 list_del_init(head
->next
);
5178 spin_unlock(&rq
->lock
);
5179 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5182 complete(&req
->done
);
5184 __set_current_state(TASK_RUNNING
);
5188 /* Wait for kthread_stop */
5189 set_current_state(TASK_INTERRUPTIBLE
);
5190 while (!kthread_should_stop()) {
5192 set_current_state(TASK_INTERRUPTIBLE
);
5194 __set_current_state(TASK_RUNNING
);
5198 #ifdef CONFIG_HOTPLUG_CPU
5200 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5204 local_irq_disable();
5205 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5211 * Figure out where task on dead CPU should go, use force if necessary.
5212 * NOTE: interrupts should be disabled by the caller
5214 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5216 unsigned long flags
;
5223 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5224 cpus_and(mask
, mask
, p
->cpus_allowed
);
5225 dest_cpu
= any_online_cpu(mask
);
5227 /* On any allowed CPU? */
5228 if (dest_cpu
== NR_CPUS
)
5229 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5231 /* No more Mr. Nice Guy. */
5232 if (dest_cpu
== NR_CPUS
) {
5233 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5235 * Try to stay on the same cpuset, where the
5236 * current cpuset may be a subset of all cpus.
5237 * The cpuset_cpus_allowed_locked() variant of
5238 * cpuset_cpus_allowed() will not block. It must be
5239 * called within calls to cpuset_lock/cpuset_unlock.
5241 rq
= task_rq_lock(p
, &flags
);
5242 p
->cpus_allowed
= cpus_allowed
;
5243 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5244 task_rq_unlock(rq
, &flags
);
5247 * Don't tell them about moving exiting tasks or
5248 * kernel threads (both mm NULL), since they never
5251 if (p
->mm
&& printk_ratelimit()) {
5252 printk(KERN_INFO
"process %d (%s) no "
5253 "longer affine to cpu%d\n",
5254 task_pid_nr(p
), p
->comm
, dead_cpu
);
5257 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5261 * While a dead CPU has no uninterruptible tasks queued at this point,
5262 * it might still have a nonzero ->nr_uninterruptible counter, because
5263 * for performance reasons the counter is not stricly tracking tasks to
5264 * their home CPUs. So we just add the counter to another CPU's counter,
5265 * to keep the global sum constant after CPU-down:
5267 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5269 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5270 unsigned long flags
;
5272 local_irq_save(flags
);
5273 double_rq_lock(rq_src
, rq_dest
);
5274 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5275 rq_src
->nr_uninterruptible
= 0;
5276 double_rq_unlock(rq_src
, rq_dest
);
5277 local_irq_restore(flags
);
5280 /* Run through task list and migrate tasks from the dead cpu. */
5281 static void migrate_live_tasks(int src_cpu
)
5283 struct task_struct
*p
, *t
;
5285 read_lock(&tasklist_lock
);
5287 do_each_thread(t
, p
) {
5291 if (task_cpu(p
) == src_cpu
)
5292 move_task_off_dead_cpu(src_cpu
, p
);
5293 } while_each_thread(t
, p
);
5295 read_unlock(&tasklist_lock
);
5299 * Schedules idle task to be the next runnable task on current CPU.
5300 * It does so by boosting its priority to highest possible.
5301 * Used by CPU offline code.
5303 void sched_idle_next(void)
5305 int this_cpu
= smp_processor_id();
5306 struct rq
*rq
= cpu_rq(this_cpu
);
5307 struct task_struct
*p
= rq
->idle
;
5308 unsigned long flags
;
5310 /* cpu has to be offline */
5311 BUG_ON(cpu_online(this_cpu
));
5314 * Strictly not necessary since rest of the CPUs are stopped by now
5315 * and interrupts disabled on the current cpu.
5317 spin_lock_irqsave(&rq
->lock
, flags
);
5319 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5321 update_rq_clock(rq
);
5322 activate_task(rq
, p
, 0);
5324 spin_unlock_irqrestore(&rq
->lock
, flags
);
5328 * Ensures that the idle task is using init_mm right before its cpu goes
5331 void idle_task_exit(void)
5333 struct mm_struct
*mm
= current
->active_mm
;
5335 BUG_ON(cpu_online(smp_processor_id()));
5338 switch_mm(mm
, &init_mm
, current
);
5342 /* called under rq->lock with disabled interrupts */
5343 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5345 struct rq
*rq
= cpu_rq(dead_cpu
);
5347 /* Must be exiting, otherwise would be on tasklist. */
5348 BUG_ON(!p
->exit_state
);
5350 /* Cannot have done final schedule yet: would have vanished. */
5351 BUG_ON(p
->state
== TASK_DEAD
);
5356 * Drop lock around migration; if someone else moves it,
5357 * that's OK. No task can be added to this CPU, so iteration is
5360 spin_unlock_irq(&rq
->lock
);
5361 move_task_off_dead_cpu(dead_cpu
, p
);
5362 spin_lock_irq(&rq
->lock
);
5367 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5368 static void migrate_dead_tasks(unsigned int dead_cpu
)
5370 struct rq
*rq
= cpu_rq(dead_cpu
);
5371 struct task_struct
*next
;
5374 if (!rq
->nr_running
)
5376 update_rq_clock(rq
);
5377 next
= pick_next_task(rq
, rq
->curr
);
5380 migrate_dead(dead_cpu
, next
);
5384 #endif /* CONFIG_HOTPLUG_CPU */
5386 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5388 static struct ctl_table sd_ctl_dir
[] = {
5390 .procname
= "sched_domain",
5396 static struct ctl_table sd_ctl_root
[] = {
5398 .ctl_name
= CTL_KERN
,
5399 .procname
= "kernel",
5401 .child
= sd_ctl_dir
,
5406 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5408 struct ctl_table
*entry
=
5409 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5414 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5416 struct ctl_table
*entry
;
5419 * In the intermediate directories, both the child directory and
5420 * procname are dynamically allocated and could fail but the mode
5421 * will always be set. In the lowest directory the names are
5422 * static strings and all have proc handlers.
5424 for (entry
= *tablep
; entry
->mode
; entry
++) {
5426 sd_free_ctl_entry(&entry
->child
);
5427 if (entry
->proc_handler
== NULL
)
5428 kfree(entry
->procname
);
5436 set_table_entry(struct ctl_table
*entry
,
5437 const char *procname
, void *data
, int maxlen
,
5438 mode_t mode
, proc_handler
*proc_handler
)
5440 entry
->procname
= procname
;
5442 entry
->maxlen
= maxlen
;
5444 entry
->proc_handler
= proc_handler
;
5447 static struct ctl_table
*
5448 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5450 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5455 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5456 sizeof(long), 0644, proc_doulongvec_minmax
);
5457 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5458 sizeof(long), 0644, proc_doulongvec_minmax
);
5459 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5460 sizeof(int), 0644, proc_dointvec_minmax
);
5461 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5462 sizeof(int), 0644, proc_dointvec_minmax
);
5463 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5464 sizeof(int), 0644, proc_dointvec_minmax
);
5465 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5466 sizeof(int), 0644, proc_dointvec_minmax
);
5467 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5468 sizeof(int), 0644, proc_dointvec_minmax
);
5469 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5470 sizeof(int), 0644, proc_dointvec_minmax
);
5471 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5472 sizeof(int), 0644, proc_dointvec_minmax
);
5473 set_table_entry(&table
[9], "cache_nice_tries",
5474 &sd
->cache_nice_tries
,
5475 sizeof(int), 0644, proc_dointvec_minmax
);
5476 set_table_entry(&table
[10], "flags", &sd
->flags
,
5477 sizeof(int), 0644, proc_dointvec_minmax
);
5478 /* &table[11] is terminator */
5483 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5485 struct ctl_table
*entry
, *table
;
5486 struct sched_domain
*sd
;
5487 int domain_num
= 0, i
;
5490 for_each_domain(cpu
, sd
)
5492 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5497 for_each_domain(cpu
, sd
) {
5498 snprintf(buf
, 32, "domain%d", i
);
5499 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5501 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5508 static struct ctl_table_header
*sd_sysctl_header
;
5509 static void register_sched_domain_sysctl(void)
5511 int i
, cpu_num
= num_online_cpus();
5512 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5515 WARN_ON(sd_ctl_dir
[0].child
);
5516 sd_ctl_dir
[0].child
= entry
;
5521 for_each_online_cpu(i
) {
5522 snprintf(buf
, 32, "cpu%d", i
);
5523 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5525 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5529 WARN_ON(sd_sysctl_header
);
5530 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5533 /* may be called multiple times per register */
5534 static void unregister_sched_domain_sysctl(void)
5536 if (sd_sysctl_header
)
5537 unregister_sysctl_table(sd_sysctl_header
);
5538 sd_sysctl_header
= NULL
;
5539 if (sd_ctl_dir
[0].child
)
5540 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5543 static void register_sched_domain_sysctl(void)
5546 static void unregister_sched_domain_sysctl(void)
5552 * migration_call - callback that gets triggered when a CPU is added.
5553 * Here we can start up the necessary migration thread for the new CPU.
5555 static int __cpuinit
5556 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5558 struct task_struct
*p
;
5559 int cpu
= (long)hcpu
;
5560 unsigned long flags
;
5565 case CPU_UP_PREPARE
:
5566 case CPU_UP_PREPARE_FROZEN
:
5567 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5570 kthread_bind(p
, cpu
);
5571 /* Must be high prio: stop_machine expects to yield to it. */
5572 rq
= task_rq_lock(p
, &flags
);
5573 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5574 task_rq_unlock(rq
, &flags
);
5575 cpu_rq(cpu
)->migration_thread
= p
;
5579 case CPU_ONLINE_FROZEN
:
5580 /* Strictly unnecessary, as first user will wake it. */
5581 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5583 /* Update our root-domain */
5585 spin_lock_irqsave(&rq
->lock
, flags
);
5587 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5588 cpu_set(cpu
, rq
->rd
->online
);
5590 spin_unlock_irqrestore(&rq
->lock
, flags
);
5593 #ifdef CONFIG_HOTPLUG_CPU
5594 case CPU_UP_CANCELED
:
5595 case CPU_UP_CANCELED_FROZEN
:
5596 if (!cpu_rq(cpu
)->migration_thread
)
5598 /* Unbind it from offline cpu so it can run. Fall thru. */
5599 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5600 any_online_cpu(cpu_online_map
));
5601 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5602 cpu_rq(cpu
)->migration_thread
= NULL
;
5606 case CPU_DEAD_FROZEN
:
5607 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5608 migrate_live_tasks(cpu
);
5610 kthread_stop(rq
->migration_thread
);
5611 rq
->migration_thread
= NULL
;
5612 /* Idle task back to normal (off runqueue, low prio) */
5613 spin_lock_irq(&rq
->lock
);
5614 update_rq_clock(rq
);
5615 deactivate_task(rq
, rq
->idle
, 0);
5616 rq
->idle
->static_prio
= MAX_PRIO
;
5617 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5618 rq
->idle
->sched_class
= &idle_sched_class
;
5619 migrate_dead_tasks(cpu
);
5620 spin_unlock_irq(&rq
->lock
);
5622 migrate_nr_uninterruptible(rq
);
5623 BUG_ON(rq
->nr_running
!= 0);
5626 * No need to migrate the tasks: it was best-effort if
5627 * they didn't take sched_hotcpu_mutex. Just wake up
5630 spin_lock_irq(&rq
->lock
);
5631 while (!list_empty(&rq
->migration_queue
)) {
5632 struct migration_req
*req
;
5634 req
= list_entry(rq
->migration_queue
.next
,
5635 struct migration_req
, list
);
5636 list_del_init(&req
->list
);
5637 complete(&req
->done
);
5639 spin_unlock_irq(&rq
->lock
);
5642 case CPU_DOWN_PREPARE
:
5643 /* Update our root-domain */
5645 spin_lock_irqsave(&rq
->lock
, flags
);
5647 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5648 cpu_clear(cpu
, rq
->rd
->online
);
5650 spin_unlock_irqrestore(&rq
->lock
, flags
);
5657 /* Register at highest priority so that task migration (migrate_all_tasks)
5658 * happens before everything else.
5660 static struct notifier_block __cpuinitdata migration_notifier
= {
5661 .notifier_call
= migration_call
,
5665 void __init
migration_init(void)
5667 void *cpu
= (void *)(long)smp_processor_id();
5670 /* Start one for the boot CPU: */
5671 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5672 BUG_ON(err
== NOTIFY_BAD
);
5673 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5674 register_cpu_notifier(&migration_notifier
);
5680 /* Number of possible processor ids */
5681 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5682 EXPORT_SYMBOL(nr_cpu_ids
);
5684 #ifdef CONFIG_SCHED_DEBUG
5686 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
5688 struct sched_group
*group
= sd
->groups
;
5689 cpumask_t groupmask
;
5692 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5693 cpus_clear(groupmask
);
5695 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5697 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5698 printk("does not load-balance\n");
5700 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5705 printk(KERN_CONT
"span %s\n", str
);
5707 if (!cpu_isset(cpu
, sd
->span
)) {
5708 printk(KERN_ERR
"ERROR: domain->span does not contain "
5711 if (!cpu_isset(cpu
, group
->cpumask
)) {
5712 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5716 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5720 printk(KERN_ERR
"ERROR: group is NULL\n");
5724 if (!group
->__cpu_power
) {
5725 printk(KERN_CONT
"\n");
5726 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5731 if (!cpus_weight(group
->cpumask
)) {
5732 printk(KERN_CONT
"\n");
5733 printk(KERN_ERR
"ERROR: empty group\n");
5737 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5738 printk(KERN_CONT
"\n");
5739 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5743 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5745 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5746 printk(KERN_CONT
" %s", str
);
5748 group
= group
->next
;
5749 } while (group
!= sd
->groups
);
5750 printk(KERN_CONT
"\n");
5752 if (!cpus_equal(sd
->span
, groupmask
))
5753 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5755 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
5756 printk(KERN_ERR
"ERROR: parent span is not a superset "
5757 "of domain->span\n");
5761 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5766 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5770 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5773 if (sched_domain_debug_one(sd
, cpu
, level
))
5782 # define sched_domain_debug(sd, cpu) do { } while (0)
5785 static int sd_degenerate(struct sched_domain
*sd
)
5787 if (cpus_weight(sd
->span
) == 1)
5790 /* Following flags need at least 2 groups */
5791 if (sd
->flags
& (SD_LOAD_BALANCE
|
5792 SD_BALANCE_NEWIDLE
|
5796 SD_SHARE_PKG_RESOURCES
)) {
5797 if (sd
->groups
!= sd
->groups
->next
)
5801 /* Following flags don't use groups */
5802 if (sd
->flags
& (SD_WAKE_IDLE
|
5811 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5813 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5815 if (sd_degenerate(parent
))
5818 if (!cpus_equal(sd
->span
, parent
->span
))
5821 /* Does parent contain flags not in child? */
5822 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5823 if (cflags
& SD_WAKE_AFFINE
)
5824 pflags
&= ~SD_WAKE_BALANCE
;
5825 /* Flags needing groups don't count if only 1 group in parent */
5826 if (parent
->groups
== parent
->groups
->next
) {
5827 pflags
&= ~(SD_LOAD_BALANCE
|
5828 SD_BALANCE_NEWIDLE
|
5832 SD_SHARE_PKG_RESOURCES
);
5834 if (~cflags
& pflags
)
5840 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5842 unsigned long flags
;
5843 const struct sched_class
*class;
5845 spin_lock_irqsave(&rq
->lock
, flags
);
5848 struct root_domain
*old_rd
= rq
->rd
;
5850 for (class = sched_class_highest
; class; class = class->next
) {
5851 if (class->leave_domain
)
5852 class->leave_domain(rq
);
5855 if (atomic_dec_and_test(&old_rd
->refcount
))
5859 atomic_inc(&rd
->refcount
);
5862 for (class = sched_class_highest
; class; class = class->next
) {
5863 if (class->join_domain
)
5864 class->join_domain(rq
);
5867 spin_unlock_irqrestore(&rq
->lock
, flags
);
5870 static void init_rootdomain(struct root_domain
*rd
, const cpumask_t
*map
)
5872 memset(rd
, 0, sizeof(*rd
));
5875 cpus_and(rd
->online
, rd
->span
, cpu_online_map
);
5878 static void init_defrootdomain(void)
5880 cpumask_t cpus
= CPU_MASK_ALL
;
5882 init_rootdomain(&def_root_domain
, &cpus
);
5883 atomic_set(&def_root_domain
.refcount
, 1);
5886 static struct root_domain
*alloc_rootdomain(const cpumask_t
*map
)
5888 struct root_domain
*rd
;
5890 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5894 init_rootdomain(rd
, map
);
5900 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5901 * hold the hotplug lock.
5904 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5906 struct rq
*rq
= cpu_rq(cpu
);
5907 struct sched_domain
*tmp
;
5909 /* Remove the sched domains which do not contribute to scheduling. */
5910 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5911 struct sched_domain
*parent
= tmp
->parent
;
5914 if (sd_parent_degenerate(tmp
, parent
)) {
5915 tmp
->parent
= parent
->parent
;
5917 parent
->parent
->child
= tmp
;
5921 if (sd
&& sd_degenerate(sd
)) {
5927 sched_domain_debug(sd
, cpu
);
5929 rq_attach_root(rq
, rd
);
5930 rcu_assign_pointer(rq
->sd
, sd
);
5933 /* cpus with isolated domains */
5934 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5936 /* Setup the mask of cpus configured for isolated domains */
5937 static int __init
isolated_cpu_setup(char *str
)
5939 int ints
[NR_CPUS
], i
;
5941 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5942 cpus_clear(cpu_isolated_map
);
5943 for (i
= 1; i
<= ints
[0]; i
++)
5944 if (ints
[i
] < NR_CPUS
)
5945 cpu_set(ints
[i
], cpu_isolated_map
);
5949 __setup("isolcpus=", isolated_cpu_setup
);
5952 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5953 * to a function which identifies what group(along with sched group) a CPU
5954 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5955 * (due to the fact that we keep track of groups covered with a cpumask_t).
5957 * init_sched_build_groups will build a circular linked list of the groups
5958 * covered by the given span, and will set each group's ->cpumask correctly,
5959 * and ->cpu_power to 0.
5962 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5963 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5964 struct sched_group
**sg
))
5966 struct sched_group
*first
= NULL
, *last
= NULL
;
5967 cpumask_t covered
= CPU_MASK_NONE
;
5970 for_each_cpu_mask(i
, span
) {
5971 struct sched_group
*sg
;
5972 int group
= group_fn(i
, cpu_map
, &sg
);
5975 if (cpu_isset(i
, covered
))
5978 sg
->cpumask
= CPU_MASK_NONE
;
5979 sg
->__cpu_power
= 0;
5981 for_each_cpu_mask(j
, span
) {
5982 if (group_fn(j
, cpu_map
, NULL
) != group
)
5985 cpu_set(j
, covered
);
5986 cpu_set(j
, sg
->cpumask
);
5997 #define SD_NODES_PER_DOMAIN 16
6002 * find_next_best_node - find the next node to include in a sched_domain
6003 * @node: node whose sched_domain we're building
6004 * @used_nodes: nodes already in the sched_domain
6006 * Find the next node to include in a given scheduling domain. Simply
6007 * finds the closest node not already in the @used_nodes map.
6009 * Should use nodemask_t.
6011 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6013 int i
, n
, val
, min_val
, best_node
= 0;
6017 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6018 /* Start at @node */
6019 n
= (node
+ i
) % MAX_NUMNODES
;
6021 if (!nr_cpus_node(n
))
6024 /* Skip already used nodes */
6025 if (test_bit(n
, used_nodes
))
6028 /* Simple min distance search */
6029 val
= node_distance(node
, n
);
6031 if (val
< min_val
) {
6037 set_bit(best_node
, used_nodes
);
6042 * sched_domain_node_span - get a cpumask for a node's sched_domain
6043 * @node: node whose cpumask we're constructing
6044 * @size: number of nodes to include in this span
6046 * Given a node, construct a good cpumask for its sched_domain to span. It
6047 * should be one that prevents unnecessary balancing, but also spreads tasks
6050 static cpumask_t
sched_domain_node_span(int node
)
6052 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6053 cpumask_t span
, nodemask
;
6057 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6059 nodemask
= node_to_cpumask(node
);
6060 cpus_or(span
, span
, nodemask
);
6061 set_bit(node
, used_nodes
);
6063 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6064 int next_node
= find_next_best_node(node
, used_nodes
);
6066 nodemask
= node_to_cpumask(next_node
);
6067 cpus_or(span
, span
, nodemask
);
6074 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6077 * SMT sched-domains:
6079 #ifdef CONFIG_SCHED_SMT
6080 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6081 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6084 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6087 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6093 * multi-core sched-domains:
6095 #ifdef CONFIG_SCHED_MC
6096 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6097 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6100 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6102 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6105 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6106 cpus_and(mask
, mask
, *cpu_map
);
6107 group
= first_cpu(mask
);
6109 *sg
= &per_cpu(sched_group_core
, group
);
6112 #elif defined(CONFIG_SCHED_MC)
6114 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6117 *sg
= &per_cpu(sched_group_core
, cpu
);
6122 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6123 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6126 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6129 #ifdef CONFIG_SCHED_MC
6130 cpumask_t mask
= cpu_coregroup_map(cpu
);
6131 cpus_and(mask
, mask
, *cpu_map
);
6132 group
= first_cpu(mask
);
6133 #elif defined(CONFIG_SCHED_SMT)
6134 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6135 cpus_and(mask
, mask
, *cpu_map
);
6136 group
= first_cpu(mask
);
6141 *sg
= &per_cpu(sched_group_phys
, group
);
6147 * The init_sched_build_groups can't handle what we want to do with node
6148 * groups, so roll our own. Now each node has its own list of groups which
6149 * gets dynamically allocated.
6151 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6152 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6154 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6155 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6157 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6158 struct sched_group
**sg
)
6160 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6163 cpus_and(nodemask
, nodemask
, *cpu_map
);
6164 group
= first_cpu(nodemask
);
6167 *sg
= &per_cpu(sched_group_allnodes
, group
);
6171 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6173 struct sched_group
*sg
= group_head
;
6179 for_each_cpu_mask(j
, sg
->cpumask
) {
6180 struct sched_domain
*sd
;
6182 sd
= &per_cpu(phys_domains
, j
);
6183 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6185 * Only add "power" once for each
6191 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6194 } while (sg
!= group_head
);
6199 /* Free memory allocated for various sched_group structures */
6200 static void free_sched_groups(const cpumask_t
*cpu_map
)
6204 for_each_cpu_mask(cpu
, *cpu_map
) {
6205 struct sched_group
**sched_group_nodes
6206 = sched_group_nodes_bycpu
[cpu
];
6208 if (!sched_group_nodes
)
6211 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6212 cpumask_t nodemask
= node_to_cpumask(i
);
6213 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6215 cpus_and(nodemask
, nodemask
, *cpu_map
);
6216 if (cpus_empty(nodemask
))
6226 if (oldsg
!= sched_group_nodes
[i
])
6229 kfree(sched_group_nodes
);
6230 sched_group_nodes_bycpu
[cpu
] = NULL
;
6234 static void free_sched_groups(const cpumask_t
*cpu_map
)
6240 * Initialize sched groups cpu_power.
6242 * cpu_power indicates the capacity of sched group, which is used while
6243 * distributing the load between different sched groups in a sched domain.
6244 * Typically cpu_power for all the groups in a sched domain will be same unless
6245 * there are asymmetries in the topology. If there are asymmetries, group
6246 * having more cpu_power will pickup more load compared to the group having
6249 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6250 * the maximum number of tasks a group can handle in the presence of other idle
6251 * or lightly loaded groups in the same sched domain.
6253 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6255 struct sched_domain
*child
;
6256 struct sched_group
*group
;
6258 WARN_ON(!sd
|| !sd
->groups
);
6260 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6265 sd
->groups
->__cpu_power
= 0;
6268 * For perf policy, if the groups in child domain share resources
6269 * (for example cores sharing some portions of the cache hierarchy
6270 * or SMT), then set this domain groups cpu_power such that each group
6271 * can handle only one task, when there are other idle groups in the
6272 * same sched domain.
6274 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6276 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6277 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6282 * add cpu_power of each child group to this groups cpu_power
6284 group
= child
->groups
;
6286 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6287 group
= group
->next
;
6288 } while (group
!= child
->groups
);
6292 * Build sched domains for a given set of cpus and attach the sched domains
6293 * to the individual cpus
6295 static int build_sched_domains(const cpumask_t
*cpu_map
)
6298 struct root_domain
*rd
;
6300 struct sched_group
**sched_group_nodes
= NULL
;
6301 int sd_allnodes
= 0;
6304 * Allocate the per-node list of sched groups
6306 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6308 if (!sched_group_nodes
) {
6309 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6312 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6315 rd
= alloc_rootdomain(cpu_map
);
6317 printk(KERN_WARNING
"Cannot alloc root domain\n");
6322 * Set up domains for cpus specified by the cpu_map.
6324 for_each_cpu_mask(i
, *cpu_map
) {
6325 struct sched_domain
*sd
= NULL
, *p
;
6326 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6328 cpus_and(nodemask
, nodemask
, *cpu_map
);
6331 if (cpus_weight(*cpu_map
) >
6332 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6333 sd
= &per_cpu(allnodes_domains
, i
);
6334 *sd
= SD_ALLNODES_INIT
;
6335 sd
->span
= *cpu_map
;
6336 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6342 sd
= &per_cpu(node_domains
, i
);
6344 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6348 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6352 sd
= &per_cpu(phys_domains
, i
);
6354 sd
->span
= nodemask
;
6358 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6360 #ifdef CONFIG_SCHED_MC
6362 sd
= &per_cpu(core_domains
, i
);
6364 sd
->span
= cpu_coregroup_map(i
);
6365 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6368 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6371 #ifdef CONFIG_SCHED_SMT
6373 sd
= &per_cpu(cpu_domains
, i
);
6374 *sd
= SD_SIBLING_INIT
;
6375 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6376 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6379 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6383 #ifdef CONFIG_SCHED_SMT
6384 /* Set up CPU (sibling) groups */
6385 for_each_cpu_mask(i
, *cpu_map
) {
6386 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6387 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6388 if (i
!= first_cpu(this_sibling_map
))
6391 init_sched_build_groups(this_sibling_map
, cpu_map
,
6396 #ifdef CONFIG_SCHED_MC
6397 /* Set up multi-core groups */
6398 for_each_cpu_mask(i
, *cpu_map
) {
6399 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6400 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6401 if (i
!= first_cpu(this_core_map
))
6403 init_sched_build_groups(this_core_map
, cpu_map
,
6404 &cpu_to_core_group
);
6408 /* Set up physical groups */
6409 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6410 cpumask_t nodemask
= node_to_cpumask(i
);
6412 cpus_and(nodemask
, nodemask
, *cpu_map
);
6413 if (cpus_empty(nodemask
))
6416 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6420 /* Set up node groups */
6422 init_sched_build_groups(*cpu_map
, cpu_map
,
6423 &cpu_to_allnodes_group
);
6425 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6426 /* Set up node groups */
6427 struct sched_group
*sg
, *prev
;
6428 cpumask_t nodemask
= node_to_cpumask(i
);
6429 cpumask_t domainspan
;
6430 cpumask_t covered
= CPU_MASK_NONE
;
6433 cpus_and(nodemask
, nodemask
, *cpu_map
);
6434 if (cpus_empty(nodemask
)) {
6435 sched_group_nodes
[i
] = NULL
;
6439 domainspan
= sched_domain_node_span(i
);
6440 cpus_and(domainspan
, domainspan
, *cpu_map
);
6442 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6444 printk(KERN_WARNING
"Can not alloc domain group for "
6448 sched_group_nodes
[i
] = sg
;
6449 for_each_cpu_mask(j
, nodemask
) {
6450 struct sched_domain
*sd
;
6452 sd
= &per_cpu(node_domains
, j
);
6455 sg
->__cpu_power
= 0;
6456 sg
->cpumask
= nodemask
;
6458 cpus_or(covered
, covered
, nodemask
);
6461 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6462 cpumask_t tmp
, notcovered
;
6463 int n
= (i
+ j
) % MAX_NUMNODES
;
6465 cpus_complement(notcovered
, covered
);
6466 cpus_and(tmp
, notcovered
, *cpu_map
);
6467 cpus_and(tmp
, tmp
, domainspan
);
6468 if (cpus_empty(tmp
))
6471 nodemask
= node_to_cpumask(n
);
6472 cpus_and(tmp
, tmp
, nodemask
);
6473 if (cpus_empty(tmp
))
6476 sg
= kmalloc_node(sizeof(struct sched_group
),
6480 "Can not alloc domain group for node %d\n", j
);
6483 sg
->__cpu_power
= 0;
6485 sg
->next
= prev
->next
;
6486 cpus_or(covered
, covered
, tmp
);
6493 /* Calculate CPU power for physical packages and nodes */
6494 #ifdef CONFIG_SCHED_SMT
6495 for_each_cpu_mask(i
, *cpu_map
) {
6496 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6498 init_sched_groups_power(i
, sd
);
6501 #ifdef CONFIG_SCHED_MC
6502 for_each_cpu_mask(i
, *cpu_map
) {
6503 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6505 init_sched_groups_power(i
, sd
);
6509 for_each_cpu_mask(i
, *cpu_map
) {
6510 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6512 init_sched_groups_power(i
, sd
);
6516 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6517 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6520 struct sched_group
*sg
;
6522 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6523 init_numa_sched_groups_power(sg
);
6527 /* Attach the domains */
6528 for_each_cpu_mask(i
, *cpu_map
) {
6529 struct sched_domain
*sd
;
6530 #ifdef CONFIG_SCHED_SMT
6531 sd
= &per_cpu(cpu_domains
, i
);
6532 #elif defined(CONFIG_SCHED_MC)
6533 sd
= &per_cpu(core_domains
, i
);
6535 sd
= &per_cpu(phys_domains
, i
);
6537 cpu_attach_domain(sd
, rd
, i
);
6544 free_sched_groups(cpu_map
);
6549 static cpumask_t
*doms_cur
; /* current sched domains */
6550 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6553 * Special case: If a kmalloc of a doms_cur partition (array of
6554 * cpumask_t) fails, then fallback to a single sched domain,
6555 * as determined by the single cpumask_t fallback_doms.
6557 static cpumask_t fallback_doms
;
6560 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6561 * For now this just excludes isolated cpus, but could be used to
6562 * exclude other special cases in the future.
6564 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6569 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6571 doms_cur
= &fallback_doms
;
6572 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6573 err
= build_sched_domains(doms_cur
);
6574 register_sched_domain_sysctl();
6579 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6581 free_sched_groups(cpu_map
);
6585 * Detach sched domains from a group of cpus specified in cpu_map
6586 * These cpus will now be attached to the NULL domain
6588 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6592 unregister_sched_domain_sysctl();
6594 for_each_cpu_mask(i
, *cpu_map
)
6595 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6596 synchronize_sched();
6597 arch_destroy_sched_domains(cpu_map
);
6601 * Partition sched domains as specified by the 'ndoms_new'
6602 * cpumasks in the array doms_new[] of cpumasks. This compares
6603 * doms_new[] to the current sched domain partitioning, doms_cur[].
6604 * It destroys each deleted domain and builds each new domain.
6606 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6607 * The masks don't intersect (don't overlap.) We should setup one
6608 * sched domain for each mask. CPUs not in any of the cpumasks will
6609 * not be load balanced. If the same cpumask appears both in the
6610 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6613 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6614 * ownership of it and will kfree it when done with it. If the caller
6615 * failed the kmalloc call, then it can pass in doms_new == NULL,
6616 * and partition_sched_domains() will fallback to the single partition
6619 * Call with hotplug lock held
6621 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6627 /* always unregister in case we don't destroy any domains */
6628 unregister_sched_domain_sysctl();
6630 if (doms_new
== NULL
) {
6632 doms_new
= &fallback_doms
;
6633 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6636 /* Destroy deleted domains */
6637 for (i
= 0; i
< ndoms_cur
; i
++) {
6638 for (j
= 0; j
< ndoms_new
; j
++) {
6639 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
6642 /* no match - a current sched domain not in new doms_new[] */
6643 detach_destroy_domains(doms_cur
+ i
);
6648 /* Build new domains */
6649 for (i
= 0; i
< ndoms_new
; i
++) {
6650 for (j
= 0; j
< ndoms_cur
; j
++) {
6651 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
6654 /* no match - add a new doms_new */
6655 build_sched_domains(doms_new
+ i
);
6660 /* Remember the new sched domains */
6661 if (doms_cur
!= &fallback_doms
)
6663 doms_cur
= doms_new
;
6664 ndoms_cur
= ndoms_new
;
6666 register_sched_domain_sysctl();
6671 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6672 static int arch_reinit_sched_domains(void)
6677 detach_destroy_domains(&cpu_online_map
);
6678 err
= arch_init_sched_domains(&cpu_online_map
);
6684 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6688 if (buf
[0] != '0' && buf
[0] != '1')
6692 sched_smt_power_savings
= (buf
[0] == '1');
6694 sched_mc_power_savings
= (buf
[0] == '1');
6696 ret
= arch_reinit_sched_domains();
6698 return ret
? ret
: count
;
6701 #ifdef CONFIG_SCHED_MC
6702 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6704 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6706 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6707 const char *buf
, size_t count
)
6709 return sched_power_savings_store(buf
, count
, 0);
6711 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6712 sched_mc_power_savings_store
);
6715 #ifdef CONFIG_SCHED_SMT
6716 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6718 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6720 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6721 const char *buf
, size_t count
)
6723 return sched_power_savings_store(buf
, count
, 1);
6725 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6726 sched_smt_power_savings_store
);
6729 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6733 #ifdef CONFIG_SCHED_SMT
6735 err
= sysfs_create_file(&cls
->kset
.kobj
,
6736 &attr_sched_smt_power_savings
.attr
);
6738 #ifdef CONFIG_SCHED_MC
6739 if (!err
&& mc_capable())
6740 err
= sysfs_create_file(&cls
->kset
.kobj
,
6741 &attr_sched_mc_power_savings
.attr
);
6748 * Force a reinitialization of the sched domains hierarchy. The domains
6749 * and groups cannot be updated in place without racing with the balancing
6750 * code, so we temporarily attach all running cpus to the NULL domain
6751 * which will prevent rebalancing while the sched domains are recalculated.
6753 static int update_sched_domains(struct notifier_block
*nfb
,
6754 unsigned long action
, void *hcpu
)
6757 case CPU_UP_PREPARE
:
6758 case CPU_UP_PREPARE_FROZEN
:
6759 case CPU_DOWN_PREPARE
:
6760 case CPU_DOWN_PREPARE_FROZEN
:
6761 detach_destroy_domains(&cpu_online_map
);
6764 case CPU_UP_CANCELED
:
6765 case CPU_UP_CANCELED_FROZEN
:
6766 case CPU_DOWN_FAILED
:
6767 case CPU_DOWN_FAILED_FROZEN
:
6769 case CPU_ONLINE_FROZEN
:
6771 case CPU_DEAD_FROZEN
:
6773 * Fall through and re-initialise the domains.
6780 /* The hotplug lock is already held by cpu_up/cpu_down */
6781 arch_init_sched_domains(&cpu_online_map
);
6786 void __init
sched_init_smp(void)
6788 cpumask_t non_isolated_cpus
;
6791 arch_init_sched_domains(&cpu_online_map
);
6792 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6793 if (cpus_empty(non_isolated_cpus
))
6794 cpu_set(smp_processor_id(), non_isolated_cpus
);
6796 /* XXX: Theoretical race here - CPU may be hotplugged now */
6797 hotcpu_notifier(update_sched_domains
, 0);
6799 /* Move init over to a non-isolated CPU */
6800 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6802 sched_init_granularity();
6804 #ifdef CONFIG_FAIR_GROUP_SCHED
6805 if (nr_cpu_ids
== 1)
6808 lb_monitor_task
= kthread_create(load_balance_monitor
, NULL
,
6810 if (!IS_ERR(lb_monitor_task
)) {
6811 lb_monitor_task
->flags
|= PF_NOFREEZE
;
6812 wake_up_process(lb_monitor_task
);
6814 printk(KERN_ERR
"Could not create load balance monitor thread"
6815 "(error = %ld) \n", PTR_ERR(lb_monitor_task
));
6820 void __init
sched_init_smp(void)
6822 sched_init_granularity();
6824 #endif /* CONFIG_SMP */
6826 int in_sched_functions(unsigned long addr
)
6828 return in_lock_functions(addr
) ||
6829 (addr
>= (unsigned long)__sched_text_start
6830 && addr
< (unsigned long)__sched_text_end
);
6833 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6835 cfs_rq
->tasks_timeline
= RB_ROOT
;
6836 #ifdef CONFIG_FAIR_GROUP_SCHED
6839 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6842 void __init
sched_init(void)
6844 int highest_cpu
= 0;
6848 init_defrootdomain();
6851 for_each_possible_cpu(i
) {
6852 struct rt_prio_array
*array
;
6856 spin_lock_init(&rq
->lock
);
6857 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6860 init_cfs_rq(&rq
->cfs
, rq
);
6861 #ifdef CONFIG_FAIR_GROUP_SCHED
6862 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6864 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6865 struct sched_entity
*se
=
6866 &per_cpu(init_sched_entity
, i
);
6868 init_cfs_rq_p
[i
] = cfs_rq
;
6869 init_cfs_rq(cfs_rq
, rq
);
6870 cfs_rq
->tg
= &init_task_group
;
6871 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6872 &rq
->leaf_cfs_rq_list
);
6874 init_sched_entity_p
[i
] = se
;
6875 se
->cfs_rq
= &rq
->cfs
;
6877 se
->load
.weight
= init_task_group_load
;
6878 se
->load
.inv_weight
=
6879 div64_64(1ULL<<32, init_task_group_load
);
6882 init_task_group
.shares
= init_task_group_load
;
6885 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6886 rq
->cpu_load
[j
] = 0;
6890 rq_attach_root(rq
, &def_root_domain
);
6891 rq
->active_balance
= 0;
6892 rq
->next_balance
= jiffies
;
6895 rq
->migration_thread
= NULL
;
6896 INIT_LIST_HEAD(&rq
->migration_queue
);
6897 rq
->rt
.highest_prio
= MAX_RT_PRIO
;
6898 rq
->rt
.overloaded
= 0;
6900 atomic_set(&rq
->nr_iowait
, 0);
6902 array
= &rq
->rt
.active
;
6903 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6904 INIT_LIST_HEAD(array
->queue
+ j
);
6905 __clear_bit(j
, array
->bitmap
);
6908 /* delimiter for bitsearch: */
6909 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6912 set_load_weight(&init_task
);
6914 #ifdef CONFIG_PREEMPT_NOTIFIERS
6915 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6919 nr_cpu_ids
= highest_cpu
+ 1;
6920 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6923 #ifdef CONFIG_RT_MUTEXES
6924 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6928 * The boot idle thread does lazy MMU switching as well:
6930 atomic_inc(&init_mm
.mm_count
);
6931 enter_lazy_tlb(&init_mm
, current
);
6934 * Make us the idle thread. Technically, schedule() should not be
6935 * called from this thread, however somewhere below it might be,
6936 * but because we are the idle thread, we just pick up running again
6937 * when this runqueue becomes "idle".
6939 init_idle(current
, smp_processor_id());
6941 * During early bootup we pretend to be a normal task:
6943 current
->sched_class
= &fair_sched_class
;
6946 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6947 void __might_sleep(char *file
, int line
)
6950 static unsigned long prev_jiffy
; /* ratelimiting */
6952 if ((in_atomic() || irqs_disabled()) &&
6953 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6954 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6956 prev_jiffy
= jiffies
;
6957 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6958 " context at %s:%d\n", file
, line
);
6959 printk("in_atomic():%d, irqs_disabled():%d\n",
6960 in_atomic(), irqs_disabled());
6961 debug_show_held_locks(current
);
6962 if (irqs_disabled())
6963 print_irqtrace_events(current
);
6968 EXPORT_SYMBOL(__might_sleep
);
6971 #ifdef CONFIG_MAGIC_SYSRQ
6972 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6975 update_rq_clock(rq
);
6976 on_rq
= p
->se
.on_rq
;
6978 deactivate_task(rq
, p
, 0);
6979 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6981 activate_task(rq
, p
, 0);
6982 resched_task(rq
->curr
);
6986 void normalize_rt_tasks(void)
6988 struct task_struct
*g
, *p
;
6989 unsigned long flags
;
6992 read_lock_irq(&tasklist_lock
);
6993 do_each_thread(g
, p
) {
6995 * Only normalize user tasks:
7000 p
->se
.exec_start
= 0;
7001 #ifdef CONFIG_SCHEDSTATS
7002 p
->se
.wait_start
= 0;
7003 p
->se
.sleep_start
= 0;
7004 p
->se
.block_start
= 0;
7006 task_rq(p
)->clock
= 0;
7010 * Renice negative nice level userspace
7013 if (TASK_NICE(p
) < 0 && p
->mm
)
7014 set_user_nice(p
, 0);
7018 spin_lock_irqsave(&p
->pi_lock
, flags
);
7019 rq
= __task_rq_lock(p
);
7021 normalize_task(rq
, p
);
7023 __task_rq_unlock(rq
);
7024 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
7025 } while_each_thread(g
, p
);
7027 read_unlock_irq(&tasklist_lock
);
7030 #endif /* CONFIG_MAGIC_SYSRQ */
7034 * These functions are only useful for the IA64 MCA handling.
7036 * They can only be called when the whole system has been
7037 * stopped - every CPU needs to be quiescent, and no scheduling
7038 * activity can take place. Using them for anything else would
7039 * be a serious bug, and as a result, they aren't even visible
7040 * under any other configuration.
7044 * curr_task - return the current task for a given cpu.
7045 * @cpu: the processor in question.
7047 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7049 struct task_struct
*curr_task(int cpu
)
7051 return cpu_curr(cpu
);
7055 * set_curr_task - set the current task for a given cpu.
7056 * @cpu: the processor in question.
7057 * @p: the task pointer to set.
7059 * Description: This function must only be used when non-maskable interrupts
7060 * are serviced on a separate stack. It allows the architecture to switch the
7061 * notion of the current task on a cpu in a non-blocking manner. This function
7062 * must be called with all CPU's synchronized, and interrupts disabled, the
7063 * and caller must save the original value of the current task (see
7064 * curr_task() above) and restore that value before reenabling interrupts and
7065 * re-starting the system.
7067 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7069 void set_curr_task(int cpu
, struct task_struct
*p
)
7076 #ifdef CONFIG_FAIR_GROUP_SCHED
7080 * distribute shares of all task groups among their schedulable entities,
7081 * to reflect load distrbution across cpus.
7083 static int rebalance_shares(struct sched_domain
*sd
, int this_cpu
)
7085 struct cfs_rq
*cfs_rq
;
7086 struct rq
*rq
= cpu_rq(this_cpu
);
7087 cpumask_t sdspan
= sd
->span
;
7090 /* Walk thr' all the task groups that we have */
7091 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
7093 unsigned long total_load
= 0, total_shares
;
7094 struct task_group
*tg
= cfs_rq
->tg
;
7096 /* Gather total task load of this group across cpus */
7097 for_each_cpu_mask(i
, sdspan
)
7098 total_load
+= tg
->cfs_rq
[i
]->load
.weight
;
7100 /* Nothing to do if this group has no load */
7105 * tg->shares represents the number of cpu shares the task group
7106 * is eligible to hold on a single cpu. On N cpus, it is
7107 * eligible to hold (N * tg->shares) number of cpu shares.
7109 total_shares
= tg
->shares
* cpus_weight(sdspan
);
7112 * redistribute total_shares across cpus as per the task load
7115 for_each_cpu_mask(i
, sdspan
) {
7116 unsigned long local_load
, local_shares
;
7118 local_load
= tg
->cfs_rq
[i
]->load
.weight
;
7119 local_shares
= (local_load
* total_shares
) / total_load
;
7121 local_shares
= MIN_GROUP_SHARES
;
7122 if (local_shares
== tg
->se
[i
]->load
.weight
)
7125 spin_lock_irq(&cpu_rq(i
)->lock
);
7126 set_se_shares(tg
->se
[i
], local_shares
);
7127 spin_unlock_irq(&cpu_rq(i
)->lock
);
7136 * How frequently should we rebalance_shares() across cpus?
7138 * The more frequently we rebalance shares, the more accurate is the fairness
7139 * of cpu bandwidth distribution between task groups. However higher frequency
7140 * also implies increased scheduling overhead.
7142 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7143 * consecutive calls to rebalance_shares() in the same sched domain.
7145 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7146 * consecutive calls to rebalance_shares() in the same sched domain.
7148 * These settings allows for the appropriate tradeoff between accuracy of
7149 * fairness and the associated overhead.
7153 /* default: 8ms, units: milliseconds */
7154 const_debug
unsigned int sysctl_sched_min_bal_int_shares
= 8;
7156 /* default: 128ms, units: milliseconds */
7157 const_debug
unsigned int sysctl_sched_max_bal_int_shares
= 128;
7159 /* kernel thread that runs rebalance_shares() periodically */
7160 static int load_balance_monitor(void *unused
)
7162 unsigned int timeout
= sysctl_sched_min_bal_int_shares
;
7163 struct sched_param schedparm
;
7167 * We don't want this thread's execution to be limited by the shares
7168 * assigned to default group (init_task_group). Hence make it run
7169 * as a SCHED_RR RT task at the lowest priority.
7171 schedparm
.sched_priority
= 1;
7172 ret
= sched_setscheduler(current
, SCHED_RR
, &schedparm
);
7174 printk(KERN_ERR
"Couldn't set SCHED_RR policy for load balance"
7175 " monitor thread (error = %d) \n", ret
);
7177 while (!kthread_should_stop()) {
7178 int i
, cpu
, balanced
= 1;
7180 /* Prevent cpus going down or coming up */
7182 /* lockout changes to doms_cur[] array */
7185 * Enter a rcu read-side critical section to safely walk rq->sd
7186 * chain on various cpus and to walk task group list
7187 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7191 for (i
= 0; i
< ndoms_cur
; i
++) {
7192 cpumask_t cpumap
= doms_cur
[i
];
7193 struct sched_domain
*sd
= NULL
, *sd_prev
= NULL
;
7195 cpu
= first_cpu(cpumap
);
7197 /* Find the highest domain at which to balance shares */
7198 for_each_domain(cpu
, sd
) {
7199 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7205 /* sd == NULL? No load balance reqd in this domain */
7209 balanced
&= rebalance_shares(sd
, cpu
);
7218 timeout
= sysctl_sched_min_bal_int_shares
;
7219 else if (timeout
< sysctl_sched_max_bal_int_shares
)
7222 msleep_interruptible(timeout
);
7227 #endif /* CONFIG_SMP */
7229 /* allocate runqueue etc for a new task group */
7230 struct task_group
*sched_create_group(void)
7232 struct task_group
*tg
;
7233 struct cfs_rq
*cfs_rq
;
7234 struct sched_entity
*se
;
7238 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7240 return ERR_PTR(-ENOMEM
);
7242 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
7245 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
7249 for_each_possible_cpu(i
) {
7252 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
7257 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
7262 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
7263 memset(se
, 0, sizeof(struct sched_entity
));
7265 tg
->cfs_rq
[i
] = cfs_rq
;
7266 init_cfs_rq(cfs_rq
, rq
);
7270 se
->cfs_rq
= &rq
->cfs
;
7272 se
->load
.weight
= NICE_0_LOAD
;
7273 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
7277 tg
->shares
= NICE_0_LOAD
;
7279 lock_task_group_list();
7280 for_each_possible_cpu(i
) {
7282 cfs_rq
= tg
->cfs_rq
[i
];
7283 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7285 unlock_task_group_list();
7290 for_each_possible_cpu(i
) {
7292 kfree(tg
->cfs_rq
[i
]);
7300 return ERR_PTR(-ENOMEM
);
7303 /* rcu callback to free various structures associated with a task group */
7304 static void free_sched_group(struct rcu_head
*rhp
)
7306 struct task_group
*tg
= container_of(rhp
, struct task_group
, rcu
);
7307 struct cfs_rq
*cfs_rq
;
7308 struct sched_entity
*se
;
7311 /* now it should be safe to free those cfs_rqs */
7312 for_each_possible_cpu(i
) {
7313 cfs_rq
= tg
->cfs_rq
[i
];
7325 /* Destroy runqueue etc associated with a task group */
7326 void sched_destroy_group(struct task_group
*tg
)
7328 struct cfs_rq
*cfs_rq
= NULL
;
7331 lock_task_group_list();
7332 for_each_possible_cpu(i
) {
7333 cfs_rq
= tg
->cfs_rq
[i
];
7334 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7336 unlock_task_group_list();
7340 /* wait for possible concurrent references to cfs_rqs complete */
7341 call_rcu(&tg
->rcu
, free_sched_group
);
7344 /* change task's runqueue when it moves between groups.
7345 * The caller of this function should have put the task in its new group
7346 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7347 * reflect its new group.
7349 void sched_move_task(struct task_struct
*tsk
)
7352 unsigned long flags
;
7355 rq
= task_rq_lock(tsk
, &flags
);
7357 if (tsk
->sched_class
!= &fair_sched_class
) {
7358 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7362 update_rq_clock(rq
);
7364 running
= task_current(rq
, tsk
);
7365 on_rq
= tsk
->se
.on_rq
;
7368 dequeue_task(rq
, tsk
, 0);
7369 if (unlikely(running
))
7370 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7373 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7376 if (unlikely(running
))
7377 tsk
->sched_class
->set_curr_task(rq
);
7378 enqueue_task(rq
, tsk
, 0);
7382 task_rq_unlock(rq
, &flags
);
7385 /* rq->lock to be locked by caller */
7386 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7388 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7389 struct rq
*rq
= cfs_rq
->rq
;
7393 shares
= MIN_GROUP_SHARES
;
7397 dequeue_entity(cfs_rq
, se
, 0);
7398 dec_cpu_load(rq
, se
->load
.weight
);
7401 se
->load
.weight
= shares
;
7402 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7405 enqueue_entity(cfs_rq
, se
, 0);
7406 inc_cpu_load(rq
, se
->load
.weight
);
7410 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7413 struct cfs_rq
*cfs_rq
;
7416 lock_task_group_list();
7417 if (tg
->shares
== shares
)
7420 if (shares
< MIN_GROUP_SHARES
)
7421 shares
= MIN_GROUP_SHARES
;
7424 * Prevent any load balance activity (rebalance_shares,
7425 * load_balance_fair) from referring to this group first,
7426 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7428 for_each_possible_cpu(i
) {
7429 cfs_rq
= tg
->cfs_rq
[i
];
7430 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7433 /* wait for any ongoing reference to this group to finish */
7434 synchronize_sched();
7437 * Now we are free to modify the group's share on each cpu
7438 * w/o tripping rebalance_share or load_balance_fair.
7440 tg
->shares
= shares
;
7441 for_each_possible_cpu(i
) {
7442 spin_lock_irq(&cpu_rq(i
)->lock
);
7443 set_se_shares(tg
->se
[i
], shares
);
7444 spin_unlock_irq(&cpu_rq(i
)->lock
);
7448 * Enable load balance activity on this group, by inserting it back on
7449 * each cpu's rq->leaf_cfs_rq_list.
7451 for_each_possible_cpu(i
) {
7453 cfs_rq
= tg
->cfs_rq
[i
];
7454 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7457 unlock_task_group_list();
7461 unsigned long sched_group_shares(struct task_group
*tg
)
7466 #endif /* CONFIG_FAIR_GROUP_SCHED */
7468 #ifdef CONFIG_FAIR_CGROUP_SCHED
7470 /* return corresponding task_group object of a cgroup */
7471 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7473 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7474 struct task_group
, css
);
7477 static struct cgroup_subsys_state
*
7478 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7480 struct task_group
*tg
;
7482 if (!cgrp
->parent
) {
7483 /* This is early initialization for the top cgroup */
7484 init_task_group
.css
.cgroup
= cgrp
;
7485 return &init_task_group
.css
;
7488 /* we support only 1-level deep hierarchical scheduler atm */
7489 if (cgrp
->parent
->parent
)
7490 return ERR_PTR(-EINVAL
);
7492 tg
= sched_create_group();
7494 return ERR_PTR(-ENOMEM
);
7496 /* Bind the cgroup to task_group object we just created */
7497 tg
->css
.cgroup
= cgrp
;
7503 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7505 struct task_group
*tg
= cgroup_tg(cgrp
);
7507 sched_destroy_group(tg
);
7511 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7512 struct task_struct
*tsk
)
7514 /* We don't support RT-tasks being in separate groups */
7515 if (tsk
->sched_class
!= &fair_sched_class
)
7522 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7523 struct cgroup
*old_cont
, struct task_struct
*tsk
)
7525 sched_move_task(tsk
);
7528 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7531 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7534 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7536 struct task_group
*tg
= cgroup_tg(cgrp
);
7538 return (u64
) tg
->shares
;
7541 static struct cftype cpu_files
[] = {
7544 .read_uint
= cpu_shares_read_uint
,
7545 .write_uint
= cpu_shares_write_uint
,
7549 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7551 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7554 struct cgroup_subsys cpu_cgroup_subsys
= {
7556 .create
= cpu_cgroup_create
,
7557 .destroy
= cpu_cgroup_destroy
,
7558 .can_attach
= cpu_cgroup_can_attach
,
7559 .attach
= cpu_cgroup_attach
,
7560 .populate
= cpu_cgroup_populate
,
7561 .subsys_id
= cpu_cgroup_subsys_id
,
7565 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7567 #ifdef CONFIG_CGROUP_CPUACCT
7570 * CPU accounting code for task groups.
7572 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7573 * (balbir@in.ibm.com).
7576 /* track cpu usage of a group of tasks */
7578 struct cgroup_subsys_state css
;
7579 /* cpuusage holds pointer to a u64-type object on every cpu */
7583 struct cgroup_subsys cpuacct_subsys
;
7585 /* return cpu accounting group corresponding to this container */
7586 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
7588 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
7589 struct cpuacct
, css
);
7592 /* return cpu accounting group to which this task belongs */
7593 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
7595 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
7596 struct cpuacct
, css
);
7599 /* create a new cpu accounting group */
7600 static struct cgroup_subsys_state
*cpuacct_create(
7601 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7603 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7606 return ERR_PTR(-ENOMEM
);
7608 ca
->cpuusage
= alloc_percpu(u64
);
7609 if (!ca
->cpuusage
) {
7611 return ERR_PTR(-ENOMEM
);
7617 /* destroy an existing cpu accounting group */
7619 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7621 struct cpuacct
*ca
= cgroup_ca(cont
);
7623 free_percpu(ca
->cpuusage
);
7627 /* return total cpu usage (in nanoseconds) of a group */
7628 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
7630 struct cpuacct
*ca
= cgroup_ca(cont
);
7631 u64 totalcpuusage
= 0;
7634 for_each_possible_cpu(i
) {
7635 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
7638 * Take rq->lock to make 64-bit addition safe on 32-bit
7641 spin_lock_irq(&cpu_rq(i
)->lock
);
7642 totalcpuusage
+= *cpuusage
;
7643 spin_unlock_irq(&cpu_rq(i
)->lock
);
7646 return totalcpuusage
;
7649 static struct cftype files
[] = {
7652 .read_uint
= cpuusage_read
,
7656 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7658 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
7662 * charge this task's execution time to its accounting group.
7664 * called with rq->lock held.
7666 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
7670 if (!cpuacct_subsys
.active
)
7675 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
7677 *cpuusage
+= cputime
;
7681 struct cgroup_subsys cpuacct_subsys
= {
7683 .create
= cpuacct_create
,
7684 .destroy
= cpuacct_destroy
,
7685 .populate
= cpuacct_populate
,
7686 .subsys_id
= cpuacct_subsys_id
,
7688 #endif /* CONFIG_CGROUP_CPUACCT */