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
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
70 #include <asm/irq_regs.h>
73 * Scheduler clock - returns current time in nanosec units.
74 * This is default implementation.
75 * Architectures and sub-architectures can override this.
77 unsigned long long __attribute__((weak
)) sched_clock(void)
79 return (unsigned long long)jiffies
* (NSEC_PER_SEC
/ HZ
);
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
119 * Since cpu_power is a 'constant', we can use a reciprocal divide.
121 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
123 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
127 * Each time a sched group cpu_power is changed,
128 * we must compute its reciprocal value
130 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
132 sg
->__cpu_power
+= val
;
133 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
137 static inline int rt_policy(int policy
)
139 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
144 static inline int task_has_rt_policy(struct task_struct
*p
)
146 return rt_policy(p
->policy
);
150 * This is the priority-queue data structure of the RT scheduling class:
152 struct rt_prio_array
{
153 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
154 struct list_head queue
[MAX_RT_PRIO
];
157 #ifdef CONFIG_FAIR_GROUP_SCHED
159 #include <linux/cgroup.h>
163 /* task group related information */
165 #ifdef CONFIG_FAIR_CGROUP_SCHED
166 struct cgroup_subsys_state css
;
168 /* schedulable entities of this group on each cpu */
169 struct sched_entity
**se
;
170 /* runqueue "owned" by this group on each cpu */
171 struct cfs_rq
**cfs_rq
;
174 * shares assigned to a task group governs how much of cpu bandwidth
175 * is allocated to the group. The more shares a group has, the more is
176 * the cpu bandwidth allocated to it.
178 * For ex, lets say that there are three task groups, A, B and C which
179 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
180 * cpu bandwidth allocated by the scheduler to task groups A, B and C
183 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
184 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
185 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
187 * The weight assigned to a task group's schedulable entities on every
188 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
189 * group's shares. For ex: lets say that task group A has been
190 * assigned shares of 1000 and there are two CPUs in a system. Then,
192 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
194 * Note: It's not necessary that each of a task's group schedulable
195 * entity have the same weight on all CPUs. If the group
196 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
197 * better distribution of weight could be:
199 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
200 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
202 * rebalance_shares() is responsible for distributing the shares of a
203 * task groups like this among the group's schedulable entities across
207 unsigned long shares
;
212 /* Default task group's sched entity on each cpu */
213 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
214 /* Default task group's cfs_rq on each cpu */
215 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
217 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
218 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
220 /* task_group_mutex serializes add/remove of task groups and also changes to
221 * a task group's cpu shares.
223 static DEFINE_MUTEX(task_group_mutex
);
225 /* doms_cur_mutex serializes access to doms_cur[] array */
226 static DEFINE_MUTEX(doms_cur_mutex
);
229 /* kernel thread that runs rebalance_shares() periodically */
230 static struct task_struct
*lb_monitor_task
;
231 static int load_balance_monitor(void *unused
);
234 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
);
236 /* Default task group.
237 * Every task in system belong to this group at bootup.
239 struct task_group init_task_group
= {
240 .se
= init_sched_entity_p
,
241 .cfs_rq
= init_cfs_rq_p
,
244 #ifdef CONFIG_FAIR_USER_SCHED
245 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
247 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
250 #define MIN_GROUP_SHARES 2
252 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
254 /* return group to which a task belongs */
255 static inline struct task_group
*task_group(struct task_struct
*p
)
257 struct task_group
*tg
;
259 #ifdef CONFIG_FAIR_USER_SCHED
261 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
262 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
263 struct task_group
, css
);
265 tg
= &init_task_group
;
270 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
271 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
)
273 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
274 p
->se
.parent
= task_group(p
)->se
[cpu
];
277 static inline void lock_task_group_list(void)
279 mutex_lock(&task_group_mutex
);
282 static inline void unlock_task_group_list(void)
284 mutex_unlock(&task_group_mutex
);
287 static inline void lock_doms_cur(void)
289 mutex_lock(&doms_cur_mutex
);
292 static inline void unlock_doms_cur(void)
294 mutex_unlock(&doms_cur_mutex
);
299 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
) { }
300 static inline void lock_task_group_list(void) { }
301 static inline void unlock_task_group_list(void) { }
302 static inline void lock_doms_cur(void) { }
303 static inline void unlock_doms_cur(void) { }
305 #endif /* CONFIG_FAIR_GROUP_SCHED */
307 /* CFS-related fields in a runqueue */
309 struct load_weight load
;
310 unsigned long nr_running
;
315 struct rb_root tasks_timeline
;
316 struct rb_node
*rb_leftmost
;
317 struct rb_node
*rb_load_balance_curr
;
318 /* 'curr' points to currently running entity on this cfs_rq.
319 * It is set to NULL otherwise (i.e when none are currently running).
321 struct sched_entity
*curr
;
323 unsigned long nr_spread_over
;
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
329 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
330 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
331 * (like users, containers etc.)
333 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
334 * list is used during load balance.
336 struct list_head leaf_cfs_rq_list
;
337 struct task_group
*tg
; /* group that "owns" this runqueue */
341 /* Real-Time classes' related field in a runqueue: */
343 struct rt_prio_array active
;
344 int rt_load_balance_idx
;
345 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
346 unsigned long rt_nr_running
;
347 unsigned long rt_nr_migratory
;
348 /* highest queued rt task prio */
356 * We add the notion of a root-domain which will be used to define per-domain
357 * variables. Each exclusive cpuset essentially defines an island domain by
358 * fully partitioning the member cpus from any other cpuset. Whenever a new
359 * exclusive cpuset is created, we also create and attach a new root-domain
362 * By default the system creates a single root-domain with all cpus as
363 * members (mimicking the global state we have today).
371 * The "RT overload" flag: it gets set if a CPU has more than
372 * one runnable RT task.
378 static struct root_domain def_root_domain
;
383 * This is the main, per-CPU runqueue data structure.
385 * Locking rule: those places that want to lock multiple runqueues
386 * (such as the load balancing or the thread migration code), lock
387 * acquire operations must be ordered by ascending &runqueue.
394 * nr_running and cpu_load should be in the same cacheline because
395 * remote CPUs use both these fields when doing load calculation.
397 unsigned long nr_running
;
398 #define CPU_LOAD_IDX_MAX 5
399 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
400 unsigned char idle_at_tick
;
402 unsigned char in_nohz_recently
;
404 /* capture load from *all* tasks on this cpu: */
405 struct load_weight load
;
406 unsigned long nr_load_updates
;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 /* list of leaf cfs_rq on this cpu: */
412 struct list_head leaf_cfs_rq_list
;
417 * This is part of a global counter where only the total sum
418 * over all CPUs matters. A task can increase this counter on
419 * one CPU and if it got migrated afterwards it may decrease
420 * it on another CPU. Always updated under the runqueue lock:
422 unsigned long nr_uninterruptible
;
424 struct task_struct
*curr
, *idle
;
425 unsigned long next_balance
;
426 struct mm_struct
*prev_mm
;
428 u64 clock
, prev_clock_raw
;
431 unsigned int clock_warps
, clock_overflows
;
433 unsigned int clock_deep_idle_events
;
439 struct root_domain
*rd
;
440 struct sched_domain
*sd
;
442 /* For active balancing */
445 /* cpu of this runqueue: */
448 struct task_struct
*migration_thread
;
449 struct list_head migration_queue
;
452 #ifdef CONFIG_SCHEDSTATS
454 struct sched_info rq_sched_info
;
456 /* sys_sched_yield() stats */
457 unsigned int yld_exp_empty
;
458 unsigned int yld_act_empty
;
459 unsigned int yld_both_empty
;
460 unsigned int yld_count
;
462 /* schedule() stats */
463 unsigned int sched_switch
;
464 unsigned int sched_count
;
465 unsigned int sched_goidle
;
467 /* try_to_wake_up() stats */
468 unsigned int ttwu_count
;
469 unsigned int ttwu_local
;
472 unsigned int bkl_count
;
474 struct lock_class_key rq_lock_key
;
477 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
479 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
481 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
484 static inline int cpu_of(struct rq
*rq
)
494 * Update the per-runqueue clock, as finegrained as the platform can give
495 * us, but without assuming monotonicity, etc.:
497 static void __update_rq_clock(struct rq
*rq
)
499 u64 prev_raw
= rq
->prev_clock_raw
;
500 u64 now
= sched_clock();
501 s64 delta
= now
- prev_raw
;
502 u64 clock
= rq
->clock
;
504 #ifdef CONFIG_SCHED_DEBUG
505 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
508 * Protect against sched_clock() occasionally going backwards:
510 if (unlikely(delta
< 0)) {
515 * Catch too large forward jumps too:
517 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
518 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
519 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
522 rq
->clock_overflows
++;
524 if (unlikely(delta
> rq
->clock_max_delta
))
525 rq
->clock_max_delta
= delta
;
530 rq
->prev_clock_raw
= now
;
534 static void update_rq_clock(struct rq
*rq
)
536 if (likely(smp_processor_id() == cpu_of(rq
)))
537 __update_rq_clock(rq
);
541 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
542 * See detach_destroy_domains: synchronize_sched for details.
544 * The domain tree of any CPU may only be accessed from within
545 * preempt-disabled sections.
547 #define for_each_domain(cpu, __sd) \
548 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
550 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
551 #define this_rq() (&__get_cpu_var(runqueues))
552 #define task_rq(p) cpu_rq(task_cpu(p))
553 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
556 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
558 #ifdef CONFIG_SCHED_DEBUG
559 # define const_debug __read_mostly
561 # define const_debug static const
565 * Debugging: various feature bits
568 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
569 SCHED_FEAT_WAKEUP_PREEMPT
= 2,
570 SCHED_FEAT_START_DEBIT
= 4,
571 SCHED_FEAT_TREE_AVG
= 8,
572 SCHED_FEAT_APPROX_AVG
= 16,
575 const_debug
unsigned int sysctl_sched_features
=
576 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
577 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
578 SCHED_FEAT_START_DEBIT
* 1 |
579 SCHED_FEAT_TREE_AVG
* 0 |
580 SCHED_FEAT_APPROX_AVG
* 0;
582 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
585 * Number of tasks to iterate in a single balance run.
586 * Limited because this is done with IRQs disabled.
588 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
591 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
592 * clock constructed from sched_clock():
594 unsigned long long cpu_clock(int cpu
)
596 unsigned long long now
;
600 local_irq_save(flags
);
603 * Only call sched_clock() if the scheduler has already been
604 * initialized (some code might call cpu_clock() very early):
609 local_irq_restore(flags
);
613 EXPORT_SYMBOL_GPL(cpu_clock
);
615 #ifndef prepare_arch_switch
616 # define prepare_arch_switch(next) do { } while (0)
618 #ifndef finish_arch_switch
619 # define finish_arch_switch(prev) do { } while (0)
622 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
624 return rq
->curr
== p
;
627 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
628 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
630 return task_current(rq
, p
);
633 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
637 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
639 #ifdef CONFIG_DEBUG_SPINLOCK
640 /* this is a valid case when another task releases the spinlock */
641 rq
->lock
.owner
= current
;
644 * If we are tracking spinlock dependencies then we have to
645 * fix up the runqueue lock - which gets 'carried over' from
648 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
650 spin_unlock_irq(&rq
->lock
);
653 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
654 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
659 return task_current(rq
, p
);
663 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
667 * We can optimise this out completely for !SMP, because the
668 * SMP rebalancing from interrupt is the only thing that cares
673 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
674 spin_unlock_irq(&rq
->lock
);
676 spin_unlock(&rq
->lock
);
680 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
684 * After ->oncpu is cleared, the task can be moved to a different CPU.
685 * We must ensure this doesn't happen until the switch is completely
691 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
695 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
698 * __task_rq_lock - lock the runqueue a given task resides on.
699 * Must be called interrupts disabled.
701 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
705 struct rq
*rq
= task_rq(p
);
706 spin_lock(&rq
->lock
);
707 if (likely(rq
== task_rq(p
)))
709 spin_unlock(&rq
->lock
);
714 * task_rq_lock - lock the runqueue a given task resides on and disable
715 * interrupts. Note the ordering: we can safely lookup the task_rq without
716 * explicitly disabling preemption.
718 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
724 local_irq_save(*flags
);
726 spin_lock(&rq
->lock
);
727 if (likely(rq
== task_rq(p
)))
729 spin_unlock_irqrestore(&rq
->lock
, *flags
);
733 static void __task_rq_unlock(struct rq
*rq
)
736 spin_unlock(&rq
->lock
);
739 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
742 spin_unlock_irqrestore(&rq
->lock
, *flags
);
746 * this_rq_lock - lock this runqueue and disable interrupts.
748 static struct rq
*this_rq_lock(void)
755 spin_lock(&rq
->lock
);
761 * We are going deep-idle (irqs are disabled):
763 void sched_clock_idle_sleep_event(void)
765 struct rq
*rq
= cpu_rq(smp_processor_id());
767 spin_lock(&rq
->lock
);
768 __update_rq_clock(rq
);
769 spin_unlock(&rq
->lock
);
770 rq
->clock_deep_idle_events
++;
772 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
775 * We just idled delta nanoseconds (called with irqs disabled):
777 void sched_clock_idle_wakeup_event(u64 delta_ns
)
779 struct rq
*rq
= cpu_rq(smp_processor_id());
780 u64 now
= sched_clock();
782 touch_softlockup_watchdog();
783 rq
->idle_clock
+= delta_ns
;
785 * Override the previous timestamp and ignore all
786 * sched_clock() deltas that occured while we idled,
787 * and use the PM-provided delta_ns to advance the
790 spin_lock(&rq
->lock
);
791 rq
->prev_clock_raw
= now
;
792 rq
->clock
+= delta_ns
;
793 spin_unlock(&rq
->lock
);
795 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
798 * resched_task - mark a task 'to be rescheduled now'.
800 * On UP this means the setting of the need_resched flag, on SMP it
801 * might also involve a cross-CPU call to trigger the scheduler on
806 #ifndef tsk_is_polling
807 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
810 static void resched_task(struct task_struct
*p
)
814 assert_spin_locked(&task_rq(p
)->lock
);
816 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
819 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
822 if (cpu
== smp_processor_id())
825 /* NEED_RESCHED must be visible before we test polling */
827 if (!tsk_is_polling(p
))
828 smp_send_reschedule(cpu
);
831 static void resched_cpu(int cpu
)
833 struct rq
*rq
= cpu_rq(cpu
);
836 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
838 resched_task(cpu_curr(cpu
));
839 spin_unlock_irqrestore(&rq
->lock
, flags
);
842 static inline void resched_task(struct task_struct
*p
)
844 assert_spin_locked(&task_rq(p
)->lock
);
845 set_tsk_need_resched(p
);
849 #if BITS_PER_LONG == 32
850 # define WMULT_CONST (~0UL)
852 # define WMULT_CONST (1UL << 32)
855 #define WMULT_SHIFT 32
858 * Shift right and round:
860 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
863 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
864 struct load_weight
*lw
)
868 if (unlikely(!lw
->inv_weight
))
869 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
871 tmp
= (u64
)delta_exec
* weight
;
873 * Check whether we'd overflow the 64-bit multiplication:
875 if (unlikely(tmp
> WMULT_CONST
))
876 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
879 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
881 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
884 static inline unsigned long
885 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
887 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
890 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
895 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
901 * To aid in avoiding the subversion of "niceness" due to uneven distribution
902 * of tasks with abnormal "nice" values across CPUs the contribution that
903 * each task makes to its run queue's load is weighted according to its
904 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
905 * scaled version of the new time slice allocation that they receive on time
909 #define WEIGHT_IDLEPRIO 2
910 #define WMULT_IDLEPRIO (1 << 31)
913 * Nice levels are multiplicative, with a gentle 10% change for every
914 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
915 * nice 1, it will get ~10% less CPU time than another CPU-bound task
916 * that remained on nice 0.
918 * The "10% effect" is relative and cumulative: from _any_ nice level,
919 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
920 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
921 * If a task goes up by ~10% and another task goes down by ~10% then
922 * the relative distance between them is ~25%.)
924 static const int prio_to_weight
[40] = {
925 /* -20 */ 88761, 71755, 56483, 46273, 36291,
926 /* -15 */ 29154, 23254, 18705, 14949, 11916,
927 /* -10 */ 9548, 7620, 6100, 4904, 3906,
928 /* -5 */ 3121, 2501, 1991, 1586, 1277,
929 /* 0 */ 1024, 820, 655, 526, 423,
930 /* 5 */ 335, 272, 215, 172, 137,
931 /* 10 */ 110, 87, 70, 56, 45,
932 /* 15 */ 36, 29, 23, 18, 15,
936 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
938 * In cases where the weight does not change often, we can use the
939 * precalculated inverse to speed up arithmetics by turning divisions
940 * into multiplications:
942 static const u32 prio_to_wmult
[40] = {
943 /* -20 */ 48388, 59856, 76040, 92818, 118348,
944 /* -15 */ 147320, 184698, 229616, 287308, 360437,
945 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
946 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
947 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
948 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
949 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
950 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
953 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
956 * runqueue iterator, to support SMP load-balancing between different
957 * scheduling classes, without having to expose their internal data
958 * structures to the load-balancing proper:
962 struct task_struct
*(*start
)(void *);
963 struct task_struct
*(*next
)(void *);
968 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
969 unsigned long max_load_move
, struct sched_domain
*sd
,
970 enum cpu_idle_type idle
, int *all_pinned
,
971 int *this_best_prio
, struct rq_iterator
*iterator
);
974 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
975 struct sched_domain
*sd
, enum cpu_idle_type idle
,
976 struct rq_iterator
*iterator
);
979 #ifdef CONFIG_CGROUP_CPUACCT
980 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
982 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
985 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
987 update_load_add(&rq
->load
, load
);
990 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
992 update_load_sub(&rq
->load
, load
);
996 static unsigned long source_load(int cpu
, int type
);
997 static unsigned long target_load(int cpu
, int type
);
998 static unsigned long cpu_avg_load_per_task(int cpu
);
999 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1000 #endif /* CONFIG_SMP */
1002 #include "sched_stats.h"
1003 #include "sched_idletask.c"
1004 #include "sched_fair.c"
1005 #include "sched_rt.c"
1006 #ifdef CONFIG_SCHED_DEBUG
1007 # include "sched_debug.c"
1010 #define sched_class_highest (&rt_sched_class)
1012 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
1017 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
1022 static void set_load_weight(struct task_struct
*p
)
1024 if (task_has_rt_policy(p
)) {
1025 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1026 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1031 * SCHED_IDLE tasks get minimal weight:
1033 if (p
->policy
== SCHED_IDLE
) {
1034 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1035 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1039 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1040 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1043 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1045 sched_info_queued(p
);
1046 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1050 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1052 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1057 * __normal_prio - return the priority that is based on the static prio
1059 static inline int __normal_prio(struct task_struct
*p
)
1061 return p
->static_prio
;
1065 * Calculate the expected normal priority: i.e. priority
1066 * without taking RT-inheritance into account. Might be
1067 * boosted by interactivity modifiers. Changes upon fork,
1068 * setprio syscalls, and whenever the interactivity
1069 * estimator recalculates.
1071 static inline int normal_prio(struct task_struct
*p
)
1075 if (task_has_rt_policy(p
))
1076 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1078 prio
= __normal_prio(p
);
1083 * Calculate the current priority, i.e. the priority
1084 * taken into account by the scheduler. This value might
1085 * be boosted by RT tasks, or might be boosted by
1086 * interactivity modifiers. Will be RT if the task got
1087 * RT-boosted. If not then it returns p->normal_prio.
1089 static int effective_prio(struct task_struct
*p
)
1091 p
->normal_prio
= normal_prio(p
);
1093 * If we are RT tasks or we were boosted to RT priority,
1094 * keep the priority unchanged. Otherwise, update priority
1095 * to the normal priority:
1097 if (!rt_prio(p
->prio
))
1098 return p
->normal_prio
;
1103 * activate_task - move a task to the runqueue.
1105 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1107 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1108 rq
->nr_uninterruptible
--;
1110 enqueue_task(rq
, p
, wakeup
);
1111 inc_nr_running(p
, rq
);
1115 * deactivate_task - remove a task from the runqueue.
1117 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1119 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1120 rq
->nr_uninterruptible
++;
1122 dequeue_task(rq
, p
, sleep
);
1123 dec_nr_running(p
, rq
);
1127 * task_curr - is this task currently executing on a CPU?
1128 * @p: the task in question.
1130 inline int task_curr(const struct task_struct
*p
)
1132 return cpu_curr(task_cpu(p
)) == p
;
1135 /* Used instead of source_load when we know the type == 0 */
1136 unsigned long weighted_cpuload(const int cpu
)
1138 return cpu_rq(cpu
)->load
.weight
;
1141 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1143 set_task_cfs_rq(p
, cpu
);
1146 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1147 * successfuly executed on another CPU. We must ensure that updates of
1148 * per-task data have been completed by this moment.
1151 task_thread_info(p
)->cpu
= cpu
;
1158 * Is this task likely cache-hot:
1161 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1165 if (p
->sched_class
!= &fair_sched_class
)
1168 if (sysctl_sched_migration_cost
== -1)
1170 if (sysctl_sched_migration_cost
== 0)
1173 delta
= now
- p
->se
.exec_start
;
1175 return delta
< (s64
)sysctl_sched_migration_cost
;
1179 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1181 int old_cpu
= task_cpu(p
);
1182 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1183 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1184 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1187 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1189 #ifdef CONFIG_SCHEDSTATS
1190 if (p
->se
.wait_start
)
1191 p
->se
.wait_start
-= clock_offset
;
1192 if (p
->se
.sleep_start
)
1193 p
->se
.sleep_start
-= clock_offset
;
1194 if (p
->se
.block_start
)
1195 p
->se
.block_start
-= clock_offset
;
1196 if (old_cpu
!= new_cpu
) {
1197 schedstat_inc(p
, se
.nr_migrations
);
1198 if (task_hot(p
, old_rq
->clock
, NULL
))
1199 schedstat_inc(p
, se
.nr_forced2_migrations
);
1202 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1203 new_cfsrq
->min_vruntime
;
1205 __set_task_cpu(p
, new_cpu
);
1208 struct migration_req
{
1209 struct list_head list
;
1211 struct task_struct
*task
;
1214 struct completion done
;
1218 * The task's runqueue lock must be held.
1219 * Returns true if you have to wait for migration thread.
1222 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1224 struct rq
*rq
= task_rq(p
);
1227 * If the task is not on a runqueue (and not running), then
1228 * it is sufficient to simply update the task's cpu field.
1230 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1231 set_task_cpu(p
, dest_cpu
);
1235 init_completion(&req
->done
);
1237 req
->dest_cpu
= dest_cpu
;
1238 list_add(&req
->list
, &rq
->migration_queue
);
1244 * wait_task_inactive - wait for a thread to unschedule.
1246 * The caller must ensure that the task *will* unschedule sometime soon,
1247 * else this function might spin for a *long* time. This function can't
1248 * be called with interrupts off, or it may introduce deadlock with
1249 * smp_call_function() if an IPI is sent by the same process we are
1250 * waiting to become inactive.
1252 void wait_task_inactive(struct task_struct
*p
)
1254 unsigned long flags
;
1260 * We do the initial early heuristics without holding
1261 * any task-queue locks at all. We'll only try to get
1262 * the runqueue lock when things look like they will
1268 * If the task is actively running on another CPU
1269 * still, just relax and busy-wait without holding
1272 * NOTE! Since we don't hold any locks, it's not
1273 * even sure that "rq" stays as the right runqueue!
1274 * But we don't care, since "task_running()" will
1275 * return false if the runqueue has changed and p
1276 * is actually now running somewhere else!
1278 while (task_running(rq
, p
))
1282 * Ok, time to look more closely! We need the rq
1283 * lock now, to be *sure*. If we're wrong, we'll
1284 * just go back and repeat.
1286 rq
= task_rq_lock(p
, &flags
);
1287 running
= task_running(rq
, p
);
1288 on_rq
= p
->se
.on_rq
;
1289 task_rq_unlock(rq
, &flags
);
1292 * Was it really running after all now that we
1293 * checked with the proper locks actually held?
1295 * Oops. Go back and try again..
1297 if (unlikely(running
)) {
1303 * It's not enough that it's not actively running,
1304 * it must be off the runqueue _entirely_, and not
1307 * So if it wa still runnable (but just not actively
1308 * running right now), it's preempted, and we should
1309 * yield - it could be a while.
1311 if (unlikely(on_rq
)) {
1312 schedule_timeout_uninterruptible(1);
1317 * Ahh, all good. It wasn't running, and it wasn't
1318 * runnable, which means that it will never become
1319 * running in the future either. We're all done!
1326 * kick_process - kick a running thread to enter/exit the kernel
1327 * @p: the to-be-kicked thread
1329 * Cause a process which is running on another CPU to enter
1330 * kernel-mode, without any delay. (to get signals handled.)
1332 * NOTE: this function doesnt have to take the runqueue lock,
1333 * because all it wants to ensure is that the remote task enters
1334 * the kernel. If the IPI races and the task has been migrated
1335 * to another CPU then no harm is done and the purpose has been
1338 void kick_process(struct task_struct
*p
)
1344 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1345 smp_send_reschedule(cpu
);
1350 * Return a low guess at the load of a migration-source cpu weighted
1351 * according to the scheduling class and "nice" value.
1353 * We want to under-estimate the load of migration sources, to
1354 * balance conservatively.
1356 static unsigned long source_load(int cpu
, int type
)
1358 struct rq
*rq
= cpu_rq(cpu
);
1359 unsigned long total
= weighted_cpuload(cpu
);
1364 return min(rq
->cpu_load
[type
-1], total
);
1368 * Return a high guess at the load of a migration-target cpu weighted
1369 * according to the scheduling class and "nice" value.
1371 static unsigned long target_load(int cpu
, int type
)
1373 struct rq
*rq
= cpu_rq(cpu
);
1374 unsigned long total
= weighted_cpuload(cpu
);
1379 return max(rq
->cpu_load
[type
-1], total
);
1383 * Return the average load per task on the cpu's run queue
1385 static unsigned long cpu_avg_load_per_task(int cpu
)
1387 struct rq
*rq
= cpu_rq(cpu
);
1388 unsigned long total
= weighted_cpuload(cpu
);
1389 unsigned long n
= rq
->nr_running
;
1391 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1395 * find_idlest_group finds and returns the least busy CPU group within the
1398 static struct sched_group
*
1399 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1401 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1402 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1403 int load_idx
= sd
->forkexec_idx
;
1404 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1407 unsigned long load
, avg_load
;
1411 /* Skip over this group if it has no CPUs allowed */
1412 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1415 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1417 /* Tally up the load of all CPUs in the group */
1420 for_each_cpu_mask(i
, group
->cpumask
) {
1421 /* Bias balancing toward cpus of our domain */
1423 load
= source_load(i
, load_idx
);
1425 load
= target_load(i
, load_idx
);
1430 /* Adjust by relative CPU power of the group */
1431 avg_load
= sg_div_cpu_power(group
,
1432 avg_load
* SCHED_LOAD_SCALE
);
1435 this_load
= avg_load
;
1437 } else if (avg_load
< min_load
) {
1438 min_load
= avg_load
;
1441 } while (group
= group
->next
, group
!= sd
->groups
);
1443 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1449 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1452 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1455 unsigned long load
, min_load
= ULONG_MAX
;
1459 /* Traverse only the allowed CPUs */
1460 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1462 for_each_cpu_mask(i
, tmp
) {
1463 load
= weighted_cpuload(i
);
1465 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1475 * sched_balance_self: balance the current task (running on cpu) in domains
1476 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1479 * Balance, ie. select the least loaded group.
1481 * Returns the target CPU number, or the same CPU if no balancing is needed.
1483 * preempt must be disabled.
1485 static int sched_balance_self(int cpu
, int flag
)
1487 struct task_struct
*t
= current
;
1488 struct sched_domain
*tmp
, *sd
= NULL
;
1490 for_each_domain(cpu
, tmp
) {
1492 * If power savings logic is enabled for a domain, stop there.
1494 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1496 if (tmp
->flags
& flag
)
1502 struct sched_group
*group
;
1503 int new_cpu
, weight
;
1505 if (!(sd
->flags
& flag
)) {
1511 group
= find_idlest_group(sd
, t
, cpu
);
1517 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1518 if (new_cpu
== -1 || new_cpu
== cpu
) {
1519 /* Now try balancing at a lower domain level of cpu */
1524 /* Now try balancing at a lower domain level of new_cpu */
1527 weight
= cpus_weight(span
);
1528 for_each_domain(cpu
, tmp
) {
1529 if (weight
<= cpus_weight(tmp
->span
))
1531 if (tmp
->flags
& flag
)
1534 /* while loop will break here if sd == NULL */
1540 #endif /* CONFIG_SMP */
1543 * try_to_wake_up - wake up a thread
1544 * @p: the to-be-woken-up thread
1545 * @state: the mask of task states that can be woken
1546 * @sync: do a synchronous wakeup?
1548 * Put it on the run-queue if it's not already there. The "current"
1549 * thread is always on the run-queue (except when the actual
1550 * re-schedule is in progress), and as such you're allowed to do
1551 * the simpler "current->state = TASK_RUNNING" to mark yourself
1552 * runnable without the overhead of this.
1554 * returns failure only if the task is already active.
1556 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1558 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1559 unsigned long flags
;
1566 rq
= task_rq_lock(p
, &flags
);
1567 old_state
= p
->state
;
1568 if (!(old_state
& state
))
1576 this_cpu
= smp_processor_id();
1579 if (unlikely(task_running(rq
, p
)))
1582 new_cpu
= p
->sched_class
->select_task_rq(p
, sync
);
1583 if (new_cpu
!= cpu
) {
1584 set_task_cpu(p
, new_cpu
);
1585 task_rq_unlock(rq
, &flags
);
1586 /* might preempt at this point */
1587 rq
= task_rq_lock(p
, &flags
);
1588 old_state
= p
->state
;
1589 if (!(old_state
& state
))
1594 this_cpu
= smp_processor_id();
1598 #ifdef CONFIG_SCHEDSTATS
1599 schedstat_inc(rq
, ttwu_count
);
1600 if (cpu
== this_cpu
)
1601 schedstat_inc(rq
, ttwu_local
);
1603 struct sched_domain
*sd
;
1604 for_each_domain(this_cpu
, sd
) {
1605 if (cpu_isset(cpu
, sd
->span
)) {
1606 schedstat_inc(sd
, ttwu_wake_remote
);
1616 #endif /* CONFIG_SMP */
1617 schedstat_inc(p
, se
.nr_wakeups
);
1619 schedstat_inc(p
, se
.nr_wakeups_sync
);
1620 if (orig_cpu
!= cpu
)
1621 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1622 if (cpu
== this_cpu
)
1623 schedstat_inc(p
, se
.nr_wakeups_local
);
1625 schedstat_inc(p
, se
.nr_wakeups_remote
);
1626 update_rq_clock(rq
);
1627 activate_task(rq
, p
, 1);
1628 check_preempt_curr(rq
, p
);
1632 p
->state
= TASK_RUNNING
;
1633 wakeup_balance_rt(rq
, p
);
1635 task_rq_unlock(rq
, &flags
);
1640 int fastcall
wake_up_process(struct task_struct
*p
)
1642 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1643 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1645 EXPORT_SYMBOL(wake_up_process
);
1647 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1649 return try_to_wake_up(p
, state
, 0);
1653 * Perform scheduler related setup for a newly forked process p.
1654 * p is forked by current.
1656 * __sched_fork() is basic setup used by init_idle() too:
1658 static void __sched_fork(struct task_struct
*p
)
1660 p
->se
.exec_start
= 0;
1661 p
->se
.sum_exec_runtime
= 0;
1662 p
->se
.prev_sum_exec_runtime
= 0;
1664 #ifdef CONFIG_SCHEDSTATS
1665 p
->se
.wait_start
= 0;
1666 p
->se
.sum_sleep_runtime
= 0;
1667 p
->se
.sleep_start
= 0;
1668 p
->se
.block_start
= 0;
1669 p
->se
.sleep_max
= 0;
1670 p
->se
.block_max
= 0;
1672 p
->se
.slice_max
= 0;
1676 INIT_LIST_HEAD(&p
->run_list
);
1679 #ifdef CONFIG_PREEMPT_NOTIFIERS
1680 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1684 * We mark the process as running here, but have not actually
1685 * inserted it onto the runqueue yet. This guarantees that
1686 * nobody will actually run it, and a signal or other external
1687 * event cannot wake it up and insert it on the runqueue either.
1689 p
->state
= TASK_RUNNING
;
1693 * fork()/clone()-time setup:
1695 void sched_fork(struct task_struct
*p
, int clone_flags
)
1697 int cpu
= get_cpu();
1702 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1704 set_task_cpu(p
, cpu
);
1707 * Make sure we do not leak PI boosting priority to the child:
1709 p
->prio
= current
->normal_prio
;
1710 if (!rt_prio(p
->prio
))
1711 p
->sched_class
= &fair_sched_class
;
1713 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1714 if (likely(sched_info_on()))
1715 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1717 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1720 #ifdef CONFIG_PREEMPT
1721 /* Want to start with kernel preemption disabled. */
1722 task_thread_info(p
)->preempt_count
= 1;
1728 * wake_up_new_task - wake up a newly created task for the first time.
1730 * This function will do some initial scheduler statistics housekeeping
1731 * that must be done for every newly created context, then puts the task
1732 * on the runqueue and wakes it.
1734 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1736 unsigned long flags
;
1739 rq
= task_rq_lock(p
, &flags
);
1740 BUG_ON(p
->state
!= TASK_RUNNING
);
1741 update_rq_clock(rq
);
1743 p
->prio
= effective_prio(p
);
1745 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
1746 activate_task(rq
, p
, 0);
1749 * Let the scheduling class do new task startup
1750 * management (if any):
1752 p
->sched_class
->task_new(rq
, p
);
1753 inc_nr_running(p
, rq
);
1755 check_preempt_curr(rq
, p
);
1756 wakeup_balance_rt(rq
, p
);
1757 task_rq_unlock(rq
, &flags
);
1760 #ifdef CONFIG_PREEMPT_NOTIFIERS
1763 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1764 * @notifier: notifier struct to register
1766 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1768 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1770 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1773 * preempt_notifier_unregister - no longer interested in preemption notifications
1774 * @notifier: notifier struct to unregister
1776 * This is safe to call from within a preemption notifier.
1778 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1780 hlist_del(¬ifier
->link
);
1782 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1784 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1786 struct preempt_notifier
*notifier
;
1787 struct hlist_node
*node
;
1789 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1790 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1794 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1795 struct task_struct
*next
)
1797 struct preempt_notifier
*notifier
;
1798 struct hlist_node
*node
;
1800 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1801 notifier
->ops
->sched_out(notifier
, next
);
1806 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1811 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1812 struct task_struct
*next
)
1819 * prepare_task_switch - prepare to switch tasks
1820 * @rq: the runqueue preparing to switch
1821 * @prev: the current task that is being switched out
1822 * @next: the task we are going to switch to.
1824 * This is called with the rq lock held and interrupts off. It must
1825 * be paired with a subsequent finish_task_switch after the context
1828 * prepare_task_switch sets up locking and calls architecture specific
1832 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1833 struct task_struct
*next
)
1835 fire_sched_out_preempt_notifiers(prev
, next
);
1836 prepare_lock_switch(rq
, next
);
1837 prepare_arch_switch(next
);
1841 * finish_task_switch - clean up after a task-switch
1842 * @rq: runqueue associated with task-switch
1843 * @prev: the thread we just switched away from.
1845 * finish_task_switch must be called after the context switch, paired
1846 * with a prepare_task_switch call before the context switch.
1847 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1848 * and do any other architecture-specific cleanup actions.
1850 * Note that we may have delayed dropping an mm in context_switch(). If
1851 * so, we finish that here outside of the runqueue lock. (Doing it
1852 * with the lock held can cause deadlocks; see schedule() for
1855 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1856 __releases(rq
->lock
)
1858 struct mm_struct
*mm
= rq
->prev_mm
;
1864 * A task struct has one reference for the use as "current".
1865 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1866 * schedule one last time. The schedule call will never return, and
1867 * the scheduled task must drop that reference.
1868 * The test for TASK_DEAD must occur while the runqueue locks are
1869 * still held, otherwise prev could be scheduled on another cpu, die
1870 * there before we look at prev->state, and then the reference would
1872 * Manfred Spraul <manfred@colorfullife.com>
1874 prev_state
= prev
->state
;
1875 finish_arch_switch(prev
);
1876 finish_lock_switch(rq
, prev
);
1877 schedule_tail_balance_rt(rq
);
1879 fire_sched_in_preempt_notifiers(current
);
1882 if (unlikely(prev_state
== TASK_DEAD
)) {
1884 * Remove function-return probe instances associated with this
1885 * task and put them back on the free list.
1887 kprobe_flush_task(prev
);
1888 put_task_struct(prev
);
1893 * schedule_tail - first thing a freshly forked thread must call.
1894 * @prev: the thread we just switched away from.
1896 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1897 __releases(rq
->lock
)
1899 struct rq
*rq
= this_rq();
1901 finish_task_switch(rq
, prev
);
1902 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1903 /* In this case, finish_task_switch does not reenable preemption */
1906 if (current
->set_child_tid
)
1907 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1911 * context_switch - switch to the new MM and the new
1912 * thread's register state.
1915 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1916 struct task_struct
*next
)
1918 struct mm_struct
*mm
, *oldmm
;
1920 prepare_task_switch(rq
, prev
, next
);
1922 oldmm
= prev
->active_mm
;
1924 * For paravirt, this is coupled with an exit in switch_to to
1925 * combine the page table reload and the switch backend into
1928 arch_enter_lazy_cpu_mode();
1930 if (unlikely(!mm
)) {
1931 next
->active_mm
= oldmm
;
1932 atomic_inc(&oldmm
->mm_count
);
1933 enter_lazy_tlb(oldmm
, next
);
1935 switch_mm(oldmm
, mm
, next
);
1937 if (unlikely(!prev
->mm
)) {
1938 prev
->active_mm
= NULL
;
1939 rq
->prev_mm
= oldmm
;
1942 * Since the runqueue lock will be released by the next
1943 * task (which is an invalid locking op but in the case
1944 * of the scheduler it's an obvious special-case), so we
1945 * do an early lockdep release here:
1947 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1948 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1951 /* Here we just switch the register state and the stack. */
1952 switch_to(prev
, next
, prev
);
1956 * this_rq must be evaluated again because prev may have moved
1957 * CPUs since it called schedule(), thus the 'rq' on its stack
1958 * frame will be invalid.
1960 finish_task_switch(this_rq(), prev
);
1964 * nr_running, nr_uninterruptible and nr_context_switches:
1966 * externally visible scheduler statistics: current number of runnable
1967 * threads, current number of uninterruptible-sleeping threads, total
1968 * number of context switches performed since bootup.
1970 unsigned long nr_running(void)
1972 unsigned long i
, sum
= 0;
1974 for_each_online_cpu(i
)
1975 sum
+= cpu_rq(i
)->nr_running
;
1980 unsigned long nr_uninterruptible(void)
1982 unsigned long i
, sum
= 0;
1984 for_each_possible_cpu(i
)
1985 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1988 * Since we read the counters lockless, it might be slightly
1989 * inaccurate. Do not allow it to go below zero though:
1991 if (unlikely((long)sum
< 0))
1997 unsigned long long nr_context_switches(void)
2000 unsigned long long sum
= 0;
2002 for_each_possible_cpu(i
)
2003 sum
+= cpu_rq(i
)->nr_switches
;
2008 unsigned long nr_iowait(void)
2010 unsigned long i
, sum
= 0;
2012 for_each_possible_cpu(i
)
2013 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2018 unsigned long nr_active(void)
2020 unsigned long i
, running
= 0, uninterruptible
= 0;
2022 for_each_online_cpu(i
) {
2023 running
+= cpu_rq(i
)->nr_running
;
2024 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2027 if (unlikely((long)uninterruptible
< 0))
2028 uninterruptible
= 0;
2030 return running
+ uninterruptible
;
2034 * Update rq->cpu_load[] statistics. This function is usually called every
2035 * scheduler tick (TICK_NSEC).
2037 static void update_cpu_load(struct rq
*this_rq
)
2039 unsigned long this_load
= this_rq
->load
.weight
;
2042 this_rq
->nr_load_updates
++;
2044 /* Update our load: */
2045 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2046 unsigned long old_load
, new_load
;
2048 /* scale is effectively 1 << i now, and >> i divides by scale */
2050 old_load
= this_rq
->cpu_load
[i
];
2051 new_load
= this_load
;
2053 * Round up the averaging division if load is increasing. This
2054 * prevents us from getting stuck on 9 if the load is 10, for
2057 if (new_load
> old_load
)
2058 new_load
+= scale
-1;
2059 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2066 * double_rq_lock - safely lock two runqueues
2068 * Note this does not disable interrupts like task_rq_lock,
2069 * you need to do so manually before calling.
2071 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2072 __acquires(rq1
->lock
)
2073 __acquires(rq2
->lock
)
2075 BUG_ON(!irqs_disabled());
2077 spin_lock(&rq1
->lock
);
2078 __acquire(rq2
->lock
); /* Fake it out ;) */
2081 spin_lock(&rq1
->lock
);
2082 spin_lock(&rq2
->lock
);
2084 spin_lock(&rq2
->lock
);
2085 spin_lock(&rq1
->lock
);
2088 update_rq_clock(rq1
);
2089 update_rq_clock(rq2
);
2093 * double_rq_unlock - safely unlock two runqueues
2095 * Note this does not restore interrupts like task_rq_unlock,
2096 * you need to do so manually after calling.
2098 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2099 __releases(rq1
->lock
)
2100 __releases(rq2
->lock
)
2102 spin_unlock(&rq1
->lock
);
2104 spin_unlock(&rq2
->lock
);
2106 __release(rq2
->lock
);
2110 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2112 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2113 __releases(this_rq
->lock
)
2114 __acquires(busiest
->lock
)
2115 __acquires(this_rq
->lock
)
2119 if (unlikely(!irqs_disabled())) {
2120 /* printk() doesn't work good under rq->lock */
2121 spin_unlock(&this_rq
->lock
);
2124 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2125 if (busiest
< this_rq
) {
2126 spin_unlock(&this_rq
->lock
);
2127 spin_lock(&busiest
->lock
);
2128 spin_lock(&this_rq
->lock
);
2131 spin_lock(&busiest
->lock
);
2137 * If dest_cpu is allowed for this process, migrate the task to it.
2138 * This is accomplished by forcing the cpu_allowed mask to only
2139 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2140 * the cpu_allowed mask is restored.
2142 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2144 struct migration_req req
;
2145 unsigned long flags
;
2148 rq
= task_rq_lock(p
, &flags
);
2149 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2150 || unlikely(cpu_is_offline(dest_cpu
)))
2153 /* force the process onto the specified CPU */
2154 if (migrate_task(p
, dest_cpu
, &req
)) {
2155 /* Need to wait for migration thread (might exit: take ref). */
2156 struct task_struct
*mt
= rq
->migration_thread
;
2158 get_task_struct(mt
);
2159 task_rq_unlock(rq
, &flags
);
2160 wake_up_process(mt
);
2161 put_task_struct(mt
);
2162 wait_for_completion(&req
.done
);
2167 task_rq_unlock(rq
, &flags
);
2171 * sched_exec - execve() is a valuable balancing opportunity, because at
2172 * this point the task has the smallest effective memory and cache footprint.
2174 void sched_exec(void)
2176 int new_cpu
, this_cpu
= get_cpu();
2177 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2179 if (new_cpu
!= this_cpu
)
2180 sched_migrate_task(current
, new_cpu
);
2184 * pull_task - move a task from a remote runqueue to the local runqueue.
2185 * Both runqueues must be locked.
2187 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2188 struct rq
*this_rq
, int this_cpu
)
2190 deactivate_task(src_rq
, p
, 0);
2191 set_task_cpu(p
, this_cpu
);
2192 activate_task(this_rq
, p
, 0);
2194 * Note that idle threads have a prio of MAX_PRIO, for this test
2195 * to be always true for them.
2197 check_preempt_curr(this_rq
, p
);
2201 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2204 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2205 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2209 * We do not migrate tasks that are:
2210 * 1) running (obviously), or
2211 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2212 * 3) are cache-hot on their current CPU.
2214 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2215 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2220 if (task_running(rq
, p
)) {
2221 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2226 * Aggressive migration if:
2227 * 1) task is cache cold, or
2228 * 2) too many balance attempts have failed.
2231 if (!task_hot(p
, rq
->clock
, sd
) ||
2232 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2233 #ifdef CONFIG_SCHEDSTATS
2234 if (task_hot(p
, rq
->clock
, sd
)) {
2235 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2236 schedstat_inc(p
, se
.nr_forced_migrations
);
2242 if (task_hot(p
, rq
->clock
, sd
)) {
2243 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2249 static unsigned long
2250 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2251 unsigned long max_load_move
, struct sched_domain
*sd
,
2252 enum cpu_idle_type idle
, int *all_pinned
,
2253 int *this_best_prio
, struct rq_iterator
*iterator
)
2255 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2256 struct task_struct
*p
;
2257 long rem_load_move
= max_load_move
;
2259 if (max_load_move
== 0)
2265 * Start the load-balancing iterator:
2267 p
= iterator
->start(iterator
->arg
);
2269 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2272 * To help distribute high priority tasks across CPUs we don't
2273 * skip a task if it will be the highest priority task (i.e. smallest
2274 * prio value) on its new queue regardless of its load weight
2276 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2277 SCHED_LOAD_SCALE_FUZZ
;
2278 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2279 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2280 p
= iterator
->next(iterator
->arg
);
2284 pull_task(busiest
, p
, this_rq
, this_cpu
);
2286 rem_load_move
-= p
->se
.load
.weight
;
2289 * We only want to steal up to the prescribed amount of weighted load.
2291 if (rem_load_move
> 0) {
2292 if (p
->prio
< *this_best_prio
)
2293 *this_best_prio
= p
->prio
;
2294 p
= iterator
->next(iterator
->arg
);
2299 * Right now, this is one of only two places pull_task() is called,
2300 * so we can safely collect pull_task() stats here rather than
2301 * inside pull_task().
2303 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2306 *all_pinned
= pinned
;
2308 return max_load_move
- rem_load_move
;
2312 * move_tasks tries to move up to max_load_move weighted load from busiest to
2313 * this_rq, as part of a balancing operation within domain "sd".
2314 * Returns 1 if successful and 0 otherwise.
2316 * Called with both runqueues locked.
2318 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2319 unsigned long max_load_move
,
2320 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2323 const struct sched_class
*class = sched_class_highest
;
2324 unsigned long total_load_moved
= 0;
2325 int this_best_prio
= this_rq
->curr
->prio
;
2329 class->load_balance(this_rq
, this_cpu
, busiest
,
2330 max_load_move
- total_load_moved
,
2331 sd
, idle
, all_pinned
, &this_best_prio
);
2332 class = class->next
;
2333 } while (class && max_load_move
> total_load_moved
);
2335 return total_load_moved
> 0;
2339 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2340 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2341 struct rq_iterator
*iterator
)
2343 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2347 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2348 pull_task(busiest
, p
, this_rq
, this_cpu
);
2350 * Right now, this is only the second place pull_task()
2351 * is called, so we can safely collect pull_task()
2352 * stats here rather than inside pull_task().
2354 schedstat_inc(sd
, lb_gained
[idle
]);
2358 p
= iterator
->next(iterator
->arg
);
2365 * move_one_task tries to move exactly one task from busiest to this_rq, as
2366 * part of active balancing operations within "domain".
2367 * Returns 1 if successful and 0 otherwise.
2369 * Called with both runqueues locked.
2371 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2372 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2374 const struct sched_class
*class;
2376 for (class = sched_class_highest
; class; class = class->next
)
2377 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2384 * find_busiest_group finds and returns the busiest CPU group within the
2385 * domain. It calculates and returns the amount of weighted load which
2386 * should be moved to restore balance via the imbalance parameter.
2388 static struct sched_group
*
2389 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2390 unsigned long *imbalance
, enum cpu_idle_type idle
,
2391 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2393 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2394 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2395 unsigned long max_pull
;
2396 unsigned long busiest_load_per_task
, busiest_nr_running
;
2397 unsigned long this_load_per_task
, this_nr_running
;
2398 int load_idx
, group_imb
= 0;
2399 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2400 int power_savings_balance
= 1;
2401 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2402 unsigned long min_nr_running
= ULONG_MAX
;
2403 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2406 max_load
= this_load
= total_load
= total_pwr
= 0;
2407 busiest_load_per_task
= busiest_nr_running
= 0;
2408 this_load_per_task
= this_nr_running
= 0;
2409 if (idle
== CPU_NOT_IDLE
)
2410 load_idx
= sd
->busy_idx
;
2411 else if (idle
== CPU_NEWLY_IDLE
)
2412 load_idx
= sd
->newidle_idx
;
2414 load_idx
= sd
->idle_idx
;
2417 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2420 int __group_imb
= 0;
2421 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2422 unsigned long sum_nr_running
, sum_weighted_load
;
2424 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2427 balance_cpu
= first_cpu(group
->cpumask
);
2429 /* Tally up the load of all CPUs in the group */
2430 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2432 min_cpu_load
= ~0UL;
2434 for_each_cpu_mask(i
, group
->cpumask
) {
2437 if (!cpu_isset(i
, *cpus
))
2442 if (*sd_idle
&& rq
->nr_running
)
2445 /* Bias balancing toward cpus of our domain */
2447 if (idle_cpu(i
) && !first_idle_cpu
) {
2452 load
= target_load(i
, load_idx
);
2454 load
= source_load(i
, load_idx
);
2455 if (load
> max_cpu_load
)
2456 max_cpu_load
= load
;
2457 if (min_cpu_load
> load
)
2458 min_cpu_load
= load
;
2462 sum_nr_running
+= rq
->nr_running
;
2463 sum_weighted_load
+= weighted_cpuload(i
);
2467 * First idle cpu or the first cpu(busiest) in this sched group
2468 * is eligible for doing load balancing at this and above
2469 * domains. In the newly idle case, we will allow all the cpu's
2470 * to do the newly idle load balance.
2472 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2473 balance_cpu
!= this_cpu
&& balance
) {
2478 total_load
+= avg_load
;
2479 total_pwr
+= group
->__cpu_power
;
2481 /* Adjust by relative CPU power of the group */
2482 avg_load
= sg_div_cpu_power(group
,
2483 avg_load
* SCHED_LOAD_SCALE
);
2485 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2488 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2491 this_load
= avg_load
;
2493 this_nr_running
= sum_nr_running
;
2494 this_load_per_task
= sum_weighted_load
;
2495 } else if (avg_load
> max_load
&&
2496 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2497 max_load
= avg_load
;
2499 busiest_nr_running
= sum_nr_running
;
2500 busiest_load_per_task
= sum_weighted_load
;
2501 group_imb
= __group_imb
;
2504 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2506 * Busy processors will not participate in power savings
2509 if (idle
== CPU_NOT_IDLE
||
2510 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2514 * If the local group is idle or completely loaded
2515 * no need to do power savings balance at this domain
2517 if (local_group
&& (this_nr_running
>= group_capacity
||
2519 power_savings_balance
= 0;
2522 * If a group is already running at full capacity or idle,
2523 * don't include that group in power savings calculations
2525 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2530 * Calculate the group which has the least non-idle load.
2531 * This is the group from where we need to pick up the load
2534 if ((sum_nr_running
< min_nr_running
) ||
2535 (sum_nr_running
== min_nr_running
&&
2536 first_cpu(group
->cpumask
) <
2537 first_cpu(group_min
->cpumask
))) {
2539 min_nr_running
= sum_nr_running
;
2540 min_load_per_task
= sum_weighted_load
/
2545 * Calculate the group which is almost near its
2546 * capacity but still has some space to pick up some load
2547 * from other group and save more power
2549 if (sum_nr_running
<= group_capacity
- 1) {
2550 if (sum_nr_running
> leader_nr_running
||
2551 (sum_nr_running
== leader_nr_running
&&
2552 first_cpu(group
->cpumask
) >
2553 first_cpu(group_leader
->cpumask
))) {
2554 group_leader
= group
;
2555 leader_nr_running
= sum_nr_running
;
2560 group
= group
->next
;
2561 } while (group
!= sd
->groups
);
2563 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2566 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2568 if (this_load
>= avg_load
||
2569 100*max_load
<= sd
->imbalance_pct
*this_load
)
2572 busiest_load_per_task
/= busiest_nr_running
;
2574 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2577 * We're trying to get all the cpus to the average_load, so we don't
2578 * want to push ourselves above the average load, nor do we wish to
2579 * reduce the max loaded cpu below the average load, as either of these
2580 * actions would just result in more rebalancing later, and ping-pong
2581 * tasks around. Thus we look for the minimum possible imbalance.
2582 * Negative imbalances (*we* are more loaded than anyone else) will
2583 * be counted as no imbalance for these purposes -- we can't fix that
2584 * by pulling tasks to us. Be careful of negative numbers as they'll
2585 * appear as very large values with unsigned longs.
2587 if (max_load
<= busiest_load_per_task
)
2591 * In the presence of smp nice balancing, certain scenarios can have
2592 * max load less than avg load(as we skip the groups at or below
2593 * its cpu_power, while calculating max_load..)
2595 if (max_load
< avg_load
) {
2597 goto small_imbalance
;
2600 /* Don't want to pull so many tasks that a group would go idle */
2601 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2603 /* How much load to actually move to equalise the imbalance */
2604 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2605 (avg_load
- this_load
) * this->__cpu_power
)
2609 * if *imbalance is less than the average load per runnable task
2610 * there is no gaurantee that any tasks will be moved so we'll have
2611 * a think about bumping its value to force at least one task to be
2614 if (*imbalance
< busiest_load_per_task
) {
2615 unsigned long tmp
, pwr_now
, pwr_move
;
2619 pwr_move
= pwr_now
= 0;
2621 if (this_nr_running
) {
2622 this_load_per_task
/= this_nr_running
;
2623 if (busiest_load_per_task
> this_load_per_task
)
2626 this_load_per_task
= SCHED_LOAD_SCALE
;
2628 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2629 busiest_load_per_task
* imbn
) {
2630 *imbalance
= busiest_load_per_task
;
2635 * OK, we don't have enough imbalance to justify moving tasks,
2636 * however we may be able to increase total CPU power used by
2640 pwr_now
+= busiest
->__cpu_power
*
2641 min(busiest_load_per_task
, max_load
);
2642 pwr_now
+= this->__cpu_power
*
2643 min(this_load_per_task
, this_load
);
2644 pwr_now
/= SCHED_LOAD_SCALE
;
2646 /* Amount of load we'd subtract */
2647 tmp
= sg_div_cpu_power(busiest
,
2648 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2650 pwr_move
+= busiest
->__cpu_power
*
2651 min(busiest_load_per_task
, max_load
- tmp
);
2653 /* Amount of load we'd add */
2654 if (max_load
* busiest
->__cpu_power
<
2655 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2656 tmp
= sg_div_cpu_power(this,
2657 max_load
* busiest
->__cpu_power
);
2659 tmp
= sg_div_cpu_power(this,
2660 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2661 pwr_move
+= this->__cpu_power
*
2662 min(this_load_per_task
, this_load
+ tmp
);
2663 pwr_move
/= SCHED_LOAD_SCALE
;
2665 /* Move if we gain throughput */
2666 if (pwr_move
> pwr_now
)
2667 *imbalance
= busiest_load_per_task
;
2673 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2674 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2677 if (this == group_leader
&& group_leader
!= group_min
) {
2678 *imbalance
= min_load_per_task
;
2688 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2691 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2692 unsigned long imbalance
, cpumask_t
*cpus
)
2694 struct rq
*busiest
= NULL
, *rq
;
2695 unsigned long max_load
= 0;
2698 for_each_cpu_mask(i
, group
->cpumask
) {
2701 if (!cpu_isset(i
, *cpus
))
2705 wl
= weighted_cpuload(i
);
2707 if (rq
->nr_running
== 1 && wl
> imbalance
)
2710 if (wl
> max_load
) {
2720 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2721 * so long as it is large enough.
2723 #define MAX_PINNED_INTERVAL 512
2726 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2727 * tasks if there is an imbalance.
2729 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2730 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2733 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2734 struct sched_group
*group
;
2735 unsigned long imbalance
;
2737 cpumask_t cpus
= CPU_MASK_ALL
;
2738 unsigned long flags
;
2741 * When power savings policy is enabled for the parent domain, idle
2742 * sibling can pick up load irrespective of busy siblings. In this case,
2743 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2744 * portraying it as CPU_NOT_IDLE.
2746 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2747 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2750 schedstat_inc(sd
, lb_count
[idle
]);
2753 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2760 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2764 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2766 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2770 BUG_ON(busiest
== this_rq
);
2772 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2775 if (busiest
->nr_running
> 1) {
2777 * Attempt to move tasks. If find_busiest_group has found
2778 * an imbalance but busiest->nr_running <= 1, the group is
2779 * still unbalanced. ld_moved simply stays zero, so it is
2780 * correctly treated as an imbalance.
2782 local_irq_save(flags
);
2783 double_rq_lock(this_rq
, busiest
);
2784 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2785 imbalance
, sd
, idle
, &all_pinned
);
2786 double_rq_unlock(this_rq
, busiest
);
2787 local_irq_restore(flags
);
2790 * some other cpu did the load balance for us.
2792 if (ld_moved
&& this_cpu
!= smp_processor_id())
2793 resched_cpu(this_cpu
);
2795 /* All tasks on this runqueue were pinned by CPU affinity */
2796 if (unlikely(all_pinned
)) {
2797 cpu_clear(cpu_of(busiest
), cpus
);
2798 if (!cpus_empty(cpus
))
2805 schedstat_inc(sd
, lb_failed
[idle
]);
2806 sd
->nr_balance_failed
++;
2808 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2810 spin_lock_irqsave(&busiest
->lock
, flags
);
2812 /* don't kick the migration_thread, if the curr
2813 * task on busiest cpu can't be moved to this_cpu
2815 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2816 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2818 goto out_one_pinned
;
2821 if (!busiest
->active_balance
) {
2822 busiest
->active_balance
= 1;
2823 busiest
->push_cpu
= this_cpu
;
2826 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2828 wake_up_process(busiest
->migration_thread
);
2831 * We've kicked active balancing, reset the failure
2834 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2837 sd
->nr_balance_failed
= 0;
2839 if (likely(!active_balance
)) {
2840 /* We were unbalanced, so reset the balancing interval */
2841 sd
->balance_interval
= sd
->min_interval
;
2844 * If we've begun active balancing, start to back off. This
2845 * case may not be covered by the all_pinned logic if there
2846 * is only 1 task on the busy runqueue (because we don't call
2849 if (sd
->balance_interval
< sd
->max_interval
)
2850 sd
->balance_interval
*= 2;
2853 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2854 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2859 schedstat_inc(sd
, lb_balanced
[idle
]);
2861 sd
->nr_balance_failed
= 0;
2864 /* tune up the balancing interval */
2865 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2866 (sd
->balance_interval
< sd
->max_interval
))
2867 sd
->balance_interval
*= 2;
2869 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2870 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2876 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2877 * tasks if there is an imbalance.
2879 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2880 * this_rq is locked.
2883 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2885 struct sched_group
*group
;
2886 struct rq
*busiest
= NULL
;
2887 unsigned long imbalance
;
2891 cpumask_t cpus
= CPU_MASK_ALL
;
2894 * When power savings policy is enabled for the parent domain, idle
2895 * sibling can pick up load irrespective of busy siblings. In this case,
2896 * let the state of idle sibling percolate up as IDLE, instead of
2897 * portraying it as CPU_NOT_IDLE.
2899 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2900 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2903 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
2905 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2906 &sd_idle
, &cpus
, NULL
);
2908 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2912 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2915 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2919 BUG_ON(busiest
== this_rq
);
2921 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2924 if (busiest
->nr_running
> 1) {
2925 /* Attempt to move tasks */
2926 double_lock_balance(this_rq
, busiest
);
2927 /* this_rq->clock is already updated */
2928 update_rq_clock(busiest
);
2929 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2930 imbalance
, sd
, CPU_NEWLY_IDLE
,
2932 spin_unlock(&busiest
->lock
);
2934 if (unlikely(all_pinned
)) {
2935 cpu_clear(cpu_of(busiest
), cpus
);
2936 if (!cpus_empty(cpus
))
2942 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2943 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2944 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2947 sd
->nr_balance_failed
= 0;
2952 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2953 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2954 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2956 sd
->nr_balance_failed
= 0;
2962 * idle_balance is called by schedule() if this_cpu is about to become
2963 * idle. Attempts to pull tasks from other CPUs.
2965 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2967 struct sched_domain
*sd
;
2968 int pulled_task
= -1;
2969 unsigned long next_balance
= jiffies
+ HZ
;
2971 for_each_domain(this_cpu
, sd
) {
2972 unsigned long interval
;
2974 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2977 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2978 /* If we've pulled tasks over stop searching: */
2979 pulled_task
= load_balance_newidle(this_cpu
,
2982 interval
= msecs_to_jiffies(sd
->balance_interval
);
2983 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2984 next_balance
= sd
->last_balance
+ interval
;
2988 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2990 * We are going idle. next_balance may be set based on
2991 * a busy processor. So reset next_balance.
2993 this_rq
->next_balance
= next_balance
;
2998 * active_load_balance is run by migration threads. It pushes running tasks
2999 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3000 * running on each physical CPU where possible, and avoids physical /
3001 * logical imbalances.
3003 * Called with busiest_rq locked.
3005 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3007 int target_cpu
= busiest_rq
->push_cpu
;
3008 struct sched_domain
*sd
;
3009 struct rq
*target_rq
;
3011 /* Is there any task to move? */
3012 if (busiest_rq
->nr_running
<= 1)
3015 target_rq
= cpu_rq(target_cpu
);
3018 * This condition is "impossible", if it occurs
3019 * we need to fix it. Originally reported by
3020 * Bjorn Helgaas on a 128-cpu setup.
3022 BUG_ON(busiest_rq
== target_rq
);
3024 /* move a task from busiest_rq to target_rq */
3025 double_lock_balance(busiest_rq
, target_rq
);
3026 update_rq_clock(busiest_rq
);
3027 update_rq_clock(target_rq
);
3029 /* Search for an sd spanning us and the target CPU. */
3030 for_each_domain(target_cpu
, sd
) {
3031 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3032 cpu_isset(busiest_cpu
, sd
->span
))
3037 schedstat_inc(sd
, alb_count
);
3039 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3041 schedstat_inc(sd
, alb_pushed
);
3043 schedstat_inc(sd
, alb_failed
);
3045 spin_unlock(&target_rq
->lock
);
3050 atomic_t load_balancer
;
3052 } nohz ____cacheline_aligned
= {
3053 .load_balancer
= ATOMIC_INIT(-1),
3054 .cpu_mask
= CPU_MASK_NONE
,
3058 * This routine will try to nominate the ilb (idle load balancing)
3059 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3060 * load balancing on behalf of all those cpus. If all the cpus in the system
3061 * go into this tickless mode, then there will be no ilb owner (as there is
3062 * no need for one) and all the cpus will sleep till the next wakeup event
3065 * For the ilb owner, tick is not stopped. And this tick will be used
3066 * for idle load balancing. ilb owner will still be part of
3069 * While stopping the tick, this cpu will become the ilb owner if there
3070 * is no other owner. And will be the owner till that cpu becomes busy
3071 * or if all cpus in the system stop their ticks at which point
3072 * there is no need for ilb owner.
3074 * When the ilb owner becomes busy, it nominates another owner, during the
3075 * next busy scheduler_tick()
3077 int select_nohz_load_balancer(int stop_tick
)
3079 int cpu
= smp_processor_id();
3082 cpu_set(cpu
, nohz
.cpu_mask
);
3083 cpu_rq(cpu
)->in_nohz_recently
= 1;
3086 * If we are going offline and still the leader, give up!
3088 if (cpu_is_offline(cpu
) &&
3089 atomic_read(&nohz
.load_balancer
) == cpu
) {
3090 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3095 /* time for ilb owner also to sleep */
3096 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3097 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3098 atomic_set(&nohz
.load_balancer
, -1);
3102 if (atomic_read(&nohz
.load_balancer
) == -1) {
3103 /* make me the ilb owner */
3104 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3106 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3109 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3112 cpu_clear(cpu
, nohz
.cpu_mask
);
3114 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3115 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3122 static DEFINE_SPINLOCK(balancing
);
3125 * It checks each scheduling domain to see if it is due to be balanced,
3126 * and initiates a balancing operation if so.
3128 * Balancing parameters are set up in arch_init_sched_domains.
3130 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3133 struct rq
*rq
= cpu_rq(cpu
);
3134 unsigned long interval
;
3135 struct sched_domain
*sd
;
3136 /* Earliest time when we have to do rebalance again */
3137 unsigned long next_balance
= jiffies
+ 60*HZ
;
3138 int update_next_balance
= 0;
3140 for_each_domain(cpu
, sd
) {
3141 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3144 interval
= sd
->balance_interval
;
3145 if (idle
!= CPU_IDLE
)
3146 interval
*= sd
->busy_factor
;
3148 /* scale ms to jiffies */
3149 interval
= msecs_to_jiffies(interval
);
3150 if (unlikely(!interval
))
3152 if (interval
> HZ
*NR_CPUS
/10)
3153 interval
= HZ
*NR_CPUS
/10;
3156 if (sd
->flags
& SD_SERIALIZE
) {
3157 if (!spin_trylock(&balancing
))
3161 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3162 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3164 * We've pulled tasks over so either we're no
3165 * longer idle, or one of our SMT siblings is
3168 idle
= CPU_NOT_IDLE
;
3170 sd
->last_balance
= jiffies
;
3172 if (sd
->flags
& SD_SERIALIZE
)
3173 spin_unlock(&balancing
);
3175 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3176 next_balance
= sd
->last_balance
+ interval
;
3177 update_next_balance
= 1;
3181 * Stop the load balance at this level. There is another
3182 * CPU in our sched group which is doing load balancing more
3190 * next_balance will be updated only when there is a need.
3191 * When the cpu is attached to null domain for ex, it will not be
3194 if (likely(update_next_balance
))
3195 rq
->next_balance
= next_balance
;
3199 * run_rebalance_domains is triggered when needed from the scheduler tick.
3200 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3201 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3203 static void run_rebalance_domains(struct softirq_action
*h
)
3205 int this_cpu
= smp_processor_id();
3206 struct rq
*this_rq
= cpu_rq(this_cpu
);
3207 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3208 CPU_IDLE
: CPU_NOT_IDLE
;
3210 rebalance_domains(this_cpu
, idle
);
3214 * If this cpu is the owner for idle load balancing, then do the
3215 * balancing on behalf of the other idle cpus whose ticks are
3218 if (this_rq
->idle_at_tick
&&
3219 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3220 cpumask_t cpus
= nohz
.cpu_mask
;
3224 cpu_clear(this_cpu
, cpus
);
3225 for_each_cpu_mask(balance_cpu
, cpus
) {
3227 * If this cpu gets work to do, stop the load balancing
3228 * work being done for other cpus. Next load
3229 * balancing owner will pick it up.
3234 rebalance_domains(balance_cpu
, CPU_IDLE
);
3236 rq
= cpu_rq(balance_cpu
);
3237 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3238 this_rq
->next_balance
= rq
->next_balance
;
3245 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3247 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3248 * idle load balancing owner or decide to stop the periodic load balancing,
3249 * if the whole system is idle.
3251 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3255 * If we were in the nohz mode recently and busy at the current
3256 * scheduler tick, then check if we need to nominate new idle
3259 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3260 rq
->in_nohz_recently
= 0;
3262 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3263 cpu_clear(cpu
, nohz
.cpu_mask
);
3264 atomic_set(&nohz
.load_balancer
, -1);
3267 if (atomic_read(&nohz
.load_balancer
) == -1) {
3269 * simple selection for now: Nominate the
3270 * first cpu in the nohz list to be the next
3273 * TBD: Traverse the sched domains and nominate
3274 * the nearest cpu in the nohz.cpu_mask.
3276 int ilb
= first_cpu(nohz
.cpu_mask
);
3284 * If this cpu is idle and doing idle load balancing for all the
3285 * cpus with ticks stopped, is it time for that to stop?
3287 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3288 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3294 * If this cpu is idle and the idle load balancing is done by
3295 * someone else, then no need raise the SCHED_SOFTIRQ
3297 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3298 cpu_isset(cpu
, nohz
.cpu_mask
))
3301 if (time_after_eq(jiffies
, rq
->next_balance
))
3302 raise_softirq(SCHED_SOFTIRQ
);
3305 #else /* CONFIG_SMP */
3308 * on UP we do not need to balance between CPUs:
3310 static inline void idle_balance(int cpu
, struct rq
*rq
)
3316 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3318 EXPORT_PER_CPU_SYMBOL(kstat
);
3321 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3322 * that have not yet been banked in case the task is currently running.
3324 unsigned long long task_sched_runtime(struct task_struct
*p
)
3326 unsigned long flags
;
3330 rq
= task_rq_lock(p
, &flags
);
3331 ns
= p
->se
.sum_exec_runtime
;
3332 if (task_current(rq
, p
)) {
3333 update_rq_clock(rq
);
3334 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3335 if ((s64
)delta_exec
> 0)
3338 task_rq_unlock(rq
, &flags
);
3344 * Account user cpu time to a process.
3345 * @p: the process that the cpu time gets accounted to
3346 * @cputime: the cpu time spent in user space since the last update
3348 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3350 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3353 p
->utime
= cputime_add(p
->utime
, cputime
);
3355 /* Add user time to cpustat. */
3356 tmp
= cputime_to_cputime64(cputime
);
3357 if (TASK_NICE(p
) > 0)
3358 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3360 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3364 * Account guest cpu time to a process.
3365 * @p: the process that the cpu time gets accounted to
3366 * @cputime: the cpu time spent in virtual machine since the last update
3368 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3371 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3373 tmp
= cputime_to_cputime64(cputime
);
3375 p
->utime
= cputime_add(p
->utime
, cputime
);
3376 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3378 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3379 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3383 * Account scaled user cpu time to a process.
3384 * @p: the process that the cpu time gets accounted to
3385 * @cputime: the cpu time spent in user space since the last update
3387 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3389 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3393 * Account system cpu time to a process.
3394 * @p: the process that the cpu time gets accounted to
3395 * @hardirq_offset: the offset to subtract from hardirq_count()
3396 * @cputime: the cpu time spent in kernel space since the last update
3398 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3401 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3402 struct rq
*rq
= this_rq();
3405 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3406 return account_guest_time(p
, cputime
);
3408 p
->stime
= cputime_add(p
->stime
, cputime
);
3410 /* Add system time to cpustat. */
3411 tmp
= cputime_to_cputime64(cputime
);
3412 if (hardirq_count() - hardirq_offset
)
3413 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3414 else if (softirq_count())
3415 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3416 else if (p
!= rq
->idle
)
3417 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3418 else if (atomic_read(&rq
->nr_iowait
) > 0)
3419 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3421 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3422 /* Account for system time used */
3423 acct_update_integrals(p
);
3427 * Account scaled system cpu time to a process.
3428 * @p: the process that the cpu time gets accounted to
3429 * @hardirq_offset: the offset to subtract from hardirq_count()
3430 * @cputime: the cpu time spent in kernel space since the last update
3432 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3434 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3438 * Account for involuntary wait time.
3439 * @p: the process from which the cpu time has been stolen
3440 * @steal: the cpu time spent in involuntary wait
3442 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3444 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3445 cputime64_t tmp
= cputime_to_cputime64(steal
);
3446 struct rq
*rq
= this_rq();
3448 if (p
== rq
->idle
) {
3449 p
->stime
= cputime_add(p
->stime
, steal
);
3450 if (atomic_read(&rq
->nr_iowait
) > 0)
3451 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3453 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3455 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3459 * This function gets called by the timer code, with HZ frequency.
3460 * We call it with interrupts disabled.
3462 * It also gets called by the fork code, when changing the parent's
3465 void scheduler_tick(void)
3467 int cpu
= smp_processor_id();
3468 struct rq
*rq
= cpu_rq(cpu
);
3469 struct task_struct
*curr
= rq
->curr
;
3470 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3472 spin_lock(&rq
->lock
);
3473 __update_rq_clock(rq
);
3475 * Let rq->clock advance by at least TICK_NSEC:
3477 if (unlikely(rq
->clock
< next_tick
))
3478 rq
->clock
= next_tick
;
3479 rq
->tick_timestamp
= rq
->clock
;
3480 update_cpu_load(rq
);
3481 if (curr
!= rq
->idle
) /* FIXME: needed? */
3482 curr
->sched_class
->task_tick(rq
, curr
);
3483 spin_unlock(&rq
->lock
);
3486 rq
->idle_at_tick
= idle_cpu(cpu
);
3487 trigger_load_balance(rq
, cpu
);
3491 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3493 void fastcall
add_preempt_count(int val
)
3498 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3500 preempt_count() += val
;
3502 * Spinlock count overflowing soon?
3504 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3507 EXPORT_SYMBOL(add_preempt_count
);
3509 void fastcall
sub_preempt_count(int val
)
3514 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3517 * Is the spinlock portion underflowing?
3519 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3520 !(preempt_count() & PREEMPT_MASK
)))
3523 preempt_count() -= val
;
3525 EXPORT_SYMBOL(sub_preempt_count
);
3530 * Print scheduling while atomic bug:
3532 static noinline
void __schedule_bug(struct task_struct
*prev
)
3534 struct pt_regs
*regs
= get_irq_regs();
3536 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3537 prev
->comm
, prev
->pid
, preempt_count());
3539 debug_show_held_locks(prev
);
3540 if (irqs_disabled())
3541 print_irqtrace_events(prev
);
3550 * Various schedule()-time debugging checks and statistics:
3552 static inline void schedule_debug(struct task_struct
*prev
)
3555 * Test if we are atomic. Since do_exit() needs to call into
3556 * schedule() atomically, we ignore that path for now.
3557 * Otherwise, whine if we are scheduling when we should not be.
3559 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3560 __schedule_bug(prev
);
3562 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3564 schedstat_inc(this_rq(), sched_count
);
3565 #ifdef CONFIG_SCHEDSTATS
3566 if (unlikely(prev
->lock_depth
>= 0)) {
3567 schedstat_inc(this_rq(), bkl_count
);
3568 schedstat_inc(prev
, sched_info
.bkl_count
);
3574 * Pick up the highest-prio task:
3576 static inline struct task_struct
*
3577 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3579 const struct sched_class
*class;
3580 struct task_struct
*p
;
3583 * Optimization: we know that if all tasks are in
3584 * the fair class we can call that function directly:
3586 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3587 p
= fair_sched_class
.pick_next_task(rq
);
3592 class = sched_class_highest
;
3594 p
= class->pick_next_task(rq
);
3598 * Will never be NULL as the idle class always
3599 * returns a non-NULL p:
3601 class = class->next
;
3606 * schedule() is the main scheduler function.
3608 asmlinkage
void __sched
schedule(void)
3610 struct task_struct
*prev
, *next
;
3617 cpu
= smp_processor_id();
3621 switch_count
= &prev
->nivcsw
;
3623 release_kernel_lock(prev
);
3624 need_resched_nonpreemptible
:
3626 schedule_debug(prev
);
3629 * Do the rq-clock update outside the rq lock:
3631 local_irq_disable();
3632 __update_rq_clock(rq
);
3633 spin_lock(&rq
->lock
);
3634 clear_tsk_need_resched(prev
);
3636 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3637 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3638 unlikely(signal_pending(prev
)))) {
3639 prev
->state
= TASK_RUNNING
;
3641 deactivate_task(rq
, prev
, 1);
3643 switch_count
= &prev
->nvcsw
;
3646 schedule_balance_rt(rq
, prev
);
3648 if (unlikely(!rq
->nr_running
))
3649 idle_balance(cpu
, rq
);
3651 prev
->sched_class
->put_prev_task(rq
, prev
);
3652 next
= pick_next_task(rq
, prev
);
3654 sched_info_switch(prev
, next
);
3656 if (likely(prev
!= next
)) {
3661 context_switch(rq
, prev
, next
); /* unlocks the rq */
3663 spin_unlock_irq(&rq
->lock
);
3665 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3666 cpu
= smp_processor_id();
3668 goto need_resched_nonpreemptible
;
3670 preempt_enable_no_resched();
3671 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3674 EXPORT_SYMBOL(schedule
);
3676 #ifdef CONFIG_PREEMPT
3678 * this is the entry point to schedule() from in-kernel preemption
3679 * off of preempt_enable. Kernel preemptions off return from interrupt
3680 * occur there and call schedule directly.
3682 asmlinkage
void __sched
preempt_schedule(void)
3684 struct thread_info
*ti
= current_thread_info();
3685 #ifdef CONFIG_PREEMPT_BKL
3686 struct task_struct
*task
= current
;
3687 int saved_lock_depth
;
3690 * If there is a non-zero preempt_count or interrupts are disabled,
3691 * we do not want to preempt the current task. Just return..
3693 if (likely(ti
->preempt_count
|| irqs_disabled()))
3697 add_preempt_count(PREEMPT_ACTIVE
);
3700 * We keep the big kernel semaphore locked, but we
3701 * clear ->lock_depth so that schedule() doesnt
3702 * auto-release the semaphore:
3704 #ifdef CONFIG_PREEMPT_BKL
3705 saved_lock_depth
= task
->lock_depth
;
3706 task
->lock_depth
= -1;
3709 #ifdef CONFIG_PREEMPT_BKL
3710 task
->lock_depth
= saved_lock_depth
;
3712 sub_preempt_count(PREEMPT_ACTIVE
);
3715 * Check again in case we missed a preemption opportunity
3716 * between schedule and now.
3719 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3721 EXPORT_SYMBOL(preempt_schedule
);
3724 * this is the entry point to schedule() from kernel preemption
3725 * off of irq context.
3726 * Note, that this is called and return with irqs disabled. This will
3727 * protect us against recursive calling from irq.
3729 asmlinkage
void __sched
preempt_schedule_irq(void)
3731 struct thread_info
*ti
= current_thread_info();
3732 #ifdef CONFIG_PREEMPT_BKL
3733 struct task_struct
*task
= current
;
3734 int saved_lock_depth
;
3736 /* Catch callers which need to be fixed */
3737 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3740 add_preempt_count(PREEMPT_ACTIVE
);
3743 * We keep the big kernel semaphore locked, but we
3744 * clear ->lock_depth so that schedule() doesnt
3745 * auto-release the semaphore:
3747 #ifdef CONFIG_PREEMPT_BKL
3748 saved_lock_depth
= task
->lock_depth
;
3749 task
->lock_depth
= -1;
3753 local_irq_disable();
3754 #ifdef CONFIG_PREEMPT_BKL
3755 task
->lock_depth
= saved_lock_depth
;
3757 sub_preempt_count(PREEMPT_ACTIVE
);
3760 * Check again in case we missed a preemption opportunity
3761 * between schedule and now.
3764 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3767 #endif /* CONFIG_PREEMPT */
3769 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3772 return try_to_wake_up(curr
->private, mode
, sync
);
3774 EXPORT_SYMBOL(default_wake_function
);
3777 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3778 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3779 * number) then we wake all the non-exclusive tasks and one exclusive task.
3781 * There are circumstances in which we can try to wake a task which has already
3782 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3783 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3785 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3786 int nr_exclusive
, int sync
, void *key
)
3788 wait_queue_t
*curr
, *next
;
3790 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3791 unsigned flags
= curr
->flags
;
3793 if (curr
->func(curr
, mode
, sync
, key
) &&
3794 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3800 * __wake_up - wake up threads blocked on a waitqueue.
3802 * @mode: which threads
3803 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3804 * @key: is directly passed to the wakeup function
3806 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3807 int nr_exclusive
, void *key
)
3809 unsigned long flags
;
3811 spin_lock_irqsave(&q
->lock
, flags
);
3812 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3813 spin_unlock_irqrestore(&q
->lock
, flags
);
3815 EXPORT_SYMBOL(__wake_up
);
3818 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3820 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3822 __wake_up_common(q
, mode
, 1, 0, NULL
);
3826 * __wake_up_sync - wake up threads blocked on a waitqueue.
3828 * @mode: which threads
3829 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3831 * The sync wakeup differs that the waker knows that it will schedule
3832 * away soon, so while the target thread will be woken up, it will not
3833 * be migrated to another CPU - ie. the two threads are 'synchronized'
3834 * with each other. This can prevent needless bouncing between CPUs.
3836 * On UP it can prevent extra preemption.
3839 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3841 unsigned long flags
;
3847 if (unlikely(!nr_exclusive
))
3850 spin_lock_irqsave(&q
->lock
, flags
);
3851 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3852 spin_unlock_irqrestore(&q
->lock
, flags
);
3854 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3856 void complete(struct completion
*x
)
3858 unsigned long flags
;
3860 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3862 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3864 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3866 EXPORT_SYMBOL(complete
);
3868 void complete_all(struct completion
*x
)
3870 unsigned long flags
;
3872 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3873 x
->done
+= UINT_MAX
/2;
3874 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3876 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3878 EXPORT_SYMBOL(complete_all
);
3880 static inline long __sched
3881 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3884 DECLARE_WAITQUEUE(wait
, current
);
3886 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3887 __add_wait_queue_tail(&x
->wait
, &wait
);
3889 if (state
== TASK_INTERRUPTIBLE
&&
3890 signal_pending(current
)) {
3891 __remove_wait_queue(&x
->wait
, &wait
);
3892 return -ERESTARTSYS
;
3894 __set_current_state(state
);
3895 spin_unlock_irq(&x
->wait
.lock
);
3896 timeout
= schedule_timeout(timeout
);
3897 spin_lock_irq(&x
->wait
.lock
);
3899 __remove_wait_queue(&x
->wait
, &wait
);
3903 __remove_wait_queue(&x
->wait
, &wait
);
3910 wait_for_common(struct completion
*x
, long timeout
, int state
)
3914 spin_lock_irq(&x
->wait
.lock
);
3915 timeout
= do_wait_for_common(x
, timeout
, state
);
3916 spin_unlock_irq(&x
->wait
.lock
);
3920 void __sched
wait_for_completion(struct completion
*x
)
3922 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3924 EXPORT_SYMBOL(wait_for_completion
);
3926 unsigned long __sched
3927 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3929 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3931 EXPORT_SYMBOL(wait_for_completion_timeout
);
3933 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3935 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3936 if (t
== -ERESTARTSYS
)
3940 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3942 unsigned long __sched
3943 wait_for_completion_interruptible_timeout(struct completion
*x
,
3944 unsigned long timeout
)
3946 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3948 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3951 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3953 unsigned long flags
;
3956 init_waitqueue_entry(&wait
, current
);
3958 __set_current_state(state
);
3960 spin_lock_irqsave(&q
->lock
, flags
);
3961 __add_wait_queue(q
, &wait
);
3962 spin_unlock(&q
->lock
);
3963 timeout
= schedule_timeout(timeout
);
3964 spin_lock_irq(&q
->lock
);
3965 __remove_wait_queue(q
, &wait
);
3966 spin_unlock_irqrestore(&q
->lock
, flags
);
3971 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3973 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3975 EXPORT_SYMBOL(interruptible_sleep_on
);
3978 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3980 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3982 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3984 void __sched
sleep_on(wait_queue_head_t
*q
)
3986 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3988 EXPORT_SYMBOL(sleep_on
);
3990 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3992 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3994 EXPORT_SYMBOL(sleep_on_timeout
);
3996 #ifdef CONFIG_RT_MUTEXES
3999 * rt_mutex_setprio - set the current priority of a task
4001 * @prio: prio value (kernel-internal form)
4003 * This function changes the 'effective' priority of a task. It does
4004 * not touch ->normal_prio like __setscheduler().
4006 * Used by the rt_mutex code to implement priority inheritance logic.
4008 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4010 unsigned long flags
;
4011 int oldprio
, on_rq
, running
;
4014 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4016 rq
= task_rq_lock(p
, &flags
);
4017 update_rq_clock(rq
);
4020 on_rq
= p
->se
.on_rq
;
4021 running
= task_current(rq
, p
);
4023 dequeue_task(rq
, p
, 0);
4025 p
->sched_class
->put_prev_task(rq
, p
);
4029 p
->sched_class
= &rt_sched_class
;
4031 p
->sched_class
= &fair_sched_class
;
4037 p
->sched_class
->set_curr_task(rq
);
4038 enqueue_task(rq
, p
, 0);
4040 * Reschedule if we are currently running on this runqueue and
4041 * our priority decreased, or if we are not currently running on
4042 * this runqueue and our priority is higher than the current's
4045 if (p
->prio
> oldprio
)
4046 resched_task(rq
->curr
);
4048 check_preempt_curr(rq
, p
);
4051 task_rq_unlock(rq
, &flags
);
4056 void set_user_nice(struct task_struct
*p
, long nice
)
4058 int old_prio
, delta
, on_rq
;
4059 unsigned long flags
;
4062 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4065 * We have to be careful, if called from sys_setpriority(),
4066 * the task might be in the middle of scheduling on another CPU.
4068 rq
= task_rq_lock(p
, &flags
);
4069 update_rq_clock(rq
);
4071 * The RT priorities are set via sched_setscheduler(), but we still
4072 * allow the 'normal' nice value to be set - but as expected
4073 * it wont have any effect on scheduling until the task is
4074 * SCHED_FIFO/SCHED_RR:
4076 if (task_has_rt_policy(p
)) {
4077 p
->static_prio
= NICE_TO_PRIO(nice
);
4080 on_rq
= p
->se
.on_rq
;
4082 dequeue_task(rq
, p
, 0);
4084 p
->static_prio
= NICE_TO_PRIO(nice
);
4087 p
->prio
= effective_prio(p
);
4088 delta
= p
->prio
- old_prio
;
4091 enqueue_task(rq
, p
, 0);
4093 * If the task increased its priority or is running and
4094 * lowered its priority, then reschedule its CPU:
4096 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4097 resched_task(rq
->curr
);
4100 task_rq_unlock(rq
, &flags
);
4102 EXPORT_SYMBOL(set_user_nice
);
4105 * can_nice - check if a task can reduce its nice value
4109 int can_nice(const struct task_struct
*p
, const int nice
)
4111 /* convert nice value [19,-20] to rlimit style value [1,40] */
4112 int nice_rlim
= 20 - nice
;
4114 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4115 capable(CAP_SYS_NICE
));
4118 #ifdef __ARCH_WANT_SYS_NICE
4121 * sys_nice - change the priority of the current process.
4122 * @increment: priority increment
4124 * sys_setpriority is a more generic, but much slower function that
4125 * does similar things.
4127 asmlinkage
long sys_nice(int increment
)
4132 * Setpriority might change our priority at the same moment.
4133 * We don't have to worry. Conceptually one call occurs first
4134 * and we have a single winner.
4136 if (increment
< -40)
4141 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4147 if (increment
< 0 && !can_nice(current
, nice
))
4150 retval
= security_task_setnice(current
, nice
);
4154 set_user_nice(current
, nice
);
4161 * task_prio - return the priority value of a given task.
4162 * @p: the task in question.
4164 * This is the priority value as seen by users in /proc.
4165 * RT tasks are offset by -200. Normal tasks are centered
4166 * around 0, value goes from -16 to +15.
4168 int task_prio(const struct task_struct
*p
)
4170 return p
->prio
- MAX_RT_PRIO
;
4174 * task_nice - return the nice value of a given task.
4175 * @p: the task in question.
4177 int task_nice(const struct task_struct
*p
)
4179 return TASK_NICE(p
);
4181 EXPORT_SYMBOL_GPL(task_nice
);
4184 * idle_cpu - is a given cpu idle currently?
4185 * @cpu: the processor in question.
4187 int idle_cpu(int cpu
)
4189 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4193 * idle_task - return the idle task for a given cpu.
4194 * @cpu: the processor in question.
4196 struct task_struct
*idle_task(int cpu
)
4198 return cpu_rq(cpu
)->idle
;
4202 * find_process_by_pid - find a process with a matching PID value.
4203 * @pid: the pid in question.
4205 static struct task_struct
*find_process_by_pid(pid_t pid
)
4207 return pid
? find_task_by_vpid(pid
) : current
;
4210 /* Actually do priority change: must hold rq lock. */
4212 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4214 BUG_ON(p
->se
.on_rq
);
4217 switch (p
->policy
) {
4221 p
->sched_class
= &fair_sched_class
;
4225 p
->sched_class
= &rt_sched_class
;
4229 p
->rt_priority
= prio
;
4230 p
->normal_prio
= normal_prio(p
);
4231 /* we are holding p->pi_lock already */
4232 p
->prio
= rt_mutex_getprio(p
);
4237 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4238 * @p: the task in question.
4239 * @policy: new policy.
4240 * @param: structure containing the new RT priority.
4242 * NOTE that the task may be already dead.
4244 int sched_setscheduler(struct task_struct
*p
, int policy
,
4245 struct sched_param
*param
)
4247 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4248 unsigned long flags
;
4251 /* may grab non-irq protected spin_locks */
4252 BUG_ON(in_interrupt());
4254 /* double check policy once rq lock held */
4256 policy
= oldpolicy
= p
->policy
;
4257 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4258 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4259 policy
!= SCHED_IDLE
)
4262 * Valid priorities for SCHED_FIFO and SCHED_RR are
4263 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4264 * SCHED_BATCH and SCHED_IDLE is 0.
4266 if (param
->sched_priority
< 0 ||
4267 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4268 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4270 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4274 * Allow unprivileged RT tasks to decrease priority:
4276 if (!capable(CAP_SYS_NICE
)) {
4277 if (rt_policy(policy
)) {
4278 unsigned long rlim_rtprio
;
4280 if (!lock_task_sighand(p
, &flags
))
4282 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4283 unlock_task_sighand(p
, &flags
);
4285 /* can't set/change the rt policy */
4286 if (policy
!= p
->policy
&& !rlim_rtprio
)
4289 /* can't increase priority */
4290 if (param
->sched_priority
> p
->rt_priority
&&
4291 param
->sched_priority
> rlim_rtprio
)
4295 * Like positive nice levels, dont allow tasks to
4296 * move out of SCHED_IDLE either:
4298 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4301 /* can't change other user's priorities */
4302 if ((current
->euid
!= p
->euid
) &&
4303 (current
->euid
!= p
->uid
))
4307 retval
= security_task_setscheduler(p
, policy
, param
);
4311 * make sure no PI-waiters arrive (or leave) while we are
4312 * changing the priority of the task:
4314 spin_lock_irqsave(&p
->pi_lock
, flags
);
4316 * To be able to change p->policy safely, the apropriate
4317 * runqueue lock must be held.
4319 rq
= __task_rq_lock(p
);
4320 /* recheck policy now with rq lock held */
4321 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4322 policy
= oldpolicy
= -1;
4323 __task_rq_unlock(rq
);
4324 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4327 update_rq_clock(rq
);
4328 on_rq
= p
->se
.on_rq
;
4329 running
= task_current(rq
, p
);
4331 deactivate_task(rq
, p
, 0);
4333 p
->sched_class
->put_prev_task(rq
, p
);
4337 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4341 p
->sched_class
->set_curr_task(rq
);
4342 activate_task(rq
, p
, 0);
4344 * Reschedule if we are currently running on this runqueue and
4345 * our priority decreased, or if we are not currently running on
4346 * this runqueue and our priority is higher than the current's
4349 if (p
->prio
> oldprio
)
4350 resched_task(rq
->curr
);
4352 check_preempt_curr(rq
, p
);
4355 __task_rq_unlock(rq
);
4356 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4358 rt_mutex_adjust_pi(p
);
4362 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4365 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4367 struct sched_param lparam
;
4368 struct task_struct
*p
;
4371 if (!param
|| pid
< 0)
4373 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4378 p
= find_process_by_pid(pid
);
4380 retval
= sched_setscheduler(p
, policy
, &lparam
);
4387 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4388 * @pid: the pid in question.
4389 * @policy: new policy.
4390 * @param: structure containing the new RT priority.
4393 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4395 /* negative values for policy are not valid */
4399 return do_sched_setscheduler(pid
, policy
, param
);
4403 * sys_sched_setparam - set/change the RT priority of a thread
4404 * @pid: the pid in question.
4405 * @param: structure containing the new RT priority.
4407 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4409 return do_sched_setscheduler(pid
, -1, param
);
4413 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4414 * @pid: the pid in question.
4416 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4418 struct task_struct
*p
;
4425 read_lock(&tasklist_lock
);
4426 p
= find_process_by_pid(pid
);
4428 retval
= security_task_getscheduler(p
);
4432 read_unlock(&tasklist_lock
);
4437 * sys_sched_getscheduler - get the RT priority of a thread
4438 * @pid: the pid in question.
4439 * @param: structure containing the RT priority.
4441 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4443 struct sched_param lp
;
4444 struct task_struct
*p
;
4447 if (!param
|| pid
< 0)
4450 read_lock(&tasklist_lock
);
4451 p
= find_process_by_pid(pid
);
4456 retval
= security_task_getscheduler(p
);
4460 lp
.sched_priority
= p
->rt_priority
;
4461 read_unlock(&tasklist_lock
);
4464 * This one might sleep, we cannot do it with a spinlock held ...
4466 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4471 read_unlock(&tasklist_lock
);
4475 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4477 cpumask_t cpus_allowed
;
4478 struct task_struct
*p
;
4482 read_lock(&tasklist_lock
);
4484 p
= find_process_by_pid(pid
);
4486 read_unlock(&tasklist_lock
);
4492 * It is not safe to call set_cpus_allowed with the
4493 * tasklist_lock held. We will bump the task_struct's
4494 * usage count and then drop tasklist_lock.
4497 read_unlock(&tasklist_lock
);
4500 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4501 !capable(CAP_SYS_NICE
))
4504 retval
= security_task_setscheduler(p
, 0, NULL
);
4508 cpus_allowed
= cpuset_cpus_allowed(p
);
4509 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4511 retval
= set_cpus_allowed(p
, new_mask
);
4514 cpus_allowed
= cpuset_cpus_allowed(p
);
4515 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4517 * We must have raced with a concurrent cpuset
4518 * update. Just reset the cpus_allowed to the
4519 * cpuset's cpus_allowed
4521 new_mask
= cpus_allowed
;
4531 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4532 cpumask_t
*new_mask
)
4534 if (len
< sizeof(cpumask_t
)) {
4535 memset(new_mask
, 0, sizeof(cpumask_t
));
4536 } else if (len
> sizeof(cpumask_t
)) {
4537 len
= sizeof(cpumask_t
);
4539 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4543 * sys_sched_setaffinity - set the cpu affinity of a process
4544 * @pid: pid of the process
4545 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4546 * @user_mask_ptr: user-space pointer to the new cpu mask
4548 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4549 unsigned long __user
*user_mask_ptr
)
4554 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4558 return sched_setaffinity(pid
, new_mask
);
4562 * Represents all cpu's present in the system
4563 * In systems capable of hotplug, this map could dynamically grow
4564 * as new cpu's are detected in the system via any platform specific
4565 * method, such as ACPI for e.g.
4568 cpumask_t cpu_present_map __read_mostly
;
4569 EXPORT_SYMBOL(cpu_present_map
);
4572 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4573 EXPORT_SYMBOL(cpu_online_map
);
4575 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4576 EXPORT_SYMBOL(cpu_possible_map
);
4579 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4581 struct task_struct
*p
;
4585 read_lock(&tasklist_lock
);
4588 p
= find_process_by_pid(pid
);
4592 retval
= security_task_getscheduler(p
);
4596 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4599 read_unlock(&tasklist_lock
);
4606 * sys_sched_getaffinity - get the cpu affinity of a process
4607 * @pid: pid of the process
4608 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4609 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4611 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4612 unsigned long __user
*user_mask_ptr
)
4617 if (len
< sizeof(cpumask_t
))
4620 ret
= sched_getaffinity(pid
, &mask
);
4624 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4627 return sizeof(cpumask_t
);
4631 * sys_sched_yield - yield the current processor to other threads.
4633 * This function yields the current CPU to other tasks. If there are no
4634 * other threads running on this CPU then this function will return.
4636 asmlinkage
long sys_sched_yield(void)
4638 struct rq
*rq
= this_rq_lock();
4640 schedstat_inc(rq
, yld_count
);
4641 current
->sched_class
->yield_task(rq
);
4644 * Since we are going to call schedule() anyway, there's
4645 * no need to preempt or enable interrupts:
4647 __release(rq
->lock
);
4648 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4649 _raw_spin_unlock(&rq
->lock
);
4650 preempt_enable_no_resched();
4657 static void __cond_resched(void)
4659 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4660 __might_sleep(__FILE__
, __LINE__
);
4663 * The BKS might be reacquired before we have dropped
4664 * PREEMPT_ACTIVE, which could trigger a second
4665 * cond_resched() call.
4668 add_preempt_count(PREEMPT_ACTIVE
);
4670 sub_preempt_count(PREEMPT_ACTIVE
);
4671 } while (need_resched());
4674 int __sched
cond_resched(void)
4676 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4677 system_state
== SYSTEM_RUNNING
) {
4683 EXPORT_SYMBOL(cond_resched
);
4686 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4687 * call schedule, and on return reacquire the lock.
4689 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4690 * operations here to prevent schedule() from being called twice (once via
4691 * spin_unlock(), once by hand).
4693 int cond_resched_lock(spinlock_t
*lock
)
4697 if (need_lockbreak(lock
)) {
4703 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4704 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4705 _raw_spin_unlock(lock
);
4706 preempt_enable_no_resched();
4713 EXPORT_SYMBOL(cond_resched_lock
);
4715 int __sched
cond_resched_softirq(void)
4717 BUG_ON(!in_softirq());
4719 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4727 EXPORT_SYMBOL(cond_resched_softirq
);
4730 * yield - yield the current processor to other threads.
4732 * This is a shortcut for kernel-space yielding - it marks the
4733 * thread runnable and calls sys_sched_yield().
4735 void __sched
yield(void)
4737 set_current_state(TASK_RUNNING
);
4740 EXPORT_SYMBOL(yield
);
4743 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4744 * that process accounting knows that this is a task in IO wait state.
4746 * But don't do that if it is a deliberate, throttling IO wait (this task
4747 * has set its backing_dev_info: the queue against which it should throttle)
4749 void __sched
io_schedule(void)
4751 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4753 delayacct_blkio_start();
4754 atomic_inc(&rq
->nr_iowait
);
4756 atomic_dec(&rq
->nr_iowait
);
4757 delayacct_blkio_end();
4759 EXPORT_SYMBOL(io_schedule
);
4761 long __sched
io_schedule_timeout(long timeout
)
4763 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4766 delayacct_blkio_start();
4767 atomic_inc(&rq
->nr_iowait
);
4768 ret
= schedule_timeout(timeout
);
4769 atomic_dec(&rq
->nr_iowait
);
4770 delayacct_blkio_end();
4775 * sys_sched_get_priority_max - return maximum RT priority.
4776 * @policy: scheduling class.
4778 * this syscall returns the maximum rt_priority that can be used
4779 * by a given scheduling class.
4781 asmlinkage
long sys_sched_get_priority_max(int policy
)
4788 ret
= MAX_USER_RT_PRIO
-1;
4800 * sys_sched_get_priority_min - return minimum RT priority.
4801 * @policy: scheduling class.
4803 * this syscall returns the minimum rt_priority that can be used
4804 * by a given scheduling class.
4806 asmlinkage
long sys_sched_get_priority_min(int policy
)
4824 * sys_sched_rr_get_interval - return the default timeslice of a process.
4825 * @pid: pid of the process.
4826 * @interval: userspace pointer to the timeslice value.
4828 * this syscall writes the default timeslice value of a given process
4829 * into the user-space timespec buffer. A value of '0' means infinity.
4832 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4834 struct task_struct
*p
;
4835 unsigned int time_slice
;
4843 read_lock(&tasklist_lock
);
4844 p
= find_process_by_pid(pid
);
4848 retval
= security_task_getscheduler(p
);
4853 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4854 * tasks that are on an otherwise idle runqueue:
4857 if (p
->policy
== SCHED_RR
) {
4858 time_slice
= DEF_TIMESLICE
;
4860 struct sched_entity
*se
= &p
->se
;
4861 unsigned long flags
;
4864 rq
= task_rq_lock(p
, &flags
);
4865 if (rq
->cfs
.load
.weight
)
4866 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
4867 task_rq_unlock(rq
, &flags
);
4869 read_unlock(&tasklist_lock
);
4870 jiffies_to_timespec(time_slice
, &t
);
4871 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4875 read_unlock(&tasklist_lock
);
4879 static const char stat_nam
[] = "RSDTtZX";
4881 void sched_show_task(struct task_struct
*p
)
4883 unsigned long free
= 0;
4886 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4887 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
4888 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4889 #if BITS_PER_LONG == 32
4890 if (state
== TASK_RUNNING
)
4891 printk(KERN_CONT
" running ");
4893 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4895 if (state
== TASK_RUNNING
)
4896 printk(KERN_CONT
" running task ");
4898 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4900 #ifdef CONFIG_DEBUG_STACK_USAGE
4902 unsigned long *n
= end_of_stack(p
);
4905 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4908 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
4909 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
4911 if (state
!= TASK_RUNNING
)
4912 show_stack(p
, NULL
);
4915 void show_state_filter(unsigned long state_filter
)
4917 struct task_struct
*g
, *p
;
4919 #if BITS_PER_LONG == 32
4921 " task PC stack pid father\n");
4924 " task PC stack pid father\n");
4926 read_lock(&tasklist_lock
);
4927 do_each_thread(g
, p
) {
4929 * reset the NMI-timeout, listing all files on a slow
4930 * console might take alot of time:
4932 touch_nmi_watchdog();
4933 if (!state_filter
|| (p
->state
& state_filter
))
4935 } while_each_thread(g
, p
);
4937 touch_all_softlockup_watchdogs();
4939 #ifdef CONFIG_SCHED_DEBUG
4940 sysrq_sched_debug_show();
4942 read_unlock(&tasklist_lock
);
4944 * Only show locks if all tasks are dumped:
4946 if (state_filter
== -1)
4947 debug_show_all_locks();
4950 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4952 idle
->sched_class
= &idle_sched_class
;
4956 * init_idle - set up an idle thread for a given CPU
4957 * @idle: task in question
4958 * @cpu: cpu the idle task belongs to
4960 * NOTE: this function does not set the idle thread's NEED_RESCHED
4961 * flag, to make booting more robust.
4963 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4965 struct rq
*rq
= cpu_rq(cpu
);
4966 unsigned long flags
;
4969 idle
->se
.exec_start
= sched_clock();
4971 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4972 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4973 __set_task_cpu(idle
, cpu
);
4975 spin_lock_irqsave(&rq
->lock
, flags
);
4976 rq
->curr
= rq
->idle
= idle
;
4977 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4980 spin_unlock_irqrestore(&rq
->lock
, flags
);
4982 /* Set the preempt count _outside_ the spinlocks! */
4983 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4984 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4986 task_thread_info(idle
)->preempt_count
= 0;
4989 * The idle tasks have their own, simple scheduling class:
4991 idle
->sched_class
= &idle_sched_class
;
4995 * In a system that switches off the HZ timer nohz_cpu_mask
4996 * indicates which cpus entered this state. This is used
4997 * in the rcu update to wait only for active cpus. For system
4998 * which do not switch off the HZ timer nohz_cpu_mask should
4999 * always be CPU_MASK_NONE.
5001 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5004 * Increase the granularity value when there are more CPUs,
5005 * because with more CPUs the 'effective latency' as visible
5006 * to users decreases. But the relationship is not linear,
5007 * so pick a second-best guess by going with the log2 of the
5010 * This idea comes from the SD scheduler of Con Kolivas:
5012 static inline void sched_init_granularity(void)
5014 unsigned int factor
= 1 + ilog2(num_online_cpus());
5015 const unsigned long limit
= 200000000;
5017 sysctl_sched_min_granularity
*= factor
;
5018 if (sysctl_sched_min_granularity
> limit
)
5019 sysctl_sched_min_granularity
= limit
;
5021 sysctl_sched_latency
*= factor
;
5022 if (sysctl_sched_latency
> limit
)
5023 sysctl_sched_latency
= limit
;
5025 sysctl_sched_wakeup_granularity
*= factor
;
5026 sysctl_sched_batch_wakeup_granularity
*= factor
;
5031 * This is how migration works:
5033 * 1) we queue a struct migration_req structure in the source CPU's
5034 * runqueue and wake up that CPU's migration thread.
5035 * 2) we down() the locked semaphore => thread blocks.
5036 * 3) migration thread wakes up (implicitly it forces the migrated
5037 * thread off the CPU)
5038 * 4) it gets the migration request and checks whether the migrated
5039 * task is still in the wrong runqueue.
5040 * 5) if it's in the wrong runqueue then the migration thread removes
5041 * it and puts it into the right queue.
5042 * 6) migration thread up()s the semaphore.
5043 * 7) we wake up and the migration is done.
5047 * Change a given task's CPU affinity. Migrate the thread to a
5048 * proper CPU and schedule it away if the CPU it's executing on
5049 * is removed from the allowed bitmask.
5051 * NOTE: the caller must have a valid reference to the task, the
5052 * task must not exit() & deallocate itself prematurely. The
5053 * call is not atomic; no spinlocks may be held.
5055 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5057 struct migration_req req
;
5058 unsigned long flags
;
5062 rq
= task_rq_lock(p
, &flags
);
5063 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5068 if (p
->sched_class
->set_cpus_allowed
)
5069 p
->sched_class
->set_cpus_allowed(p
, &new_mask
);
5071 p
->cpus_allowed
= new_mask
;
5072 p
->nr_cpus_allowed
= cpus_weight(new_mask
);
5075 /* Can the task run on the task's current CPU? If so, we're done */
5076 if (cpu_isset(task_cpu(p
), new_mask
))
5079 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5080 /* Need help from migration thread: drop lock and wait. */
5081 task_rq_unlock(rq
, &flags
);
5082 wake_up_process(rq
->migration_thread
);
5083 wait_for_completion(&req
.done
);
5084 tlb_migrate_finish(p
->mm
);
5088 task_rq_unlock(rq
, &flags
);
5092 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5095 * Move (not current) task off this cpu, onto dest cpu. We're doing
5096 * this because either it can't run here any more (set_cpus_allowed()
5097 * away from this CPU, or CPU going down), or because we're
5098 * attempting to rebalance this task on exec (sched_exec).
5100 * So we race with normal scheduler movements, but that's OK, as long
5101 * as the task is no longer on this CPU.
5103 * Returns non-zero if task was successfully migrated.
5105 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5107 struct rq
*rq_dest
, *rq_src
;
5110 if (unlikely(cpu_is_offline(dest_cpu
)))
5113 rq_src
= cpu_rq(src_cpu
);
5114 rq_dest
= cpu_rq(dest_cpu
);
5116 double_rq_lock(rq_src
, rq_dest
);
5117 /* Already moved. */
5118 if (task_cpu(p
) != src_cpu
)
5120 /* Affinity changed (again). */
5121 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5124 on_rq
= p
->se
.on_rq
;
5126 deactivate_task(rq_src
, p
, 0);
5128 set_task_cpu(p
, dest_cpu
);
5130 activate_task(rq_dest
, p
, 0);
5131 check_preempt_curr(rq_dest
, p
);
5135 double_rq_unlock(rq_src
, rq_dest
);
5140 * migration_thread - this is a highprio system thread that performs
5141 * thread migration by bumping thread off CPU then 'pushing' onto
5144 static int migration_thread(void *data
)
5146 int cpu
= (long)data
;
5150 BUG_ON(rq
->migration_thread
!= current
);
5152 set_current_state(TASK_INTERRUPTIBLE
);
5153 while (!kthread_should_stop()) {
5154 struct migration_req
*req
;
5155 struct list_head
*head
;
5157 spin_lock_irq(&rq
->lock
);
5159 if (cpu_is_offline(cpu
)) {
5160 spin_unlock_irq(&rq
->lock
);
5164 if (rq
->active_balance
) {
5165 active_load_balance(rq
, cpu
);
5166 rq
->active_balance
= 0;
5169 head
= &rq
->migration_queue
;
5171 if (list_empty(head
)) {
5172 spin_unlock_irq(&rq
->lock
);
5174 set_current_state(TASK_INTERRUPTIBLE
);
5177 req
= list_entry(head
->next
, struct migration_req
, list
);
5178 list_del_init(head
->next
);
5180 spin_unlock(&rq
->lock
);
5181 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5184 complete(&req
->done
);
5186 __set_current_state(TASK_RUNNING
);
5190 /* Wait for kthread_stop */
5191 set_current_state(TASK_INTERRUPTIBLE
);
5192 while (!kthread_should_stop()) {
5194 set_current_state(TASK_INTERRUPTIBLE
);
5196 __set_current_state(TASK_RUNNING
);
5200 #ifdef CONFIG_HOTPLUG_CPU
5202 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5206 local_irq_disable();
5207 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5213 * Figure out where task on dead CPU should go, use force if necessary.
5214 * NOTE: interrupts should be disabled by the caller
5216 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5218 unsigned long flags
;
5225 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5226 cpus_and(mask
, mask
, p
->cpus_allowed
);
5227 dest_cpu
= any_online_cpu(mask
);
5229 /* On any allowed CPU? */
5230 if (dest_cpu
== NR_CPUS
)
5231 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5233 /* No more Mr. Nice Guy. */
5234 if (dest_cpu
== NR_CPUS
) {
5235 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5237 * Try to stay on the same cpuset, where the
5238 * current cpuset may be a subset of all cpus.
5239 * The cpuset_cpus_allowed_locked() variant of
5240 * cpuset_cpus_allowed() will not block. It must be
5241 * called within calls to cpuset_lock/cpuset_unlock.
5243 rq
= task_rq_lock(p
, &flags
);
5244 p
->cpus_allowed
= cpus_allowed
;
5245 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5246 task_rq_unlock(rq
, &flags
);
5249 * Don't tell them about moving exiting tasks or
5250 * kernel threads (both mm NULL), since they never
5253 if (p
->mm
&& printk_ratelimit()) {
5254 printk(KERN_INFO
"process %d (%s) no "
5255 "longer affine to cpu%d\n",
5256 task_pid_nr(p
), p
->comm
, dead_cpu
);
5259 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5263 * While a dead CPU has no uninterruptible tasks queued at this point,
5264 * it might still have a nonzero ->nr_uninterruptible counter, because
5265 * for performance reasons the counter is not stricly tracking tasks to
5266 * their home CPUs. So we just add the counter to another CPU's counter,
5267 * to keep the global sum constant after CPU-down:
5269 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5271 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5272 unsigned long flags
;
5274 local_irq_save(flags
);
5275 double_rq_lock(rq_src
, rq_dest
);
5276 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5277 rq_src
->nr_uninterruptible
= 0;
5278 double_rq_unlock(rq_src
, rq_dest
);
5279 local_irq_restore(flags
);
5282 /* Run through task list and migrate tasks from the dead cpu. */
5283 static void migrate_live_tasks(int src_cpu
)
5285 struct task_struct
*p
, *t
;
5287 read_lock(&tasklist_lock
);
5289 do_each_thread(t
, p
) {
5293 if (task_cpu(p
) == src_cpu
)
5294 move_task_off_dead_cpu(src_cpu
, p
);
5295 } while_each_thread(t
, p
);
5297 read_unlock(&tasklist_lock
);
5301 * Schedules idle task to be the next runnable task on current CPU.
5302 * It does so by boosting its priority to highest possible.
5303 * Used by CPU offline code.
5305 void sched_idle_next(void)
5307 int this_cpu
= smp_processor_id();
5308 struct rq
*rq
= cpu_rq(this_cpu
);
5309 struct task_struct
*p
= rq
->idle
;
5310 unsigned long flags
;
5312 /* cpu has to be offline */
5313 BUG_ON(cpu_online(this_cpu
));
5316 * Strictly not necessary since rest of the CPUs are stopped by now
5317 * and interrupts disabled on the current cpu.
5319 spin_lock_irqsave(&rq
->lock
, flags
);
5321 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5323 update_rq_clock(rq
);
5324 activate_task(rq
, p
, 0);
5326 spin_unlock_irqrestore(&rq
->lock
, flags
);
5330 * Ensures that the idle task is using init_mm right before its cpu goes
5333 void idle_task_exit(void)
5335 struct mm_struct
*mm
= current
->active_mm
;
5337 BUG_ON(cpu_online(smp_processor_id()));
5340 switch_mm(mm
, &init_mm
, current
);
5344 /* called under rq->lock with disabled interrupts */
5345 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5347 struct rq
*rq
= cpu_rq(dead_cpu
);
5349 /* Must be exiting, otherwise would be on tasklist. */
5350 BUG_ON(!p
->exit_state
);
5352 /* Cannot have done final schedule yet: would have vanished. */
5353 BUG_ON(p
->state
== TASK_DEAD
);
5358 * Drop lock around migration; if someone else moves it,
5359 * that's OK. No task can be added to this CPU, so iteration is
5362 spin_unlock_irq(&rq
->lock
);
5363 move_task_off_dead_cpu(dead_cpu
, p
);
5364 spin_lock_irq(&rq
->lock
);
5369 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5370 static void migrate_dead_tasks(unsigned int dead_cpu
)
5372 struct rq
*rq
= cpu_rq(dead_cpu
);
5373 struct task_struct
*next
;
5376 if (!rq
->nr_running
)
5378 update_rq_clock(rq
);
5379 next
= pick_next_task(rq
, rq
->curr
);
5382 migrate_dead(dead_cpu
, next
);
5386 #endif /* CONFIG_HOTPLUG_CPU */
5388 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5390 static struct ctl_table sd_ctl_dir
[] = {
5392 .procname
= "sched_domain",
5398 static struct ctl_table sd_ctl_root
[] = {
5400 .ctl_name
= CTL_KERN
,
5401 .procname
= "kernel",
5403 .child
= sd_ctl_dir
,
5408 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5410 struct ctl_table
*entry
=
5411 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5416 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5418 struct ctl_table
*entry
;
5421 * In the intermediate directories, both the child directory and
5422 * procname are dynamically allocated and could fail but the mode
5423 * will always be set. In the lowest directory the names are
5424 * static strings and all have proc handlers.
5426 for (entry
= *tablep
; entry
->mode
; entry
++) {
5428 sd_free_ctl_entry(&entry
->child
);
5429 if (entry
->proc_handler
== NULL
)
5430 kfree(entry
->procname
);
5438 set_table_entry(struct ctl_table
*entry
,
5439 const char *procname
, void *data
, int maxlen
,
5440 mode_t mode
, proc_handler
*proc_handler
)
5442 entry
->procname
= procname
;
5444 entry
->maxlen
= maxlen
;
5446 entry
->proc_handler
= proc_handler
;
5449 static struct ctl_table
*
5450 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5452 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5457 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5458 sizeof(long), 0644, proc_doulongvec_minmax
);
5459 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5460 sizeof(long), 0644, proc_doulongvec_minmax
);
5461 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5462 sizeof(int), 0644, proc_dointvec_minmax
);
5463 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5464 sizeof(int), 0644, proc_dointvec_minmax
);
5465 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5466 sizeof(int), 0644, proc_dointvec_minmax
);
5467 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5468 sizeof(int), 0644, proc_dointvec_minmax
);
5469 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5470 sizeof(int), 0644, proc_dointvec_minmax
);
5471 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5472 sizeof(int), 0644, proc_dointvec_minmax
);
5473 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5474 sizeof(int), 0644, proc_dointvec_minmax
);
5475 set_table_entry(&table
[9], "cache_nice_tries",
5476 &sd
->cache_nice_tries
,
5477 sizeof(int), 0644, proc_dointvec_minmax
);
5478 set_table_entry(&table
[10], "flags", &sd
->flags
,
5479 sizeof(int), 0644, proc_dointvec_minmax
);
5480 /* &table[11] is terminator */
5485 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5487 struct ctl_table
*entry
, *table
;
5488 struct sched_domain
*sd
;
5489 int domain_num
= 0, i
;
5492 for_each_domain(cpu
, sd
)
5494 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5499 for_each_domain(cpu
, sd
) {
5500 snprintf(buf
, 32, "domain%d", i
);
5501 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5503 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5510 static struct ctl_table_header
*sd_sysctl_header
;
5511 static void register_sched_domain_sysctl(void)
5513 int i
, cpu_num
= num_online_cpus();
5514 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5517 WARN_ON(sd_ctl_dir
[0].child
);
5518 sd_ctl_dir
[0].child
= entry
;
5523 for_each_online_cpu(i
) {
5524 snprintf(buf
, 32, "cpu%d", i
);
5525 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5527 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5531 WARN_ON(sd_sysctl_header
);
5532 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5535 /* may be called multiple times per register */
5536 static void unregister_sched_domain_sysctl(void)
5538 if (sd_sysctl_header
)
5539 unregister_sysctl_table(sd_sysctl_header
);
5540 sd_sysctl_header
= NULL
;
5541 if (sd_ctl_dir
[0].child
)
5542 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5545 static void register_sched_domain_sysctl(void)
5548 static void unregister_sched_domain_sysctl(void)
5554 * migration_call - callback that gets triggered when a CPU is added.
5555 * Here we can start up the necessary migration thread for the new CPU.
5557 static int __cpuinit
5558 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5560 struct task_struct
*p
;
5561 int cpu
= (long)hcpu
;
5562 unsigned long flags
;
5567 case CPU_UP_PREPARE
:
5568 case CPU_UP_PREPARE_FROZEN
:
5569 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5572 kthread_bind(p
, cpu
);
5573 /* Must be high prio: stop_machine expects to yield to it. */
5574 rq
= task_rq_lock(p
, &flags
);
5575 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5576 task_rq_unlock(rq
, &flags
);
5577 cpu_rq(cpu
)->migration_thread
= p
;
5581 case CPU_ONLINE_FROZEN
:
5582 /* Strictly unnecessary, as first user will wake it. */
5583 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5585 /* Update our root-domain */
5587 spin_lock_irqsave(&rq
->lock
, flags
);
5589 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5590 cpu_set(cpu
, rq
->rd
->online
);
5592 spin_unlock_irqrestore(&rq
->lock
, flags
);
5595 #ifdef CONFIG_HOTPLUG_CPU
5596 case CPU_UP_CANCELED
:
5597 case CPU_UP_CANCELED_FROZEN
:
5598 if (!cpu_rq(cpu
)->migration_thread
)
5600 /* Unbind it from offline cpu so it can run. Fall thru. */
5601 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5602 any_online_cpu(cpu_online_map
));
5603 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5604 cpu_rq(cpu
)->migration_thread
= NULL
;
5608 case CPU_DEAD_FROZEN
:
5609 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5610 migrate_live_tasks(cpu
);
5612 kthread_stop(rq
->migration_thread
);
5613 rq
->migration_thread
= NULL
;
5614 /* Idle task back to normal (off runqueue, low prio) */
5615 spin_lock_irq(&rq
->lock
);
5616 update_rq_clock(rq
);
5617 deactivate_task(rq
, rq
->idle
, 0);
5618 rq
->idle
->static_prio
= MAX_PRIO
;
5619 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5620 rq
->idle
->sched_class
= &idle_sched_class
;
5621 migrate_dead_tasks(cpu
);
5622 spin_unlock_irq(&rq
->lock
);
5624 migrate_nr_uninterruptible(rq
);
5625 BUG_ON(rq
->nr_running
!= 0);
5628 * No need to migrate the tasks: it was best-effort if
5629 * they didn't take sched_hotcpu_mutex. Just wake up
5632 spin_lock_irq(&rq
->lock
);
5633 while (!list_empty(&rq
->migration_queue
)) {
5634 struct migration_req
*req
;
5636 req
= list_entry(rq
->migration_queue
.next
,
5637 struct migration_req
, list
);
5638 list_del_init(&req
->list
);
5639 complete(&req
->done
);
5641 spin_unlock_irq(&rq
->lock
);
5644 case CPU_DOWN_PREPARE
:
5645 /* Update our root-domain */
5647 spin_lock_irqsave(&rq
->lock
, flags
);
5649 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5650 cpu_clear(cpu
, rq
->rd
->online
);
5652 spin_unlock_irqrestore(&rq
->lock
, flags
);
5659 /* Register at highest priority so that task migration (migrate_all_tasks)
5660 * happens before everything else.
5662 static struct notifier_block __cpuinitdata migration_notifier
= {
5663 .notifier_call
= migration_call
,
5667 void __init
migration_init(void)
5669 void *cpu
= (void *)(long)smp_processor_id();
5672 /* Start one for the boot CPU: */
5673 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5674 BUG_ON(err
== NOTIFY_BAD
);
5675 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5676 register_cpu_notifier(&migration_notifier
);
5682 /* Number of possible processor ids */
5683 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5684 EXPORT_SYMBOL(nr_cpu_ids
);
5686 #ifdef CONFIG_SCHED_DEBUG
5688 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
5690 struct sched_group
*group
= sd
->groups
;
5691 cpumask_t groupmask
;
5694 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5695 cpus_clear(groupmask
);
5697 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5699 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5700 printk("does not load-balance\n");
5702 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5707 printk(KERN_CONT
"span %s\n", str
);
5709 if (!cpu_isset(cpu
, sd
->span
)) {
5710 printk(KERN_ERR
"ERROR: domain->span does not contain "
5713 if (!cpu_isset(cpu
, group
->cpumask
)) {
5714 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5718 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5722 printk(KERN_ERR
"ERROR: group is NULL\n");
5726 if (!group
->__cpu_power
) {
5727 printk(KERN_CONT
"\n");
5728 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5733 if (!cpus_weight(group
->cpumask
)) {
5734 printk(KERN_CONT
"\n");
5735 printk(KERN_ERR
"ERROR: empty group\n");
5739 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5740 printk(KERN_CONT
"\n");
5741 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5745 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5747 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5748 printk(KERN_CONT
" %s", str
);
5750 group
= group
->next
;
5751 } while (group
!= sd
->groups
);
5752 printk(KERN_CONT
"\n");
5754 if (!cpus_equal(sd
->span
, groupmask
))
5755 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5757 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
5758 printk(KERN_ERR
"ERROR: parent span is not a superset "
5759 "of domain->span\n");
5763 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5768 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5772 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5775 if (sched_domain_debug_one(sd
, cpu
, level
))
5784 # define sched_domain_debug(sd, cpu) do { } while (0)
5787 static int sd_degenerate(struct sched_domain
*sd
)
5789 if (cpus_weight(sd
->span
) == 1)
5792 /* Following flags need at least 2 groups */
5793 if (sd
->flags
& (SD_LOAD_BALANCE
|
5794 SD_BALANCE_NEWIDLE
|
5798 SD_SHARE_PKG_RESOURCES
)) {
5799 if (sd
->groups
!= sd
->groups
->next
)
5803 /* Following flags don't use groups */
5804 if (sd
->flags
& (SD_WAKE_IDLE
|
5813 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5815 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5817 if (sd_degenerate(parent
))
5820 if (!cpus_equal(sd
->span
, parent
->span
))
5823 /* Does parent contain flags not in child? */
5824 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5825 if (cflags
& SD_WAKE_AFFINE
)
5826 pflags
&= ~SD_WAKE_BALANCE
;
5827 /* Flags needing groups don't count if only 1 group in parent */
5828 if (parent
->groups
== parent
->groups
->next
) {
5829 pflags
&= ~(SD_LOAD_BALANCE
|
5830 SD_BALANCE_NEWIDLE
|
5834 SD_SHARE_PKG_RESOURCES
);
5836 if (~cflags
& pflags
)
5842 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5844 unsigned long flags
;
5845 const struct sched_class
*class;
5847 spin_lock_irqsave(&rq
->lock
, flags
);
5850 struct root_domain
*old_rd
= rq
->rd
;
5852 for (class = sched_class_highest
; class; class = class->next
) {
5853 if (class->leave_domain
)
5854 class->leave_domain(rq
);
5857 if (atomic_dec_and_test(&old_rd
->refcount
))
5861 atomic_inc(&rd
->refcount
);
5864 for (class = sched_class_highest
; class; class = class->next
) {
5865 if (class->join_domain
)
5866 class->join_domain(rq
);
5869 spin_unlock_irqrestore(&rq
->lock
, flags
);
5872 static void init_rootdomain(struct root_domain
*rd
, const cpumask_t
*map
)
5874 memset(rd
, 0, sizeof(*rd
));
5877 cpus_and(rd
->online
, rd
->span
, cpu_online_map
);
5880 static void init_defrootdomain(void)
5882 cpumask_t cpus
= CPU_MASK_ALL
;
5884 init_rootdomain(&def_root_domain
, &cpus
);
5885 atomic_set(&def_root_domain
.refcount
, 1);
5888 static struct root_domain
*alloc_rootdomain(const cpumask_t
*map
)
5890 struct root_domain
*rd
;
5892 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5896 init_rootdomain(rd
, map
);
5902 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5903 * hold the hotplug lock.
5906 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5908 struct rq
*rq
= cpu_rq(cpu
);
5909 struct sched_domain
*tmp
;
5911 /* Remove the sched domains which do not contribute to scheduling. */
5912 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5913 struct sched_domain
*parent
= tmp
->parent
;
5916 if (sd_parent_degenerate(tmp
, parent
)) {
5917 tmp
->parent
= parent
->parent
;
5919 parent
->parent
->child
= tmp
;
5923 if (sd
&& sd_degenerate(sd
)) {
5929 sched_domain_debug(sd
, cpu
);
5931 rq_attach_root(rq
, rd
);
5932 rcu_assign_pointer(rq
->sd
, sd
);
5935 /* cpus with isolated domains */
5936 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5938 /* Setup the mask of cpus configured for isolated domains */
5939 static int __init
isolated_cpu_setup(char *str
)
5941 int ints
[NR_CPUS
], i
;
5943 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5944 cpus_clear(cpu_isolated_map
);
5945 for (i
= 1; i
<= ints
[0]; i
++)
5946 if (ints
[i
] < NR_CPUS
)
5947 cpu_set(ints
[i
], cpu_isolated_map
);
5951 __setup("isolcpus=", isolated_cpu_setup
);
5954 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5955 * to a function which identifies what group(along with sched group) a CPU
5956 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5957 * (due to the fact that we keep track of groups covered with a cpumask_t).
5959 * init_sched_build_groups will build a circular linked list of the groups
5960 * covered by the given span, and will set each group's ->cpumask correctly,
5961 * and ->cpu_power to 0.
5964 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5965 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5966 struct sched_group
**sg
))
5968 struct sched_group
*first
= NULL
, *last
= NULL
;
5969 cpumask_t covered
= CPU_MASK_NONE
;
5972 for_each_cpu_mask(i
, span
) {
5973 struct sched_group
*sg
;
5974 int group
= group_fn(i
, cpu_map
, &sg
);
5977 if (cpu_isset(i
, covered
))
5980 sg
->cpumask
= CPU_MASK_NONE
;
5981 sg
->__cpu_power
= 0;
5983 for_each_cpu_mask(j
, span
) {
5984 if (group_fn(j
, cpu_map
, NULL
) != group
)
5987 cpu_set(j
, covered
);
5988 cpu_set(j
, sg
->cpumask
);
5999 #define SD_NODES_PER_DOMAIN 16
6004 * find_next_best_node - find the next node to include in a sched_domain
6005 * @node: node whose sched_domain we're building
6006 * @used_nodes: nodes already in the sched_domain
6008 * Find the next node to include in a given scheduling domain. Simply
6009 * finds the closest node not already in the @used_nodes map.
6011 * Should use nodemask_t.
6013 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6015 int i
, n
, val
, min_val
, best_node
= 0;
6019 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6020 /* Start at @node */
6021 n
= (node
+ i
) % MAX_NUMNODES
;
6023 if (!nr_cpus_node(n
))
6026 /* Skip already used nodes */
6027 if (test_bit(n
, used_nodes
))
6030 /* Simple min distance search */
6031 val
= node_distance(node
, n
);
6033 if (val
< min_val
) {
6039 set_bit(best_node
, used_nodes
);
6044 * sched_domain_node_span - get a cpumask for a node's sched_domain
6045 * @node: node whose cpumask we're constructing
6046 * @size: number of nodes to include in this span
6048 * Given a node, construct a good cpumask for its sched_domain to span. It
6049 * should be one that prevents unnecessary balancing, but also spreads tasks
6052 static cpumask_t
sched_domain_node_span(int node
)
6054 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6055 cpumask_t span
, nodemask
;
6059 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6061 nodemask
= node_to_cpumask(node
);
6062 cpus_or(span
, span
, nodemask
);
6063 set_bit(node
, used_nodes
);
6065 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6066 int next_node
= find_next_best_node(node
, used_nodes
);
6068 nodemask
= node_to_cpumask(next_node
);
6069 cpus_or(span
, span
, nodemask
);
6076 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6079 * SMT sched-domains:
6081 #ifdef CONFIG_SCHED_SMT
6082 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6083 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6086 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6089 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6095 * multi-core sched-domains:
6097 #ifdef CONFIG_SCHED_MC
6098 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6099 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6102 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6104 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6107 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6108 cpus_and(mask
, mask
, *cpu_map
);
6109 group
= first_cpu(mask
);
6111 *sg
= &per_cpu(sched_group_core
, group
);
6114 #elif defined(CONFIG_SCHED_MC)
6116 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6119 *sg
= &per_cpu(sched_group_core
, cpu
);
6124 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6125 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6128 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6131 #ifdef CONFIG_SCHED_MC
6132 cpumask_t mask
= cpu_coregroup_map(cpu
);
6133 cpus_and(mask
, mask
, *cpu_map
);
6134 group
= first_cpu(mask
);
6135 #elif defined(CONFIG_SCHED_SMT)
6136 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6137 cpus_and(mask
, mask
, *cpu_map
);
6138 group
= first_cpu(mask
);
6143 *sg
= &per_cpu(sched_group_phys
, group
);
6149 * The init_sched_build_groups can't handle what we want to do with node
6150 * groups, so roll our own. Now each node has its own list of groups which
6151 * gets dynamically allocated.
6153 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6154 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6156 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6157 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6159 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6160 struct sched_group
**sg
)
6162 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6165 cpus_and(nodemask
, nodemask
, *cpu_map
);
6166 group
= first_cpu(nodemask
);
6169 *sg
= &per_cpu(sched_group_allnodes
, group
);
6173 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6175 struct sched_group
*sg
= group_head
;
6181 for_each_cpu_mask(j
, sg
->cpumask
) {
6182 struct sched_domain
*sd
;
6184 sd
= &per_cpu(phys_domains
, j
);
6185 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6187 * Only add "power" once for each
6193 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6196 } while (sg
!= group_head
);
6201 /* Free memory allocated for various sched_group structures */
6202 static void free_sched_groups(const cpumask_t
*cpu_map
)
6206 for_each_cpu_mask(cpu
, *cpu_map
) {
6207 struct sched_group
**sched_group_nodes
6208 = sched_group_nodes_bycpu
[cpu
];
6210 if (!sched_group_nodes
)
6213 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6214 cpumask_t nodemask
= node_to_cpumask(i
);
6215 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6217 cpus_and(nodemask
, nodemask
, *cpu_map
);
6218 if (cpus_empty(nodemask
))
6228 if (oldsg
!= sched_group_nodes
[i
])
6231 kfree(sched_group_nodes
);
6232 sched_group_nodes_bycpu
[cpu
] = NULL
;
6236 static void free_sched_groups(const cpumask_t
*cpu_map
)
6242 * Initialize sched groups cpu_power.
6244 * cpu_power indicates the capacity of sched group, which is used while
6245 * distributing the load between different sched groups in a sched domain.
6246 * Typically cpu_power for all the groups in a sched domain will be same unless
6247 * there are asymmetries in the topology. If there are asymmetries, group
6248 * having more cpu_power will pickup more load compared to the group having
6251 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6252 * the maximum number of tasks a group can handle in the presence of other idle
6253 * or lightly loaded groups in the same sched domain.
6255 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6257 struct sched_domain
*child
;
6258 struct sched_group
*group
;
6260 WARN_ON(!sd
|| !sd
->groups
);
6262 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6267 sd
->groups
->__cpu_power
= 0;
6270 * For perf policy, if the groups in child domain share resources
6271 * (for example cores sharing some portions of the cache hierarchy
6272 * or SMT), then set this domain groups cpu_power such that each group
6273 * can handle only one task, when there are other idle groups in the
6274 * same sched domain.
6276 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6278 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6279 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6284 * add cpu_power of each child group to this groups cpu_power
6286 group
= child
->groups
;
6288 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6289 group
= group
->next
;
6290 } while (group
!= child
->groups
);
6294 * Build sched domains for a given set of cpus and attach the sched domains
6295 * to the individual cpus
6297 static int build_sched_domains(const cpumask_t
*cpu_map
)
6300 struct root_domain
*rd
;
6302 struct sched_group
**sched_group_nodes
= NULL
;
6303 int sd_allnodes
= 0;
6306 * Allocate the per-node list of sched groups
6308 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6310 if (!sched_group_nodes
) {
6311 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6314 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6317 rd
= alloc_rootdomain(cpu_map
);
6319 printk(KERN_WARNING
"Cannot alloc root domain\n");
6324 * Set up domains for cpus specified by the cpu_map.
6326 for_each_cpu_mask(i
, *cpu_map
) {
6327 struct sched_domain
*sd
= NULL
, *p
;
6328 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6330 cpus_and(nodemask
, nodemask
, *cpu_map
);
6333 if (cpus_weight(*cpu_map
) >
6334 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6335 sd
= &per_cpu(allnodes_domains
, i
);
6336 *sd
= SD_ALLNODES_INIT
;
6337 sd
->span
= *cpu_map
;
6338 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6344 sd
= &per_cpu(node_domains
, i
);
6346 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6350 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6354 sd
= &per_cpu(phys_domains
, i
);
6356 sd
->span
= nodemask
;
6360 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6362 #ifdef CONFIG_SCHED_MC
6364 sd
= &per_cpu(core_domains
, i
);
6366 sd
->span
= cpu_coregroup_map(i
);
6367 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6370 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6373 #ifdef CONFIG_SCHED_SMT
6375 sd
= &per_cpu(cpu_domains
, i
);
6376 *sd
= SD_SIBLING_INIT
;
6377 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6378 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6381 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6385 #ifdef CONFIG_SCHED_SMT
6386 /* Set up CPU (sibling) groups */
6387 for_each_cpu_mask(i
, *cpu_map
) {
6388 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6389 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6390 if (i
!= first_cpu(this_sibling_map
))
6393 init_sched_build_groups(this_sibling_map
, cpu_map
,
6398 #ifdef CONFIG_SCHED_MC
6399 /* Set up multi-core groups */
6400 for_each_cpu_mask(i
, *cpu_map
) {
6401 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6402 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6403 if (i
!= first_cpu(this_core_map
))
6405 init_sched_build_groups(this_core_map
, cpu_map
,
6406 &cpu_to_core_group
);
6410 /* Set up physical groups */
6411 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6412 cpumask_t nodemask
= node_to_cpumask(i
);
6414 cpus_and(nodemask
, nodemask
, *cpu_map
);
6415 if (cpus_empty(nodemask
))
6418 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6422 /* Set up node groups */
6424 init_sched_build_groups(*cpu_map
, cpu_map
,
6425 &cpu_to_allnodes_group
);
6427 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6428 /* Set up node groups */
6429 struct sched_group
*sg
, *prev
;
6430 cpumask_t nodemask
= node_to_cpumask(i
);
6431 cpumask_t domainspan
;
6432 cpumask_t covered
= CPU_MASK_NONE
;
6435 cpus_and(nodemask
, nodemask
, *cpu_map
);
6436 if (cpus_empty(nodemask
)) {
6437 sched_group_nodes
[i
] = NULL
;
6441 domainspan
= sched_domain_node_span(i
);
6442 cpus_and(domainspan
, domainspan
, *cpu_map
);
6444 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6446 printk(KERN_WARNING
"Can not alloc domain group for "
6450 sched_group_nodes
[i
] = sg
;
6451 for_each_cpu_mask(j
, nodemask
) {
6452 struct sched_domain
*sd
;
6454 sd
= &per_cpu(node_domains
, j
);
6457 sg
->__cpu_power
= 0;
6458 sg
->cpumask
= nodemask
;
6460 cpus_or(covered
, covered
, nodemask
);
6463 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6464 cpumask_t tmp
, notcovered
;
6465 int n
= (i
+ j
) % MAX_NUMNODES
;
6467 cpus_complement(notcovered
, covered
);
6468 cpus_and(tmp
, notcovered
, *cpu_map
);
6469 cpus_and(tmp
, tmp
, domainspan
);
6470 if (cpus_empty(tmp
))
6473 nodemask
= node_to_cpumask(n
);
6474 cpus_and(tmp
, tmp
, nodemask
);
6475 if (cpus_empty(tmp
))
6478 sg
= kmalloc_node(sizeof(struct sched_group
),
6482 "Can not alloc domain group for node %d\n", j
);
6485 sg
->__cpu_power
= 0;
6487 sg
->next
= prev
->next
;
6488 cpus_or(covered
, covered
, tmp
);
6495 /* Calculate CPU power for physical packages and nodes */
6496 #ifdef CONFIG_SCHED_SMT
6497 for_each_cpu_mask(i
, *cpu_map
) {
6498 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6500 init_sched_groups_power(i
, sd
);
6503 #ifdef CONFIG_SCHED_MC
6504 for_each_cpu_mask(i
, *cpu_map
) {
6505 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6507 init_sched_groups_power(i
, sd
);
6511 for_each_cpu_mask(i
, *cpu_map
) {
6512 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6514 init_sched_groups_power(i
, sd
);
6518 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6519 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6522 struct sched_group
*sg
;
6524 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6525 init_numa_sched_groups_power(sg
);
6529 /* Attach the domains */
6530 for_each_cpu_mask(i
, *cpu_map
) {
6531 struct sched_domain
*sd
;
6532 #ifdef CONFIG_SCHED_SMT
6533 sd
= &per_cpu(cpu_domains
, i
);
6534 #elif defined(CONFIG_SCHED_MC)
6535 sd
= &per_cpu(core_domains
, i
);
6537 sd
= &per_cpu(phys_domains
, i
);
6539 cpu_attach_domain(sd
, rd
, i
);
6546 free_sched_groups(cpu_map
);
6551 static cpumask_t
*doms_cur
; /* current sched domains */
6552 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6555 * Special case: If a kmalloc of a doms_cur partition (array of
6556 * cpumask_t) fails, then fallback to a single sched domain,
6557 * as determined by the single cpumask_t fallback_doms.
6559 static cpumask_t fallback_doms
;
6562 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6563 * For now this just excludes isolated cpus, but could be used to
6564 * exclude other special cases in the future.
6566 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6571 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6573 doms_cur
= &fallback_doms
;
6574 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6575 err
= build_sched_domains(doms_cur
);
6576 register_sched_domain_sysctl();
6581 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6583 free_sched_groups(cpu_map
);
6587 * Detach sched domains from a group of cpus specified in cpu_map
6588 * These cpus will now be attached to the NULL domain
6590 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6594 unregister_sched_domain_sysctl();
6596 for_each_cpu_mask(i
, *cpu_map
)
6597 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6598 synchronize_sched();
6599 arch_destroy_sched_domains(cpu_map
);
6603 * Partition sched domains as specified by the 'ndoms_new'
6604 * cpumasks in the array doms_new[] of cpumasks. This compares
6605 * doms_new[] to the current sched domain partitioning, doms_cur[].
6606 * It destroys each deleted domain and builds each new domain.
6608 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6609 * The masks don't intersect (don't overlap.) We should setup one
6610 * sched domain for each mask. CPUs not in any of the cpumasks will
6611 * not be load balanced. If the same cpumask appears both in the
6612 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6615 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6616 * ownership of it and will kfree it when done with it. If the caller
6617 * failed the kmalloc call, then it can pass in doms_new == NULL,
6618 * and partition_sched_domains() will fallback to the single partition
6621 * Call with hotplug lock held
6623 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6629 /* always unregister in case we don't destroy any domains */
6630 unregister_sched_domain_sysctl();
6632 if (doms_new
== NULL
) {
6634 doms_new
= &fallback_doms
;
6635 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6638 /* Destroy deleted domains */
6639 for (i
= 0; i
< ndoms_cur
; i
++) {
6640 for (j
= 0; j
< ndoms_new
; j
++) {
6641 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
6644 /* no match - a current sched domain not in new doms_new[] */
6645 detach_destroy_domains(doms_cur
+ i
);
6650 /* Build new domains */
6651 for (i
= 0; i
< ndoms_new
; i
++) {
6652 for (j
= 0; j
< ndoms_cur
; j
++) {
6653 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
6656 /* no match - add a new doms_new */
6657 build_sched_domains(doms_new
+ i
);
6662 /* Remember the new sched domains */
6663 if (doms_cur
!= &fallback_doms
)
6665 doms_cur
= doms_new
;
6666 ndoms_cur
= ndoms_new
;
6668 register_sched_domain_sysctl();
6673 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6674 static int arch_reinit_sched_domains(void)
6679 detach_destroy_domains(&cpu_online_map
);
6680 err
= arch_init_sched_domains(&cpu_online_map
);
6686 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6690 if (buf
[0] != '0' && buf
[0] != '1')
6694 sched_smt_power_savings
= (buf
[0] == '1');
6696 sched_mc_power_savings
= (buf
[0] == '1');
6698 ret
= arch_reinit_sched_domains();
6700 return ret
? ret
: count
;
6703 #ifdef CONFIG_SCHED_MC
6704 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6706 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6708 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6709 const char *buf
, size_t count
)
6711 return sched_power_savings_store(buf
, count
, 0);
6713 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6714 sched_mc_power_savings_store
);
6717 #ifdef CONFIG_SCHED_SMT
6718 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6720 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6722 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6723 const char *buf
, size_t count
)
6725 return sched_power_savings_store(buf
, count
, 1);
6727 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6728 sched_smt_power_savings_store
);
6731 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6735 #ifdef CONFIG_SCHED_SMT
6737 err
= sysfs_create_file(&cls
->kset
.kobj
,
6738 &attr_sched_smt_power_savings
.attr
);
6740 #ifdef CONFIG_SCHED_MC
6741 if (!err
&& mc_capable())
6742 err
= sysfs_create_file(&cls
->kset
.kobj
,
6743 &attr_sched_mc_power_savings
.attr
);
6750 * Force a reinitialization of the sched domains hierarchy. The domains
6751 * and groups cannot be updated in place without racing with the balancing
6752 * code, so we temporarily attach all running cpus to the NULL domain
6753 * which will prevent rebalancing while the sched domains are recalculated.
6755 static int update_sched_domains(struct notifier_block
*nfb
,
6756 unsigned long action
, void *hcpu
)
6759 case CPU_UP_PREPARE
:
6760 case CPU_UP_PREPARE_FROZEN
:
6761 case CPU_DOWN_PREPARE
:
6762 case CPU_DOWN_PREPARE_FROZEN
:
6763 detach_destroy_domains(&cpu_online_map
);
6766 case CPU_UP_CANCELED
:
6767 case CPU_UP_CANCELED_FROZEN
:
6768 case CPU_DOWN_FAILED
:
6769 case CPU_DOWN_FAILED_FROZEN
:
6771 case CPU_ONLINE_FROZEN
:
6773 case CPU_DEAD_FROZEN
:
6775 * Fall through and re-initialise the domains.
6782 /* The hotplug lock is already held by cpu_up/cpu_down */
6783 arch_init_sched_domains(&cpu_online_map
);
6788 void __init
sched_init_smp(void)
6790 cpumask_t non_isolated_cpus
;
6793 arch_init_sched_domains(&cpu_online_map
);
6794 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6795 if (cpus_empty(non_isolated_cpus
))
6796 cpu_set(smp_processor_id(), non_isolated_cpus
);
6798 /* XXX: Theoretical race here - CPU may be hotplugged now */
6799 hotcpu_notifier(update_sched_domains
, 0);
6801 /* Move init over to a non-isolated CPU */
6802 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6804 sched_init_granularity();
6806 #ifdef CONFIG_FAIR_GROUP_SCHED
6807 if (nr_cpu_ids
== 1)
6810 lb_monitor_task
= kthread_create(load_balance_monitor
, NULL
,
6812 if (!IS_ERR(lb_monitor_task
)) {
6813 lb_monitor_task
->flags
|= PF_NOFREEZE
;
6814 wake_up_process(lb_monitor_task
);
6816 printk(KERN_ERR
"Could not create load balance monitor thread"
6817 "(error = %ld) \n", PTR_ERR(lb_monitor_task
));
6822 void __init
sched_init_smp(void)
6824 sched_init_granularity();
6826 #endif /* CONFIG_SMP */
6828 int in_sched_functions(unsigned long addr
)
6830 return in_lock_functions(addr
) ||
6831 (addr
>= (unsigned long)__sched_text_start
6832 && addr
< (unsigned long)__sched_text_end
);
6835 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6837 cfs_rq
->tasks_timeline
= RB_ROOT
;
6838 #ifdef CONFIG_FAIR_GROUP_SCHED
6841 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6844 void __init
sched_init(void)
6846 int highest_cpu
= 0;
6850 init_defrootdomain();
6853 for_each_possible_cpu(i
) {
6854 struct rt_prio_array
*array
;
6858 spin_lock_init(&rq
->lock
);
6859 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6862 init_cfs_rq(&rq
->cfs
, rq
);
6863 #ifdef CONFIG_FAIR_GROUP_SCHED
6864 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6866 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6867 struct sched_entity
*se
=
6868 &per_cpu(init_sched_entity
, i
);
6870 init_cfs_rq_p
[i
] = cfs_rq
;
6871 init_cfs_rq(cfs_rq
, rq
);
6872 cfs_rq
->tg
= &init_task_group
;
6873 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6874 &rq
->leaf_cfs_rq_list
);
6876 init_sched_entity_p
[i
] = se
;
6877 se
->cfs_rq
= &rq
->cfs
;
6879 se
->load
.weight
= init_task_group_load
;
6880 se
->load
.inv_weight
=
6881 div64_64(1ULL<<32, init_task_group_load
);
6884 init_task_group
.shares
= init_task_group_load
;
6887 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6888 rq
->cpu_load
[j
] = 0;
6892 rq_attach_root(rq
, &def_root_domain
);
6893 rq
->active_balance
= 0;
6894 rq
->next_balance
= jiffies
;
6897 rq
->migration_thread
= NULL
;
6898 INIT_LIST_HEAD(&rq
->migration_queue
);
6899 rq
->rt
.highest_prio
= MAX_RT_PRIO
;
6900 rq
->rt
.overloaded
= 0;
6902 atomic_set(&rq
->nr_iowait
, 0);
6904 array
= &rq
->rt
.active
;
6905 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6906 INIT_LIST_HEAD(array
->queue
+ j
);
6907 __clear_bit(j
, array
->bitmap
);
6910 /* delimiter for bitsearch: */
6911 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6914 set_load_weight(&init_task
);
6916 #ifdef CONFIG_PREEMPT_NOTIFIERS
6917 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6921 nr_cpu_ids
= highest_cpu
+ 1;
6922 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6925 #ifdef CONFIG_RT_MUTEXES
6926 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6930 * The boot idle thread does lazy MMU switching as well:
6932 atomic_inc(&init_mm
.mm_count
);
6933 enter_lazy_tlb(&init_mm
, current
);
6936 * Make us the idle thread. Technically, schedule() should not be
6937 * called from this thread, however somewhere below it might be,
6938 * but because we are the idle thread, we just pick up running again
6939 * when this runqueue becomes "idle".
6941 init_idle(current
, smp_processor_id());
6943 * During early bootup we pretend to be a normal task:
6945 current
->sched_class
= &fair_sched_class
;
6948 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6949 void __might_sleep(char *file
, int line
)
6952 static unsigned long prev_jiffy
; /* ratelimiting */
6954 if ((in_atomic() || irqs_disabled()) &&
6955 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6956 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6958 prev_jiffy
= jiffies
;
6959 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6960 " context at %s:%d\n", file
, line
);
6961 printk("in_atomic():%d, irqs_disabled():%d\n",
6962 in_atomic(), irqs_disabled());
6963 debug_show_held_locks(current
);
6964 if (irqs_disabled())
6965 print_irqtrace_events(current
);
6970 EXPORT_SYMBOL(__might_sleep
);
6973 #ifdef CONFIG_MAGIC_SYSRQ
6974 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6977 update_rq_clock(rq
);
6978 on_rq
= p
->se
.on_rq
;
6980 deactivate_task(rq
, p
, 0);
6981 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6983 activate_task(rq
, p
, 0);
6984 resched_task(rq
->curr
);
6988 void normalize_rt_tasks(void)
6990 struct task_struct
*g
, *p
;
6991 unsigned long flags
;
6994 read_lock_irq(&tasklist_lock
);
6995 do_each_thread(g
, p
) {
6997 * Only normalize user tasks:
7002 p
->se
.exec_start
= 0;
7003 #ifdef CONFIG_SCHEDSTATS
7004 p
->se
.wait_start
= 0;
7005 p
->se
.sleep_start
= 0;
7006 p
->se
.block_start
= 0;
7008 task_rq(p
)->clock
= 0;
7012 * Renice negative nice level userspace
7015 if (TASK_NICE(p
) < 0 && p
->mm
)
7016 set_user_nice(p
, 0);
7020 spin_lock_irqsave(&p
->pi_lock
, flags
);
7021 rq
= __task_rq_lock(p
);
7023 normalize_task(rq
, p
);
7025 __task_rq_unlock(rq
);
7026 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
7027 } while_each_thread(g
, p
);
7029 read_unlock_irq(&tasklist_lock
);
7032 #endif /* CONFIG_MAGIC_SYSRQ */
7036 * These functions are only useful for the IA64 MCA handling.
7038 * They can only be called when the whole system has been
7039 * stopped - every CPU needs to be quiescent, and no scheduling
7040 * activity can take place. Using them for anything else would
7041 * be a serious bug, and as a result, they aren't even visible
7042 * under any other configuration.
7046 * curr_task - return the current task for a given cpu.
7047 * @cpu: the processor in question.
7049 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7051 struct task_struct
*curr_task(int cpu
)
7053 return cpu_curr(cpu
);
7057 * set_curr_task - set the current task for a given cpu.
7058 * @cpu: the processor in question.
7059 * @p: the task pointer to set.
7061 * Description: This function must only be used when non-maskable interrupts
7062 * are serviced on a separate stack. It allows the architecture to switch the
7063 * notion of the current task on a cpu in a non-blocking manner. This function
7064 * must be called with all CPU's synchronized, and interrupts disabled, the
7065 * and caller must save the original value of the current task (see
7066 * curr_task() above) and restore that value before reenabling interrupts and
7067 * re-starting the system.
7069 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7071 void set_curr_task(int cpu
, struct task_struct
*p
)
7078 #ifdef CONFIG_FAIR_GROUP_SCHED
7082 * distribute shares of all task groups among their schedulable entities,
7083 * to reflect load distrbution across cpus.
7085 static int rebalance_shares(struct sched_domain
*sd
, int this_cpu
)
7087 struct cfs_rq
*cfs_rq
;
7088 struct rq
*rq
= cpu_rq(this_cpu
);
7089 cpumask_t sdspan
= sd
->span
;
7092 /* Walk thr' all the task groups that we have */
7093 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
7095 unsigned long total_load
= 0, total_shares
;
7096 struct task_group
*tg
= cfs_rq
->tg
;
7098 /* Gather total task load of this group across cpus */
7099 for_each_cpu_mask(i
, sdspan
)
7100 total_load
+= tg
->cfs_rq
[i
]->load
.weight
;
7102 /* Nothing to do if this group has no load */
7107 * tg->shares represents the number of cpu shares the task group
7108 * is eligible to hold on a single cpu. On N cpus, it is
7109 * eligible to hold (N * tg->shares) number of cpu shares.
7111 total_shares
= tg
->shares
* cpus_weight(sdspan
);
7114 * redistribute total_shares across cpus as per the task load
7117 for_each_cpu_mask(i
, sdspan
) {
7118 unsigned long local_load
, local_shares
;
7120 local_load
= tg
->cfs_rq
[i
]->load
.weight
;
7121 local_shares
= (local_load
* total_shares
) / total_load
;
7123 local_shares
= MIN_GROUP_SHARES
;
7124 if (local_shares
== tg
->se
[i
]->load
.weight
)
7127 spin_lock_irq(&cpu_rq(i
)->lock
);
7128 set_se_shares(tg
->se
[i
], local_shares
);
7129 spin_unlock_irq(&cpu_rq(i
)->lock
);
7138 * How frequently should we rebalance_shares() across cpus?
7140 * The more frequently we rebalance shares, the more accurate is the fairness
7141 * of cpu bandwidth distribution between task groups. However higher frequency
7142 * also implies increased scheduling overhead.
7144 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7145 * consecutive calls to rebalance_shares() in the same sched domain.
7147 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7148 * consecutive calls to rebalance_shares() in the same sched domain.
7150 * These settings allows for the appropriate tradeoff between accuracy of
7151 * fairness and the associated overhead.
7155 /* default: 8ms, units: milliseconds */
7156 const_debug
unsigned int sysctl_sched_min_bal_int_shares
= 8;
7158 /* default: 128ms, units: milliseconds */
7159 const_debug
unsigned int sysctl_sched_max_bal_int_shares
= 128;
7161 /* kernel thread that runs rebalance_shares() periodically */
7162 static int load_balance_monitor(void *unused
)
7164 unsigned int timeout
= sysctl_sched_min_bal_int_shares
;
7165 struct sched_param schedparm
;
7169 * We don't want this thread's execution to be limited by the shares
7170 * assigned to default group (init_task_group). Hence make it run
7171 * as a SCHED_RR RT task at the lowest priority.
7173 schedparm
.sched_priority
= 1;
7174 ret
= sched_setscheduler(current
, SCHED_RR
, &schedparm
);
7176 printk(KERN_ERR
"Couldn't set SCHED_RR policy for load balance"
7177 " monitor thread (error = %d) \n", ret
);
7179 while (!kthread_should_stop()) {
7180 int i
, cpu
, balanced
= 1;
7182 /* Prevent cpus going down or coming up */
7184 /* lockout changes to doms_cur[] array */
7187 * Enter a rcu read-side critical section to safely walk rq->sd
7188 * chain on various cpus and to walk task group list
7189 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7193 for (i
= 0; i
< ndoms_cur
; i
++) {
7194 cpumask_t cpumap
= doms_cur
[i
];
7195 struct sched_domain
*sd
= NULL
, *sd_prev
= NULL
;
7197 cpu
= first_cpu(cpumap
);
7199 /* Find the highest domain at which to balance shares */
7200 for_each_domain(cpu
, sd
) {
7201 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7207 /* sd == NULL? No load balance reqd in this domain */
7211 balanced
&= rebalance_shares(sd
, cpu
);
7220 timeout
= sysctl_sched_min_bal_int_shares
;
7221 else if (timeout
< sysctl_sched_max_bal_int_shares
)
7224 msleep_interruptible(timeout
);
7229 #endif /* CONFIG_SMP */
7231 /* allocate runqueue etc for a new task group */
7232 struct task_group
*sched_create_group(void)
7234 struct task_group
*tg
;
7235 struct cfs_rq
*cfs_rq
;
7236 struct sched_entity
*se
;
7240 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7242 return ERR_PTR(-ENOMEM
);
7244 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
7247 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
7251 for_each_possible_cpu(i
) {
7254 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
7259 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
7264 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
7265 memset(se
, 0, sizeof(struct sched_entity
));
7267 tg
->cfs_rq
[i
] = cfs_rq
;
7268 init_cfs_rq(cfs_rq
, rq
);
7272 se
->cfs_rq
= &rq
->cfs
;
7274 se
->load
.weight
= NICE_0_LOAD
;
7275 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
7279 tg
->shares
= NICE_0_LOAD
;
7281 lock_task_group_list();
7282 for_each_possible_cpu(i
) {
7284 cfs_rq
= tg
->cfs_rq
[i
];
7285 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7287 unlock_task_group_list();
7292 for_each_possible_cpu(i
) {
7294 kfree(tg
->cfs_rq
[i
]);
7302 return ERR_PTR(-ENOMEM
);
7305 /* rcu callback to free various structures associated with a task group */
7306 static void free_sched_group(struct rcu_head
*rhp
)
7308 struct task_group
*tg
= container_of(rhp
, struct task_group
, rcu
);
7309 struct cfs_rq
*cfs_rq
;
7310 struct sched_entity
*se
;
7313 /* now it should be safe to free those cfs_rqs */
7314 for_each_possible_cpu(i
) {
7315 cfs_rq
= tg
->cfs_rq
[i
];
7327 /* Destroy runqueue etc associated with a task group */
7328 void sched_destroy_group(struct task_group
*tg
)
7330 struct cfs_rq
*cfs_rq
= NULL
;
7333 lock_task_group_list();
7334 for_each_possible_cpu(i
) {
7335 cfs_rq
= tg
->cfs_rq
[i
];
7336 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7338 unlock_task_group_list();
7342 /* wait for possible concurrent references to cfs_rqs complete */
7343 call_rcu(&tg
->rcu
, free_sched_group
);
7346 /* change task's runqueue when it moves between groups.
7347 * The caller of this function should have put the task in its new group
7348 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7349 * reflect its new group.
7351 void sched_move_task(struct task_struct
*tsk
)
7354 unsigned long flags
;
7357 rq
= task_rq_lock(tsk
, &flags
);
7359 if (tsk
->sched_class
!= &fair_sched_class
) {
7360 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7364 update_rq_clock(rq
);
7366 running
= task_current(rq
, tsk
);
7367 on_rq
= tsk
->se
.on_rq
;
7370 dequeue_task(rq
, tsk
, 0);
7371 if (unlikely(running
))
7372 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7375 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7378 if (unlikely(running
))
7379 tsk
->sched_class
->set_curr_task(rq
);
7380 enqueue_task(rq
, tsk
, 0);
7384 task_rq_unlock(rq
, &flags
);
7387 /* rq->lock to be locked by caller */
7388 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7390 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7391 struct rq
*rq
= cfs_rq
->rq
;
7395 shares
= MIN_GROUP_SHARES
;
7399 dequeue_entity(cfs_rq
, se
, 0);
7400 dec_cpu_load(rq
, se
->load
.weight
);
7403 se
->load
.weight
= shares
;
7404 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7407 enqueue_entity(cfs_rq
, se
, 0);
7408 inc_cpu_load(rq
, se
->load
.weight
);
7412 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7415 struct cfs_rq
*cfs_rq
;
7418 lock_task_group_list();
7419 if (tg
->shares
== shares
)
7422 if (shares
< MIN_GROUP_SHARES
)
7423 shares
= MIN_GROUP_SHARES
;
7426 * Prevent any load balance activity (rebalance_shares,
7427 * load_balance_fair) from referring to this group first,
7428 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7430 for_each_possible_cpu(i
) {
7431 cfs_rq
= tg
->cfs_rq
[i
];
7432 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7435 /* wait for any ongoing reference to this group to finish */
7436 synchronize_sched();
7439 * Now we are free to modify the group's share on each cpu
7440 * w/o tripping rebalance_share or load_balance_fair.
7442 tg
->shares
= shares
;
7443 for_each_possible_cpu(i
) {
7444 spin_lock_irq(&cpu_rq(i
)->lock
);
7445 set_se_shares(tg
->se
[i
], shares
);
7446 spin_unlock_irq(&cpu_rq(i
)->lock
);
7450 * Enable load balance activity on this group, by inserting it back on
7451 * each cpu's rq->leaf_cfs_rq_list.
7453 for_each_possible_cpu(i
) {
7455 cfs_rq
= tg
->cfs_rq
[i
];
7456 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7459 unlock_task_group_list();
7463 unsigned long sched_group_shares(struct task_group
*tg
)
7468 #endif /* CONFIG_FAIR_GROUP_SCHED */
7470 #ifdef CONFIG_FAIR_CGROUP_SCHED
7472 /* return corresponding task_group object of a cgroup */
7473 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7475 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7476 struct task_group
, css
);
7479 static struct cgroup_subsys_state
*
7480 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7482 struct task_group
*tg
;
7484 if (!cgrp
->parent
) {
7485 /* This is early initialization for the top cgroup */
7486 init_task_group
.css
.cgroup
= cgrp
;
7487 return &init_task_group
.css
;
7490 /* we support only 1-level deep hierarchical scheduler atm */
7491 if (cgrp
->parent
->parent
)
7492 return ERR_PTR(-EINVAL
);
7494 tg
= sched_create_group();
7496 return ERR_PTR(-ENOMEM
);
7498 /* Bind the cgroup to task_group object we just created */
7499 tg
->css
.cgroup
= cgrp
;
7505 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7507 struct task_group
*tg
= cgroup_tg(cgrp
);
7509 sched_destroy_group(tg
);
7513 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7514 struct task_struct
*tsk
)
7516 /* We don't support RT-tasks being in separate groups */
7517 if (tsk
->sched_class
!= &fair_sched_class
)
7524 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7525 struct cgroup
*old_cont
, struct task_struct
*tsk
)
7527 sched_move_task(tsk
);
7530 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7533 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7536 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7538 struct task_group
*tg
= cgroup_tg(cgrp
);
7540 return (u64
) tg
->shares
;
7543 static struct cftype cpu_files
[] = {
7546 .read_uint
= cpu_shares_read_uint
,
7547 .write_uint
= cpu_shares_write_uint
,
7551 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7553 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7556 struct cgroup_subsys cpu_cgroup_subsys
= {
7558 .create
= cpu_cgroup_create
,
7559 .destroy
= cpu_cgroup_destroy
,
7560 .can_attach
= cpu_cgroup_can_attach
,
7561 .attach
= cpu_cgroup_attach
,
7562 .populate
= cpu_cgroup_populate
,
7563 .subsys_id
= cpu_cgroup_subsys_id
,
7567 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7569 #ifdef CONFIG_CGROUP_CPUACCT
7572 * CPU accounting code for task groups.
7574 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7575 * (balbir@in.ibm.com).
7578 /* track cpu usage of a group of tasks */
7580 struct cgroup_subsys_state css
;
7581 /* cpuusage holds pointer to a u64-type object on every cpu */
7585 struct cgroup_subsys cpuacct_subsys
;
7587 /* return cpu accounting group corresponding to this container */
7588 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
7590 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
7591 struct cpuacct
, css
);
7594 /* return cpu accounting group to which this task belongs */
7595 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
7597 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
7598 struct cpuacct
, css
);
7601 /* create a new cpu accounting group */
7602 static struct cgroup_subsys_state
*cpuacct_create(
7603 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7605 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7608 return ERR_PTR(-ENOMEM
);
7610 ca
->cpuusage
= alloc_percpu(u64
);
7611 if (!ca
->cpuusage
) {
7613 return ERR_PTR(-ENOMEM
);
7619 /* destroy an existing cpu accounting group */
7621 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7623 struct cpuacct
*ca
= cgroup_ca(cont
);
7625 free_percpu(ca
->cpuusage
);
7629 /* return total cpu usage (in nanoseconds) of a group */
7630 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
7632 struct cpuacct
*ca
= cgroup_ca(cont
);
7633 u64 totalcpuusage
= 0;
7636 for_each_possible_cpu(i
) {
7637 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
7640 * Take rq->lock to make 64-bit addition safe on 32-bit
7643 spin_lock_irq(&cpu_rq(i
)->lock
);
7644 totalcpuusage
+= *cpuusage
;
7645 spin_unlock_irq(&cpu_rq(i
)->lock
);
7648 return totalcpuusage
;
7651 static struct cftype files
[] = {
7654 .read_uint
= cpuusage_read
,
7658 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7660 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
7664 * charge this task's execution time to its accounting group.
7666 * called with rq->lock held.
7668 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
7672 if (!cpuacct_subsys
.active
)
7677 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
7679 *cpuusage
+= cputime
;
7683 struct cgroup_subsys cpuacct_subsys
= {
7685 .create
= cpuacct_create
,
7686 .destroy
= cpuacct_destroy
,
7687 .populate
= cpuacct_populate
,
7688 .subsys_id
= cpuacct_subsys_id
,
7690 #endif /* CONFIG_CGROUP_CPUACCT */