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>
68 #include <linux/hrtimer.h>
71 #include <asm/irq_regs.h>
74 * Scheduler clock - returns current time in nanosec units.
75 * This is default implementation.
76 * Architectures and sub-architectures can override this.
78 unsigned long long __attribute__((weak
)) sched_clock(void)
80 return (unsigned long long)jiffies
* (NSEC_PER_SEC
/ HZ
);
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
122 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
124 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
128 * Each time a sched group cpu_power is changed,
129 * we must compute its reciprocal value
131 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
133 sg
->__cpu_power
+= val
;
134 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
138 static inline int rt_policy(int policy
)
140 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
145 static inline int task_has_rt_policy(struct task_struct
*p
)
147 return rt_policy(p
->policy
);
151 * This is the priority-queue data structure of the RT scheduling class:
153 struct rt_prio_array
{
154 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
155 struct list_head queue
[MAX_RT_PRIO
];
158 #ifdef CONFIG_FAIR_GROUP_SCHED
160 #include <linux/cgroup.h>
164 static LIST_HEAD(task_groups
);
166 /* task group related information */
168 #ifdef CONFIG_FAIR_CGROUP_SCHED
169 struct cgroup_subsys_state css
;
171 /* schedulable entities of this group on each cpu */
172 struct sched_entity
**se
;
173 /* runqueue "owned" by this group on each cpu */
174 struct cfs_rq
**cfs_rq
;
176 struct sched_rt_entity
**rt_se
;
177 struct rt_rq
**rt_rq
;
182 * shares assigned to a task group governs how much of cpu bandwidth
183 * is allocated to the group. The more shares a group has, the more is
184 * the cpu bandwidth allocated to it.
186 * For ex, lets say that there are three task groups, A, B and C which
187 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
188 * cpu bandwidth allocated by the scheduler to task groups A, B and C
191 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
192 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
193 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
195 * The weight assigned to a task group's schedulable entities on every
196 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
197 * group's shares. For ex: lets say that task group A has been
198 * assigned shares of 1000 and there are two CPUs in a system. Then,
200 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
202 * Note: It's not necessary that each of a task's group schedulable
203 * entity have the same weight on all CPUs. If the group
204 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
205 * better distribution of weight could be:
207 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
208 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
210 * rebalance_shares() is responsible for distributing the shares of a
211 * task groups like this among the group's schedulable entities across
215 unsigned long shares
;
218 struct list_head list
;
221 /* Default task group's sched entity on each cpu */
222 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
223 /* Default task group's cfs_rq on each cpu */
224 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
226 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
227 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
229 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
230 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
232 static struct sched_rt_entity
*init_sched_rt_entity_p
[NR_CPUS
];
233 static struct rt_rq
*init_rt_rq_p
[NR_CPUS
];
235 /* task_group_lock serializes add/remove of task groups and also changes to
236 * a task group's cpu shares.
238 static DEFINE_SPINLOCK(task_group_lock
);
240 /* doms_cur_mutex serializes access to doms_cur[] array */
241 static DEFINE_MUTEX(doms_cur_mutex
);
244 /* kernel thread that runs rebalance_shares() periodically */
245 static struct task_struct
*lb_monitor_task
;
246 static int load_balance_monitor(void *unused
);
249 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
);
251 /* Default task group.
252 * Every task in system belong to this group at bootup.
254 struct task_group init_task_group
= {
255 .se
= init_sched_entity_p
,
256 .cfs_rq
= init_cfs_rq_p
,
258 .rt_se
= init_sched_rt_entity_p
,
259 .rt_rq
= init_rt_rq_p
,
262 #ifdef CONFIG_FAIR_USER_SCHED
263 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
265 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
268 #define MIN_GROUP_SHARES 2
270 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
272 /* return group to which a task belongs */
273 static inline struct task_group
*task_group(struct task_struct
*p
)
275 struct task_group
*tg
;
277 #ifdef CONFIG_FAIR_USER_SCHED
279 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
280 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
281 struct task_group
, css
);
283 tg
= &init_task_group
;
288 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
289 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
291 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
292 p
->se
.parent
= task_group(p
)->se
[cpu
];
294 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
295 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
298 static inline void lock_doms_cur(void)
300 mutex_lock(&doms_cur_mutex
);
303 static inline void unlock_doms_cur(void)
305 mutex_unlock(&doms_cur_mutex
);
310 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
311 static inline void lock_doms_cur(void) { }
312 static inline void unlock_doms_cur(void) { }
314 #endif /* CONFIG_FAIR_GROUP_SCHED */
316 /* CFS-related fields in a runqueue */
318 struct load_weight load
;
319 unsigned long nr_running
;
324 struct rb_root tasks_timeline
;
325 struct rb_node
*rb_leftmost
;
326 struct rb_node
*rb_load_balance_curr
;
327 /* 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity
*curr
;
332 unsigned long nr_spread_over
;
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
338 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
339 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
340 * (like users, containers etc.)
342 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
343 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list
;
346 struct task_group
*tg
; /* group that "owns" this runqueue */
350 /* Real-Time classes' related field in a runqueue: */
352 struct rt_prio_array active
;
353 unsigned long rt_nr_running
;
354 #if defined CONFIG_SMP || defined CONFIG_FAIR_GROUP_SCHED
355 int highest_prio
; /* highest queued rt task prio */
358 unsigned long rt_nr_migratory
;
364 #ifdef CONFIG_FAIR_GROUP_SCHED
365 unsigned long rt_nr_boosted
;
368 struct list_head leaf_rt_rq_list
;
369 struct task_group
*tg
;
370 struct sched_rt_entity
*rt_se
;
377 * We add the notion of a root-domain which will be used to define per-domain
378 * variables. Each exclusive cpuset essentially defines an island domain by
379 * fully partitioning the member cpus from any other cpuset. Whenever a new
380 * exclusive cpuset is created, we also create and attach a new root-domain
390 * The "RT overload" flag: it gets set if a CPU has more than
391 * one runnable RT task.
398 * By default the system creates a single root-domain with all cpus as
399 * members (mimicking the global state we have today).
401 static struct root_domain def_root_domain
;
406 * This is the main, per-CPU runqueue data structure.
408 * Locking rule: those places that want to lock multiple runqueues
409 * (such as the load balancing or the thread migration code), lock
410 * acquire operations must be ordered by ascending &runqueue.
417 * nr_running and cpu_load should be in the same cacheline because
418 * remote CPUs use both these fields when doing load calculation.
420 unsigned long nr_running
;
421 #define CPU_LOAD_IDX_MAX 5
422 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
423 unsigned char idle_at_tick
;
425 unsigned char in_nohz_recently
;
427 /* capture load from *all* tasks on this cpu: */
428 struct load_weight load
;
429 unsigned long nr_load_updates
;
434 u64 rt_period_expire
;
437 #ifdef CONFIG_FAIR_GROUP_SCHED
438 /* list of leaf cfs_rq on this cpu: */
439 struct list_head leaf_cfs_rq_list
;
440 struct list_head leaf_rt_rq_list
;
444 * This is part of a global counter where only the total sum
445 * over all CPUs matters. A task can increase this counter on
446 * one CPU and if it got migrated afterwards it may decrease
447 * it on another CPU. Always updated under the runqueue lock:
449 unsigned long nr_uninterruptible
;
451 struct task_struct
*curr
, *idle
;
452 unsigned long next_balance
;
453 struct mm_struct
*prev_mm
;
455 u64 clock
, prev_clock_raw
;
458 unsigned int clock_warps
, clock_overflows
, clock_underflows
;
460 unsigned int clock_deep_idle_events
;
466 struct root_domain
*rd
;
467 struct sched_domain
*sd
;
469 /* For active balancing */
472 /* cpu of this runqueue: */
475 struct task_struct
*migration_thread
;
476 struct list_head migration_queue
;
479 #ifdef CONFIG_SCHED_HRTICK
480 unsigned long hrtick_flags
;
481 ktime_t hrtick_expire
;
482 struct hrtimer hrtick_timer
;
485 #ifdef CONFIG_SCHEDSTATS
487 struct sched_info rq_sched_info
;
489 /* sys_sched_yield() stats */
490 unsigned int yld_exp_empty
;
491 unsigned int yld_act_empty
;
492 unsigned int yld_both_empty
;
493 unsigned int yld_count
;
495 /* schedule() stats */
496 unsigned int sched_switch
;
497 unsigned int sched_count
;
498 unsigned int sched_goidle
;
500 /* try_to_wake_up() stats */
501 unsigned int ttwu_count
;
502 unsigned int ttwu_local
;
505 unsigned int bkl_count
;
507 struct lock_class_key rq_lock_key
;
510 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
512 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
514 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
517 static inline int cpu_of(struct rq
*rq
)
527 * Update the per-runqueue clock, as finegrained as the platform can give
528 * us, but without assuming monotonicity, etc.:
530 static void __update_rq_clock(struct rq
*rq
)
532 u64 prev_raw
= rq
->prev_clock_raw
;
533 u64 now
= sched_clock();
534 s64 delta
= now
- prev_raw
;
535 u64 clock
= rq
->clock
;
537 #ifdef CONFIG_SCHED_DEBUG
538 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
541 * Protect against sched_clock() occasionally going backwards:
543 if (unlikely(delta
< 0)) {
548 * Catch too large forward jumps too:
550 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
551 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
552 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
555 rq
->clock_overflows
++;
557 if (unlikely(delta
> rq
->clock_max_delta
))
558 rq
->clock_max_delta
= delta
;
563 rq
->prev_clock_raw
= now
;
567 static void update_rq_clock(struct rq
*rq
)
569 if (likely(smp_processor_id() == cpu_of(rq
)))
570 __update_rq_clock(rq
);
574 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
575 * See detach_destroy_domains: synchronize_sched for details.
577 * The domain tree of any CPU may only be accessed from within
578 * preempt-disabled sections.
580 #define for_each_domain(cpu, __sd) \
581 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
583 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
584 #define this_rq() (&__get_cpu_var(runqueues))
585 #define task_rq(p) cpu_rq(task_cpu(p))
586 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
588 unsigned long rt_needs_cpu(int cpu
)
590 struct rq
*rq
= cpu_rq(cpu
);
593 if (!rq
->rt_throttled
)
596 if (rq
->clock
> rq
->rt_period_expire
)
599 delta
= rq
->rt_period_expire
- rq
->clock
;
600 do_div(delta
, NSEC_PER_SEC
/ HZ
);
602 return (unsigned long)delta
;
606 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
608 #ifdef CONFIG_SCHED_DEBUG
609 # define const_debug __read_mostly
611 # define const_debug static const
615 * Debugging: various feature bits
618 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
619 SCHED_FEAT_WAKEUP_PREEMPT
= 2,
620 SCHED_FEAT_START_DEBIT
= 4,
621 SCHED_FEAT_TREE_AVG
= 8,
622 SCHED_FEAT_APPROX_AVG
= 16,
623 SCHED_FEAT_HRTICK
= 32,
624 SCHED_FEAT_DOUBLE_TICK
= 64,
627 const_debug
unsigned int sysctl_sched_features
=
628 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
629 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
630 SCHED_FEAT_START_DEBIT
* 1 |
631 SCHED_FEAT_TREE_AVG
* 0 |
632 SCHED_FEAT_APPROX_AVG
* 0 |
633 SCHED_FEAT_HRTICK
* 1 |
634 SCHED_FEAT_DOUBLE_TICK
* 0;
636 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
639 * Number of tasks to iterate in a single balance run.
640 * Limited because this is done with IRQs disabled.
642 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
645 * period over which we measure -rt task cpu usage in us.
648 unsigned int sysctl_sched_rt_period
= 1000000;
651 * part of the period that we allow rt tasks to run in us.
654 int sysctl_sched_rt_runtime
= 950000;
657 * single value that denotes runtime == period, ie unlimited time.
659 #define RUNTIME_INF ((u64)~0ULL)
662 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
663 * clock constructed from sched_clock():
665 unsigned long long cpu_clock(int cpu
)
667 unsigned long long now
;
671 local_irq_save(flags
);
674 * Only call sched_clock() if the scheduler has already been
675 * initialized (some code might call cpu_clock() very early):
680 local_irq_restore(flags
);
684 EXPORT_SYMBOL_GPL(cpu_clock
);
686 #ifndef prepare_arch_switch
687 # define prepare_arch_switch(next) do { } while (0)
689 #ifndef finish_arch_switch
690 # define finish_arch_switch(prev) do { } while (0)
693 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
695 return rq
->curr
== p
;
698 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
699 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
701 return task_current(rq
, p
);
704 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
708 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
710 #ifdef CONFIG_DEBUG_SPINLOCK
711 /* this is a valid case when another task releases the spinlock */
712 rq
->lock
.owner
= current
;
715 * If we are tracking spinlock dependencies then we have to
716 * fix up the runqueue lock - which gets 'carried over' from
719 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
721 spin_unlock_irq(&rq
->lock
);
724 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
725 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
730 return task_current(rq
, p
);
734 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
738 * We can optimise this out completely for !SMP, because the
739 * SMP rebalancing from interrupt is the only thing that cares
744 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
745 spin_unlock_irq(&rq
->lock
);
747 spin_unlock(&rq
->lock
);
751 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
755 * After ->oncpu is cleared, the task can be moved to a different CPU.
756 * We must ensure this doesn't happen until the switch is completely
762 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
766 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
769 * __task_rq_lock - lock the runqueue a given task resides on.
770 * Must be called interrupts disabled.
772 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
776 struct rq
*rq
= task_rq(p
);
777 spin_lock(&rq
->lock
);
778 if (likely(rq
== task_rq(p
)))
780 spin_unlock(&rq
->lock
);
785 * task_rq_lock - lock the runqueue a given task resides on and disable
786 * interrupts. Note the ordering: we can safely lookup the task_rq without
787 * explicitly disabling preemption.
789 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
795 local_irq_save(*flags
);
797 spin_lock(&rq
->lock
);
798 if (likely(rq
== task_rq(p
)))
800 spin_unlock_irqrestore(&rq
->lock
, *flags
);
804 static void __task_rq_unlock(struct rq
*rq
)
807 spin_unlock(&rq
->lock
);
810 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
813 spin_unlock_irqrestore(&rq
->lock
, *flags
);
817 * this_rq_lock - lock this runqueue and disable interrupts.
819 static struct rq
*this_rq_lock(void)
826 spin_lock(&rq
->lock
);
832 * We are going deep-idle (irqs are disabled):
834 void sched_clock_idle_sleep_event(void)
836 struct rq
*rq
= cpu_rq(smp_processor_id());
838 spin_lock(&rq
->lock
);
839 __update_rq_clock(rq
);
840 spin_unlock(&rq
->lock
);
841 rq
->clock_deep_idle_events
++;
843 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
846 * We just idled delta nanoseconds (called with irqs disabled):
848 void sched_clock_idle_wakeup_event(u64 delta_ns
)
850 struct rq
*rq
= cpu_rq(smp_processor_id());
851 u64 now
= sched_clock();
853 rq
->idle_clock
+= delta_ns
;
855 * Override the previous timestamp and ignore all
856 * sched_clock() deltas that occured while we idled,
857 * and use the PM-provided delta_ns to advance the
860 spin_lock(&rq
->lock
);
861 rq
->prev_clock_raw
= now
;
862 rq
->clock
+= delta_ns
;
863 spin_unlock(&rq
->lock
);
864 touch_softlockup_watchdog();
866 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
868 static void __resched_task(struct task_struct
*p
, int tif_bit
);
870 static inline void resched_task(struct task_struct
*p
)
872 __resched_task(p
, TIF_NEED_RESCHED
);
875 #ifdef CONFIG_SCHED_HRTICK
877 * Use HR-timers to deliver accurate preemption points.
879 * Its all a bit involved since we cannot program an hrt while holding the
880 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
883 * When we get rescheduled we reprogram the hrtick_timer outside of the
886 static inline void resched_hrt(struct task_struct
*p
)
888 __resched_task(p
, TIF_HRTICK_RESCHED
);
891 static inline void resched_rq(struct rq
*rq
)
895 spin_lock_irqsave(&rq
->lock
, flags
);
896 resched_task(rq
->curr
);
897 spin_unlock_irqrestore(&rq
->lock
, flags
);
901 HRTICK_SET
, /* re-programm hrtick_timer */
902 HRTICK_RESET
, /* not a new slice */
907 * - enabled by features
908 * - hrtimer is actually high res
910 static inline int hrtick_enabled(struct rq
*rq
)
912 if (!sched_feat(HRTICK
))
914 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
918 * Called to set the hrtick timer state.
920 * called with rq->lock held and irqs disabled
922 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
924 assert_spin_locked(&rq
->lock
);
927 * preempt at: now + delay
930 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
932 * indicate we need to program the timer
934 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
936 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
939 * New slices are called from the schedule path and don't need a
943 resched_hrt(rq
->curr
);
946 static void hrtick_clear(struct rq
*rq
)
948 if (hrtimer_active(&rq
->hrtick_timer
))
949 hrtimer_cancel(&rq
->hrtick_timer
);
953 * Update the timer from the possible pending state.
955 static void hrtick_set(struct rq
*rq
)
961 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
963 spin_lock_irqsave(&rq
->lock
, flags
);
964 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
965 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
966 time
= rq
->hrtick_expire
;
967 clear_thread_flag(TIF_HRTICK_RESCHED
);
968 spin_unlock_irqrestore(&rq
->lock
, flags
);
971 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
972 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
979 * High-resolution timer tick.
980 * Runs from hardirq context with interrupts disabled.
982 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
984 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
986 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
988 spin_lock(&rq
->lock
);
989 __update_rq_clock(rq
);
990 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
991 spin_unlock(&rq
->lock
);
993 return HRTIMER_NORESTART
;
996 static inline void init_rq_hrtick(struct rq
*rq
)
998 rq
->hrtick_flags
= 0;
999 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1000 rq
->hrtick_timer
.function
= hrtick
;
1001 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1004 void hrtick_resched(void)
1007 unsigned long flags
;
1009 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
1012 local_irq_save(flags
);
1013 rq
= cpu_rq(smp_processor_id());
1015 local_irq_restore(flags
);
1018 static inline void hrtick_clear(struct rq
*rq
)
1022 static inline void hrtick_set(struct rq
*rq
)
1026 static inline void init_rq_hrtick(struct rq
*rq
)
1030 void hrtick_resched(void)
1036 * resched_task - mark a task 'to be rescheduled now'.
1038 * On UP this means the setting of the need_resched flag, on SMP it
1039 * might also involve a cross-CPU call to trigger the scheduler on
1044 #ifndef tsk_is_polling
1045 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1048 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1052 assert_spin_locked(&task_rq(p
)->lock
);
1054 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1057 set_tsk_thread_flag(p
, tif_bit
);
1060 if (cpu
== smp_processor_id())
1063 /* NEED_RESCHED must be visible before we test polling */
1065 if (!tsk_is_polling(p
))
1066 smp_send_reschedule(cpu
);
1069 static void resched_cpu(int cpu
)
1071 struct rq
*rq
= cpu_rq(cpu
);
1072 unsigned long flags
;
1074 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1076 resched_task(cpu_curr(cpu
));
1077 spin_unlock_irqrestore(&rq
->lock
, flags
);
1080 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1082 assert_spin_locked(&task_rq(p
)->lock
);
1083 set_tsk_thread_flag(p
, tif_bit
);
1087 #if BITS_PER_LONG == 32
1088 # define WMULT_CONST (~0UL)
1090 # define WMULT_CONST (1UL << 32)
1093 #define WMULT_SHIFT 32
1096 * Shift right and round:
1098 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1100 static unsigned long
1101 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1102 struct load_weight
*lw
)
1106 if (unlikely(!lw
->inv_weight
))
1107 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
1109 tmp
= (u64
)delta_exec
* weight
;
1111 * Check whether we'd overflow the 64-bit multiplication:
1113 if (unlikely(tmp
> WMULT_CONST
))
1114 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1117 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1119 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1122 static inline unsigned long
1123 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
1125 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
1128 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1133 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1139 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1140 * of tasks with abnormal "nice" values across CPUs the contribution that
1141 * each task makes to its run queue's load is weighted according to its
1142 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1143 * scaled version of the new time slice allocation that they receive on time
1147 #define WEIGHT_IDLEPRIO 2
1148 #define WMULT_IDLEPRIO (1 << 31)
1151 * Nice levels are multiplicative, with a gentle 10% change for every
1152 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1153 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1154 * that remained on nice 0.
1156 * The "10% effect" is relative and cumulative: from _any_ nice level,
1157 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1158 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1159 * If a task goes up by ~10% and another task goes down by ~10% then
1160 * the relative distance between them is ~25%.)
1162 static const int prio_to_weight
[40] = {
1163 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1164 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1165 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1166 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1167 /* 0 */ 1024, 820, 655, 526, 423,
1168 /* 5 */ 335, 272, 215, 172, 137,
1169 /* 10 */ 110, 87, 70, 56, 45,
1170 /* 15 */ 36, 29, 23, 18, 15,
1174 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1176 * In cases where the weight does not change often, we can use the
1177 * precalculated inverse to speed up arithmetics by turning divisions
1178 * into multiplications:
1180 static const u32 prio_to_wmult
[40] = {
1181 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1182 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1183 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1184 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1185 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1186 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1187 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1188 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1191 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1194 * runqueue iterator, to support SMP load-balancing between different
1195 * scheduling classes, without having to expose their internal data
1196 * structures to the load-balancing proper:
1198 struct rq_iterator
{
1200 struct task_struct
*(*start
)(void *);
1201 struct task_struct
*(*next
)(void *);
1205 static unsigned long
1206 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1207 unsigned long max_load_move
, struct sched_domain
*sd
,
1208 enum cpu_idle_type idle
, int *all_pinned
,
1209 int *this_best_prio
, struct rq_iterator
*iterator
);
1212 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1213 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1214 struct rq_iterator
*iterator
);
1217 #ifdef CONFIG_CGROUP_CPUACCT
1218 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1220 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1223 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1225 update_load_add(&rq
->load
, load
);
1228 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1230 update_load_sub(&rq
->load
, load
);
1234 static unsigned long source_load(int cpu
, int type
);
1235 static unsigned long target_load(int cpu
, int type
);
1236 static unsigned long cpu_avg_load_per_task(int cpu
);
1237 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1238 #endif /* CONFIG_SMP */
1240 #include "sched_stats.h"
1241 #include "sched_idletask.c"
1242 #include "sched_fair.c"
1243 #include "sched_rt.c"
1244 #ifdef CONFIG_SCHED_DEBUG
1245 # include "sched_debug.c"
1248 #define sched_class_highest (&rt_sched_class)
1250 static void inc_nr_running(struct rq
*rq
)
1255 static void dec_nr_running(struct rq
*rq
)
1260 static void set_load_weight(struct task_struct
*p
)
1262 if (task_has_rt_policy(p
)) {
1263 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1264 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1269 * SCHED_IDLE tasks get minimal weight:
1271 if (p
->policy
== SCHED_IDLE
) {
1272 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1273 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1277 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1278 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1281 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1283 sched_info_queued(p
);
1284 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1288 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1290 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1295 * __normal_prio - return the priority that is based on the static prio
1297 static inline int __normal_prio(struct task_struct
*p
)
1299 return p
->static_prio
;
1303 * Calculate the expected normal priority: i.e. priority
1304 * without taking RT-inheritance into account. Might be
1305 * boosted by interactivity modifiers. Changes upon fork,
1306 * setprio syscalls, and whenever the interactivity
1307 * estimator recalculates.
1309 static inline int normal_prio(struct task_struct
*p
)
1313 if (task_has_rt_policy(p
))
1314 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1316 prio
= __normal_prio(p
);
1321 * Calculate the current priority, i.e. the priority
1322 * taken into account by the scheduler. This value might
1323 * be boosted by RT tasks, or might be boosted by
1324 * interactivity modifiers. Will be RT if the task got
1325 * RT-boosted. If not then it returns p->normal_prio.
1327 static int effective_prio(struct task_struct
*p
)
1329 p
->normal_prio
= normal_prio(p
);
1331 * If we are RT tasks or we were boosted to RT priority,
1332 * keep the priority unchanged. Otherwise, update priority
1333 * to the normal priority:
1335 if (!rt_prio(p
->prio
))
1336 return p
->normal_prio
;
1341 * activate_task - move a task to the runqueue.
1343 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1345 if (task_contributes_to_load(p
))
1346 rq
->nr_uninterruptible
--;
1348 enqueue_task(rq
, p
, wakeup
);
1353 * deactivate_task - remove a task from the runqueue.
1355 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1357 if (task_contributes_to_load(p
))
1358 rq
->nr_uninterruptible
++;
1360 dequeue_task(rq
, p
, sleep
);
1365 * task_curr - is this task currently executing on a CPU?
1366 * @p: the task in question.
1368 inline int task_curr(const struct task_struct
*p
)
1370 return cpu_curr(task_cpu(p
)) == p
;
1373 /* Used instead of source_load when we know the type == 0 */
1374 unsigned long weighted_cpuload(const int cpu
)
1376 return cpu_rq(cpu
)->load
.weight
;
1379 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1381 set_task_rq(p
, cpu
);
1384 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1385 * successfuly executed on another CPU. We must ensure that updates of
1386 * per-task data have been completed by this moment.
1389 task_thread_info(p
)->cpu
= cpu
;
1393 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1394 const struct sched_class
*prev_class
,
1395 int oldprio
, int running
)
1397 if (prev_class
!= p
->sched_class
) {
1398 if (prev_class
->switched_from
)
1399 prev_class
->switched_from(rq
, p
, running
);
1400 p
->sched_class
->switched_to(rq
, p
, running
);
1402 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1408 * Is this task likely cache-hot:
1411 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1415 if (p
->sched_class
!= &fair_sched_class
)
1418 if (sysctl_sched_migration_cost
== -1)
1420 if (sysctl_sched_migration_cost
== 0)
1423 delta
= now
- p
->se
.exec_start
;
1425 return delta
< (s64
)sysctl_sched_migration_cost
;
1429 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1431 int old_cpu
= task_cpu(p
);
1432 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1433 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1434 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1437 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1439 #ifdef CONFIG_SCHEDSTATS
1440 if (p
->se
.wait_start
)
1441 p
->se
.wait_start
-= clock_offset
;
1442 if (p
->se
.sleep_start
)
1443 p
->se
.sleep_start
-= clock_offset
;
1444 if (p
->se
.block_start
)
1445 p
->se
.block_start
-= clock_offset
;
1446 if (old_cpu
!= new_cpu
) {
1447 schedstat_inc(p
, se
.nr_migrations
);
1448 if (task_hot(p
, old_rq
->clock
, NULL
))
1449 schedstat_inc(p
, se
.nr_forced2_migrations
);
1452 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1453 new_cfsrq
->min_vruntime
;
1455 __set_task_cpu(p
, new_cpu
);
1458 struct migration_req
{
1459 struct list_head list
;
1461 struct task_struct
*task
;
1464 struct completion done
;
1468 * The task's runqueue lock must be held.
1469 * Returns true if you have to wait for migration thread.
1472 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1474 struct rq
*rq
= task_rq(p
);
1477 * If the task is not on a runqueue (and not running), then
1478 * it is sufficient to simply update the task's cpu field.
1480 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1481 set_task_cpu(p
, dest_cpu
);
1485 init_completion(&req
->done
);
1487 req
->dest_cpu
= dest_cpu
;
1488 list_add(&req
->list
, &rq
->migration_queue
);
1494 * wait_task_inactive - wait for a thread to unschedule.
1496 * The caller must ensure that the task *will* unschedule sometime soon,
1497 * else this function might spin for a *long* time. This function can't
1498 * be called with interrupts off, or it may introduce deadlock with
1499 * smp_call_function() if an IPI is sent by the same process we are
1500 * waiting to become inactive.
1502 void wait_task_inactive(struct task_struct
*p
)
1504 unsigned long flags
;
1510 * We do the initial early heuristics without holding
1511 * any task-queue locks at all. We'll only try to get
1512 * the runqueue lock when things look like they will
1518 * If the task is actively running on another CPU
1519 * still, just relax and busy-wait without holding
1522 * NOTE! Since we don't hold any locks, it's not
1523 * even sure that "rq" stays as the right runqueue!
1524 * But we don't care, since "task_running()" will
1525 * return false if the runqueue has changed and p
1526 * is actually now running somewhere else!
1528 while (task_running(rq
, p
))
1532 * Ok, time to look more closely! We need the rq
1533 * lock now, to be *sure*. If we're wrong, we'll
1534 * just go back and repeat.
1536 rq
= task_rq_lock(p
, &flags
);
1537 running
= task_running(rq
, p
);
1538 on_rq
= p
->se
.on_rq
;
1539 task_rq_unlock(rq
, &flags
);
1542 * Was it really running after all now that we
1543 * checked with the proper locks actually held?
1545 * Oops. Go back and try again..
1547 if (unlikely(running
)) {
1553 * It's not enough that it's not actively running,
1554 * it must be off the runqueue _entirely_, and not
1557 * So if it wa still runnable (but just not actively
1558 * running right now), it's preempted, and we should
1559 * yield - it could be a while.
1561 if (unlikely(on_rq
)) {
1562 schedule_timeout_uninterruptible(1);
1567 * Ahh, all good. It wasn't running, and it wasn't
1568 * runnable, which means that it will never become
1569 * running in the future either. We're all done!
1576 * kick_process - kick a running thread to enter/exit the kernel
1577 * @p: the to-be-kicked thread
1579 * Cause a process which is running on another CPU to enter
1580 * kernel-mode, without any delay. (to get signals handled.)
1582 * NOTE: this function doesnt have to take the runqueue lock,
1583 * because all it wants to ensure is that the remote task enters
1584 * the kernel. If the IPI races and the task has been migrated
1585 * to another CPU then no harm is done and the purpose has been
1588 void kick_process(struct task_struct
*p
)
1594 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1595 smp_send_reschedule(cpu
);
1600 * Return a low guess at the load of a migration-source cpu weighted
1601 * according to the scheduling class and "nice" value.
1603 * We want to under-estimate the load of migration sources, to
1604 * balance conservatively.
1606 static unsigned long source_load(int cpu
, int type
)
1608 struct rq
*rq
= cpu_rq(cpu
);
1609 unsigned long total
= weighted_cpuload(cpu
);
1614 return min(rq
->cpu_load
[type
-1], total
);
1618 * Return a high guess at the load of a migration-target cpu weighted
1619 * according to the scheduling class and "nice" value.
1621 static unsigned long target_load(int cpu
, int type
)
1623 struct rq
*rq
= cpu_rq(cpu
);
1624 unsigned long total
= weighted_cpuload(cpu
);
1629 return max(rq
->cpu_load
[type
-1], total
);
1633 * Return the average load per task on the cpu's run queue
1635 static unsigned long cpu_avg_load_per_task(int cpu
)
1637 struct rq
*rq
= cpu_rq(cpu
);
1638 unsigned long total
= weighted_cpuload(cpu
);
1639 unsigned long n
= rq
->nr_running
;
1641 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1645 * find_idlest_group finds and returns the least busy CPU group within the
1648 static struct sched_group
*
1649 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1651 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1652 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1653 int load_idx
= sd
->forkexec_idx
;
1654 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1657 unsigned long load
, avg_load
;
1661 /* Skip over this group if it has no CPUs allowed */
1662 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1665 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1667 /* Tally up the load of all CPUs in the group */
1670 for_each_cpu_mask(i
, group
->cpumask
) {
1671 /* Bias balancing toward cpus of our domain */
1673 load
= source_load(i
, load_idx
);
1675 load
= target_load(i
, load_idx
);
1680 /* Adjust by relative CPU power of the group */
1681 avg_load
= sg_div_cpu_power(group
,
1682 avg_load
* SCHED_LOAD_SCALE
);
1685 this_load
= avg_load
;
1687 } else if (avg_load
< min_load
) {
1688 min_load
= avg_load
;
1691 } while (group
= group
->next
, group
!= sd
->groups
);
1693 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1699 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1702 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1705 unsigned long load
, min_load
= ULONG_MAX
;
1709 /* Traverse only the allowed CPUs */
1710 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1712 for_each_cpu_mask(i
, tmp
) {
1713 load
= weighted_cpuload(i
);
1715 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1725 * sched_balance_self: balance the current task (running on cpu) in domains
1726 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1729 * Balance, ie. select the least loaded group.
1731 * Returns the target CPU number, or the same CPU if no balancing is needed.
1733 * preempt must be disabled.
1735 static int sched_balance_self(int cpu
, int flag
)
1737 struct task_struct
*t
= current
;
1738 struct sched_domain
*tmp
, *sd
= NULL
;
1740 for_each_domain(cpu
, tmp
) {
1742 * If power savings logic is enabled for a domain, stop there.
1744 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1746 if (tmp
->flags
& flag
)
1752 struct sched_group
*group
;
1753 int new_cpu
, weight
;
1755 if (!(sd
->flags
& flag
)) {
1761 group
= find_idlest_group(sd
, t
, cpu
);
1767 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1768 if (new_cpu
== -1 || new_cpu
== cpu
) {
1769 /* Now try balancing at a lower domain level of cpu */
1774 /* Now try balancing at a lower domain level of new_cpu */
1777 weight
= cpus_weight(span
);
1778 for_each_domain(cpu
, tmp
) {
1779 if (weight
<= cpus_weight(tmp
->span
))
1781 if (tmp
->flags
& flag
)
1784 /* while loop will break here if sd == NULL */
1790 #endif /* CONFIG_SMP */
1793 * try_to_wake_up - wake up a thread
1794 * @p: the to-be-woken-up thread
1795 * @state: the mask of task states that can be woken
1796 * @sync: do a synchronous wakeup?
1798 * Put it on the run-queue if it's not already there. The "current"
1799 * thread is always on the run-queue (except when the actual
1800 * re-schedule is in progress), and as such you're allowed to do
1801 * the simpler "current->state = TASK_RUNNING" to mark yourself
1802 * runnable without the overhead of this.
1804 * returns failure only if the task is already active.
1806 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1808 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1809 unsigned long flags
;
1813 rq
= task_rq_lock(p
, &flags
);
1814 old_state
= p
->state
;
1815 if (!(old_state
& state
))
1823 this_cpu
= smp_processor_id();
1826 if (unlikely(task_running(rq
, p
)))
1829 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
1830 if (cpu
!= orig_cpu
) {
1831 set_task_cpu(p
, cpu
);
1832 task_rq_unlock(rq
, &flags
);
1833 /* might preempt at this point */
1834 rq
= task_rq_lock(p
, &flags
);
1835 old_state
= p
->state
;
1836 if (!(old_state
& state
))
1841 this_cpu
= smp_processor_id();
1845 #ifdef CONFIG_SCHEDSTATS
1846 schedstat_inc(rq
, ttwu_count
);
1847 if (cpu
== this_cpu
)
1848 schedstat_inc(rq
, ttwu_local
);
1850 struct sched_domain
*sd
;
1851 for_each_domain(this_cpu
, sd
) {
1852 if (cpu_isset(cpu
, sd
->span
)) {
1853 schedstat_inc(sd
, ttwu_wake_remote
);
1861 #endif /* CONFIG_SMP */
1862 schedstat_inc(p
, se
.nr_wakeups
);
1864 schedstat_inc(p
, se
.nr_wakeups_sync
);
1865 if (orig_cpu
!= cpu
)
1866 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1867 if (cpu
== this_cpu
)
1868 schedstat_inc(p
, se
.nr_wakeups_local
);
1870 schedstat_inc(p
, se
.nr_wakeups_remote
);
1871 update_rq_clock(rq
);
1872 activate_task(rq
, p
, 1);
1873 check_preempt_curr(rq
, p
);
1877 p
->state
= TASK_RUNNING
;
1879 if (p
->sched_class
->task_wake_up
)
1880 p
->sched_class
->task_wake_up(rq
, p
);
1883 task_rq_unlock(rq
, &flags
);
1888 int wake_up_process(struct task_struct
*p
)
1890 return try_to_wake_up(p
, TASK_ALL
, 0);
1892 EXPORT_SYMBOL(wake_up_process
);
1894 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1896 return try_to_wake_up(p
, state
, 0);
1900 * Perform scheduler related setup for a newly forked process p.
1901 * p is forked by current.
1903 * __sched_fork() is basic setup used by init_idle() too:
1905 static void __sched_fork(struct task_struct
*p
)
1907 p
->se
.exec_start
= 0;
1908 p
->se
.sum_exec_runtime
= 0;
1909 p
->se
.prev_sum_exec_runtime
= 0;
1911 #ifdef CONFIG_SCHEDSTATS
1912 p
->se
.wait_start
= 0;
1913 p
->se
.sum_sleep_runtime
= 0;
1914 p
->se
.sleep_start
= 0;
1915 p
->se
.block_start
= 0;
1916 p
->se
.sleep_max
= 0;
1917 p
->se
.block_max
= 0;
1919 p
->se
.slice_max
= 0;
1923 INIT_LIST_HEAD(&p
->rt
.run_list
);
1926 #ifdef CONFIG_PREEMPT_NOTIFIERS
1927 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1931 * We mark the process as running here, but have not actually
1932 * inserted it onto the runqueue yet. This guarantees that
1933 * nobody will actually run it, and a signal or other external
1934 * event cannot wake it up and insert it on the runqueue either.
1936 p
->state
= TASK_RUNNING
;
1940 * fork()/clone()-time setup:
1942 void sched_fork(struct task_struct
*p
, int clone_flags
)
1944 int cpu
= get_cpu();
1949 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1951 set_task_cpu(p
, cpu
);
1954 * Make sure we do not leak PI boosting priority to the child:
1956 p
->prio
= current
->normal_prio
;
1957 if (!rt_prio(p
->prio
))
1958 p
->sched_class
= &fair_sched_class
;
1960 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1961 if (likely(sched_info_on()))
1962 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1964 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1967 #ifdef CONFIG_PREEMPT
1968 /* Want to start with kernel preemption disabled. */
1969 task_thread_info(p
)->preempt_count
= 1;
1975 * wake_up_new_task - wake up a newly created task for the first time.
1977 * This function will do some initial scheduler statistics housekeeping
1978 * that must be done for every newly created context, then puts the task
1979 * on the runqueue and wakes it.
1981 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1983 unsigned long flags
;
1986 rq
= task_rq_lock(p
, &flags
);
1987 BUG_ON(p
->state
!= TASK_RUNNING
);
1988 update_rq_clock(rq
);
1990 p
->prio
= effective_prio(p
);
1992 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
1993 activate_task(rq
, p
, 0);
1996 * Let the scheduling class do new task startup
1997 * management (if any):
1999 p
->sched_class
->task_new(rq
, p
);
2002 check_preempt_curr(rq
, p
);
2004 if (p
->sched_class
->task_wake_up
)
2005 p
->sched_class
->task_wake_up(rq
, p
);
2007 task_rq_unlock(rq
, &flags
);
2010 #ifdef CONFIG_PREEMPT_NOTIFIERS
2013 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2014 * @notifier: notifier struct to register
2016 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2018 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2020 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2023 * preempt_notifier_unregister - no longer interested in preemption notifications
2024 * @notifier: notifier struct to unregister
2026 * This is safe to call from within a preemption notifier.
2028 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2030 hlist_del(¬ifier
->link
);
2032 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2034 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2036 struct preempt_notifier
*notifier
;
2037 struct hlist_node
*node
;
2039 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2040 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2044 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2045 struct task_struct
*next
)
2047 struct preempt_notifier
*notifier
;
2048 struct hlist_node
*node
;
2050 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2051 notifier
->ops
->sched_out(notifier
, next
);
2056 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2061 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2062 struct task_struct
*next
)
2069 * prepare_task_switch - prepare to switch tasks
2070 * @rq: the runqueue preparing to switch
2071 * @prev: the current task that is being switched out
2072 * @next: the task we are going to switch to.
2074 * This is called with the rq lock held and interrupts off. It must
2075 * be paired with a subsequent finish_task_switch after the context
2078 * prepare_task_switch sets up locking and calls architecture specific
2082 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2083 struct task_struct
*next
)
2085 fire_sched_out_preempt_notifiers(prev
, next
);
2086 prepare_lock_switch(rq
, next
);
2087 prepare_arch_switch(next
);
2091 * finish_task_switch - clean up after a task-switch
2092 * @rq: runqueue associated with task-switch
2093 * @prev: the thread we just switched away from.
2095 * finish_task_switch must be called after the context switch, paired
2096 * with a prepare_task_switch call before the context switch.
2097 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2098 * and do any other architecture-specific cleanup actions.
2100 * Note that we may have delayed dropping an mm in context_switch(). If
2101 * so, we finish that here outside of the runqueue lock. (Doing it
2102 * with the lock held can cause deadlocks; see schedule() for
2105 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2106 __releases(rq
->lock
)
2108 struct mm_struct
*mm
= rq
->prev_mm
;
2114 * A task struct has one reference for the use as "current".
2115 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2116 * schedule one last time. The schedule call will never return, and
2117 * the scheduled task must drop that reference.
2118 * The test for TASK_DEAD must occur while the runqueue locks are
2119 * still held, otherwise prev could be scheduled on another cpu, die
2120 * there before we look at prev->state, and then the reference would
2122 * Manfred Spraul <manfred@colorfullife.com>
2124 prev_state
= prev
->state
;
2125 finish_arch_switch(prev
);
2126 finish_lock_switch(rq
, prev
);
2128 if (current
->sched_class
->post_schedule
)
2129 current
->sched_class
->post_schedule(rq
);
2132 fire_sched_in_preempt_notifiers(current
);
2135 if (unlikely(prev_state
== TASK_DEAD
)) {
2137 * Remove function-return probe instances associated with this
2138 * task and put them back on the free list.
2140 kprobe_flush_task(prev
);
2141 put_task_struct(prev
);
2146 * schedule_tail - first thing a freshly forked thread must call.
2147 * @prev: the thread we just switched away from.
2149 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2150 __releases(rq
->lock
)
2152 struct rq
*rq
= this_rq();
2154 finish_task_switch(rq
, prev
);
2155 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2156 /* In this case, finish_task_switch does not reenable preemption */
2159 if (current
->set_child_tid
)
2160 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2164 * context_switch - switch to the new MM and the new
2165 * thread's register state.
2168 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2169 struct task_struct
*next
)
2171 struct mm_struct
*mm
, *oldmm
;
2173 prepare_task_switch(rq
, prev
, next
);
2175 oldmm
= prev
->active_mm
;
2177 * For paravirt, this is coupled with an exit in switch_to to
2178 * combine the page table reload and the switch backend into
2181 arch_enter_lazy_cpu_mode();
2183 if (unlikely(!mm
)) {
2184 next
->active_mm
= oldmm
;
2185 atomic_inc(&oldmm
->mm_count
);
2186 enter_lazy_tlb(oldmm
, next
);
2188 switch_mm(oldmm
, mm
, next
);
2190 if (unlikely(!prev
->mm
)) {
2191 prev
->active_mm
= NULL
;
2192 rq
->prev_mm
= oldmm
;
2195 * Since the runqueue lock will be released by the next
2196 * task (which is an invalid locking op but in the case
2197 * of the scheduler it's an obvious special-case), so we
2198 * do an early lockdep release here:
2200 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2201 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2204 /* Here we just switch the register state and the stack. */
2205 switch_to(prev
, next
, prev
);
2209 * this_rq must be evaluated again because prev may have moved
2210 * CPUs since it called schedule(), thus the 'rq' on its stack
2211 * frame will be invalid.
2213 finish_task_switch(this_rq(), prev
);
2217 * nr_running, nr_uninterruptible and nr_context_switches:
2219 * externally visible scheduler statistics: current number of runnable
2220 * threads, current number of uninterruptible-sleeping threads, total
2221 * number of context switches performed since bootup.
2223 unsigned long nr_running(void)
2225 unsigned long i
, sum
= 0;
2227 for_each_online_cpu(i
)
2228 sum
+= cpu_rq(i
)->nr_running
;
2233 unsigned long nr_uninterruptible(void)
2235 unsigned long i
, sum
= 0;
2237 for_each_possible_cpu(i
)
2238 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2241 * Since we read the counters lockless, it might be slightly
2242 * inaccurate. Do not allow it to go below zero though:
2244 if (unlikely((long)sum
< 0))
2250 unsigned long long nr_context_switches(void)
2253 unsigned long long sum
= 0;
2255 for_each_possible_cpu(i
)
2256 sum
+= cpu_rq(i
)->nr_switches
;
2261 unsigned long nr_iowait(void)
2263 unsigned long i
, sum
= 0;
2265 for_each_possible_cpu(i
)
2266 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2271 unsigned long nr_active(void)
2273 unsigned long i
, running
= 0, uninterruptible
= 0;
2275 for_each_online_cpu(i
) {
2276 running
+= cpu_rq(i
)->nr_running
;
2277 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2280 if (unlikely((long)uninterruptible
< 0))
2281 uninterruptible
= 0;
2283 return running
+ uninterruptible
;
2287 * Update rq->cpu_load[] statistics. This function is usually called every
2288 * scheduler tick (TICK_NSEC).
2290 static void update_cpu_load(struct rq
*this_rq
)
2292 unsigned long this_load
= this_rq
->load
.weight
;
2295 this_rq
->nr_load_updates
++;
2297 /* Update our load: */
2298 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2299 unsigned long old_load
, new_load
;
2301 /* scale is effectively 1 << i now, and >> i divides by scale */
2303 old_load
= this_rq
->cpu_load
[i
];
2304 new_load
= this_load
;
2306 * Round up the averaging division if load is increasing. This
2307 * prevents us from getting stuck on 9 if the load is 10, for
2310 if (new_load
> old_load
)
2311 new_load
+= scale
-1;
2312 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2319 * double_rq_lock - safely lock two runqueues
2321 * Note this does not disable interrupts like task_rq_lock,
2322 * you need to do so manually before calling.
2324 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2325 __acquires(rq1
->lock
)
2326 __acquires(rq2
->lock
)
2328 BUG_ON(!irqs_disabled());
2330 spin_lock(&rq1
->lock
);
2331 __acquire(rq2
->lock
); /* Fake it out ;) */
2334 spin_lock(&rq1
->lock
);
2335 spin_lock(&rq2
->lock
);
2337 spin_lock(&rq2
->lock
);
2338 spin_lock(&rq1
->lock
);
2341 update_rq_clock(rq1
);
2342 update_rq_clock(rq2
);
2346 * double_rq_unlock - safely unlock two runqueues
2348 * Note this does not restore interrupts like task_rq_unlock,
2349 * you need to do so manually after calling.
2351 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2352 __releases(rq1
->lock
)
2353 __releases(rq2
->lock
)
2355 spin_unlock(&rq1
->lock
);
2357 spin_unlock(&rq2
->lock
);
2359 __release(rq2
->lock
);
2363 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2365 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2366 __releases(this_rq
->lock
)
2367 __acquires(busiest
->lock
)
2368 __acquires(this_rq
->lock
)
2372 if (unlikely(!irqs_disabled())) {
2373 /* printk() doesn't work good under rq->lock */
2374 spin_unlock(&this_rq
->lock
);
2377 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2378 if (busiest
< this_rq
) {
2379 spin_unlock(&this_rq
->lock
);
2380 spin_lock(&busiest
->lock
);
2381 spin_lock(&this_rq
->lock
);
2384 spin_lock(&busiest
->lock
);
2390 * If dest_cpu is allowed for this process, migrate the task to it.
2391 * This is accomplished by forcing the cpu_allowed mask to only
2392 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2393 * the cpu_allowed mask is restored.
2395 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2397 struct migration_req req
;
2398 unsigned long flags
;
2401 rq
= task_rq_lock(p
, &flags
);
2402 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2403 || unlikely(cpu_is_offline(dest_cpu
)))
2406 /* force the process onto the specified CPU */
2407 if (migrate_task(p
, dest_cpu
, &req
)) {
2408 /* Need to wait for migration thread (might exit: take ref). */
2409 struct task_struct
*mt
= rq
->migration_thread
;
2411 get_task_struct(mt
);
2412 task_rq_unlock(rq
, &flags
);
2413 wake_up_process(mt
);
2414 put_task_struct(mt
);
2415 wait_for_completion(&req
.done
);
2420 task_rq_unlock(rq
, &flags
);
2424 * sched_exec - execve() is a valuable balancing opportunity, because at
2425 * this point the task has the smallest effective memory and cache footprint.
2427 void sched_exec(void)
2429 int new_cpu
, this_cpu
= get_cpu();
2430 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2432 if (new_cpu
!= this_cpu
)
2433 sched_migrate_task(current
, new_cpu
);
2437 * pull_task - move a task from a remote runqueue to the local runqueue.
2438 * Both runqueues must be locked.
2440 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2441 struct rq
*this_rq
, int this_cpu
)
2443 deactivate_task(src_rq
, p
, 0);
2444 set_task_cpu(p
, this_cpu
);
2445 activate_task(this_rq
, p
, 0);
2447 * Note that idle threads have a prio of MAX_PRIO, for this test
2448 * to be always true for them.
2450 check_preempt_curr(this_rq
, p
);
2454 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2457 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2458 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2462 * We do not migrate tasks that are:
2463 * 1) running (obviously), or
2464 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2465 * 3) are cache-hot on their current CPU.
2467 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2468 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2473 if (task_running(rq
, p
)) {
2474 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2479 * Aggressive migration if:
2480 * 1) task is cache cold, or
2481 * 2) too many balance attempts have failed.
2484 if (!task_hot(p
, rq
->clock
, sd
) ||
2485 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2486 #ifdef CONFIG_SCHEDSTATS
2487 if (task_hot(p
, rq
->clock
, sd
)) {
2488 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2489 schedstat_inc(p
, se
.nr_forced_migrations
);
2495 if (task_hot(p
, rq
->clock
, sd
)) {
2496 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2502 static unsigned long
2503 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2504 unsigned long max_load_move
, struct sched_domain
*sd
,
2505 enum cpu_idle_type idle
, int *all_pinned
,
2506 int *this_best_prio
, struct rq_iterator
*iterator
)
2508 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2509 struct task_struct
*p
;
2510 long rem_load_move
= max_load_move
;
2512 if (max_load_move
== 0)
2518 * Start the load-balancing iterator:
2520 p
= iterator
->start(iterator
->arg
);
2522 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2525 * To help distribute high priority tasks across CPUs we don't
2526 * skip a task if it will be the highest priority task (i.e. smallest
2527 * prio value) on its new queue regardless of its load weight
2529 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2530 SCHED_LOAD_SCALE_FUZZ
;
2531 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2532 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2533 p
= iterator
->next(iterator
->arg
);
2537 pull_task(busiest
, p
, this_rq
, this_cpu
);
2539 rem_load_move
-= p
->se
.load
.weight
;
2542 * We only want to steal up to the prescribed amount of weighted load.
2544 if (rem_load_move
> 0) {
2545 if (p
->prio
< *this_best_prio
)
2546 *this_best_prio
= p
->prio
;
2547 p
= iterator
->next(iterator
->arg
);
2552 * Right now, this is one of only two places pull_task() is called,
2553 * so we can safely collect pull_task() stats here rather than
2554 * inside pull_task().
2556 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2559 *all_pinned
= pinned
;
2561 return max_load_move
- rem_load_move
;
2565 * move_tasks tries to move up to max_load_move weighted load from busiest to
2566 * this_rq, as part of a balancing operation within domain "sd".
2567 * Returns 1 if successful and 0 otherwise.
2569 * Called with both runqueues locked.
2571 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2572 unsigned long max_load_move
,
2573 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2576 const struct sched_class
*class = sched_class_highest
;
2577 unsigned long total_load_moved
= 0;
2578 int this_best_prio
= this_rq
->curr
->prio
;
2582 class->load_balance(this_rq
, this_cpu
, busiest
,
2583 max_load_move
- total_load_moved
,
2584 sd
, idle
, all_pinned
, &this_best_prio
);
2585 class = class->next
;
2586 } while (class && max_load_move
> total_load_moved
);
2588 return total_load_moved
> 0;
2592 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2593 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2594 struct rq_iterator
*iterator
)
2596 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2600 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2601 pull_task(busiest
, p
, this_rq
, this_cpu
);
2603 * Right now, this is only the second place pull_task()
2604 * is called, so we can safely collect pull_task()
2605 * stats here rather than inside pull_task().
2607 schedstat_inc(sd
, lb_gained
[idle
]);
2611 p
= iterator
->next(iterator
->arg
);
2618 * move_one_task tries to move exactly one task from busiest to this_rq, as
2619 * part of active balancing operations within "domain".
2620 * Returns 1 if successful and 0 otherwise.
2622 * Called with both runqueues locked.
2624 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2625 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2627 const struct sched_class
*class;
2629 for (class = sched_class_highest
; class; class = class->next
)
2630 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2637 * find_busiest_group finds and returns the busiest CPU group within the
2638 * domain. It calculates and returns the amount of weighted load which
2639 * should be moved to restore balance via the imbalance parameter.
2641 static struct sched_group
*
2642 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2643 unsigned long *imbalance
, enum cpu_idle_type idle
,
2644 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2646 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2647 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2648 unsigned long max_pull
;
2649 unsigned long busiest_load_per_task
, busiest_nr_running
;
2650 unsigned long this_load_per_task
, this_nr_running
;
2651 int load_idx
, group_imb
= 0;
2652 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2653 int power_savings_balance
= 1;
2654 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2655 unsigned long min_nr_running
= ULONG_MAX
;
2656 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2659 max_load
= this_load
= total_load
= total_pwr
= 0;
2660 busiest_load_per_task
= busiest_nr_running
= 0;
2661 this_load_per_task
= this_nr_running
= 0;
2662 if (idle
== CPU_NOT_IDLE
)
2663 load_idx
= sd
->busy_idx
;
2664 else if (idle
== CPU_NEWLY_IDLE
)
2665 load_idx
= sd
->newidle_idx
;
2667 load_idx
= sd
->idle_idx
;
2670 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2673 int __group_imb
= 0;
2674 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2675 unsigned long sum_nr_running
, sum_weighted_load
;
2677 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2680 balance_cpu
= first_cpu(group
->cpumask
);
2682 /* Tally up the load of all CPUs in the group */
2683 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2685 min_cpu_load
= ~0UL;
2687 for_each_cpu_mask(i
, group
->cpumask
) {
2690 if (!cpu_isset(i
, *cpus
))
2695 if (*sd_idle
&& rq
->nr_running
)
2698 /* Bias balancing toward cpus of our domain */
2700 if (idle_cpu(i
) && !first_idle_cpu
) {
2705 load
= target_load(i
, load_idx
);
2707 load
= source_load(i
, load_idx
);
2708 if (load
> max_cpu_load
)
2709 max_cpu_load
= load
;
2710 if (min_cpu_load
> load
)
2711 min_cpu_load
= load
;
2715 sum_nr_running
+= rq
->nr_running
;
2716 sum_weighted_load
+= weighted_cpuload(i
);
2720 * First idle cpu or the first cpu(busiest) in this sched group
2721 * is eligible for doing load balancing at this and above
2722 * domains. In the newly idle case, we will allow all the cpu's
2723 * to do the newly idle load balance.
2725 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2726 balance_cpu
!= this_cpu
&& balance
) {
2731 total_load
+= avg_load
;
2732 total_pwr
+= group
->__cpu_power
;
2734 /* Adjust by relative CPU power of the group */
2735 avg_load
= sg_div_cpu_power(group
,
2736 avg_load
* SCHED_LOAD_SCALE
);
2738 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2741 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2744 this_load
= avg_load
;
2746 this_nr_running
= sum_nr_running
;
2747 this_load_per_task
= sum_weighted_load
;
2748 } else if (avg_load
> max_load
&&
2749 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2750 max_load
= avg_load
;
2752 busiest_nr_running
= sum_nr_running
;
2753 busiest_load_per_task
= sum_weighted_load
;
2754 group_imb
= __group_imb
;
2757 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2759 * Busy processors will not participate in power savings
2762 if (idle
== CPU_NOT_IDLE
||
2763 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2767 * If the local group is idle or completely loaded
2768 * no need to do power savings balance at this domain
2770 if (local_group
&& (this_nr_running
>= group_capacity
||
2772 power_savings_balance
= 0;
2775 * If a group is already running at full capacity or idle,
2776 * don't include that group in power savings calculations
2778 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2783 * Calculate the group which has the least non-idle load.
2784 * This is the group from where we need to pick up the load
2787 if ((sum_nr_running
< min_nr_running
) ||
2788 (sum_nr_running
== min_nr_running
&&
2789 first_cpu(group
->cpumask
) <
2790 first_cpu(group_min
->cpumask
))) {
2792 min_nr_running
= sum_nr_running
;
2793 min_load_per_task
= sum_weighted_load
/
2798 * Calculate the group which is almost near its
2799 * capacity but still has some space to pick up some load
2800 * from other group and save more power
2802 if (sum_nr_running
<= group_capacity
- 1) {
2803 if (sum_nr_running
> leader_nr_running
||
2804 (sum_nr_running
== leader_nr_running
&&
2805 first_cpu(group
->cpumask
) >
2806 first_cpu(group_leader
->cpumask
))) {
2807 group_leader
= group
;
2808 leader_nr_running
= sum_nr_running
;
2813 group
= group
->next
;
2814 } while (group
!= sd
->groups
);
2816 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2819 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2821 if (this_load
>= avg_load
||
2822 100*max_load
<= sd
->imbalance_pct
*this_load
)
2825 busiest_load_per_task
/= busiest_nr_running
;
2827 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2830 * We're trying to get all the cpus to the average_load, so we don't
2831 * want to push ourselves above the average load, nor do we wish to
2832 * reduce the max loaded cpu below the average load, as either of these
2833 * actions would just result in more rebalancing later, and ping-pong
2834 * tasks around. Thus we look for the minimum possible imbalance.
2835 * Negative imbalances (*we* are more loaded than anyone else) will
2836 * be counted as no imbalance for these purposes -- we can't fix that
2837 * by pulling tasks to us. Be careful of negative numbers as they'll
2838 * appear as very large values with unsigned longs.
2840 if (max_load
<= busiest_load_per_task
)
2844 * In the presence of smp nice balancing, certain scenarios can have
2845 * max load less than avg load(as we skip the groups at or below
2846 * its cpu_power, while calculating max_load..)
2848 if (max_load
< avg_load
) {
2850 goto small_imbalance
;
2853 /* Don't want to pull so many tasks that a group would go idle */
2854 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2856 /* How much load to actually move to equalise the imbalance */
2857 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2858 (avg_load
- this_load
) * this->__cpu_power
)
2862 * if *imbalance is less than the average load per runnable task
2863 * there is no gaurantee that any tasks will be moved so we'll have
2864 * a think about bumping its value to force at least one task to be
2867 if (*imbalance
< busiest_load_per_task
) {
2868 unsigned long tmp
, pwr_now
, pwr_move
;
2872 pwr_move
= pwr_now
= 0;
2874 if (this_nr_running
) {
2875 this_load_per_task
/= this_nr_running
;
2876 if (busiest_load_per_task
> this_load_per_task
)
2879 this_load_per_task
= SCHED_LOAD_SCALE
;
2881 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2882 busiest_load_per_task
* imbn
) {
2883 *imbalance
= busiest_load_per_task
;
2888 * OK, we don't have enough imbalance to justify moving tasks,
2889 * however we may be able to increase total CPU power used by
2893 pwr_now
+= busiest
->__cpu_power
*
2894 min(busiest_load_per_task
, max_load
);
2895 pwr_now
+= this->__cpu_power
*
2896 min(this_load_per_task
, this_load
);
2897 pwr_now
/= SCHED_LOAD_SCALE
;
2899 /* Amount of load we'd subtract */
2900 tmp
= sg_div_cpu_power(busiest
,
2901 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2903 pwr_move
+= busiest
->__cpu_power
*
2904 min(busiest_load_per_task
, max_load
- tmp
);
2906 /* Amount of load we'd add */
2907 if (max_load
* busiest
->__cpu_power
<
2908 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2909 tmp
= sg_div_cpu_power(this,
2910 max_load
* busiest
->__cpu_power
);
2912 tmp
= sg_div_cpu_power(this,
2913 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2914 pwr_move
+= this->__cpu_power
*
2915 min(this_load_per_task
, this_load
+ tmp
);
2916 pwr_move
/= SCHED_LOAD_SCALE
;
2918 /* Move if we gain throughput */
2919 if (pwr_move
> pwr_now
)
2920 *imbalance
= busiest_load_per_task
;
2926 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2927 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2930 if (this == group_leader
&& group_leader
!= group_min
) {
2931 *imbalance
= min_load_per_task
;
2941 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2944 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2945 unsigned long imbalance
, cpumask_t
*cpus
)
2947 struct rq
*busiest
= NULL
, *rq
;
2948 unsigned long max_load
= 0;
2951 for_each_cpu_mask(i
, group
->cpumask
) {
2954 if (!cpu_isset(i
, *cpus
))
2958 wl
= weighted_cpuload(i
);
2960 if (rq
->nr_running
== 1 && wl
> imbalance
)
2963 if (wl
> max_load
) {
2973 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2974 * so long as it is large enough.
2976 #define MAX_PINNED_INTERVAL 512
2979 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2980 * tasks if there is an imbalance.
2982 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2983 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2986 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2987 struct sched_group
*group
;
2988 unsigned long imbalance
;
2990 cpumask_t cpus
= CPU_MASK_ALL
;
2991 unsigned long flags
;
2994 * When power savings policy is enabled for the parent domain, idle
2995 * sibling can pick up load irrespective of busy siblings. In this case,
2996 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2997 * portraying it as CPU_NOT_IDLE.
2999 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3000 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3003 schedstat_inc(sd
, lb_count
[idle
]);
3006 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3013 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3017 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
3019 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3023 BUG_ON(busiest
== this_rq
);
3025 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3028 if (busiest
->nr_running
> 1) {
3030 * Attempt to move tasks. If find_busiest_group has found
3031 * an imbalance but busiest->nr_running <= 1, the group is
3032 * still unbalanced. ld_moved simply stays zero, so it is
3033 * correctly treated as an imbalance.
3035 local_irq_save(flags
);
3036 double_rq_lock(this_rq
, busiest
);
3037 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3038 imbalance
, sd
, idle
, &all_pinned
);
3039 double_rq_unlock(this_rq
, busiest
);
3040 local_irq_restore(flags
);
3043 * some other cpu did the load balance for us.
3045 if (ld_moved
&& this_cpu
!= smp_processor_id())
3046 resched_cpu(this_cpu
);
3048 /* All tasks on this runqueue were pinned by CPU affinity */
3049 if (unlikely(all_pinned
)) {
3050 cpu_clear(cpu_of(busiest
), cpus
);
3051 if (!cpus_empty(cpus
))
3058 schedstat_inc(sd
, lb_failed
[idle
]);
3059 sd
->nr_balance_failed
++;
3061 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3063 spin_lock_irqsave(&busiest
->lock
, flags
);
3065 /* don't kick the migration_thread, if the curr
3066 * task on busiest cpu can't be moved to this_cpu
3068 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3069 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3071 goto out_one_pinned
;
3074 if (!busiest
->active_balance
) {
3075 busiest
->active_balance
= 1;
3076 busiest
->push_cpu
= this_cpu
;
3079 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3081 wake_up_process(busiest
->migration_thread
);
3084 * We've kicked active balancing, reset the failure
3087 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3090 sd
->nr_balance_failed
= 0;
3092 if (likely(!active_balance
)) {
3093 /* We were unbalanced, so reset the balancing interval */
3094 sd
->balance_interval
= sd
->min_interval
;
3097 * If we've begun active balancing, start to back off. This
3098 * case may not be covered by the all_pinned logic if there
3099 * is only 1 task on the busy runqueue (because we don't call
3102 if (sd
->balance_interval
< sd
->max_interval
)
3103 sd
->balance_interval
*= 2;
3106 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3107 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3112 schedstat_inc(sd
, lb_balanced
[idle
]);
3114 sd
->nr_balance_failed
= 0;
3117 /* tune up the balancing interval */
3118 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3119 (sd
->balance_interval
< sd
->max_interval
))
3120 sd
->balance_interval
*= 2;
3122 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3123 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3129 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3130 * tasks if there is an imbalance.
3132 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3133 * this_rq is locked.
3136 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
3138 struct sched_group
*group
;
3139 struct rq
*busiest
= NULL
;
3140 unsigned long imbalance
;
3144 cpumask_t cpus
= CPU_MASK_ALL
;
3147 * When power savings policy is enabled for the parent domain, idle
3148 * sibling can pick up load irrespective of busy siblings. In this case,
3149 * let the state of idle sibling percolate up as IDLE, instead of
3150 * portraying it as CPU_NOT_IDLE.
3152 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3153 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3156 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3158 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3159 &sd_idle
, &cpus
, NULL
);
3161 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3165 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
3168 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3172 BUG_ON(busiest
== this_rq
);
3174 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3177 if (busiest
->nr_running
> 1) {
3178 /* Attempt to move tasks */
3179 double_lock_balance(this_rq
, busiest
);
3180 /* this_rq->clock is already updated */
3181 update_rq_clock(busiest
);
3182 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3183 imbalance
, sd
, CPU_NEWLY_IDLE
,
3185 spin_unlock(&busiest
->lock
);
3187 if (unlikely(all_pinned
)) {
3188 cpu_clear(cpu_of(busiest
), cpus
);
3189 if (!cpus_empty(cpus
))
3195 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3196 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3197 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3200 sd
->nr_balance_failed
= 0;
3205 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3206 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3207 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3209 sd
->nr_balance_failed
= 0;
3215 * idle_balance is called by schedule() if this_cpu is about to become
3216 * idle. Attempts to pull tasks from other CPUs.
3218 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3220 struct sched_domain
*sd
;
3221 int pulled_task
= -1;
3222 unsigned long next_balance
= jiffies
+ HZ
;
3224 for_each_domain(this_cpu
, sd
) {
3225 unsigned long interval
;
3227 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3230 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3231 /* If we've pulled tasks over stop searching: */
3232 pulled_task
= load_balance_newidle(this_cpu
,
3235 interval
= msecs_to_jiffies(sd
->balance_interval
);
3236 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3237 next_balance
= sd
->last_balance
+ interval
;
3241 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3243 * We are going idle. next_balance may be set based on
3244 * a busy processor. So reset next_balance.
3246 this_rq
->next_balance
= next_balance
;
3251 * active_load_balance is run by migration threads. It pushes running tasks
3252 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3253 * running on each physical CPU where possible, and avoids physical /
3254 * logical imbalances.
3256 * Called with busiest_rq locked.
3258 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3260 int target_cpu
= busiest_rq
->push_cpu
;
3261 struct sched_domain
*sd
;
3262 struct rq
*target_rq
;
3264 /* Is there any task to move? */
3265 if (busiest_rq
->nr_running
<= 1)
3268 target_rq
= cpu_rq(target_cpu
);
3271 * This condition is "impossible", if it occurs
3272 * we need to fix it. Originally reported by
3273 * Bjorn Helgaas on a 128-cpu setup.
3275 BUG_ON(busiest_rq
== target_rq
);
3277 /* move a task from busiest_rq to target_rq */
3278 double_lock_balance(busiest_rq
, target_rq
);
3279 update_rq_clock(busiest_rq
);
3280 update_rq_clock(target_rq
);
3282 /* Search for an sd spanning us and the target CPU. */
3283 for_each_domain(target_cpu
, sd
) {
3284 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3285 cpu_isset(busiest_cpu
, sd
->span
))
3290 schedstat_inc(sd
, alb_count
);
3292 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3294 schedstat_inc(sd
, alb_pushed
);
3296 schedstat_inc(sd
, alb_failed
);
3298 spin_unlock(&target_rq
->lock
);
3303 atomic_t load_balancer
;
3305 } nohz ____cacheline_aligned
= {
3306 .load_balancer
= ATOMIC_INIT(-1),
3307 .cpu_mask
= CPU_MASK_NONE
,
3311 * This routine will try to nominate the ilb (idle load balancing)
3312 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3313 * load balancing on behalf of all those cpus. If all the cpus in the system
3314 * go into this tickless mode, then there will be no ilb owner (as there is
3315 * no need for one) and all the cpus will sleep till the next wakeup event
3318 * For the ilb owner, tick is not stopped. And this tick will be used
3319 * for idle load balancing. ilb owner will still be part of
3322 * While stopping the tick, this cpu will become the ilb owner if there
3323 * is no other owner. And will be the owner till that cpu becomes busy
3324 * or if all cpus in the system stop their ticks at which point
3325 * there is no need for ilb owner.
3327 * When the ilb owner becomes busy, it nominates another owner, during the
3328 * next busy scheduler_tick()
3330 int select_nohz_load_balancer(int stop_tick
)
3332 int cpu
= smp_processor_id();
3335 cpu_set(cpu
, nohz
.cpu_mask
);
3336 cpu_rq(cpu
)->in_nohz_recently
= 1;
3339 * If we are going offline and still the leader, give up!
3341 if (cpu_is_offline(cpu
) &&
3342 atomic_read(&nohz
.load_balancer
) == cpu
) {
3343 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3348 /* time for ilb owner also to sleep */
3349 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3350 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3351 atomic_set(&nohz
.load_balancer
, -1);
3355 if (atomic_read(&nohz
.load_balancer
) == -1) {
3356 /* make me the ilb owner */
3357 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3359 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3362 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3365 cpu_clear(cpu
, nohz
.cpu_mask
);
3367 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3368 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3375 static DEFINE_SPINLOCK(balancing
);
3378 * It checks each scheduling domain to see if it is due to be balanced,
3379 * and initiates a balancing operation if so.
3381 * Balancing parameters are set up in arch_init_sched_domains.
3383 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3386 struct rq
*rq
= cpu_rq(cpu
);
3387 unsigned long interval
;
3388 struct sched_domain
*sd
;
3389 /* Earliest time when we have to do rebalance again */
3390 unsigned long next_balance
= jiffies
+ 60*HZ
;
3391 int update_next_balance
= 0;
3393 for_each_domain(cpu
, sd
) {
3394 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3397 interval
= sd
->balance_interval
;
3398 if (idle
!= CPU_IDLE
)
3399 interval
*= sd
->busy_factor
;
3401 /* scale ms to jiffies */
3402 interval
= msecs_to_jiffies(interval
);
3403 if (unlikely(!interval
))
3405 if (interval
> HZ
*NR_CPUS
/10)
3406 interval
= HZ
*NR_CPUS
/10;
3409 if (sd
->flags
& SD_SERIALIZE
) {
3410 if (!spin_trylock(&balancing
))
3414 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3415 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3417 * We've pulled tasks over so either we're no
3418 * longer idle, or one of our SMT siblings is
3421 idle
= CPU_NOT_IDLE
;
3423 sd
->last_balance
= jiffies
;
3425 if (sd
->flags
& SD_SERIALIZE
)
3426 spin_unlock(&balancing
);
3428 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3429 next_balance
= sd
->last_balance
+ interval
;
3430 update_next_balance
= 1;
3434 * Stop the load balance at this level. There is another
3435 * CPU in our sched group which is doing load balancing more
3443 * next_balance will be updated only when there is a need.
3444 * When the cpu is attached to null domain for ex, it will not be
3447 if (likely(update_next_balance
))
3448 rq
->next_balance
= next_balance
;
3452 * run_rebalance_domains is triggered when needed from the scheduler tick.
3453 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3454 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3456 static void run_rebalance_domains(struct softirq_action
*h
)
3458 int this_cpu
= smp_processor_id();
3459 struct rq
*this_rq
= cpu_rq(this_cpu
);
3460 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3461 CPU_IDLE
: CPU_NOT_IDLE
;
3463 rebalance_domains(this_cpu
, idle
);
3467 * If this cpu is the owner for idle load balancing, then do the
3468 * balancing on behalf of the other idle cpus whose ticks are
3471 if (this_rq
->idle_at_tick
&&
3472 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3473 cpumask_t cpus
= nohz
.cpu_mask
;
3477 cpu_clear(this_cpu
, cpus
);
3478 for_each_cpu_mask(balance_cpu
, cpus
) {
3480 * If this cpu gets work to do, stop the load balancing
3481 * work being done for other cpus. Next load
3482 * balancing owner will pick it up.
3487 rebalance_domains(balance_cpu
, CPU_IDLE
);
3489 rq
= cpu_rq(balance_cpu
);
3490 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3491 this_rq
->next_balance
= rq
->next_balance
;
3498 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3500 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3501 * idle load balancing owner or decide to stop the periodic load balancing,
3502 * if the whole system is idle.
3504 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3508 * If we were in the nohz mode recently and busy at the current
3509 * scheduler tick, then check if we need to nominate new idle
3512 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3513 rq
->in_nohz_recently
= 0;
3515 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3516 cpu_clear(cpu
, nohz
.cpu_mask
);
3517 atomic_set(&nohz
.load_balancer
, -1);
3520 if (atomic_read(&nohz
.load_balancer
) == -1) {
3522 * simple selection for now: Nominate the
3523 * first cpu in the nohz list to be the next
3526 * TBD: Traverse the sched domains and nominate
3527 * the nearest cpu in the nohz.cpu_mask.
3529 int ilb
= first_cpu(nohz
.cpu_mask
);
3537 * If this cpu is idle and doing idle load balancing for all the
3538 * cpus with ticks stopped, is it time for that to stop?
3540 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3541 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3547 * If this cpu is idle and the idle load balancing is done by
3548 * someone else, then no need raise the SCHED_SOFTIRQ
3550 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3551 cpu_isset(cpu
, nohz
.cpu_mask
))
3554 if (time_after_eq(jiffies
, rq
->next_balance
))
3555 raise_softirq(SCHED_SOFTIRQ
);
3558 #else /* CONFIG_SMP */
3561 * on UP we do not need to balance between CPUs:
3563 static inline void idle_balance(int cpu
, struct rq
*rq
)
3569 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3571 EXPORT_PER_CPU_SYMBOL(kstat
);
3574 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3575 * that have not yet been banked in case the task is currently running.
3577 unsigned long long task_sched_runtime(struct task_struct
*p
)
3579 unsigned long flags
;
3583 rq
= task_rq_lock(p
, &flags
);
3584 ns
= p
->se
.sum_exec_runtime
;
3585 if (task_current(rq
, p
)) {
3586 update_rq_clock(rq
);
3587 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3588 if ((s64
)delta_exec
> 0)
3591 task_rq_unlock(rq
, &flags
);
3597 * Account user cpu time to a process.
3598 * @p: the process that the cpu time gets accounted to
3599 * @cputime: the cpu time spent in user space since the last update
3601 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3603 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3606 p
->utime
= cputime_add(p
->utime
, cputime
);
3608 /* Add user time to cpustat. */
3609 tmp
= cputime_to_cputime64(cputime
);
3610 if (TASK_NICE(p
) > 0)
3611 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3613 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3617 * Account guest cpu time to a process.
3618 * @p: the process that the cpu time gets accounted to
3619 * @cputime: the cpu time spent in virtual machine since the last update
3621 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3624 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3626 tmp
= cputime_to_cputime64(cputime
);
3628 p
->utime
= cputime_add(p
->utime
, cputime
);
3629 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3631 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3632 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3636 * Account scaled user cpu time to a process.
3637 * @p: the process that the cpu time gets accounted to
3638 * @cputime: the cpu time spent in user space since the last update
3640 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3642 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3646 * Account system cpu time to a process.
3647 * @p: the process that the cpu time gets accounted to
3648 * @hardirq_offset: the offset to subtract from hardirq_count()
3649 * @cputime: the cpu time spent in kernel space since the last update
3651 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3654 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3655 struct rq
*rq
= this_rq();
3658 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3659 return account_guest_time(p
, cputime
);
3661 p
->stime
= cputime_add(p
->stime
, cputime
);
3663 /* Add system time to cpustat. */
3664 tmp
= cputime_to_cputime64(cputime
);
3665 if (hardirq_count() - hardirq_offset
)
3666 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3667 else if (softirq_count())
3668 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3669 else if (p
!= rq
->idle
)
3670 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3671 else if (atomic_read(&rq
->nr_iowait
) > 0)
3672 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3674 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3675 /* Account for system time used */
3676 acct_update_integrals(p
);
3680 * Account scaled system cpu time to a process.
3681 * @p: the process that the cpu time gets accounted to
3682 * @hardirq_offset: the offset to subtract from hardirq_count()
3683 * @cputime: the cpu time spent in kernel space since the last update
3685 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3687 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3691 * Account for involuntary wait time.
3692 * @p: the process from which the cpu time has been stolen
3693 * @steal: the cpu time spent in involuntary wait
3695 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3697 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3698 cputime64_t tmp
= cputime_to_cputime64(steal
);
3699 struct rq
*rq
= this_rq();
3701 if (p
== rq
->idle
) {
3702 p
->stime
= cputime_add(p
->stime
, steal
);
3703 if (atomic_read(&rq
->nr_iowait
) > 0)
3704 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3706 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3708 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3712 * This function gets called by the timer code, with HZ frequency.
3713 * We call it with interrupts disabled.
3715 * It also gets called by the fork code, when changing the parent's
3718 void scheduler_tick(void)
3720 int cpu
= smp_processor_id();
3721 struct rq
*rq
= cpu_rq(cpu
);
3722 struct task_struct
*curr
= rq
->curr
;
3723 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3725 spin_lock(&rq
->lock
);
3726 __update_rq_clock(rq
);
3728 * Let rq->clock advance by at least TICK_NSEC:
3730 if (unlikely(rq
->clock
< next_tick
)) {
3731 rq
->clock
= next_tick
;
3732 rq
->clock_underflows
++;
3734 rq
->tick_timestamp
= rq
->clock
;
3735 update_cpu_load(rq
);
3736 curr
->sched_class
->task_tick(rq
, curr
, 0);
3737 update_sched_rt_period(rq
);
3738 spin_unlock(&rq
->lock
);
3741 rq
->idle_at_tick
= idle_cpu(cpu
);
3742 trigger_load_balance(rq
, cpu
);
3746 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3748 void add_preempt_count(int val
)
3753 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3755 preempt_count() += val
;
3757 * Spinlock count overflowing soon?
3759 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3762 EXPORT_SYMBOL(add_preempt_count
);
3764 void sub_preempt_count(int val
)
3769 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3772 * Is the spinlock portion underflowing?
3774 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3775 !(preempt_count() & PREEMPT_MASK
)))
3778 preempt_count() -= val
;
3780 EXPORT_SYMBOL(sub_preempt_count
);
3785 * Print scheduling while atomic bug:
3787 static noinline
void __schedule_bug(struct task_struct
*prev
)
3789 struct pt_regs
*regs
= get_irq_regs();
3791 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3792 prev
->comm
, prev
->pid
, preempt_count());
3794 debug_show_held_locks(prev
);
3795 if (irqs_disabled())
3796 print_irqtrace_events(prev
);
3805 * Various schedule()-time debugging checks and statistics:
3807 static inline void schedule_debug(struct task_struct
*prev
)
3810 * Test if we are atomic. Since do_exit() needs to call into
3811 * schedule() atomically, we ignore that path for now.
3812 * Otherwise, whine if we are scheduling when we should not be.
3814 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3815 __schedule_bug(prev
);
3817 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3819 schedstat_inc(this_rq(), sched_count
);
3820 #ifdef CONFIG_SCHEDSTATS
3821 if (unlikely(prev
->lock_depth
>= 0)) {
3822 schedstat_inc(this_rq(), bkl_count
);
3823 schedstat_inc(prev
, sched_info
.bkl_count
);
3829 * Pick up the highest-prio task:
3831 static inline struct task_struct
*
3832 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3834 const struct sched_class
*class;
3835 struct task_struct
*p
;
3838 * Optimization: we know that if all tasks are in
3839 * the fair class we can call that function directly:
3841 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3842 p
= fair_sched_class
.pick_next_task(rq
);
3847 class = sched_class_highest
;
3849 p
= class->pick_next_task(rq
);
3853 * Will never be NULL as the idle class always
3854 * returns a non-NULL p:
3856 class = class->next
;
3861 * schedule() is the main scheduler function.
3863 asmlinkage
void __sched
schedule(void)
3865 struct task_struct
*prev
, *next
;
3872 cpu
= smp_processor_id();
3876 switch_count
= &prev
->nivcsw
;
3878 release_kernel_lock(prev
);
3879 need_resched_nonpreemptible
:
3881 schedule_debug(prev
);
3886 * Do the rq-clock update outside the rq lock:
3888 local_irq_disable();
3889 __update_rq_clock(rq
);
3890 spin_lock(&rq
->lock
);
3891 clear_tsk_need_resched(prev
);
3893 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3894 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3895 unlikely(signal_pending(prev
)))) {
3896 prev
->state
= TASK_RUNNING
;
3898 deactivate_task(rq
, prev
, 1);
3900 switch_count
= &prev
->nvcsw
;
3904 if (prev
->sched_class
->pre_schedule
)
3905 prev
->sched_class
->pre_schedule(rq
, prev
);
3908 if (unlikely(!rq
->nr_running
))
3909 idle_balance(cpu
, rq
);
3911 prev
->sched_class
->put_prev_task(rq
, prev
);
3912 next
= pick_next_task(rq
, prev
);
3914 sched_info_switch(prev
, next
);
3916 if (likely(prev
!= next
)) {
3921 context_switch(rq
, prev
, next
); /* unlocks the rq */
3923 * the context switch might have flipped the stack from under
3924 * us, hence refresh the local variables.
3926 cpu
= smp_processor_id();
3929 spin_unlock_irq(&rq
->lock
);
3933 if (unlikely(reacquire_kernel_lock(current
) < 0))
3934 goto need_resched_nonpreemptible
;
3936 preempt_enable_no_resched();
3937 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3940 EXPORT_SYMBOL(schedule
);
3942 #ifdef CONFIG_PREEMPT
3944 * this is the entry point to schedule() from in-kernel preemption
3945 * off of preempt_enable. Kernel preemptions off return from interrupt
3946 * occur there and call schedule directly.
3948 asmlinkage
void __sched
preempt_schedule(void)
3950 struct thread_info
*ti
= current_thread_info();
3951 struct task_struct
*task
= current
;
3952 int saved_lock_depth
;
3955 * If there is a non-zero preempt_count or interrupts are disabled,
3956 * we do not want to preempt the current task. Just return..
3958 if (likely(ti
->preempt_count
|| irqs_disabled()))
3962 add_preempt_count(PREEMPT_ACTIVE
);
3965 * We keep the big kernel semaphore locked, but we
3966 * clear ->lock_depth so that schedule() doesnt
3967 * auto-release the semaphore:
3969 saved_lock_depth
= task
->lock_depth
;
3970 task
->lock_depth
= -1;
3972 task
->lock_depth
= saved_lock_depth
;
3973 sub_preempt_count(PREEMPT_ACTIVE
);
3976 * Check again in case we missed a preemption opportunity
3977 * between schedule and now.
3980 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3982 EXPORT_SYMBOL(preempt_schedule
);
3985 * this is the entry point to schedule() from kernel preemption
3986 * off of irq context.
3987 * Note, that this is called and return with irqs disabled. This will
3988 * protect us against recursive calling from irq.
3990 asmlinkage
void __sched
preempt_schedule_irq(void)
3992 struct thread_info
*ti
= current_thread_info();
3993 struct task_struct
*task
= current
;
3994 int saved_lock_depth
;
3996 /* Catch callers which need to be fixed */
3997 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4000 add_preempt_count(PREEMPT_ACTIVE
);
4003 * We keep the big kernel semaphore locked, but we
4004 * clear ->lock_depth so that schedule() doesnt
4005 * auto-release the semaphore:
4007 saved_lock_depth
= task
->lock_depth
;
4008 task
->lock_depth
= -1;
4011 local_irq_disable();
4012 task
->lock_depth
= saved_lock_depth
;
4013 sub_preempt_count(PREEMPT_ACTIVE
);
4016 * Check again in case we missed a preemption opportunity
4017 * between schedule and now.
4020 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4023 #endif /* CONFIG_PREEMPT */
4025 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4028 return try_to_wake_up(curr
->private, mode
, sync
);
4030 EXPORT_SYMBOL(default_wake_function
);
4033 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4034 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4035 * number) then we wake all the non-exclusive tasks and one exclusive task.
4037 * There are circumstances in which we can try to wake a task which has already
4038 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4039 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4041 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4042 int nr_exclusive
, int sync
, void *key
)
4044 wait_queue_t
*curr
, *next
;
4046 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4047 unsigned flags
= curr
->flags
;
4049 if (curr
->func(curr
, mode
, sync
, key
) &&
4050 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4056 * __wake_up - wake up threads blocked on a waitqueue.
4058 * @mode: which threads
4059 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4060 * @key: is directly passed to the wakeup function
4062 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4063 int nr_exclusive
, void *key
)
4065 unsigned long flags
;
4067 spin_lock_irqsave(&q
->lock
, flags
);
4068 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4069 spin_unlock_irqrestore(&q
->lock
, flags
);
4071 EXPORT_SYMBOL(__wake_up
);
4074 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4076 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4078 __wake_up_common(q
, mode
, 1, 0, NULL
);
4082 * __wake_up_sync - wake up threads blocked on a waitqueue.
4084 * @mode: which threads
4085 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4087 * The sync wakeup differs that the waker knows that it will schedule
4088 * away soon, so while the target thread will be woken up, it will not
4089 * be migrated to another CPU - ie. the two threads are 'synchronized'
4090 * with each other. This can prevent needless bouncing between CPUs.
4092 * On UP it can prevent extra preemption.
4095 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4097 unsigned long flags
;
4103 if (unlikely(!nr_exclusive
))
4106 spin_lock_irqsave(&q
->lock
, flags
);
4107 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4108 spin_unlock_irqrestore(&q
->lock
, flags
);
4110 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4112 void complete(struct completion
*x
)
4114 unsigned long flags
;
4116 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4118 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4119 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4121 EXPORT_SYMBOL(complete
);
4123 void complete_all(struct completion
*x
)
4125 unsigned long flags
;
4127 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4128 x
->done
+= UINT_MAX
/2;
4129 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4130 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4132 EXPORT_SYMBOL(complete_all
);
4134 static inline long __sched
4135 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4138 DECLARE_WAITQUEUE(wait
, current
);
4140 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4141 __add_wait_queue_tail(&x
->wait
, &wait
);
4143 if ((state
== TASK_INTERRUPTIBLE
&&
4144 signal_pending(current
)) ||
4145 (state
== TASK_KILLABLE
&&
4146 fatal_signal_pending(current
))) {
4147 __remove_wait_queue(&x
->wait
, &wait
);
4148 return -ERESTARTSYS
;
4150 __set_current_state(state
);
4151 spin_unlock_irq(&x
->wait
.lock
);
4152 timeout
= schedule_timeout(timeout
);
4153 spin_lock_irq(&x
->wait
.lock
);
4155 __remove_wait_queue(&x
->wait
, &wait
);
4159 __remove_wait_queue(&x
->wait
, &wait
);
4166 wait_for_common(struct completion
*x
, long timeout
, int state
)
4170 spin_lock_irq(&x
->wait
.lock
);
4171 timeout
= do_wait_for_common(x
, timeout
, state
);
4172 spin_unlock_irq(&x
->wait
.lock
);
4176 void __sched
wait_for_completion(struct completion
*x
)
4178 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4180 EXPORT_SYMBOL(wait_for_completion
);
4182 unsigned long __sched
4183 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4185 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4187 EXPORT_SYMBOL(wait_for_completion_timeout
);
4189 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4191 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4192 if (t
== -ERESTARTSYS
)
4196 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4198 unsigned long __sched
4199 wait_for_completion_interruptible_timeout(struct completion
*x
,
4200 unsigned long timeout
)
4202 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4204 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4206 int __sched
wait_for_completion_killable(struct completion
*x
)
4208 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4209 if (t
== -ERESTARTSYS
)
4213 EXPORT_SYMBOL(wait_for_completion_killable
);
4216 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4218 unsigned long flags
;
4221 init_waitqueue_entry(&wait
, current
);
4223 __set_current_state(state
);
4225 spin_lock_irqsave(&q
->lock
, flags
);
4226 __add_wait_queue(q
, &wait
);
4227 spin_unlock(&q
->lock
);
4228 timeout
= schedule_timeout(timeout
);
4229 spin_lock_irq(&q
->lock
);
4230 __remove_wait_queue(q
, &wait
);
4231 spin_unlock_irqrestore(&q
->lock
, flags
);
4236 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4238 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4240 EXPORT_SYMBOL(interruptible_sleep_on
);
4243 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4245 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4247 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4249 void __sched
sleep_on(wait_queue_head_t
*q
)
4251 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4253 EXPORT_SYMBOL(sleep_on
);
4255 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4257 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4259 EXPORT_SYMBOL(sleep_on_timeout
);
4261 #ifdef CONFIG_RT_MUTEXES
4264 * rt_mutex_setprio - set the current priority of a task
4266 * @prio: prio value (kernel-internal form)
4268 * This function changes the 'effective' priority of a task. It does
4269 * not touch ->normal_prio like __setscheduler().
4271 * Used by the rt_mutex code to implement priority inheritance logic.
4273 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4275 unsigned long flags
;
4276 int oldprio
, on_rq
, running
;
4278 const struct sched_class
*prev_class
= p
->sched_class
;
4280 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4282 rq
= task_rq_lock(p
, &flags
);
4283 update_rq_clock(rq
);
4286 on_rq
= p
->se
.on_rq
;
4287 running
= task_current(rq
, p
);
4289 dequeue_task(rq
, p
, 0);
4291 p
->sched_class
->put_prev_task(rq
, p
);
4295 p
->sched_class
= &rt_sched_class
;
4297 p
->sched_class
= &fair_sched_class
;
4303 p
->sched_class
->set_curr_task(rq
);
4305 enqueue_task(rq
, p
, 0);
4307 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4309 task_rq_unlock(rq
, &flags
);
4314 void set_user_nice(struct task_struct
*p
, long nice
)
4316 int old_prio
, delta
, on_rq
;
4317 unsigned long flags
;
4320 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4323 * We have to be careful, if called from sys_setpriority(),
4324 * the task might be in the middle of scheduling on another CPU.
4326 rq
= task_rq_lock(p
, &flags
);
4327 update_rq_clock(rq
);
4329 * The RT priorities are set via sched_setscheduler(), but we still
4330 * allow the 'normal' nice value to be set - but as expected
4331 * it wont have any effect on scheduling until the task is
4332 * SCHED_FIFO/SCHED_RR:
4334 if (task_has_rt_policy(p
)) {
4335 p
->static_prio
= NICE_TO_PRIO(nice
);
4338 on_rq
= p
->se
.on_rq
;
4340 dequeue_task(rq
, p
, 0);
4342 p
->static_prio
= NICE_TO_PRIO(nice
);
4345 p
->prio
= effective_prio(p
);
4346 delta
= p
->prio
- old_prio
;
4349 enqueue_task(rq
, p
, 0);
4351 * If the task increased its priority or is running and
4352 * lowered its priority, then reschedule its CPU:
4354 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4355 resched_task(rq
->curr
);
4358 task_rq_unlock(rq
, &flags
);
4360 EXPORT_SYMBOL(set_user_nice
);
4363 * can_nice - check if a task can reduce its nice value
4367 int can_nice(const struct task_struct
*p
, const int nice
)
4369 /* convert nice value [19,-20] to rlimit style value [1,40] */
4370 int nice_rlim
= 20 - nice
;
4372 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4373 capable(CAP_SYS_NICE
));
4376 #ifdef __ARCH_WANT_SYS_NICE
4379 * sys_nice - change the priority of the current process.
4380 * @increment: priority increment
4382 * sys_setpriority is a more generic, but much slower function that
4383 * does similar things.
4385 asmlinkage
long sys_nice(int increment
)
4390 * Setpriority might change our priority at the same moment.
4391 * We don't have to worry. Conceptually one call occurs first
4392 * and we have a single winner.
4394 if (increment
< -40)
4399 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4405 if (increment
< 0 && !can_nice(current
, nice
))
4408 retval
= security_task_setnice(current
, nice
);
4412 set_user_nice(current
, nice
);
4419 * task_prio - return the priority value of a given task.
4420 * @p: the task in question.
4422 * This is the priority value as seen by users in /proc.
4423 * RT tasks are offset by -200. Normal tasks are centered
4424 * around 0, value goes from -16 to +15.
4426 int task_prio(const struct task_struct
*p
)
4428 return p
->prio
- MAX_RT_PRIO
;
4432 * task_nice - return the nice value of a given task.
4433 * @p: the task in question.
4435 int task_nice(const struct task_struct
*p
)
4437 return TASK_NICE(p
);
4439 EXPORT_SYMBOL_GPL(task_nice
);
4442 * idle_cpu - is a given cpu idle currently?
4443 * @cpu: the processor in question.
4445 int idle_cpu(int cpu
)
4447 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4451 * idle_task - return the idle task for a given cpu.
4452 * @cpu: the processor in question.
4454 struct task_struct
*idle_task(int cpu
)
4456 return cpu_rq(cpu
)->idle
;
4460 * find_process_by_pid - find a process with a matching PID value.
4461 * @pid: the pid in question.
4463 static struct task_struct
*find_process_by_pid(pid_t pid
)
4465 return pid
? find_task_by_vpid(pid
) : current
;
4468 /* Actually do priority change: must hold rq lock. */
4470 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4472 BUG_ON(p
->se
.on_rq
);
4475 switch (p
->policy
) {
4479 p
->sched_class
= &fair_sched_class
;
4483 p
->sched_class
= &rt_sched_class
;
4487 p
->rt_priority
= prio
;
4488 p
->normal_prio
= normal_prio(p
);
4489 /* we are holding p->pi_lock already */
4490 p
->prio
= rt_mutex_getprio(p
);
4495 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4496 * @p: the task in question.
4497 * @policy: new policy.
4498 * @param: structure containing the new RT priority.
4500 * NOTE that the task may be already dead.
4502 int sched_setscheduler(struct task_struct
*p
, int policy
,
4503 struct sched_param
*param
)
4505 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4506 unsigned long flags
;
4507 const struct sched_class
*prev_class
= p
->sched_class
;
4510 /* may grab non-irq protected spin_locks */
4511 BUG_ON(in_interrupt());
4513 /* double check policy once rq lock held */
4515 policy
= oldpolicy
= p
->policy
;
4516 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4517 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4518 policy
!= SCHED_IDLE
)
4521 * Valid priorities for SCHED_FIFO and SCHED_RR are
4522 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4523 * SCHED_BATCH and SCHED_IDLE is 0.
4525 if (param
->sched_priority
< 0 ||
4526 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4527 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4529 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4533 * Allow unprivileged RT tasks to decrease priority:
4535 if (!capable(CAP_SYS_NICE
)) {
4536 if (rt_policy(policy
)) {
4537 unsigned long rlim_rtprio
;
4539 if (!lock_task_sighand(p
, &flags
))
4541 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4542 unlock_task_sighand(p
, &flags
);
4544 /* can't set/change the rt policy */
4545 if (policy
!= p
->policy
&& !rlim_rtprio
)
4548 /* can't increase priority */
4549 if (param
->sched_priority
> p
->rt_priority
&&
4550 param
->sched_priority
> rlim_rtprio
)
4554 * Like positive nice levels, dont allow tasks to
4555 * move out of SCHED_IDLE either:
4557 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4560 /* can't change other user's priorities */
4561 if ((current
->euid
!= p
->euid
) &&
4562 (current
->euid
!= p
->uid
))
4566 retval
= security_task_setscheduler(p
, policy
, param
);
4570 * make sure no PI-waiters arrive (or leave) while we are
4571 * changing the priority of the task:
4573 spin_lock_irqsave(&p
->pi_lock
, flags
);
4575 * To be able to change p->policy safely, the apropriate
4576 * runqueue lock must be held.
4578 rq
= __task_rq_lock(p
);
4579 /* recheck policy now with rq lock held */
4580 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4581 policy
= oldpolicy
= -1;
4582 __task_rq_unlock(rq
);
4583 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4586 update_rq_clock(rq
);
4587 on_rq
= p
->se
.on_rq
;
4588 running
= task_current(rq
, p
);
4590 deactivate_task(rq
, p
, 0);
4592 p
->sched_class
->put_prev_task(rq
, p
);
4596 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4600 p
->sched_class
->set_curr_task(rq
);
4602 activate_task(rq
, p
, 0);
4604 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4606 __task_rq_unlock(rq
);
4607 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4609 rt_mutex_adjust_pi(p
);
4613 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4616 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4618 struct sched_param lparam
;
4619 struct task_struct
*p
;
4622 if (!param
|| pid
< 0)
4624 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4629 p
= find_process_by_pid(pid
);
4631 retval
= sched_setscheduler(p
, policy
, &lparam
);
4638 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4639 * @pid: the pid in question.
4640 * @policy: new policy.
4641 * @param: structure containing the new RT priority.
4644 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4646 /* negative values for policy are not valid */
4650 return do_sched_setscheduler(pid
, policy
, param
);
4654 * sys_sched_setparam - set/change the RT priority of a thread
4655 * @pid: the pid in question.
4656 * @param: structure containing the new RT priority.
4658 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4660 return do_sched_setscheduler(pid
, -1, param
);
4664 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4665 * @pid: the pid in question.
4667 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4669 struct task_struct
*p
;
4676 read_lock(&tasklist_lock
);
4677 p
= find_process_by_pid(pid
);
4679 retval
= security_task_getscheduler(p
);
4683 read_unlock(&tasklist_lock
);
4688 * sys_sched_getscheduler - get the RT priority of a thread
4689 * @pid: the pid in question.
4690 * @param: structure containing the RT priority.
4692 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4694 struct sched_param lp
;
4695 struct task_struct
*p
;
4698 if (!param
|| pid
< 0)
4701 read_lock(&tasklist_lock
);
4702 p
= find_process_by_pid(pid
);
4707 retval
= security_task_getscheduler(p
);
4711 lp
.sched_priority
= p
->rt_priority
;
4712 read_unlock(&tasklist_lock
);
4715 * This one might sleep, we cannot do it with a spinlock held ...
4717 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4722 read_unlock(&tasklist_lock
);
4726 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4728 cpumask_t cpus_allowed
;
4729 struct task_struct
*p
;
4733 read_lock(&tasklist_lock
);
4735 p
= find_process_by_pid(pid
);
4737 read_unlock(&tasklist_lock
);
4743 * It is not safe to call set_cpus_allowed with the
4744 * tasklist_lock held. We will bump the task_struct's
4745 * usage count and then drop tasklist_lock.
4748 read_unlock(&tasklist_lock
);
4751 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4752 !capable(CAP_SYS_NICE
))
4755 retval
= security_task_setscheduler(p
, 0, NULL
);
4759 cpus_allowed
= cpuset_cpus_allowed(p
);
4760 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4762 retval
= set_cpus_allowed(p
, new_mask
);
4765 cpus_allowed
= cpuset_cpus_allowed(p
);
4766 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4768 * We must have raced with a concurrent cpuset
4769 * update. Just reset the cpus_allowed to the
4770 * cpuset's cpus_allowed
4772 new_mask
= cpus_allowed
;
4782 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4783 cpumask_t
*new_mask
)
4785 if (len
< sizeof(cpumask_t
)) {
4786 memset(new_mask
, 0, sizeof(cpumask_t
));
4787 } else if (len
> sizeof(cpumask_t
)) {
4788 len
= sizeof(cpumask_t
);
4790 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4794 * sys_sched_setaffinity - set the cpu affinity of a process
4795 * @pid: pid of the process
4796 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4797 * @user_mask_ptr: user-space pointer to the new cpu mask
4799 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4800 unsigned long __user
*user_mask_ptr
)
4805 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4809 return sched_setaffinity(pid
, new_mask
);
4813 * Represents all cpu's present in the system
4814 * In systems capable of hotplug, this map could dynamically grow
4815 * as new cpu's are detected in the system via any platform specific
4816 * method, such as ACPI for e.g.
4819 cpumask_t cpu_present_map __read_mostly
;
4820 EXPORT_SYMBOL(cpu_present_map
);
4823 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4824 EXPORT_SYMBOL(cpu_online_map
);
4826 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4827 EXPORT_SYMBOL(cpu_possible_map
);
4830 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4832 struct task_struct
*p
;
4836 read_lock(&tasklist_lock
);
4839 p
= find_process_by_pid(pid
);
4843 retval
= security_task_getscheduler(p
);
4847 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4850 read_unlock(&tasklist_lock
);
4857 * sys_sched_getaffinity - get the cpu affinity of a process
4858 * @pid: pid of the process
4859 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4860 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4862 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4863 unsigned long __user
*user_mask_ptr
)
4868 if (len
< sizeof(cpumask_t
))
4871 ret
= sched_getaffinity(pid
, &mask
);
4875 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4878 return sizeof(cpumask_t
);
4882 * sys_sched_yield - yield the current processor to other threads.
4884 * This function yields the current CPU to other tasks. If there are no
4885 * other threads running on this CPU then this function will return.
4887 asmlinkage
long sys_sched_yield(void)
4889 struct rq
*rq
= this_rq_lock();
4891 schedstat_inc(rq
, yld_count
);
4892 current
->sched_class
->yield_task(rq
);
4895 * Since we are going to call schedule() anyway, there's
4896 * no need to preempt or enable interrupts:
4898 __release(rq
->lock
);
4899 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4900 _raw_spin_unlock(&rq
->lock
);
4901 preempt_enable_no_resched();
4908 static void __cond_resched(void)
4910 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4911 __might_sleep(__FILE__
, __LINE__
);
4914 * The BKS might be reacquired before we have dropped
4915 * PREEMPT_ACTIVE, which could trigger a second
4916 * cond_resched() call.
4919 add_preempt_count(PREEMPT_ACTIVE
);
4921 sub_preempt_count(PREEMPT_ACTIVE
);
4922 } while (need_resched());
4925 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
4926 int __sched
_cond_resched(void)
4928 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4929 system_state
== SYSTEM_RUNNING
) {
4935 EXPORT_SYMBOL(_cond_resched
);
4939 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4940 * call schedule, and on return reacquire the lock.
4942 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4943 * operations here to prevent schedule() from being called twice (once via
4944 * spin_unlock(), once by hand).
4946 int cond_resched_lock(spinlock_t
*lock
)
4948 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
4951 if (spin_needbreak(lock
) || resched
) {
4953 if (resched
&& need_resched())
4962 EXPORT_SYMBOL(cond_resched_lock
);
4964 int __sched
cond_resched_softirq(void)
4966 BUG_ON(!in_softirq());
4968 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4976 EXPORT_SYMBOL(cond_resched_softirq
);
4979 * yield - yield the current processor to other threads.
4981 * This is a shortcut for kernel-space yielding - it marks the
4982 * thread runnable and calls sys_sched_yield().
4984 void __sched
yield(void)
4986 set_current_state(TASK_RUNNING
);
4989 EXPORT_SYMBOL(yield
);
4992 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4993 * that process accounting knows that this is a task in IO wait state.
4995 * But don't do that if it is a deliberate, throttling IO wait (this task
4996 * has set its backing_dev_info: the queue against which it should throttle)
4998 void __sched
io_schedule(void)
5000 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5002 delayacct_blkio_start();
5003 atomic_inc(&rq
->nr_iowait
);
5005 atomic_dec(&rq
->nr_iowait
);
5006 delayacct_blkio_end();
5008 EXPORT_SYMBOL(io_schedule
);
5010 long __sched
io_schedule_timeout(long timeout
)
5012 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5015 delayacct_blkio_start();
5016 atomic_inc(&rq
->nr_iowait
);
5017 ret
= schedule_timeout(timeout
);
5018 atomic_dec(&rq
->nr_iowait
);
5019 delayacct_blkio_end();
5024 * sys_sched_get_priority_max - return maximum RT priority.
5025 * @policy: scheduling class.
5027 * this syscall returns the maximum rt_priority that can be used
5028 * by a given scheduling class.
5030 asmlinkage
long sys_sched_get_priority_max(int policy
)
5037 ret
= MAX_USER_RT_PRIO
-1;
5049 * sys_sched_get_priority_min - return minimum RT priority.
5050 * @policy: scheduling class.
5052 * this syscall returns the minimum rt_priority that can be used
5053 * by a given scheduling class.
5055 asmlinkage
long sys_sched_get_priority_min(int policy
)
5073 * sys_sched_rr_get_interval - return the default timeslice of a process.
5074 * @pid: pid of the process.
5075 * @interval: userspace pointer to the timeslice value.
5077 * this syscall writes the default timeslice value of a given process
5078 * into the user-space timespec buffer. A value of '0' means infinity.
5081 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5083 struct task_struct
*p
;
5084 unsigned int time_slice
;
5092 read_lock(&tasklist_lock
);
5093 p
= find_process_by_pid(pid
);
5097 retval
= security_task_getscheduler(p
);
5102 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5103 * tasks that are on an otherwise idle runqueue:
5106 if (p
->policy
== SCHED_RR
) {
5107 time_slice
= DEF_TIMESLICE
;
5109 struct sched_entity
*se
= &p
->se
;
5110 unsigned long flags
;
5113 rq
= task_rq_lock(p
, &flags
);
5114 if (rq
->cfs
.load
.weight
)
5115 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5116 task_rq_unlock(rq
, &flags
);
5118 read_unlock(&tasklist_lock
);
5119 jiffies_to_timespec(time_slice
, &t
);
5120 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5124 read_unlock(&tasklist_lock
);
5128 static const char stat_nam
[] = "RSDTtZX";
5130 void sched_show_task(struct task_struct
*p
)
5132 unsigned long free
= 0;
5135 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5136 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5137 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5138 #if BITS_PER_LONG == 32
5139 if (state
== TASK_RUNNING
)
5140 printk(KERN_CONT
" running ");
5142 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5144 if (state
== TASK_RUNNING
)
5145 printk(KERN_CONT
" running task ");
5147 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5149 #ifdef CONFIG_DEBUG_STACK_USAGE
5151 unsigned long *n
= end_of_stack(p
);
5154 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5157 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5158 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5160 show_stack(p
, NULL
);
5163 void show_state_filter(unsigned long state_filter
)
5165 struct task_struct
*g
, *p
;
5167 #if BITS_PER_LONG == 32
5169 " task PC stack pid father\n");
5172 " task PC stack pid father\n");
5174 read_lock(&tasklist_lock
);
5175 do_each_thread(g
, p
) {
5177 * reset the NMI-timeout, listing all files on a slow
5178 * console might take alot of time:
5180 touch_nmi_watchdog();
5181 if (!state_filter
|| (p
->state
& state_filter
))
5183 } while_each_thread(g
, p
);
5185 touch_all_softlockup_watchdogs();
5187 #ifdef CONFIG_SCHED_DEBUG
5188 sysrq_sched_debug_show();
5190 read_unlock(&tasklist_lock
);
5192 * Only show locks if all tasks are dumped:
5194 if (state_filter
== -1)
5195 debug_show_all_locks();
5198 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5200 idle
->sched_class
= &idle_sched_class
;
5204 * init_idle - set up an idle thread for a given CPU
5205 * @idle: task in question
5206 * @cpu: cpu the idle task belongs to
5208 * NOTE: this function does not set the idle thread's NEED_RESCHED
5209 * flag, to make booting more robust.
5211 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5213 struct rq
*rq
= cpu_rq(cpu
);
5214 unsigned long flags
;
5217 idle
->se
.exec_start
= sched_clock();
5219 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5220 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5221 __set_task_cpu(idle
, cpu
);
5223 spin_lock_irqsave(&rq
->lock
, flags
);
5224 rq
->curr
= rq
->idle
= idle
;
5225 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5228 spin_unlock_irqrestore(&rq
->lock
, flags
);
5230 /* Set the preempt count _outside_ the spinlocks! */
5231 task_thread_info(idle
)->preempt_count
= 0;
5234 * The idle tasks have their own, simple scheduling class:
5236 idle
->sched_class
= &idle_sched_class
;
5240 * In a system that switches off the HZ timer nohz_cpu_mask
5241 * indicates which cpus entered this state. This is used
5242 * in the rcu update to wait only for active cpus. For system
5243 * which do not switch off the HZ timer nohz_cpu_mask should
5244 * always be CPU_MASK_NONE.
5246 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5249 * Increase the granularity value when there are more CPUs,
5250 * because with more CPUs the 'effective latency' as visible
5251 * to users decreases. But the relationship is not linear,
5252 * so pick a second-best guess by going with the log2 of the
5255 * This idea comes from the SD scheduler of Con Kolivas:
5257 static inline void sched_init_granularity(void)
5259 unsigned int factor
= 1 + ilog2(num_online_cpus());
5260 const unsigned long limit
= 200000000;
5262 sysctl_sched_min_granularity
*= factor
;
5263 if (sysctl_sched_min_granularity
> limit
)
5264 sysctl_sched_min_granularity
= limit
;
5266 sysctl_sched_latency
*= factor
;
5267 if (sysctl_sched_latency
> limit
)
5268 sysctl_sched_latency
= limit
;
5270 sysctl_sched_wakeup_granularity
*= factor
;
5271 sysctl_sched_batch_wakeup_granularity
*= factor
;
5276 * This is how migration works:
5278 * 1) we queue a struct migration_req structure in the source CPU's
5279 * runqueue and wake up that CPU's migration thread.
5280 * 2) we down() the locked semaphore => thread blocks.
5281 * 3) migration thread wakes up (implicitly it forces the migrated
5282 * thread off the CPU)
5283 * 4) it gets the migration request and checks whether the migrated
5284 * task is still in the wrong runqueue.
5285 * 5) if it's in the wrong runqueue then the migration thread removes
5286 * it and puts it into the right queue.
5287 * 6) migration thread up()s the semaphore.
5288 * 7) we wake up and the migration is done.
5292 * Change a given task's CPU affinity. Migrate the thread to a
5293 * proper CPU and schedule it away if the CPU it's executing on
5294 * is removed from the allowed bitmask.
5296 * NOTE: the caller must have a valid reference to the task, the
5297 * task must not exit() & deallocate itself prematurely. The
5298 * call is not atomic; no spinlocks may be held.
5300 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5302 struct migration_req req
;
5303 unsigned long flags
;
5307 rq
= task_rq_lock(p
, &flags
);
5308 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5313 if (p
->sched_class
->set_cpus_allowed
)
5314 p
->sched_class
->set_cpus_allowed(p
, &new_mask
);
5316 p
->cpus_allowed
= new_mask
;
5317 p
->rt
.nr_cpus_allowed
= cpus_weight(new_mask
);
5320 /* Can the task run on the task's current CPU? If so, we're done */
5321 if (cpu_isset(task_cpu(p
), new_mask
))
5324 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5325 /* Need help from migration thread: drop lock and wait. */
5326 task_rq_unlock(rq
, &flags
);
5327 wake_up_process(rq
->migration_thread
);
5328 wait_for_completion(&req
.done
);
5329 tlb_migrate_finish(p
->mm
);
5333 task_rq_unlock(rq
, &flags
);
5337 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5340 * Move (not current) task off this cpu, onto dest cpu. We're doing
5341 * this because either it can't run here any more (set_cpus_allowed()
5342 * away from this CPU, or CPU going down), or because we're
5343 * attempting to rebalance this task on exec (sched_exec).
5345 * So we race with normal scheduler movements, but that's OK, as long
5346 * as the task is no longer on this CPU.
5348 * Returns non-zero if task was successfully migrated.
5350 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5352 struct rq
*rq_dest
, *rq_src
;
5355 if (unlikely(cpu_is_offline(dest_cpu
)))
5358 rq_src
= cpu_rq(src_cpu
);
5359 rq_dest
= cpu_rq(dest_cpu
);
5361 double_rq_lock(rq_src
, rq_dest
);
5362 /* Already moved. */
5363 if (task_cpu(p
) != src_cpu
)
5365 /* Affinity changed (again). */
5366 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5369 on_rq
= p
->se
.on_rq
;
5371 deactivate_task(rq_src
, p
, 0);
5373 set_task_cpu(p
, dest_cpu
);
5375 activate_task(rq_dest
, p
, 0);
5376 check_preempt_curr(rq_dest
, p
);
5380 double_rq_unlock(rq_src
, rq_dest
);
5385 * migration_thread - this is a highprio system thread that performs
5386 * thread migration by bumping thread off CPU then 'pushing' onto
5389 static int migration_thread(void *data
)
5391 int cpu
= (long)data
;
5395 BUG_ON(rq
->migration_thread
!= current
);
5397 set_current_state(TASK_INTERRUPTIBLE
);
5398 while (!kthread_should_stop()) {
5399 struct migration_req
*req
;
5400 struct list_head
*head
;
5402 spin_lock_irq(&rq
->lock
);
5404 if (cpu_is_offline(cpu
)) {
5405 spin_unlock_irq(&rq
->lock
);
5409 if (rq
->active_balance
) {
5410 active_load_balance(rq
, cpu
);
5411 rq
->active_balance
= 0;
5414 head
= &rq
->migration_queue
;
5416 if (list_empty(head
)) {
5417 spin_unlock_irq(&rq
->lock
);
5419 set_current_state(TASK_INTERRUPTIBLE
);
5422 req
= list_entry(head
->next
, struct migration_req
, list
);
5423 list_del_init(head
->next
);
5425 spin_unlock(&rq
->lock
);
5426 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5429 complete(&req
->done
);
5431 __set_current_state(TASK_RUNNING
);
5435 /* Wait for kthread_stop */
5436 set_current_state(TASK_INTERRUPTIBLE
);
5437 while (!kthread_should_stop()) {
5439 set_current_state(TASK_INTERRUPTIBLE
);
5441 __set_current_state(TASK_RUNNING
);
5445 #ifdef CONFIG_HOTPLUG_CPU
5447 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5451 local_irq_disable();
5452 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5458 * Figure out where task on dead CPU should go, use force if necessary.
5459 * NOTE: interrupts should be disabled by the caller
5461 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5463 unsigned long flags
;
5470 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5471 cpus_and(mask
, mask
, p
->cpus_allowed
);
5472 dest_cpu
= any_online_cpu(mask
);
5474 /* On any allowed CPU? */
5475 if (dest_cpu
== NR_CPUS
)
5476 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5478 /* No more Mr. Nice Guy. */
5479 if (dest_cpu
== NR_CPUS
) {
5480 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5482 * Try to stay on the same cpuset, where the
5483 * current cpuset may be a subset of all cpus.
5484 * The cpuset_cpus_allowed_locked() variant of
5485 * cpuset_cpus_allowed() will not block. It must be
5486 * called within calls to cpuset_lock/cpuset_unlock.
5488 rq
= task_rq_lock(p
, &flags
);
5489 p
->cpus_allowed
= cpus_allowed
;
5490 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5491 task_rq_unlock(rq
, &flags
);
5494 * Don't tell them about moving exiting tasks or
5495 * kernel threads (both mm NULL), since they never
5498 if (p
->mm
&& printk_ratelimit()) {
5499 printk(KERN_INFO
"process %d (%s) no "
5500 "longer affine to cpu%d\n",
5501 task_pid_nr(p
), p
->comm
, dead_cpu
);
5504 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5508 * While a dead CPU has no uninterruptible tasks queued at this point,
5509 * it might still have a nonzero ->nr_uninterruptible counter, because
5510 * for performance reasons the counter is not stricly tracking tasks to
5511 * their home CPUs. So we just add the counter to another CPU's counter,
5512 * to keep the global sum constant after CPU-down:
5514 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5516 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5517 unsigned long flags
;
5519 local_irq_save(flags
);
5520 double_rq_lock(rq_src
, rq_dest
);
5521 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5522 rq_src
->nr_uninterruptible
= 0;
5523 double_rq_unlock(rq_src
, rq_dest
);
5524 local_irq_restore(flags
);
5527 /* Run through task list and migrate tasks from the dead cpu. */
5528 static void migrate_live_tasks(int src_cpu
)
5530 struct task_struct
*p
, *t
;
5532 read_lock(&tasklist_lock
);
5534 do_each_thread(t
, p
) {
5538 if (task_cpu(p
) == src_cpu
)
5539 move_task_off_dead_cpu(src_cpu
, p
);
5540 } while_each_thread(t
, p
);
5542 read_unlock(&tasklist_lock
);
5546 * Schedules idle task to be the next runnable task on current CPU.
5547 * It does so by boosting its priority to highest possible.
5548 * Used by CPU offline code.
5550 void sched_idle_next(void)
5552 int this_cpu
= smp_processor_id();
5553 struct rq
*rq
= cpu_rq(this_cpu
);
5554 struct task_struct
*p
= rq
->idle
;
5555 unsigned long flags
;
5557 /* cpu has to be offline */
5558 BUG_ON(cpu_online(this_cpu
));
5561 * Strictly not necessary since rest of the CPUs are stopped by now
5562 * and interrupts disabled on the current cpu.
5564 spin_lock_irqsave(&rq
->lock
, flags
);
5566 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5568 update_rq_clock(rq
);
5569 activate_task(rq
, p
, 0);
5571 spin_unlock_irqrestore(&rq
->lock
, flags
);
5575 * Ensures that the idle task is using init_mm right before its cpu goes
5578 void idle_task_exit(void)
5580 struct mm_struct
*mm
= current
->active_mm
;
5582 BUG_ON(cpu_online(smp_processor_id()));
5585 switch_mm(mm
, &init_mm
, current
);
5589 /* called under rq->lock with disabled interrupts */
5590 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5592 struct rq
*rq
= cpu_rq(dead_cpu
);
5594 /* Must be exiting, otherwise would be on tasklist. */
5595 BUG_ON(!p
->exit_state
);
5597 /* Cannot have done final schedule yet: would have vanished. */
5598 BUG_ON(p
->state
== TASK_DEAD
);
5603 * Drop lock around migration; if someone else moves it,
5604 * that's OK. No task can be added to this CPU, so iteration is
5607 spin_unlock_irq(&rq
->lock
);
5608 move_task_off_dead_cpu(dead_cpu
, p
);
5609 spin_lock_irq(&rq
->lock
);
5614 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5615 static void migrate_dead_tasks(unsigned int dead_cpu
)
5617 struct rq
*rq
= cpu_rq(dead_cpu
);
5618 struct task_struct
*next
;
5621 if (!rq
->nr_running
)
5623 update_rq_clock(rq
);
5624 next
= pick_next_task(rq
, rq
->curr
);
5627 migrate_dead(dead_cpu
, next
);
5631 #endif /* CONFIG_HOTPLUG_CPU */
5633 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5635 static struct ctl_table sd_ctl_dir
[] = {
5637 .procname
= "sched_domain",
5643 static struct ctl_table sd_ctl_root
[] = {
5645 .ctl_name
= CTL_KERN
,
5646 .procname
= "kernel",
5648 .child
= sd_ctl_dir
,
5653 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5655 struct ctl_table
*entry
=
5656 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5661 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5663 struct ctl_table
*entry
;
5666 * In the intermediate directories, both the child directory and
5667 * procname are dynamically allocated and could fail but the mode
5668 * will always be set. In the lowest directory the names are
5669 * static strings and all have proc handlers.
5671 for (entry
= *tablep
; entry
->mode
; entry
++) {
5673 sd_free_ctl_entry(&entry
->child
);
5674 if (entry
->proc_handler
== NULL
)
5675 kfree(entry
->procname
);
5683 set_table_entry(struct ctl_table
*entry
,
5684 const char *procname
, void *data
, int maxlen
,
5685 mode_t mode
, proc_handler
*proc_handler
)
5687 entry
->procname
= procname
;
5689 entry
->maxlen
= maxlen
;
5691 entry
->proc_handler
= proc_handler
;
5694 static struct ctl_table
*
5695 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5697 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5702 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5703 sizeof(long), 0644, proc_doulongvec_minmax
);
5704 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5705 sizeof(long), 0644, proc_doulongvec_minmax
);
5706 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5707 sizeof(int), 0644, proc_dointvec_minmax
);
5708 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5709 sizeof(int), 0644, proc_dointvec_minmax
);
5710 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5711 sizeof(int), 0644, proc_dointvec_minmax
);
5712 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5713 sizeof(int), 0644, proc_dointvec_minmax
);
5714 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5715 sizeof(int), 0644, proc_dointvec_minmax
);
5716 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5717 sizeof(int), 0644, proc_dointvec_minmax
);
5718 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5719 sizeof(int), 0644, proc_dointvec_minmax
);
5720 set_table_entry(&table
[9], "cache_nice_tries",
5721 &sd
->cache_nice_tries
,
5722 sizeof(int), 0644, proc_dointvec_minmax
);
5723 set_table_entry(&table
[10], "flags", &sd
->flags
,
5724 sizeof(int), 0644, proc_dointvec_minmax
);
5725 /* &table[11] is terminator */
5730 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5732 struct ctl_table
*entry
, *table
;
5733 struct sched_domain
*sd
;
5734 int domain_num
= 0, i
;
5737 for_each_domain(cpu
, sd
)
5739 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5744 for_each_domain(cpu
, sd
) {
5745 snprintf(buf
, 32, "domain%d", i
);
5746 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5748 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5755 static struct ctl_table_header
*sd_sysctl_header
;
5756 static void register_sched_domain_sysctl(void)
5758 int i
, cpu_num
= num_online_cpus();
5759 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5762 WARN_ON(sd_ctl_dir
[0].child
);
5763 sd_ctl_dir
[0].child
= entry
;
5768 for_each_online_cpu(i
) {
5769 snprintf(buf
, 32, "cpu%d", i
);
5770 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5772 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5776 WARN_ON(sd_sysctl_header
);
5777 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5780 /* may be called multiple times per register */
5781 static void unregister_sched_domain_sysctl(void)
5783 if (sd_sysctl_header
)
5784 unregister_sysctl_table(sd_sysctl_header
);
5785 sd_sysctl_header
= NULL
;
5786 if (sd_ctl_dir
[0].child
)
5787 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5790 static void register_sched_domain_sysctl(void)
5793 static void unregister_sched_domain_sysctl(void)
5799 * migration_call - callback that gets triggered when a CPU is added.
5800 * Here we can start up the necessary migration thread for the new CPU.
5802 static int __cpuinit
5803 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5805 struct task_struct
*p
;
5806 int cpu
= (long)hcpu
;
5807 unsigned long flags
;
5812 case CPU_UP_PREPARE
:
5813 case CPU_UP_PREPARE_FROZEN
:
5814 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5817 kthread_bind(p
, cpu
);
5818 /* Must be high prio: stop_machine expects to yield to it. */
5819 rq
= task_rq_lock(p
, &flags
);
5820 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5821 task_rq_unlock(rq
, &flags
);
5822 cpu_rq(cpu
)->migration_thread
= p
;
5826 case CPU_ONLINE_FROZEN
:
5827 /* Strictly unnecessary, as first user will wake it. */
5828 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5830 /* Update our root-domain */
5832 spin_lock_irqsave(&rq
->lock
, flags
);
5834 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5835 cpu_set(cpu
, rq
->rd
->online
);
5837 spin_unlock_irqrestore(&rq
->lock
, flags
);
5840 #ifdef CONFIG_HOTPLUG_CPU
5841 case CPU_UP_CANCELED
:
5842 case CPU_UP_CANCELED_FROZEN
:
5843 if (!cpu_rq(cpu
)->migration_thread
)
5845 /* Unbind it from offline cpu so it can run. Fall thru. */
5846 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5847 any_online_cpu(cpu_online_map
));
5848 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5849 cpu_rq(cpu
)->migration_thread
= NULL
;
5853 case CPU_DEAD_FROZEN
:
5854 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5855 migrate_live_tasks(cpu
);
5857 kthread_stop(rq
->migration_thread
);
5858 rq
->migration_thread
= NULL
;
5859 /* Idle task back to normal (off runqueue, low prio) */
5860 spin_lock_irq(&rq
->lock
);
5861 update_rq_clock(rq
);
5862 deactivate_task(rq
, rq
->idle
, 0);
5863 rq
->idle
->static_prio
= MAX_PRIO
;
5864 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5865 rq
->idle
->sched_class
= &idle_sched_class
;
5866 migrate_dead_tasks(cpu
);
5867 spin_unlock_irq(&rq
->lock
);
5869 migrate_nr_uninterruptible(rq
);
5870 BUG_ON(rq
->nr_running
!= 0);
5873 * No need to migrate the tasks: it was best-effort if
5874 * they didn't take sched_hotcpu_mutex. Just wake up
5877 spin_lock_irq(&rq
->lock
);
5878 while (!list_empty(&rq
->migration_queue
)) {
5879 struct migration_req
*req
;
5881 req
= list_entry(rq
->migration_queue
.next
,
5882 struct migration_req
, list
);
5883 list_del_init(&req
->list
);
5884 complete(&req
->done
);
5886 spin_unlock_irq(&rq
->lock
);
5889 case CPU_DOWN_PREPARE
:
5890 /* Update our root-domain */
5892 spin_lock_irqsave(&rq
->lock
, flags
);
5894 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5895 cpu_clear(cpu
, rq
->rd
->online
);
5897 spin_unlock_irqrestore(&rq
->lock
, flags
);
5904 /* Register at highest priority so that task migration (migrate_all_tasks)
5905 * happens before everything else.
5907 static struct notifier_block __cpuinitdata migration_notifier
= {
5908 .notifier_call
= migration_call
,
5912 void __init
migration_init(void)
5914 void *cpu
= (void *)(long)smp_processor_id();
5917 /* Start one for the boot CPU: */
5918 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5919 BUG_ON(err
== NOTIFY_BAD
);
5920 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5921 register_cpu_notifier(&migration_notifier
);
5927 /* Number of possible processor ids */
5928 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5929 EXPORT_SYMBOL(nr_cpu_ids
);
5931 #ifdef CONFIG_SCHED_DEBUG
5933 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
5935 struct sched_group
*group
= sd
->groups
;
5936 cpumask_t groupmask
;
5939 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5940 cpus_clear(groupmask
);
5942 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5944 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5945 printk("does not load-balance\n");
5947 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5952 printk(KERN_CONT
"span %s\n", str
);
5954 if (!cpu_isset(cpu
, sd
->span
)) {
5955 printk(KERN_ERR
"ERROR: domain->span does not contain "
5958 if (!cpu_isset(cpu
, group
->cpumask
)) {
5959 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5963 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5967 printk(KERN_ERR
"ERROR: group is NULL\n");
5971 if (!group
->__cpu_power
) {
5972 printk(KERN_CONT
"\n");
5973 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5978 if (!cpus_weight(group
->cpumask
)) {
5979 printk(KERN_CONT
"\n");
5980 printk(KERN_ERR
"ERROR: empty group\n");
5984 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5985 printk(KERN_CONT
"\n");
5986 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5990 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5992 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5993 printk(KERN_CONT
" %s", str
);
5995 group
= group
->next
;
5996 } while (group
!= sd
->groups
);
5997 printk(KERN_CONT
"\n");
5999 if (!cpus_equal(sd
->span
, groupmask
))
6000 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6002 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
6003 printk(KERN_ERR
"ERROR: parent span is not a superset "
6004 "of domain->span\n");
6008 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6013 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6017 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6020 if (sched_domain_debug_one(sd
, cpu
, level
))
6029 # define sched_domain_debug(sd, cpu) do { } while (0)
6032 static int sd_degenerate(struct sched_domain
*sd
)
6034 if (cpus_weight(sd
->span
) == 1)
6037 /* Following flags need at least 2 groups */
6038 if (sd
->flags
& (SD_LOAD_BALANCE
|
6039 SD_BALANCE_NEWIDLE
|
6043 SD_SHARE_PKG_RESOURCES
)) {
6044 if (sd
->groups
!= sd
->groups
->next
)
6048 /* Following flags don't use groups */
6049 if (sd
->flags
& (SD_WAKE_IDLE
|
6058 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6060 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6062 if (sd_degenerate(parent
))
6065 if (!cpus_equal(sd
->span
, parent
->span
))
6068 /* Does parent contain flags not in child? */
6069 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6070 if (cflags
& SD_WAKE_AFFINE
)
6071 pflags
&= ~SD_WAKE_BALANCE
;
6072 /* Flags needing groups don't count if only 1 group in parent */
6073 if (parent
->groups
== parent
->groups
->next
) {
6074 pflags
&= ~(SD_LOAD_BALANCE
|
6075 SD_BALANCE_NEWIDLE
|
6079 SD_SHARE_PKG_RESOURCES
);
6081 if (~cflags
& pflags
)
6087 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6089 unsigned long flags
;
6090 const struct sched_class
*class;
6092 spin_lock_irqsave(&rq
->lock
, flags
);
6095 struct root_domain
*old_rd
= rq
->rd
;
6097 for (class = sched_class_highest
; class; class = class->next
) {
6098 if (class->leave_domain
)
6099 class->leave_domain(rq
);
6102 cpu_clear(rq
->cpu
, old_rd
->span
);
6103 cpu_clear(rq
->cpu
, old_rd
->online
);
6105 if (atomic_dec_and_test(&old_rd
->refcount
))
6109 atomic_inc(&rd
->refcount
);
6112 cpu_set(rq
->cpu
, rd
->span
);
6113 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6114 cpu_set(rq
->cpu
, rd
->online
);
6116 for (class = sched_class_highest
; class; class = class->next
) {
6117 if (class->join_domain
)
6118 class->join_domain(rq
);
6121 spin_unlock_irqrestore(&rq
->lock
, flags
);
6124 static void init_rootdomain(struct root_domain
*rd
)
6126 memset(rd
, 0, sizeof(*rd
));
6128 cpus_clear(rd
->span
);
6129 cpus_clear(rd
->online
);
6132 static void init_defrootdomain(void)
6134 init_rootdomain(&def_root_domain
);
6135 atomic_set(&def_root_domain
.refcount
, 1);
6138 static struct root_domain
*alloc_rootdomain(void)
6140 struct root_domain
*rd
;
6142 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6146 init_rootdomain(rd
);
6152 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6153 * hold the hotplug lock.
6156 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6158 struct rq
*rq
= cpu_rq(cpu
);
6159 struct sched_domain
*tmp
;
6161 /* Remove the sched domains which do not contribute to scheduling. */
6162 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6163 struct sched_domain
*parent
= tmp
->parent
;
6166 if (sd_parent_degenerate(tmp
, parent
)) {
6167 tmp
->parent
= parent
->parent
;
6169 parent
->parent
->child
= tmp
;
6173 if (sd
&& sd_degenerate(sd
)) {
6179 sched_domain_debug(sd
, cpu
);
6181 rq_attach_root(rq
, rd
);
6182 rcu_assign_pointer(rq
->sd
, sd
);
6185 /* cpus with isolated domains */
6186 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6188 /* Setup the mask of cpus configured for isolated domains */
6189 static int __init
isolated_cpu_setup(char *str
)
6191 int ints
[NR_CPUS
], i
;
6193 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6194 cpus_clear(cpu_isolated_map
);
6195 for (i
= 1; i
<= ints
[0]; i
++)
6196 if (ints
[i
] < NR_CPUS
)
6197 cpu_set(ints
[i
], cpu_isolated_map
);
6201 __setup("isolcpus=", isolated_cpu_setup
);
6204 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6205 * to a function which identifies what group(along with sched group) a CPU
6206 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6207 * (due to the fact that we keep track of groups covered with a cpumask_t).
6209 * init_sched_build_groups will build a circular linked list of the groups
6210 * covered by the given span, and will set each group's ->cpumask correctly,
6211 * and ->cpu_power to 0.
6214 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
6215 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6216 struct sched_group
**sg
))
6218 struct sched_group
*first
= NULL
, *last
= NULL
;
6219 cpumask_t covered
= CPU_MASK_NONE
;
6222 for_each_cpu_mask(i
, span
) {
6223 struct sched_group
*sg
;
6224 int group
= group_fn(i
, cpu_map
, &sg
);
6227 if (cpu_isset(i
, covered
))
6230 sg
->cpumask
= CPU_MASK_NONE
;
6231 sg
->__cpu_power
= 0;
6233 for_each_cpu_mask(j
, span
) {
6234 if (group_fn(j
, cpu_map
, NULL
) != group
)
6237 cpu_set(j
, covered
);
6238 cpu_set(j
, sg
->cpumask
);
6249 #define SD_NODES_PER_DOMAIN 16
6254 * find_next_best_node - find the next node to include in a sched_domain
6255 * @node: node whose sched_domain we're building
6256 * @used_nodes: nodes already in the sched_domain
6258 * Find the next node to include in a given scheduling domain. Simply
6259 * finds the closest node not already in the @used_nodes map.
6261 * Should use nodemask_t.
6263 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6265 int i
, n
, val
, min_val
, best_node
= 0;
6269 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6270 /* Start at @node */
6271 n
= (node
+ i
) % MAX_NUMNODES
;
6273 if (!nr_cpus_node(n
))
6276 /* Skip already used nodes */
6277 if (test_bit(n
, used_nodes
))
6280 /* Simple min distance search */
6281 val
= node_distance(node
, n
);
6283 if (val
< min_val
) {
6289 set_bit(best_node
, used_nodes
);
6294 * sched_domain_node_span - get a cpumask for a node's sched_domain
6295 * @node: node whose cpumask we're constructing
6296 * @size: number of nodes to include in this span
6298 * Given a node, construct a good cpumask for its sched_domain to span. It
6299 * should be one that prevents unnecessary balancing, but also spreads tasks
6302 static cpumask_t
sched_domain_node_span(int node
)
6304 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6305 cpumask_t span
, nodemask
;
6309 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6311 nodemask
= node_to_cpumask(node
);
6312 cpus_or(span
, span
, nodemask
);
6313 set_bit(node
, used_nodes
);
6315 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6316 int next_node
= find_next_best_node(node
, used_nodes
);
6318 nodemask
= node_to_cpumask(next_node
);
6319 cpus_or(span
, span
, nodemask
);
6326 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6329 * SMT sched-domains:
6331 #ifdef CONFIG_SCHED_SMT
6332 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6333 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6336 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6339 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6345 * multi-core sched-domains:
6347 #ifdef CONFIG_SCHED_MC
6348 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6349 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6352 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6354 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6357 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6358 cpus_and(mask
, mask
, *cpu_map
);
6359 group
= first_cpu(mask
);
6361 *sg
= &per_cpu(sched_group_core
, group
);
6364 #elif defined(CONFIG_SCHED_MC)
6366 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6369 *sg
= &per_cpu(sched_group_core
, cpu
);
6374 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6375 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6378 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6381 #ifdef CONFIG_SCHED_MC
6382 cpumask_t mask
= cpu_coregroup_map(cpu
);
6383 cpus_and(mask
, mask
, *cpu_map
);
6384 group
= first_cpu(mask
);
6385 #elif defined(CONFIG_SCHED_SMT)
6386 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6387 cpus_and(mask
, mask
, *cpu_map
);
6388 group
= first_cpu(mask
);
6393 *sg
= &per_cpu(sched_group_phys
, group
);
6399 * The init_sched_build_groups can't handle what we want to do with node
6400 * groups, so roll our own. Now each node has its own list of groups which
6401 * gets dynamically allocated.
6403 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6404 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6406 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6407 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6409 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6410 struct sched_group
**sg
)
6412 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6415 cpus_and(nodemask
, nodemask
, *cpu_map
);
6416 group
= first_cpu(nodemask
);
6419 *sg
= &per_cpu(sched_group_allnodes
, group
);
6423 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6425 struct sched_group
*sg
= group_head
;
6431 for_each_cpu_mask(j
, sg
->cpumask
) {
6432 struct sched_domain
*sd
;
6434 sd
= &per_cpu(phys_domains
, j
);
6435 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6437 * Only add "power" once for each
6443 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6446 } while (sg
!= group_head
);
6451 /* Free memory allocated for various sched_group structures */
6452 static void free_sched_groups(const cpumask_t
*cpu_map
)
6456 for_each_cpu_mask(cpu
, *cpu_map
) {
6457 struct sched_group
**sched_group_nodes
6458 = sched_group_nodes_bycpu
[cpu
];
6460 if (!sched_group_nodes
)
6463 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6464 cpumask_t nodemask
= node_to_cpumask(i
);
6465 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6467 cpus_and(nodemask
, nodemask
, *cpu_map
);
6468 if (cpus_empty(nodemask
))
6478 if (oldsg
!= sched_group_nodes
[i
])
6481 kfree(sched_group_nodes
);
6482 sched_group_nodes_bycpu
[cpu
] = NULL
;
6486 static void free_sched_groups(const cpumask_t
*cpu_map
)
6492 * Initialize sched groups cpu_power.
6494 * cpu_power indicates the capacity of sched group, which is used while
6495 * distributing the load between different sched groups in a sched domain.
6496 * Typically cpu_power for all the groups in a sched domain will be same unless
6497 * there are asymmetries in the topology. If there are asymmetries, group
6498 * having more cpu_power will pickup more load compared to the group having
6501 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6502 * the maximum number of tasks a group can handle in the presence of other idle
6503 * or lightly loaded groups in the same sched domain.
6505 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6507 struct sched_domain
*child
;
6508 struct sched_group
*group
;
6510 WARN_ON(!sd
|| !sd
->groups
);
6512 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6517 sd
->groups
->__cpu_power
= 0;
6520 * For perf policy, if the groups in child domain share resources
6521 * (for example cores sharing some portions of the cache hierarchy
6522 * or SMT), then set this domain groups cpu_power such that each group
6523 * can handle only one task, when there are other idle groups in the
6524 * same sched domain.
6526 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6528 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6529 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6534 * add cpu_power of each child group to this groups cpu_power
6536 group
= child
->groups
;
6538 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6539 group
= group
->next
;
6540 } while (group
!= child
->groups
);
6544 * Build sched domains for a given set of cpus and attach the sched domains
6545 * to the individual cpus
6547 static int build_sched_domains(const cpumask_t
*cpu_map
)
6550 struct root_domain
*rd
;
6552 struct sched_group
**sched_group_nodes
= NULL
;
6553 int sd_allnodes
= 0;
6556 * Allocate the per-node list of sched groups
6558 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6560 if (!sched_group_nodes
) {
6561 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6564 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6567 rd
= alloc_rootdomain();
6569 printk(KERN_WARNING
"Cannot alloc root domain\n");
6574 * Set up domains for cpus specified by the cpu_map.
6576 for_each_cpu_mask(i
, *cpu_map
) {
6577 struct sched_domain
*sd
= NULL
, *p
;
6578 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6580 cpus_and(nodemask
, nodemask
, *cpu_map
);
6583 if (cpus_weight(*cpu_map
) >
6584 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6585 sd
= &per_cpu(allnodes_domains
, i
);
6586 *sd
= SD_ALLNODES_INIT
;
6587 sd
->span
= *cpu_map
;
6588 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6594 sd
= &per_cpu(node_domains
, i
);
6596 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6600 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6604 sd
= &per_cpu(phys_domains
, i
);
6606 sd
->span
= nodemask
;
6610 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6612 #ifdef CONFIG_SCHED_MC
6614 sd
= &per_cpu(core_domains
, i
);
6616 sd
->span
= cpu_coregroup_map(i
);
6617 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6620 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6623 #ifdef CONFIG_SCHED_SMT
6625 sd
= &per_cpu(cpu_domains
, i
);
6626 *sd
= SD_SIBLING_INIT
;
6627 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6628 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6631 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6635 #ifdef CONFIG_SCHED_SMT
6636 /* Set up CPU (sibling) groups */
6637 for_each_cpu_mask(i
, *cpu_map
) {
6638 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6639 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6640 if (i
!= first_cpu(this_sibling_map
))
6643 init_sched_build_groups(this_sibling_map
, cpu_map
,
6648 #ifdef CONFIG_SCHED_MC
6649 /* Set up multi-core groups */
6650 for_each_cpu_mask(i
, *cpu_map
) {
6651 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6652 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6653 if (i
!= first_cpu(this_core_map
))
6655 init_sched_build_groups(this_core_map
, cpu_map
,
6656 &cpu_to_core_group
);
6660 /* Set up physical groups */
6661 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6662 cpumask_t nodemask
= node_to_cpumask(i
);
6664 cpus_and(nodemask
, nodemask
, *cpu_map
);
6665 if (cpus_empty(nodemask
))
6668 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6672 /* Set up node groups */
6674 init_sched_build_groups(*cpu_map
, cpu_map
,
6675 &cpu_to_allnodes_group
);
6677 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6678 /* Set up node groups */
6679 struct sched_group
*sg
, *prev
;
6680 cpumask_t nodemask
= node_to_cpumask(i
);
6681 cpumask_t domainspan
;
6682 cpumask_t covered
= CPU_MASK_NONE
;
6685 cpus_and(nodemask
, nodemask
, *cpu_map
);
6686 if (cpus_empty(nodemask
)) {
6687 sched_group_nodes
[i
] = NULL
;
6691 domainspan
= sched_domain_node_span(i
);
6692 cpus_and(domainspan
, domainspan
, *cpu_map
);
6694 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6696 printk(KERN_WARNING
"Can not alloc domain group for "
6700 sched_group_nodes
[i
] = sg
;
6701 for_each_cpu_mask(j
, nodemask
) {
6702 struct sched_domain
*sd
;
6704 sd
= &per_cpu(node_domains
, j
);
6707 sg
->__cpu_power
= 0;
6708 sg
->cpumask
= nodemask
;
6710 cpus_or(covered
, covered
, nodemask
);
6713 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6714 cpumask_t tmp
, notcovered
;
6715 int n
= (i
+ j
) % MAX_NUMNODES
;
6717 cpus_complement(notcovered
, covered
);
6718 cpus_and(tmp
, notcovered
, *cpu_map
);
6719 cpus_and(tmp
, tmp
, domainspan
);
6720 if (cpus_empty(tmp
))
6723 nodemask
= node_to_cpumask(n
);
6724 cpus_and(tmp
, tmp
, nodemask
);
6725 if (cpus_empty(tmp
))
6728 sg
= kmalloc_node(sizeof(struct sched_group
),
6732 "Can not alloc domain group for node %d\n", j
);
6735 sg
->__cpu_power
= 0;
6737 sg
->next
= prev
->next
;
6738 cpus_or(covered
, covered
, tmp
);
6745 /* Calculate CPU power for physical packages and nodes */
6746 #ifdef CONFIG_SCHED_SMT
6747 for_each_cpu_mask(i
, *cpu_map
) {
6748 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6750 init_sched_groups_power(i
, sd
);
6753 #ifdef CONFIG_SCHED_MC
6754 for_each_cpu_mask(i
, *cpu_map
) {
6755 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6757 init_sched_groups_power(i
, sd
);
6761 for_each_cpu_mask(i
, *cpu_map
) {
6762 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6764 init_sched_groups_power(i
, sd
);
6768 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6769 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6772 struct sched_group
*sg
;
6774 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6775 init_numa_sched_groups_power(sg
);
6779 /* Attach the domains */
6780 for_each_cpu_mask(i
, *cpu_map
) {
6781 struct sched_domain
*sd
;
6782 #ifdef CONFIG_SCHED_SMT
6783 sd
= &per_cpu(cpu_domains
, i
);
6784 #elif defined(CONFIG_SCHED_MC)
6785 sd
= &per_cpu(core_domains
, i
);
6787 sd
= &per_cpu(phys_domains
, i
);
6789 cpu_attach_domain(sd
, rd
, i
);
6796 free_sched_groups(cpu_map
);
6801 static cpumask_t
*doms_cur
; /* current sched domains */
6802 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6805 * Special case: If a kmalloc of a doms_cur partition (array of
6806 * cpumask_t) fails, then fallback to a single sched domain,
6807 * as determined by the single cpumask_t fallback_doms.
6809 static cpumask_t fallback_doms
;
6812 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6813 * For now this just excludes isolated cpus, but could be used to
6814 * exclude other special cases in the future.
6816 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6821 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6823 doms_cur
= &fallback_doms
;
6824 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6825 err
= build_sched_domains(doms_cur
);
6826 register_sched_domain_sysctl();
6831 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6833 free_sched_groups(cpu_map
);
6837 * Detach sched domains from a group of cpus specified in cpu_map
6838 * These cpus will now be attached to the NULL domain
6840 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6844 unregister_sched_domain_sysctl();
6846 for_each_cpu_mask(i
, *cpu_map
)
6847 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6848 synchronize_sched();
6849 arch_destroy_sched_domains(cpu_map
);
6853 * Partition sched domains as specified by the 'ndoms_new'
6854 * cpumasks in the array doms_new[] of cpumasks. This compares
6855 * doms_new[] to the current sched domain partitioning, doms_cur[].
6856 * It destroys each deleted domain and builds each new domain.
6858 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6859 * The masks don't intersect (don't overlap.) We should setup one
6860 * sched domain for each mask. CPUs not in any of the cpumasks will
6861 * not be load balanced. If the same cpumask appears both in the
6862 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6865 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6866 * ownership of it and will kfree it when done with it. If the caller
6867 * failed the kmalloc call, then it can pass in doms_new == NULL,
6868 * and partition_sched_domains() will fallback to the single partition
6871 * Call with hotplug lock held
6873 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6879 /* always unregister in case we don't destroy any domains */
6880 unregister_sched_domain_sysctl();
6882 if (doms_new
== NULL
) {
6884 doms_new
= &fallback_doms
;
6885 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6888 /* Destroy deleted domains */
6889 for (i
= 0; i
< ndoms_cur
; i
++) {
6890 for (j
= 0; j
< ndoms_new
; j
++) {
6891 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
6894 /* no match - a current sched domain not in new doms_new[] */
6895 detach_destroy_domains(doms_cur
+ i
);
6900 /* Build new domains */
6901 for (i
= 0; i
< ndoms_new
; i
++) {
6902 for (j
= 0; j
< ndoms_cur
; j
++) {
6903 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
6906 /* no match - add a new doms_new */
6907 build_sched_domains(doms_new
+ i
);
6912 /* Remember the new sched domains */
6913 if (doms_cur
!= &fallback_doms
)
6915 doms_cur
= doms_new
;
6916 ndoms_cur
= ndoms_new
;
6918 register_sched_domain_sysctl();
6923 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6924 static int arch_reinit_sched_domains(void)
6929 detach_destroy_domains(&cpu_online_map
);
6930 err
= arch_init_sched_domains(&cpu_online_map
);
6936 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6940 if (buf
[0] != '0' && buf
[0] != '1')
6944 sched_smt_power_savings
= (buf
[0] == '1');
6946 sched_mc_power_savings
= (buf
[0] == '1');
6948 ret
= arch_reinit_sched_domains();
6950 return ret
? ret
: count
;
6953 #ifdef CONFIG_SCHED_MC
6954 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6956 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6958 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6959 const char *buf
, size_t count
)
6961 return sched_power_savings_store(buf
, count
, 0);
6963 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6964 sched_mc_power_savings_store
);
6967 #ifdef CONFIG_SCHED_SMT
6968 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6970 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6972 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6973 const char *buf
, size_t count
)
6975 return sched_power_savings_store(buf
, count
, 1);
6977 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6978 sched_smt_power_savings_store
);
6981 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6985 #ifdef CONFIG_SCHED_SMT
6987 err
= sysfs_create_file(&cls
->kset
.kobj
,
6988 &attr_sched_smt_power_savings
.attr
);
6990 #ifdef CONFIG_SCHED_MC
6991 if (!err
&& mc_capable())
6992 err
= sysfs_create_file(&cls
->kset
.kobj
,
6993 &attr_sched_mc_power_savings
.attr
);
7000 * Force a reinitialization of the sched domains hierarchy. The domains
7001 * and groups cannot be updated in place without racing with the balancing
7002 * code, so we temporarily attach all running cpus to the NULL domain
7003 * which will prevent rebalancing while the sched domains are recalculated.
7005 static int update_sched_domains(struct notifier_block
*nfb
,
7006 unsigned long action
, void *hcpu
)
7009 case CPU_UP_PREPARE
:
7010 case CPU_UP_PREPARE_FROZEN
:
7011 case CPU_DOWN_PREPARE
:
7012 case CPU_DOWN_PREPARE_FROZEN
:
7013 detach_destroy_domains(&cpu_online_map
);
7016 case CPU_UP_CANCELED
:
7017 case CPU_UP_CANCELED_FROZEN
:
7018 case CPU_DOWN_FAILED
:
7019 case CPU_DOWN_FAILED_FROZEN
:
7021 case CPU_ONLINE_FROZEN
:
7023 case CPU_DEAD_FROZEN
:
7025 * Fall through and re-initialise the domains.
7032 /* The hotplug lock is already held by cpu_up/cpu_down */
7033 arch_init_sched_domains(&cpu_online_map
);
7038 void __init
sched_init_smp(void)
7040 cpumask_t non_isolated_cpus
;
7043 arch_init_sched_domains(&cpu_online_map
);
7044 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7045 if (cpus_empty(non_isolated_cpus
))
7046 cpu_set(smp_processor_id(), non_isolated_cpus
);
7048 /* XXX: Theoretical race here - CPU may be hotplugged now */
7049 hotcpu_notifier(update_sched_domains
, 0);
7051 /* Move init over to a non-isolated CPU */
7052 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
7054 sched_init_granularity();
7056 #ifdef CONFIG_FAIR_GROUP_SCHED
7057 if (nr_cpu_ids
== 1)
7060 lb_monitor_task
= kthread_create(load_balance_monitor
, NULL
,
7062 if (!IS_ERR(lb_monitor_task
)) {
7063 lb_monitor_task
->flags
|= PF_NOFREEZE
;
7064 wake_up_process(lb_monitor_task
);
7066 printk(KERN_ERR
"Could not create load balance monitor thread"
7067 "(error = %ld) \n", PTR_ERR(lb_monitor_task
));
7072 void __init
sched_init_smp(void)
7074 sched_init_granularity();
7076 #endif /* CONFIG_SMP */
7078 int in_sched_functions(unsigned long addr
)
7080 return in_lock_functions(addr
) ||
7081 (addr
>= (unsigned long)__sched_text_start
7082 && addr
< (unsigned long)__sched_text_end
);
7085 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7087 cfs_rq
->tasks_timeline
= RB_ROOT
;
7088 #ifdef CONFIG_FAIR_GROUP_SCHED
7091 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7094 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7096 struct rt_prio_array
*array
;
7099 array
= &rt_rq
->active
;
7100 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7101 INIT_LIST_HEAD(array
->queue
+ i
);
7102 __clear_bit(i
, array
->bitmap
);
7104 /* delimiter for bitsearch: */
7105 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7107 #if defined CONFIG_SMP || defined CONFIG_FAIR_GROUP_SCHED
7108 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7111 rt_rq
->rt_nr_migratory
= 0;
7112 rt_rq
->overloaded
= 0;
7116 rt_rq
->rt_throttled
= 0;
7118 #ifdef CONFIG_FAIR_GROUP_SCHED
7119 rt_rq
->rt_nr_boosted
= 0;
7124 #ifdef CONFIG_FAIR_GROUP_SCHED
7125 static void init_tg_cfs_entry(struct rq
*rq
, struct task_group
*tg
,
7126 struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
7129 tg
->cfs_rq
[cpu
] = cfs_rq
;
7130 init_cfs_rq(cfs_rq
, rq
);
7133 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7136 se
->cfs_rq
= &rq
->cfs
;
7138 se
->load
.weight
= tg
->shares
;
7139 se
->load
.inv_weight
= div64_64(1ULL<<32, se
->load
.weight
);
7143 static void init_tg_rt_entry(struct rq
*rq
, struct task_group
*tg
,
7144 struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
,
7147 tg
->rt_rq
[cpu
] = rt_rq
;
7148 init_rt_rq(rt_rq
, rq
);
7150 rt_rq
->rt_se
= rt_se
;
7152 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7154 tg
->rt_se
[cpu
] = rt_se
;
7155 rt_se
->rt_rq
= &rq
->rt
;
7156 rt_se
->my_q
= rt_rq
;
7157 rt_se
->parent
= NULL
;
7158 INIT_LIST_HEAD(&rt_se
->run_list
);
7162 void __init
sched_init(void)
7164 int highest_cpu
= 0;
7168 init_defrootdomain();
7171 #ifdef CONFIG_FAIR_GROUP_SCHED
7172 list_add(&init_task_group
.list
, &task_groups
);
7175 for_each_possible_cpu(i
) {
7179 spin_lock_init(&rq
->lock
);
7180 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7183 init_cfs_rq(&rq
->cfs
, rq
);
7184 init_rt_rq(&rq
->rt
, rq
);
7185 #ifdef CONFIG_FAIR_GROUP_SCHED
7186 init_task_group
.shares
= init_task_group_load
;
7187 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7188 init_tg_cfs_entry(rq
, &init_task_group
,
7189 &per_cpu(init_cfs_rq
, i
),
7190 &per_cpu(init_sched_entity
, i
), i
, 1);
7192 init_task_group
.rt_runtime
=
7193 sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
7194 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7195 init_tg_rt_entry(rq
, &init_task_group
,
7196 &per_cpu(init_rt_rq
, i
),
7197 &per_cpu(init_sched_rt_entity
, i
), i
, 1);
7199 rq
->rt_period_expire
= 0;
7200 rq
->rt_throttled
= 0;
7202 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7203 rq
->cpu_load
[j
] = 0;
7207 rq
->active_balance
= 0;
7208 rq
->next_balance
= jiffies
;
7211 rq
->migration_thread
= NULL
;
7212 INIT_LIST_HEAD(&rq
->migration_queue
);
7213 rq_attach_root(rq
, &def_root_domain
);
7216 atomic_set(&rq
->nr_iowait
, 0);
7220 set_load_weight(&init_task
);
7222 #ifdef CONFIG_PREEMPT_NOTIFIERS
7223 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7227 nr_cpu_ids
= highest_cpu
+ 1;
7228 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7231 #ifdef CONFIG_RT_MUTEXES
7232 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7236 * The boot idle thread does lazy MMU switching as well:
7238 atomic_inc(&init_mm
.mm_count
);
7239 enter_lazy_tlb(&init_mm
, current
);
7242 * Make us the idle thread. Technically, schedule() should not be
7243 * called from this thread, however somewhere below it might be,
7244 * but because we are the idle thread, we just pick up running again
7245 * when this runqueue becomes "idle".
7247 init_idle(current
, smp_processor_id());
7249 * During early bootup we pretend to be a normal task:
7251 current
->sched_class
= &fair_sched_class
;
7254 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7255 void __might_sleep(char *file
, int line
)
7258 static unsigned long prev_jiffy
; /* ratelimiting */
7260 if ((in_atomic() || irqs_disabled()) &&
7261 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
7262 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7264 prev_jiffy
= jiffies
;
7265 printk(KERN_ERR
"BUG: sleeping function called from invalid"
7266 " context at %s:%d\n", file
, line
);
7267 printk("in_atomic():%d, irqs_disabled():%d\n",
7268 in_atomic(), irqs_disabled());
7269 debug_show_held_locks(current
);
7270 if (irqs_disabled())
7271 print_irqtrace_events(current
);
7276 EXPORT_SYMBOL(__might_sleep
);
7279 #ifdef CONFIG_MAGIC_SYSRQ
7280 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7283 update_rq_clock(rq
);
7284 on_rq
= p
->se
.on_rq
;
7286 deactivate_task(rq
, p
, 0);
7287 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7289 activate_task(rq
, p
, 0);
7290 resched_task(rq
->curr
);
7294 void normalize_rt_tasks(void)
7296 struct task_struct
*g
, *p
;
7297 unsigned long flags
;
7300 read_lock_irqsave(&tasklist_lock
, flags
);
7301 do_each_thread(g
, p
) {
7303 * Only normalize user tasks:
7308 p
->se
.exec_start
= 0;
7309 #ifdef CONFIG_SCHEDSTATS
7310 p
->se
.wait_start
= 0;
7311 p
->se
.sleep_start
= 0;
7312 p
->se
.block_start
= 0;
7314 task_rq(p
)->clock
= 0;
7318 * Renice negative nice level userspace
7321 if (TASK_NICE(p
) < 0 && p
->mm
)
7322 set_user_nice(p
, 0);
7326 spin_lock(&p
->pi_lock
);
7327 rq
= __task_rq_lock(p
);
7329 normalize_task(rq
, p
);
7331 __task_rq_unlock(rq
);
7332 spin_unlock(&p
->pi_lock
);
7333 } while_each_thread(g
, p
);
7335 read_unlock_irqrestore(&tasklist_lock
, flags
);
7338 #endif /* CONFIG_MAGIC_SYSRQ */
7342 * These functions are only useful for the IA64 MCA handling.
7344 * They can only be called when the whole system has been
7345 * stopped - every CPU needs to be quiescent, and no scheduling
7346 * activity can take place. Using them for anything else would
7347 * be a serious bug, and as a result, they aren't even visible
7348 * under any other configuration.
7352 * curr_task - return the current task for a given cpu.
7353 * @cpu: the processor in question.
7355 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7357 struct task_struct
*curr_task(int cpu
)
7359 return cpu_curr(cpu
);
7363 * set_curr_task - set the current task for a given cpu.
7364 * @cpu: the processor in question.
7365 * @p: the task pointer to set.
7367 * Description: This function must only be used when non-maskable interrupts
7368 * are serviced on a separate stack. It allows the architecture to switch the
7369 * notion of the current task on a cpu in a non-blocking manner. This function
7370 * must be called with all CPU's synchronized, and interrupts disabled, the
7371 * and caller must save the original value of the current task (see
7372 * curr_task() above) and restore that value before reenabling interrupts and
7373 * re-starting the system.
7375 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7377 void set_curr_task(int cpu
, struct task_struct
*p
)
7384 #ifdef CONFIG_FAIR_GROUP_SCHED
7388 * distribute shares of all task groups among their schedulable entities,
7389 * to reflect load distribution across cpus.
7391 static int rebalance_shares(struct sched_domain
*sd
, int this_cpu
)
7393 struct cfs_rq
*cfs_rq
;
7394 struct rq
*rq
= cpu_rq(this_cpu
);
7395 cpumask_t sdspan
= sd
->span
;
7398 /* Walk thr' all the task groups that we have */
7399 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
7401 unsigned long total_load
= 0, total_shares
;
7402 struct task_group
*tg
= cfs_rq
->tg
;
7404 /* Gather total task load of this group across cpus */
7405 for_each_cpu_mask(i
, sdspan
)
7406 total_load
+= tg
->cfs_rq
[i
]->load
.weight
;
7408 /* Nothing to do if this group has no load */
7413 * tg->shares represents the number of cpu shares the task group
7414 * is eligible to hold on a single cpu. On N cpus, it is
7415 * eligible to hold (N * tg->shares) number of cpu shares.
7417 total_shares
= tg
->shares
* cpus_weight(sdspan
);
7420 * redistribute total_shares across cpus as per the task load
7423 for_each_cpu_mask(i
, sdspan
) {
7424 unsigned long local_load
, local_shares
;
7426 local_load
= tg
->cfs_rq
[i
]->load
.weight
;
7427 local_shares
= (local_load
* total_shares
) / total_load
;
7429 local_shares
= MIN_GROUP_SHARES
;
7430 if (local_shares
== tg
->se
[i
]->load
.weight
)
7433 spin_lock_irq(&cpu_rq(i
)->lock
);
7434 set_se_shares(tg
->se
[i
], local_shares
);
7435 spin_unlock_irq(&cpu_rq(i
)->lock
);
7444 * How frequently should we rebalance_shares() across cpus?
7446 * The more frequently we rebalance shares, the more accurate is the fairness
7447 * of cpu bandwidth distribution between task groups. However higher frequency
7448 * also implies increased scheduling overhead.
7450 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7451 * consecutive calls to rebalance_shares() in the same sched domain.
7453 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7454 * consecutive calls to rebalance_shares() in the same sched domain.
7456 * These settings allows for the appropriate trade-off between accuracy of
7457 * fairness and the associated overhead.
7461 /* default: 8ms, units: milliseconds */
7462 const_debug
unsigned int sysctl_sched_min_bal_int_shares
= 8;
7464 /* default: 128ms, units: milliseconds */
7465 const_debug
unsigned int sysctl_sched_max_bal_int_shares
= 128;
7467 /* kernel thread that runs rebalance_shares() periodically */
7468 static int load_balance_monitor(void *unused
)
7470 unsigned int timeout
= sysctl_sched_min_bal_int_shares
;
7471 struct sched_param schedparm
;
7475 * We don't want this thread's execution to be limited by the shares
7476 * assigned to default group (init_task_group). Hence make it run
7477 * as a SCHED_RR RT task at the lowest priority.
7479 schedparm
.sched_priority
= 1;
7480 ret
= sched_setscheduler(current
, SCHED_RR
, &schedparm
);
7482 printk(KERN_ERR
"Couldn't set SCHED_RR policy for load balance"
7483 " monitor thread (error = %d) \n", ret
);
7485 while (!kthread_should_stop()) {
7486 int i
, cpu
, balanced
= 1;
7488 /* Prevent cpus going down or coming up */
7490 /* lockout changes to doms_cur[] array */
7493 * Enter a rcu read-side critical section to safely walk rq->sd
7494 * chain on various cpus and to walk task group list
7495 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7499 for (i
= 0; i
< ndoms_cur
; i
++) {
7500 cpumask_t cpumap
= doms_cur
[i
];
7501 struct sched_domain
*sd
= NULL
, *sd_prev
= NULL
;
7503 cpu
= first_cpu(cpumap
);
7505 /* Find the highest domain at which to balance shares */
7506 for_each_domain(cpu
, sd
) {
7507 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7513 /* sd == NULL? No load balance reqd in this domain */
7517 balanced
&= rebalance_shares(sd
, cpu
);
7526 timeout
= sysctl_sched_min_bal_int_shares
;
7527 else if (timeout
< sysctl_sched_max_bal_int_shares
)
7530 msleep_interruptible(timeout
);
7535 #endif /* CONFIG_SMP */
7537 static void free_sched_group(struct task_group
*tg
)
7541 for_each_possible_cpu(i
) {
7543 kfree(tg
->cfs_rq
[i
]);
7547 kfree(tg
->rt_rq
[i
]);
7549 kfree(tg
->rt_se
[i
]);
7559 /* allocate runqueue etc for a new task group */
7560 struct task_group
*sched_create_group(void)
7562 struct task_group
*tg
;
7563 struct cfs_rq
*cfs_rq
;
7564 struct sched_entity
*se
;
7565 struct rt_rq
*rt_rq
;
7566 struct sched_rt_entity
*rt_se
;
7568 unsigned long flags
;
7571 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7573 return ERR_PTR(-ENOMEM
);
7575 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
7578 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
7581 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * NR_CPUS
, GFP_KERNEL
);
7584 tg
->rt_se
= kzalloc(sizeof(rt_se
) * NR_CPUS
, GFP_KERNEL
);
7588 tg
->shares
= NICE_0_LOAD
;
7591 for_each_possible_cpu(i
) {
7594 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
7595 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7599 se
= kmalloc_node(sizeof(struct sched_entity
),
7600 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7604 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
7605 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7609 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
7610 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
7614 init_tg_cfs_entry(rq
, tg
, cfs_rq
, se
, i
, 0);
7615 init_tg_rt_entry(rq
, tg
, rt_rq
, rt_se
, i
, 0);
7618 spin_lock_irqsave(&task_group_lock
, flags
);
7619 for_each_possible_cpu(i
) {
7621 cfs_rq
= tg
->cfs_rq
[i
];
7622 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7623 rt_rq
= tg
->rt_rq
[i
];
7624 list_add_rcu(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7626 list_add_rcu(&tg
->list
, &task_groups
);
7627 spin_unlock_irqrestore(&task_group_lock
, flags
);
7632 free_sched_group(tg
);
7633 return ERR_PTR(-ENOMEM
);
7636 /* rcu callback to free various structures associated with a task group */
7637 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7639 /* now it should be safe to free those cfs_rqs */
7640 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7643 /* Destroy runqueue etc associated with a task group */
7644 void sched_destroy_group(struct task_group
*tg
)
7646 struct cfs_rq
*cfs_rq
= NULL
;
7647 struct rt_rq
*rt_rq
= NULL
;
7648 unsigned long flags
;
7651 spin_lock_irqsave(&task_group_lock
, flags
);
7652 for_each_possible_cpu(i
) {
7653 cfs_rq
= tg
->cfs_rq
[i
];
7654 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7655 rt_rq
= tg
->rt_rq
[i
];
7656 list_del_rcu(&rt_rq
->leaf_rt_rq_list
);
7658 list_del_rcu(&tg
->list
);
7659 spin_unlock_irqrestore(&task_group_lock
, flags
);
7663 /* wait for possible concurrent references to cfs_rqs complete */
7664 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7667 /* change task's runqueue when it moves between groups.
7668 * The caller of this function should have put the task in its new group
7669 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7670 * reflect its new group.
7672 void sched_move_task(struct task_struct
*tsk
)
7675 unsigned long flags
;
7678 rq
= task_rq_lock(tsk
, &flags
);
7680 update_rq_clock(rq
);
7682 running
= task_current(rq
, tsk
);
7683 on_rq
= tsk
->se
.on_rq
;
7686 dequeue_task(rq
, tsk
, 0);
7687 if (unlikely(running
))
7688 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7691 set_task_rq(tsk
, task_cpu(tsk
));
7694 if (unlikely(running
))
7695 tsk
->sched_class
->set_curr_task(rq
);
7696 enqueue_task(rq
, tsk
, 0);
7699 task_rq_unlock(rq
, &flags
);
7702 /* rq->lock to be locked by caller */
7703 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7705 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7706 struct rq
*rq
= cfs_rq
->rq
;
7710 shares
= MIN_GROUP_SHARES
;
7714 dequeue_entity(cfs_rq
, se
, 0);
7715 dec_cpu_load(rq
, se
->load
.weight
);
7718 se
->load
.weight
= shares
;
7719 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7722 enqueue_entity(cfs_rq
, se
, 0);
7723 inc_cpu_load(rq
, se
->load
.weight
);
7727 static DEFINE_MUTEX(shares_mutex
);
7729 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7732 struct cfs_rq
*cfs_rq
;
7734 unsigned long flags
;
7736 mutex_lock(&shares_mutex
);
7737 if (tg
->shares
== shares
)
7740 if (shares
< MIN_GROUP_SHARES
)
7741 shares
= MIN_GROUP_SHARES
;
7744 * Prevent any load balance activity (rebalance_shares,
7745 * load_balance_fair) from referring to this group first,
7746 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7748 spin_lock_irqsave(&task_group_lock
, flags
);
7749 for_each_possible_cpu(i
) {
7750 cfs_rq
= tg
->cfs_rq
[i
];
7751 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7753 spin_unlock_irqrestore(&task_group_lock
, flags
);
7755 /* wait for any ongoing reference to this group to finish */
7756 synchronize_sched();
7759 * Now we are free to modify the group's share on each cpu
7760 * w/o tripping rebalance_share or load_balance_fair.
7762 tg
->shares
= shares
;
7763 for_each_possible_cpu(i
) {
7764 spin_lock_irq(&cpu_rq(i
)->lock
);
7765 set_se_shares(tg
->se
[i
], shares
);
7766 spin_unlock_irq(&cpu_rq(i
)->lock
);
7770 * Enable load balance activity on this group, by inserting it back on
7771 * each cpu's rq->leaf_cfs_rq_list.
7773 spin_lock_irqsave(&task_group_lock
, flags
);
7774 for_each_possible_cpu(i
) {
7776 cfs_rq
= tg
->cfs_rq
[i
];
7777 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7779 spin_unlock_irqrestore(&task_group_lock
, flags
);
7781 mutex_unlock(&shares_mutex
);
7785 unsigned long sched_group_shares(struct task_group
*tg
)
7791 * Ensure that the real time constraints are schedulable.
7793 static DEFINE_MUTEX(rt_constraints_mutex
);
7795 static unsigned long to_ratio(u64 period
, u64 runtime
)
7797 if (runtime
== RUNTIME_INF
)
7800 runtime
*= (1ULL << 16);
7801 div64_64(runtime
, period
);
7805 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7807 struct task_group
*tgi
;
7808 unsigned long total
= 0;
7809 unsigned long global_ratio
=
7810 to_ratio(sysctl_sched_rt_period
,
7811 sysctl_sched_rt_runtime
< 0 ?
7812 RUNTIME_INF
: sysctl_sched_rt_runtime
);
7815 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
7819 total
+= to_ratio(period
, tgi
->rt_runtime
);
7823 return total
+ to_ratio(period
, runtime
) < global_ratio
;
7826 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7828 u64 rt_runtime
, rt_period
;
7831 rt_period
= sysctl_sched_rt_period
* NSEC_PER_USEC
;
7832 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7833 if (rt_runtime_us
== -1)
7834 rt_runtime
= rt_period
;
7836 mutex_lock(&rt_constraints_mutex
);
7837 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
7841 if (rt_runtime_us
== -1)
7842 rt_runtime
= RUNTIME_INF
;
7843 tg
->rt_runtime
= rt_runtime
;
7845 mutex_unlock(&rt_constraints_mutex
);
7850 long sched_group_rt_runtime(struct task_group
*tg
)
7854 if (tg
->rt_runtime
== RUNTIME_INF
)
7857 rt_runtime_us
= tg
->rt_runtime
;
7858 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7859 return rt_runtime_us
;
7861 #endif /* CONFIG_FAIR_GROUP_SCHED */
7863 #ifdef CONFIG_FAIR_CGROUP_SCHED
7865 /* return corresponding task_group object of a cgroup */
7866 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7868 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7869 struct task_group
, css
);
7872 static struct cgroup_subsys_state
*
7873 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7875 struct task_group
*tg
;
7877 if (!cgrp
->parent
) {
7878 /* This is early initialization for the top cgroup */
7879 init_task_group
.css
.cgroup
= cgrp
;
7880 return &init_task_group
.css
;
7883 /* we support only 1-level deep hierarchical scheduler atm */
7884 if (cgrp
->parent
->parent
)
7885 return ERR_PTR(-EINVAL
);
7887 tg
= sched_create_group();
7889 return ERR_PTR(-ENOMEM
);
7891 /* Bind the cgroup to task_group object we just created */
7892 tg
->css
.cgroup
= cgrp
;
7898 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7900 struct task_group
*tg
= cgroup_tg(cgrp
);
7902 sched_destroy_group(tg
);
7906 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7907 struct task_struct
*tsk
)
7909 /* We don't support RT-tasks being in separate groups */
7910 if (tsk
->sched_class
!= &fair_sched_class
)
7917 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7918 struct cgroup
*old_cont
, struct task_struct
*tsk
)
7920 sched_move_task(tsk
);
7923 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7926 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7929 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7931 struct task_group
*tg
= cgroup_tg(cgrp
);
7933 return (u64
) tg
->shares
;
7936 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7938 const char __user
*userbuf
,
7939 size_t nbytes
, loff_t
*unused_ppos
)
7948 if (nbytes
>= sizeof(buffer
))
7950 if (copy_from_user(buffer
, userbuf
, nbytes
))
7953 buffer
[nbytes
] = 0; /* nul-terminate */
7955 /* strip newline if necessary */
7956 if (nbytes
&& (buffer
[nbytes
-1] == '\n'))
7957 buffer
[nbytes
-1] = 0;
7958 val
= simple_strtoll(buffer
, &end
, 0);
7962 /* Pass to subsystem */
7963 retval
= sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7969 static ssize_t
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
,
7971 char __user
*buf
, size_t nbytes
,
7975 long val
= sched_group_rt_runtime(cgroup_tg(cgrp
));
7976 int len
= sprintf(tmp
, "%ld\n", val
);
7978 return simple_read_from_buffer(buf
, nbytes
, ppos
, tmp
, len
);
7981 static struct cftype cpu_files
[] = {
7984 .read_uint
= cpu_shares_read_uint
,
7985 .write_uint
= cpu_shares_write_uint
,
7988 .name
= "rt_runtime_us",
7989 .read
= cpu_rt_runtime_read
,
7990 .write
= cpu_rt_runtime_write
,
7994 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7996 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7999 struct cgroup_subsys cpu_cgroup_subsys
= {
8001 .create
= cpu_cgroup_create
,
8002 .destroy
= cpu_cgroup_destroy
,
8003 .can_attach
= cpu_cgroup_can_attach
,
8004 .attach
= cpu_cgroup_attach
,
8005 .populate
= cpu_cgroup_populate
,
8006 .subsys_id
= cpu_cgroup_subsys_id
,
8010 #endif /* CONFIG_FAIR_CGROUP_SCHED */
8012 #ifdef CONFIG_CGROUP_CPUACCT
8015 * CPU accounting code for task groups.
8017 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8018 * (balbir@in.ibm.com).
8021 /* track cpu usage of a group of tasks */
8023 struct cgroup_subsys_state css
;
8024 /* cpuusage holds pointer to a u64-type object on every cpu */
8028 struct cgroup_subsys cpuacct_subsys
;
8030 /* return cpu accounting group corresponding to this container */
8031 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
8033 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
8034 struct cpuacct
, css
);
8037 /* return cpu accounting group to which this task belongs */
8038 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8040 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8041 struct cpuacct
, css
);
8044 /* create a new cpu accounting group */
8045 static struct cgroup_subsys_state
*cpuacct_create(
8046 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8048 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8051 return ERR_PTR(-ENOMEM
);
8053 ca
->cpuusage
= alloc_percpu(u64
);
8054 if (!ca
->cpuusage
) {
8056 return ERR_PTR(-ENOMEM
);
8062 /* destroy an existing cpu accounting group */
8064 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8066 struct cpuacct
*ca
= cgroup_ca(cont
);
8068 free_percpu(ca
->cpuusage
);
8072 /* return total cpu usage (in nanoseconds) of a group */
8073 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
8075 struct cpuacct
*ca
= cgroup_ca(cont
);
8076 u64 totalcpuusage
= 0;
8079 for_each_possible_cpu(i
) {
8080 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8083 * Take rq->lock to make 64-bit addition safe on 32-bit
8086 spin_lock_irq(&cpu_rq(i
)->lock
);
8087 totalcpuusage
+= *cpuusage
;
8088 spin_unlock_irq(&cpu_rq(i
)->lock
);
8091 return totalcpuusage
;
8094 static struct cftype files
[] = {
8097 .read_uint
= cpuusage_read
,
8101 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8103 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
8107 * charge this task's execution time to its accounting group.
8109 * called with rq->lock held.
8111 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8115 if (!cpuacct_subsys
.active
)
8120 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
8122 *cpuusage
+= cputime
;
8126 struct cgroup_subsys cpuacct_subsys
= {
8128 .create
= cpuacct_create
,
8129 .destroy
= cpuacct_destroy
,
8130 .populate
= cpuacct_populate
,
8131 .subsys_id
= cpuacct_subsys_id
,
8133 #endif /* CONFIG_CGROUP_CPUACCT */