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 /* task group related information */
166 #ifdef CONFIG_FAIR_CGROUP_SCHED
167 struct cgroup_subsys_state css
;
169 /* schedulable entities of this group on each cpu */
170 struct sched_entity
**se
;
171 /* runqueue "owned" by this group on each cpu */
172 struct cfs_rq
**cfs_rq
;
175 * shares assigned to a task group governs how much of cpu bandwidth
176 * is allocated to the group. The more shares a group has, the more is
177 * the cpu bandwidth allocated to it.
179 * For ex, lets say that there are three task groups, A, B and C which
180 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
181 * cpu bandwidth allocated by the scheduler to task groups A, B and C
184 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
185 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
186 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
188 * The weight assigned to a task group's schedulable entities on every
189 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
190 * group's shares. For ex: lets say that task group A has been
191 * assigned shares of 1000 and there are two CPUs in a system. Then,
193 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
195 * Note: It's not necessary that each of a task's group schedulable
196 * entity have the same weight on all CPUs. If the group
197 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
198 * better distribution of weight could be:
200 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
201 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
203 * rebalance_shares() is responsible for distributing the shares of a
204 * task groups like this among the group's schedulable entities across
208 unsigned long shares
;
213 /* Default task group's sched entity on each cpu */
214 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
215 /* Default task group's cfs_rq on each cpu */
216 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
218 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
219 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
221 /* task_group_mutex serializes add/remove of task groups and also changes to
222 * a task group's cpu shares.
224 static DEFINE_MUTEX(task_group_mutex
);
226 /* doms_cur_mutex serializes access to doms_cur[] array */
227 static DEFINE_MUTEX(doms_cur_mutex
);
230 /* kernel thread that runs rebalance_shares() periodically */
231 static struct task_struct
*lb_monitor_task
;
232 static int load_balance_monitor(void *unused
);
235 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
);
237 /* Default task group.
238 * Every task in system belong to this group at bootup.
240 struct task_group init_task_group
= {
241 .se
= init_sched_entity_p
,
242 .cfs_rq
= init_cfs_rq_p
,
245 #ifdef CONFIG_FAIR_USER_SCHED
246 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
248 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
251 #define MIN_GROUP_SHARES 2
253 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
255 /* return group to which a task belongs */
256 static inline struct task_group
*task_group(struct task_struct
*p
)
258 struct task_group
*tg
;
260 #ifdef CONFIG_FAIR_USER_SCHED
262 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
263 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
264 struct task_group
, css
);
266 tg
= &init_task_group
;
271 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
272 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
)
274 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
275 p
->se
.parent
= task_group(p
)->se
[cpu
];
278 static inline void lock_task_group_list(void)
280 mutex_lock(&task_group_mutex
);
283 static inline void unlock_task_group_list(void)
285 mutex_unlock(&task_group_mutex
);
288 static inline void lock_doms_cur(void)
290 mutex_lock(&doms_cur_mutex
);
293 static inline void unlock_doms_cur(void)
295 mutex_unlock(&doms_cur_mutex
);
300 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
) { }
301 static inline void lock_task_group_list(void) { }
302 static inline void unlock_task_group_list(void) { }
303 static inline void lock_doms_cur(void) { }
304 static inline void unlock_doms_cur(void) { }
306 #endif /* CONFIG_FAIR_GROUP_SCHED */
308 /* CFS-related fields in a runqueue */
310 struct load_weight load
;
311 unsigned long nr_running
;
316 struct rb_root tasks_timeline
;
317 struct rb_node
*rb_leftmost
;
318 struct rb_node
*rb_load_balance_curr
;
319 /* 'curr' points to currently running entity on this cfs_rq.
320 * It is set to NULL otherwise (i.e when none are currently running).
322 struct sched_entity
*curr
;
324 unsigned long nr_spread_over
;
326 #ifdef CONFIG_FAIR_GROUP_SCHED
327 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
330 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
331 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
332 * (like users, containers etc.)
334 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
335 * list is used during load balance.
337 struct list_head leaf_cfs_rq_list
;
338 struct task_group
*tg
; /* group that "owns" this runqueue */
342 /* Real-Time classes' related field in a runqueue: */
344 struct rt_prio_array active
;
345 unsigned long rt_nr_running
;
347 unsigned long rt_nr_migratory
;
348 int highest_prio
; /* highest queued rt task prio */
358 * We add the notion of a root-domain which will be used to define per-domain
359 * variables. Each exclusive cpuset essentially defines an island domain by
360 * fully partitioning the member cpus from any other cpuset. Whenever a new
361 * exclusive cpuset is created, we also create and attach a new root-domain
371 * The "RT overload" flag: it gets set if a CPU has more than
372 * one runnable RT task.
379 * By default the system creates a single root-domain with all cpus as
380 * members (mimicking the global state we have today).
382 static struct root_domain def_root_domain
;
387 * This is the main, per-CPU runqueue data structure.
389 * Locking rule: those places that want to lock multiple runqueues
390 * (such as the load balancing or the thread migration code), lock
391 * acquire operations must be ordered by ascending &runqueue.
398 * nr_running and cpu_load should be in the same cacheline because
399 * remote CPUs use both these fields when doing load calculation.
401 unsigned long nr_running
;
402 #define CPU_LOAD_IDX_MAX 5
403 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
404 unsigned char idle_at_tick
;
406 unsigned char in_nohz_recently
;
408 /* capture load from *all* tasks on this cpu: */
409 struct load_weight load
;
410 unsigned long nr_load_updates
;
414 #ifdef CONFIG_FAIR_GROUP_SCHED
415 /* list of leaf cfs_rq on this cpu: */
416 struct list_head leaf_cfs_rq_list
;
419 u64 rt_period_expire
;
422 * This is part of a global counter where only the total sum
423 * over all CPUs matters. A task can increase this counter on
424 * one CPU and if it got migrated afterwards it may decrease
425 * it on another CPU. Always updated under the runqueue lock:
427 unsigned long nr_uninterruptible
;
429 struct task_struct
*curr
, *idle
;
430 unsigned long next_balance
;
431 struct mm_struct
*prev_mm
;
433 u64 clock
, prev_clock_raw
;
436 unsigned int clock_warps
, clock_overflows
;
438 unsigned int clock_deep_idle_events
;
444 struct root_domain
*rd
;
445 struct sched_domain
*sd
;
447 /* For active balancing */
450 /* cpu of this runqueue: */
453 struct task_struct
*migration_thread
;
454 struct list_head migration_queue
;
457 #ifdef CONFIG_SCHED_HRTICK
458 unsigned long hrtick_flags
;
459 ktime_t hrtick_expire
;
460 struct hrtimer hrtick_timer
;
463 #ifdef CONFIG_SCHEDSTATS
465 struct sched_info rq_sched_info
;
467 /* sys_sched_yield() stats */
468 unsigned int yld_exp_empty
;
469 unsigned int yld_act_empty
;
470 unsigned int yld_both_empty
;
471 unsigned int yld_count
;
473 /* schedule() stats */
474 unsigned int sched_switch
;
475 unsigned int sched_count
;
476 unsigned int sched_goidle
;
478 /* try_to_wake_up() stats */
479 unsigned int ttwu_count
;
480 unsigned int ttwu_local
;
483 unsigned int bkl_count
;
485 struct lock_class_key rq_lock_key
;
488 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
490 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
492 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
495 static inline int cpu_of(struct rq
*rq
)
505 * Update the per-runqueue clock, as finegrained as the platform can give
506 * us, but without assuming monotonicity, etc.:
508 static void __update_rq_clock(struct rq
*rq
)
510 u64 prev_raw
= rq
->prev_clock_raw
;
511 u64 now
= sched_clock();
512 s64 delta
= now
- prev_raw
;
513 u64 clock
= rq
->clock
;
515 #ifdef CONFIG_SCHED_DEBUG
516 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
519 * Protect against sched_clock() occasionally going backwards:
521 if (unlikely(delta
< 0)) {
526 * Catch too large forward jumps too:
528 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
529 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
530 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
533 rq
->clock_overflows
++;
535 if (unlikely(delta
> rq
->clock_max_delta
))
536 rq
->clock_max_delta
= delta
;
541 rq
->prev_clock_raw
= now
;
545 static void update_rq_clock(struct rq
*rq
)
547 if (likely(smp_processor_id() == cpu_of(rq
)))
548 __update_rq_clock(rq
);
552 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
553 * See detach_destroy_domains: synchronize_sched for details.
555 * The domain tree of any CPU may only be accessed from within
556 * preempt-disabled sections.
558 #define for_each_domain(cpu, __sd) \
559 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
561 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
562 #define this_rq() (&__get_cpu_var(runqueues))
563 #define task_rq(p) cpu_rq(task_cpu(p))
564 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
567 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
569 #ifdef CONFIG_SCHED_DEBUG
570 # define const_debug __read_mostly
572 # define const_debug static const
576 * Debugging: various feature bits
579 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
580 SCHED_FEAT_WAKEUP_PREEMPT
= 2,
581 SCHED_FEAT_START_DEBIT
= 4,
582 SCHED_FEAT_TREE_AVG
= 8,
583 SCHED_FEAT_APPROX_AVG
= 16,
584 SCHED_FEAT_HRTICK
= 32,
585 SCHED_FEAT_DOUBLE_TICK
= 64,
588 const_debug
unsigned int sysctl_sched_features
=
589 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
590 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
591 SCHED_FEAT_START_DEBIT
* 1 |
592 SCHED_FEAT_TREE_AVG
* 0 |
593 SCHED_FEAT_APPROX_AVG
* 0 |
594 SCHED_FEAT_HRTICK
* 1 |
595 SCHED_FEAT_DOUBLE_TICK
* 0;
597 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
600 * Number of tasks to iterate in a single balance run.
601 * Limited because this is done with IRQs disabled.
603 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
606 * period over which we measure -rt task cpu usage in ms.
609 const_debug
unsigned int sysctl_sched_rt_period
= 1000;
611 #define SCHED_RT_FRAC_SHIFT 16
612 #define SCHED_RT_FRAC (1UL << SCHED_RT_FRAC_SHIFT)
615 * ratio of time -rt tasks may consume.
618 const_debug
unsigned int sysctl_sched_rt_ratio
= SCHED_RT_FRAC
;
621 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
622 * clock constructed from sched_clock():
624 unsigned long long cpu_clock(int cpu
)
626 unsigned long long now
;
630 local_irq_save(flags
);
633 * Only call sched_clock() if the scheduler has already been
634 * initialized (some code might call cpu_clock() very early):
639 local_irq_restore(flags
);
643 EXPORT_SYMBOL_GPL(cpu_clock
);
645 #ifndef prepare_arch_switch
646 # define prepare_arch_switch(next) do { } while (0)
648 #ifndef finish_arch_switch
649 # define finish_arch_switch(prev) do { } while (0)
652 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
654 return rq
->curr
== p
;
657 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
658 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
660 return task_current(rq
, p
);
663 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
667 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
669 #ifdef CONFIG_DEBUG_SPINLOCK
670 /* this is a valid case when another task releases the spinlock */
671 rq
->lock
.owner
= current
;
674 * If we are tracking spinlock dependencies then we have to
675 * fix up the runqueue lock - which gets 'carried over' from
678 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
680 spin_unlock_irq(&rq
->lock
);
683 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
684 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
689 return task_current(rq
, p
);
693 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
697 * We can optimise this out completely for !SMP, because the
698 * SMP rebalancing from interrupt is the only thing that cares
703 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
704 spin_unlock_irq(&rq
->lock
);
706 spin_unlock(&rq
->lock
);
710 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
714 * After ->oncpu is cleared, the task can be moved to a different CPU.
715 * We must ensure this doesn't happen until the switch is completely
721 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
725 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
728 * __task_rq_lock - lock the runqueue a given task resides on.
729 * Must be called interrupts disabled.
731 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
735 struct rq
*rq
= task_rq(p
);
736 spin_lock(&rq
->lock
);
737 if (likely(rq
== task_rq(p
)))
739 spin_unlock(&rq
->lock
);
744 * task_rq_lock - lock the runqueue a given task resides on and disable
745 * interrupts. Note the ordering: we can safely lookup the task_rq without
746 * explicitly disabling preemption.
748 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
754 local_irq_save(*flags
);
756 spin_lock(&rq
->lock
);
757 if (likely(rq
== task_rq(p
)))
759 spin_unlock_irqrestore(&rq
->lock
, *flags
);
763 static void __task_rq_unlock(struct rq
*rq
)
766 spin_unlock(&rq
->lock
);
769 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
772 spin_unlock_irqrestore(&rq
->lock
, *flags
);
776 * this_rq_lock - lock this runqueue and disable interrupts.
778 static struct rq
*this_rq_lock(void)
785 spin_lock(&rq
->lock
);
791 * We are going deep-idle (irqs are disabled):
793 void sched_clock_idle_sleep_event(void)
795 struct rq
*rq
= cpu_rq(smp_processor_id());
797 spin_lock(&rq
->lock
);
798 __update_rq_clock(rq
);
799 spin_unlock(&rq
->lock
);
800 rq
->clock_deep_idle_events
++;
802 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
805 * We just idled delta nanoseconds (called with irqs disabled):
807 void sched_clock_idle_wakeup_event(u64 delta_ns
)
809 struct rq
*rq
= cpu_rq(smp_processor_id());
810 u64 now
= sched_clock();
812 touch_softlockup_watchdog();
813 rq
->idle_clock
+= delta_ns
;
815 * Override the previous timestamp and ignore all
816 * sched_clock() deltas that occured while we idled,
817 * and use the PM-provided delta_ns to advance the
820 spin_lock(&rq
->lock
);
821 rq
->prev_clock_raw
= now
;
822 rq
->clock
+= delta_ns
;
823 spin_unlock(&rq
->lock
);
825 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
827 static void __resched_task(struct task_struct
*p
, int tif_bit
);
829 static inline void resched_task(struct task_struct
*p
)
831 __resched_task(p
, TIF_NEED_RESCHED
);
834 #ifdef CONFIG_SCHED_HRTICK
836 * Use HR-timers to deliver accurate preemption points.
838 * Its all a bit involved since we cannot program an hrt while holding the
839 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
842 * When we get rescheduled we reprogram the hrtick_timer outside of the
845 static inline void resched_hrt(struct task_struct
*p
)
847 __resched_task(p
, TIF_HRTICK_RESCHED
);
850 static inline void resched_rq(struct rq
*rq
)
854 spin_lock_irqsave(&rq
->lock
, flags
);
855 resched_task(rq
->curr
);
856 spin_unlock_irqrestore(&rq
->lock
, flags
);
860 HRTICK_SET
, /* re-programm hrtick_timer */
861 HRTICK_RESET
, /* not a new slice */
866 * - enabled by features
867 * - hrtimer is actually high res
869 static inline int hrtick_enabled(struct rq
*rq
)
871 if (!sched_feat(HRTICK
))
873 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
877 * Called to set the hrtick timer state.
879 * called with rq->lock held and irqs disabled
881 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
883 assert_spin_locked(&rq
->lock
);
886 * preempt at: now + delay
889 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
891 * indicate we need to program the timer
893 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
895 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
898 * New slices are called from the schedule path and don't need a
902 resched_hrt(rq
->curr
);
905 static void hrtick_clear(struct rq
*rq
)
907 if (hrtimer_active(&rq
->hrtick_timer
))
908 hrtimer_cancel(&rq
->hrtick_timer
);
912 * Update the timer from the possible pending state.
914 static void hrtick_set(struct rq
*rq
)
920 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
922 spin_lock_irqsave(&rq
->lock
, flags
);
923 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
924 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
925 time
= rq
->hrtick_expire
;
926 clear_thread_flag(TIF_HRTICK_RESCHED
);
927 spin_unlock_irqrestore(&rq
->lock
, flags
);
930 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
931 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
938 * High-resolution timer tick.
939 * Runs from hardirq context with interrupts disabled.
941 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
943 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
945 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
947 spin_lock(&rq
->lock
);
948 __update_rq_clock(rq
);
949 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
950 spin_unlock(&rq
->lock
);
952 return HRTIMER_NORESTART
;
955 static inline void init_rq_hrtick(struct rq
*rq
)
957 rq
->hrtick_flags
= 0;
958 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
959 rq
->hrtick_timer
.function
= hrtick
;
960 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
963 void hrtick_resched(void)
968 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
971 local_irq_save(flags
);
972 rq
= cpu_rq(smp_processor_id());
974 local_irq_restore(flags
);
977 static inline void hrtick_clear(struct rq
*rq
)
981 static inline void hrtick_set(struct rq
*rq
)
985 static inline void init_rq_hrtick(struct rq
*rq
)
989 void hrtick_resched(void)
995 * resched_task - mark a task 'to be rescheduled now'.
997 * On UP this means the setting of the need_resched flag, on SMP it
998 * might also involve a cross-CPU call to trigger the scheduler on
1003 #ifndef tsk_is_polling
1004 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1007 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1011 assert_spin_locked(&task_rq(p
)->lock
);
1013 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1016 set_tsk_thread_flag(p
, tif_bit
);
1019 if (cpu
== smp_processor_id())
1022 /* NEED_RESCHED must be visible before we test polling */
1024 if (!tsk_is_polling(p
))
1025 smp_send_reschedule(cpu
);
1028 static void resched_cpu(int cpu
)
1030 struct rq
*rq
= cpu_rq(cpu
);
1031 unsigned long flags
;
1033 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1035 resched_task(cpu_curr(cpu
));
1036 spin_unlock_irqrestore(&rq
->lock
, flags
);
1039 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1041 assert_spin_locked(&task_rq(p
)->lock
);
1042 set_tsk_thread_flag(p
, tif_bit
);
1046 #if BITS_PER_LONG == 32
1047 # define WMULT_CONST (~0UL)
1049 # define WMULT_CONST (1UL << 32)
1052 #define WMULT_SHIFT 32
1055 * Shift right and round:
1057 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1059 static unsigned long
1060 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1061 struct load_weight
*lw
)
1065 if (unlikely(!lw
->inv_weight
))
1066 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
1068 tmp
= (u64
)delta_exec
* weight
;
1070 * Check whether we'd overflow the 64-bit multiplication:
1072 if (unlikely(tmp
> WMULT_CONST
))
1073 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1076 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1078 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1081 static inline unsigned long
1082 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
1084 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
1087 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1092 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1098 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1099 * of tasks with abnormal "nice" values across CPUs the contribution that
1100 * each task makes to its run queue's load is weighted according to its
1101 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1102 * scaled version of the new time slice allocation that they receive on time
1106 #define WEIGHT_IDLEPRIO 2
1107 #define WMULT_IDLEPRIO (1 << 31)
1110 * Nice levels are multiplicative, with a gentle 10% change for every
1111 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1112 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1113 * that remained on nice 0.
1115 * The "10% effect" is relative and cumulative: from _any_ nice level,
1116 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1117 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1118 * If a task goes up by ~10% and another task goes down by ~10% then
1119 * the relative distance between them is ~25%.)
1121 static const int prio_to_weight
[40] = {
1122 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1123 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1124 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1125 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1126 /* 0 */ 1024, 820, 655, 526, 423,
1127 /* 5 */ 335, 272, 215, 172, 137,
1128 /* 10 */ 110, 87, 70, 56, 45,
1129 /* 15 */ 36, 29, 23, 18, 15,
1133 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1135 * In cases where the weight does not change often, we can use the
1136 * precalculated inverse to speed up arithmetics by turning divisions
1137 * into multiplications:
1139 static const u32 prio_to_wmult
[40] = {
1140 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1141 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1142 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1143 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1144 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1145 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1146 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1147 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1150 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1153 * runqueue iterator, to support SMP load-balancing between different
1154 * scheduling classes, without having to expose their internal data
1155 * structures to the load-balancing proper:
1157 struct rq_iterator
{
1159 struct task_struct
*(*start
)(void *);
1160 struct task_struct
*(*next
)(void *);
1164 static unsigned long
1165 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1166 unsigned long max_load_move
, struct sched_domain
*sd
,
1167 enum cpu_idle_type idle
, int *all_pinned
,
1168 int *this_best_prio
, struct rq_iterator
*iterator
);
1171 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1172 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1173 struct rq_iterator
*iterator
);
1176 #ifdef CONFIG_CGROUP_CPUACCT
1177 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1179 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1182 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1184 update_load_add(&rq
->load
, load
);
1187 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1189 update_load_sub(&rq
->load
, load
);
1193 static unsigned long source_load(int cpu
, int type
);
1194 static unsigned long target_load(int cpu
, int type
);
1195 static unsigned long cpu_avg_load_per_task(int cpu
);
1196 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1197 #endif /* CONFIG_SMP */
1199 #include "sched_stats.h"
1200 #include "sched_idletask.c"
1201 #include "sched_fair.c"
1202 #include "sched_rt.c"
1203 #ifdef CONFIG_SCHED_DEBUG
1204 # include "sched_debug.c"
1207 #define sched_class_highest (&rt_sched_class)
1209 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
1214 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
1219 static void set_load_weight(struct task_struct
*p
)
1221 if (task_has_rt_policy(p
)) {
1222 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1223 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1228 * SCHED_IDLE tasks get minimal weight:
1230 if (p
->policy
== SCHED_IDLE
) {
1231 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1232 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1236 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1237 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1240 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1242 sched_info_queued(p
);
1243 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1247 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1249 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1254 * __normal_prio - return the priority that is based on the static prio
1256 static inline int __normal_prio(struct task_struct
*p
)
1258 return p
->static_prio
;
1262 * Calculate the expected normal priority: i.e. priority
1263 * without taking RT-inheritance into account. Might be
1264 * boosted by interactivity modifiers. Changes upon fork,
1265 * setprio syscalls, and whenever the interactivity
1266 * estimator recalculates.
1268 static inline int normal_prio(struct task_struct
*p
)
1272 if (task_has_rt_policy(p
))
1273 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1275 prio
= __normal_prio(p
);
1280 * Calculate the current priority, i.e. the priority
1281 * taken into account by the scheduler. This value might
1282 * be boosted by RT tasks, or might be boosted by
1283 * interactivity modifiers. Will be RT if the task got
1284 * RT-boosted. If not then it returns p->normal_prio.
1286 static int effective_prio(struct task_struct
*p
)
1288 p
->normal_prio
= normal_prio(p
);
1290 * If we are RT tasks or we were boosted to RT priority,
1291 * keep the priority unchanged. Otherwise, update priority
1292 * to the normal priority:
1294 if (!rt_prio(p
->prio
))
1295 return p
->normal_prio
;
1300 * activate_task - move a task to the runqueue.
1302 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1304 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1305 rq
->nr_uninterruptible
--;
1307 enqueue_task(rq
, p
, wakeup
);
1308 inc_nr_running(p
, rq
);
1312 * deactivate_task - remove a task from the runqueue.
1314 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1316 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1317 rq
->nr_uninterruptible
++;
1319 dequeue_task(rq
, p
, sleep
);
1320 dec_nr_running(p
, rq
);
1324 * task_curr - is this task currently executing on a CPU?
1325 * @p: the task in question.
1327 inline int task_curr(const struct task_struct
*p
)
1329 return cpu_curr(task_cpu(p
)) == p
;
1332 /* Used instead of source_load when we know the type == 0 */
1333 unsigned long weighted_cpuload(const int cpu
)
1335 return cpu_rq(cpu
)->load
.weight
;
1338 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1340 set_task_cfs_rq(p
, cpu
);
1343 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1344 * successfuly executed on another CPU. We must ensure that updates of
1345 * per-task data have been completed by this moment.
1348 task_thread_info(p
)->cpu
= cpu
;
1352 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1353 const struct sched_class
*prev_class
,
1354 int oldprio
, int running
)
1356 if (prev_class
!= p
->sched_class
) {
1357 if (prev_class
->switched_from
)
1358 prev_class
->switched_from(rq
, p
, running
);
1359 p
->sched_class
->switched_to(rq
, p
, running
);
1361 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1367 * Is this task likely cache-hot:
1370 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1374 if (p
->sched_class
!= &fair_sched_class
)
1377 if (sysctl_sched_migration_cost
== -1)
1379 if (sysctl_sched_migration_cost
== 0)
1382 delta
= now
- p
->se
.exec_start
;
1384 return delta
< (s64
)sysctl_sched_migration_cost
;
1388 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1390 int old_cpu
= task_cpu(p
);
1391 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1392 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1393 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1396 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1398 #ifdef CONFIG_SCHEDSTATS
1399 if (p
->se
.wait_start
)
1400 p
->se
.wait_start
-= clock_offset
;
1401 if (p
->se
.sleep_start
)
1402 p
->se
.sleep_start
-= clock_offset
;
1403 if (p
->se
.block_start
)
1404 p
->se
.block_start
-= clock_offset
;
1405 if (old_cpu
!= new_cpu
) {
1406 schedstat_inc(p
, se
.nr_migrations
);
1407 if (task_hot(p
, old_rq
->clock
, NULL
))
1408 schedstat_inc(p
, se
.nr_forced2_migrations
);
1411 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1412 new_cfsrq
->min_vruntime
;
1414 __set_task_cpu(p
, new_cpu
);
1417 struct migration_req
{
1418 struct list_head list
;
1420 struct task_struct
*task
;
1423 struct completion done
;
1427 * The task's runqueue lock must be held.
1428 * Returns true if you have to wait for migration thread.
1431 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1433 struct rq
*rq
= task_rq(p
);
1436 * If the task is not on a runqueue (and not running), then
1437 * it is sufficient to simply update the task's cpu field.
1439 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1440 set_task_cpu(p
, dest_cpu
);
1444 init_completion(&req
->done
);
1446 req
->dest_cpu
= dest_cpu
;
1447 list_add(&req
->list
, &rq
->migration_queue
);
1453 * wait_task_inactive - wait for a thread to unschedule.
1455 * The caller must ensure that the task *will* unschedule sometime soon,
1456 * else this function might spin for a *long* time. This function can't
1457 * be called with interrupts off, or it may introduce deadlock with
1458 * smp_call_function() if an IPI is sent by the same process we are
1459 * waiting to become inactive.
1461 void wait_task_inactive(struct task_struct
*p
)
1463 unsigned long flags
;
1469 * We do the initial early heuristics without holding
1470 * any task-queue locks at all. We'll only try to get
1471 * the runqueue lock when things look like they will
1477 * If the task is actively running on another CPU
1478 * still, just relax and busy-wait without holding
1481 * NOTE! Since we don't hold any locks, it's not
1482 * even sure that "rq" stays as the right runqueue!
1483 * But we don't care, since "task_running()" will
1484 * return false if the runqueue has changed and p
1485 * is actually now running somewhere else!
1487 while (task_running(rq
, p
))
1491 * Ok, time to look more closely! We need the rq
1492 * lock now, to be *sure*. If we're wrong, we'll
1493 * just go back and repeat.
1495 rq
= task_rq_lock(p
, &flags
);
1496 running
= task_running(rq
, p
);
1497 on_rq
= p
->se
.on_rq
;
1498 task_rq_unlock(rq
, &flags
);
1501 * Was it really running after all now that we
1502 * checked with the proper locks actually held?
1504 * Oops. Go back and try again..
1506 if (unlikely(running
)) {
1512 * It's not enough that it's not actively running,
1513 * it must be off the runqueue _entirely_, and not
1516 * So if it wa still runnable (but just not actively
1517 * running right now), it's preempted, and we should
1518 * yield - it could be a while.
1520 if (unlikely(on_rq
)) {
1521 schedule_timeout_uninterruptible(1);
1526 * Ahh, all good. It wasn't running, and it wasn't
1527 * runnable, which means that it will never become
1528 * running in the future either. We're all done!
1535 * kick_process - kick a running thread to enter/exit the kernel
1536 * @p: the to-be-kicked thread
1538 * Cause a process which is running on another CPU to enter
1539 * kernel-mode, without any delay. (to get signals handled.)
1541 * NOTE: this function doesnt have to take the runqueue lock,
1542 * because all it wants to ensure is that the remote task enters
1543 * the kernel. If the IPI races and the task has been migrated
1544 * to another CPU then no harm is done and the purpose has been
1547 void kick_process(struct task_struct
*p
)
1553 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1554 smp_send_reschedule(cpu
);
1559 * Return a low guess at the load of a migration-source cpu weighted
1560 * according to the scheduling class and "nice" value.
1562 * We want to under-estimate the load of migration sources, to
1563 * balance conservatively.
1565 static unsigned long source_load(int cpu
, int type
)
1567 struct rq
*rq
= cpu_rq(cpu
);
1568 unsigned long total
= weighted_cpuload(cpu
);
1573 return min(rq
->cpu_load
[type
-1], total
);
1577 * Return a high guess at the load of a migration-target cpu weighted
1578 * according to the scheduling class and "nice" value.
1580 static unsigned long target_load(int cpu
, int type
)
1582 struct rq
*rq
= cpu_rq(cpu
);
1583 unsigned long total
= weighted_cpuload(cpu
);
1588 return max(rq
->cpu_load
[type
-1], total
);
1592 * Return the average load per task on the cpu's run queue
1594 static unsigned long cpu_avg_load_per_task(int cpu
)
1596 struct rq
*rq
= cpu_rq(cpu
);
1597 unsigned long total
= weighted_cpuload(cpu
);
1598 unsigned long n
= rq
->nr_running
;
1600 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1604 * find_idlest_group finds and returns the least busy CPU group within the
1607 static struct sched_group
*
1608 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1610 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1611 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1612 int load_idx
= sd
->forkexec_idx
;
1613 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1616 unsigned long load
, avg_load
;
1620 /* Skip over this group if it has no CPUs allowed */
1621 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1624 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1626 /* Tally up the load of all CPUs in the group */
1629 for_each_cpu_mask(i
, group
->cpumask
) {
1630 /* Bias balancing toward cpus of our domain */
1632 load
= source_load(i
, load_idx
);
1634 load
= target_load(i
, load_idx
);
1639 /* Adjust by relative CPU power of the group */
1640 avg_load
= sg_div_cpu_power(group
,
1641 avg_load
* SCHED_LOAD_SCALE
);
1644 this_load
= avg_load
;
1646 } else if (avg_load
< min_load
) {
1647 min_load
= avg_load
;
1650 } while (group
= group
->next
, group
!= sd
->groups
);
1652 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1658 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1661 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1664 unsigned long load
, min_load
= ULONG_MAX
;
1668 /* Traverse only the allowed CPUs */
1669 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1671 for_each_cpu_mask(i
, tmp
) {
1672 load
= weighted_cpuload(i
);
1674 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1684 * sched_balance_self: balance the current task (running on cpu) in domains
1685 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1688 * Balance, ie. select the least loaded group.
1690 * Returns the target CPU number, or the same CPU if no balancing is needed.
1692 * preempt must be disabled.
1694 static int sched_balance_self(int cpu
, int flag
)
1696 struct task_struct
*t
= current
;
1697 struct sched_domain
*tmp
, *sd
= NULL
;
1699 for_each_domain(cpu
, tmp
) {
1701 * If power savings logic is enabled for a domain, stop there.
1703 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1705 if (tmp
->flags
& flag
)
1711 struct sched_group
*group
;
1712 int new_cpu
, weight
;
1714 if (!(sd
->flags
& flag
)) {
1720 group
= find_idlest_group(sd
, t
, cpu
);
1726 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1727 if (new_cpu
== -1 || new_cpu
== cpu
) {
1728 /* Now try balancing at a lower domain level of cpu */
1733 /* Now try balancing at a lower domain level of new_cpu */
1736 weight
= cpus_weight(span
);
1737 for_each_domain(cpu
, tmp
) {
1738 if (weight
<= cpus_weight(tmp
->span
))
1740 if (tmp
->flags
& flag
)
1743 /* while loop will break here if sd == NULL */
1749 #endif /* CONFIG_SMP */
1752 * try_to_wake_up - wake up a thread
1753 * @p: the to-be-woken-up thread
1754 * @state: the mask of task states that can be woken
1755 * @sync: do a synchronous wakeup?
1757 * Put it on the run-queue if it's not already there. The "current"
1758 * thread is always on the run-queue (except when the actual
1759 * re-schedule is in progress), and as such you're allowed to do
1760 * the simpler "current->state = TASK_RUNNING" to mark yourself
1761 * runnable without the overhead of this.
1763 * returns failure only if the task is already active.
1765 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1767 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1768 unsigned long flags
;
1772 rq
= task_rq_lock(p
, &flags
);
1773 old_state
= p
->state
;
1774 if (!(old_state
& state
))
1782 this_cpu
= smp_processor_id();
1785 if (unlikely(task_running(rq
, p
)))
1788 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
1789 if (cpu
!= orig_cpu
) {
1790 set_task_cpu(p
, cpu
);
1791 task_rq_unlock(rq
, &flags
);
1792 /* might preempt at this point */
1793 rq
= task_rq_lock(p
, &flags
);
1794 old_state
= p
->state
;
1795 if (!(old_state
& state
))
1800 this_cpu
= smp_processor_id();
1804 #ifdef CONFIG_SCHEDSTATS
1805 schedstat_inc(rq
, ttwu_count
);
1806 if (cpu
== this_cpu
)
1807 schedstat_inc(rq
, ttwu_local
);
1809 struct sched_domain
*sd
;
1810 for_each_domain(this_cpu
, sd
) {
1811 if (cpu_isset(cpu
, sd
->span
)) {
1812 schedstat_inc(sd
, ttwu_wake_remote
);
1820 #endif /* CONFIG_SMP */
1821 schedstat_inc(p
, se
.nr_wakeups
);
1823 schedstat_inc(p
, se
.nr_wakeups_sync
);
1824 if (orig_cpu
!= cpu
)
1825 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1826 if (cpu
== this_cpu
)
1827 schedstat_inc(p
, se
.nr_wakeups_local
);
1829 schedstat_inc(p
, se
.nr_wakeups_remote
);
1830 update_rq_clock(rq
);
1831 activate_task(rq
, p
, 1);
1832 check_preempt_curr(rq
, p
);
1836 p
->state
= TASK_RUNNING
;
1838 if (p
->sched_class
->task_wake_up
)
1839 p
->sched_class
->task_wake_up(rq
, p
);
1842 task_rq_unlock(rq
, &flags
);
1847 int fastcall
wake_up_process(struct task_struct
*p
)
1849 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1850 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1852 EXPORT_SYMBOL(wake_up_process
);
1854 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1856 return try_to_wake_up(p
, state
, 0);
1860 * Perform scheduler related setup for a newly forked process p.
1861 * p is forked by current.
1863 * __sched_fork() is basic setup used by init_idle() too:
1865 static void __sched_fork(struct task_struct
*p
)
1867 p
->se
.exec_start
= 0;
1868 p
->se
.sum_exec_runtime
= 0;
1869 p
->se
.prev_sum_exec_runtime
= 0;
1871 #ifdef CONFIG_SCHEDSTATS
1872 p
->se
.wait_start
= 0;
1873 p
->se
.sum_sleep_runtime
= 0;
1874 p
->se
.sleep_start
= 0;
1875 p
->se
.block_start
= 0;
1876 p
->se
.sleep_max
= 0;
1877 p
->se
.block_max
= 0;
1879 p
->se
.slice_max
= 0;
1883 INIT_LIST_HEAD(&p
->rt
.run_list
);
1886 #ifdef CONFIG_PREEMPT_NOTIFIERS
1887 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1891 * We mark the process as running here, but have not actually
1892 * inserted it onto the runqueue yet. This guarantees that
1893 * nobody will actually run it, and a signal or other external
1894 * event cannot wake it up and insert it on the runqueue either.
1896 p
->state
= TASK_RUNNING
;
1900 * fork()/clone()-time setup:
1902 void sched_fork(struct task_struct
*p
, int clone_flags
)
1904 int cpu
= get_cpu();
1909 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1911 set_task_cpu(p
, cpu
);
1914 * Make sure we do not leak PI boosting priority to the child:
1916 p
->prio
= current
->normal_prio
;
1917 if (!rt_prio(p
->prio
))
1918 p
->sched_class
= &fair_sched_class
;
1920 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1921 if (likely(sched_info_on()))
1922 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1924 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1927 #ifdef CONFIG_PREEMPT
1928 /* Want to start with kernel preemption disabled. */
1929 task_thread_info(p
)->preempt_count
= 1;
1935 * wake_up_new_task - wake up a newly created task for the first time.
1937 * This function will do some initial scheduler statistics housekeeping
1938 * that must be done for every newly created context, then puts the task
1939 * on the runqueue and wakes it.
1941 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1943 unsigned long flags
;
1946 rq
= task_rq_lock(p
, &flags
);
1947 BUG_ON(p
->state
!= TASK_RUNNING
);
1948 update_rq_clock(rq
);
1950 p
->prio
= effective_prio(p
);
1952 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
1953 activate_task(rq
, p
, 0);
1956 * Let the scheduling class do new task startup
1957 * management (if any):
1959 p
->sched_class
->task_new(rq
, p
);
1960 inc_nr_running(p
, rq
);
1962 check_preempt_curr(rq
, p
);
1964 if (p
->sched_class
->task_wake_up
)
1965 p
->sched_class
->task_wake_up(rq
, p
);
1967 task_rq_unlock(rq
, &flags
);
1970 #ifdef CONFIG_PREEMPT_NOTIFIERS
1973 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1974 * @notifier: notifier struct to register
1976 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1978 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1980 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1983 * preempt_notifier_unregister - no longer interested in preemption notifications
1984 * @notifier: notifier struct to unregister
1986 * This is safe to call from within a preemption notifier.
1988 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1990 hlist_del(¬ifier
->link
);
1992 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1994 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1996 struct preempt_notifier
*notifier
;
1997 struct hlist_node
*node
;
1999 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2000 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2004 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2005 struct task_struct
*next
)
2007 struct preempt_notifier
*notifier
;
2008 struct hlist_node
*node
;
2010 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2011 notifier
->ops
->sched_out(notifier
, next
);
2016 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2021 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2022 struct task_struct
*next
)
2029 * prepare_task_switch - prepare to switch tasks
2030 * @rq: the runqueue preparing to switch
2031 * @prev: the current task that is being switched out
2032 * @next: the task we are going to switch to.
2034 * This is called with the rq lock held and interrupts off. It must
2035 * be paired with a subsequent finish_task_switch after the context
2038 * prepare_task_switch sets up locking and calls architecture specific
2042 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2043 struct task_struct
*next
)
2045 fire_sched_out_preempt_notifiers(prev
, next
);
2046 prepare_lock_switch(rq
, next
);
2047 prepare_arch_switch(next
);
2051 * finish_task_switch - clean up after a task-switch
2052 * @rq: runqueue associated with task-switch
2053 * @prev: the thread we just switched away from.
2055 * finish_task_switch must be called after the context switch, paired
2056 * with a prepare_task_switch call before the context switch.
2057 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2058 * and do any other architecture-specific cleanup actions.
2060 * Note that we may have delayed dropping an mm in context_switch(). If
2061 * so, we finish that here outside of the runqueue lock. (Doing it
2062 * with the lock held can cause deadlocks; see schedule() for
2065 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2066 __releases(rq
->lock
)
2068 struct mm_struct
*mm
= rq
->prev_mm
;
2074 * A task struct has one reference for the use as "current".
2075 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2076 * schedule one last time. The schedule call will never return, and
2077 * the scheduled task must drop that reference.
2078 * The test for TASK_DEAD must occur while the runqueue locks are
2079 * still held, otherwise prev could be scheduled on another cpu, die
2080 * there before we look at prev->state, and then the reference would
2082 * Manfred Spraul <manfred@colorfullife.com>
2084 prev_state
= prev
->state
;
2085 finish_arch_switch(prev
);
2086 finish_lock_switch(rq
, prev
);
2088 if (current
->sched_class
->post_schedule
)
2089 current
->sched_class
->post_schedule(rq
);
2092 fire_sched_in_preempt_notifiers(current
);
2095 if (unlikely(prev_state
== TASK_DEAD
)) {
2097 * Remove function-return probe instances associated with this
2098 * task and put them back on the free list.
2100 kprobe_flush_task(prev
);
2101 put_task_struct(prev
);
2106 * schedule_tail - first thing a freshly forked thread must call.
2107 * @prev: the thread we just switched away from.
2109 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2110 __releases(rq
->lock
)
2112 struct rq
*rq
= this_rq();
2114 finish_task_switch(rq
, prev
);
2115 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2116 /* In this case, finish_task_switch does not reenable preemption */
2119 if (current
->set_child_tid
)
2120 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2124 * context_switch - switch to the new MM and the new
2125 * thread's register state.
2128 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2129 struct task_struct
*next
)
2131 struct mm_struct
*mm
, *oldmm
;
2133 prepare_task_switch(rq
, prev
, next
);
2135 oldmm
= prev
->active_mm
;
2137 * For paravirt, this is coupled with an exit in switch_to to
2138 * combine the page table reload and the switch backend into
2141 arch_enter_lazy_cpu_mode();
2143 if (unlikely(!mm
)) {
2144 next
->active_mm
= oldmm
;
2145 atomic_inc(&oldmm
->mm_count
);
2146 enter_lazy_tlb(oldmm
, next
);
2148 switch_mm(oldmm
, mm
, next
);
2150 if (unlikely(!prev
->mm
)) {
2151 prev
->active_mm
= NULL
;
2152 rq
->prev_mm
= oldmm
;
2155 * Since the runqueue lock will be released by the next
2156 * task (which is an invalid locking op but in the case
2157 * of the scheduler it's an obvious special-case), so we
2158 * do an early lockdep release here:
2160 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2161 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2164 /* Here we just switch the register state and the stack. */
2165 switch_to(prev
, next
, prev
);
2169 * this_rq must be evaluated again because prev may have moved
2170 * CPUs since it called schedule(), thus the 'rq' on its stack
2171 * frame will be invalid.
2173 finish_task_switch(this_rq(), prev
);
2177 * nr_running, nr_uninterruptible and nr_context_switches:
2179 * externally visible scheduler statistics: current number of runnable
2180 * threads, current number of uninterruptible-sleeping threads, total
2181 * number of context switches performed since bootup.
2183 unsigned long nr_running(void)
2185 unsigned long i
, sum
= 0;
2187 for_each_online_cpu(i
)
2188 sum
+= cpu_rq(i
)->nr_running
;
2193 unsigned long nr_uninterruptible(void)
2195 unsigned long i
, sum
= 0;
2197 for_each_possible_cpu(i
)
2198 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2201 * Since we read the counters lockless, it might be slightly
2202 * inaccurate. Do not allow it to go below zero though:
2204 if (unlikely((long)sum
< 0))
2210 unsigned long long nr_context_switches(void)
2213 unsigned long long sum
= 0;
2215 for_each_possible_cpu(i
)
2216 sum
+= cpu_rq(i
)->nr_switches
;
2221 unsigned long nr_iowait(void)
2223 unsigned long i
, sum
= 0;
2225 for_each_possible_cpu(i
)
2226 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2231 unsigned long nr_active(void)
2233 unsigned long i
, running
= 0, uninterruptible
= 0;
2235 for_each_online_cpu(i
) {
2236 running
+= cpu_rq(i
)->nr_running
;
2237 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2240 if (unlikely((long)uninterruptible
< 0))
2241 uninterruptible
= 0;
2243 return running
+ uninterruptible
;
2247 * Update rq->cpu_load[] statistics. This function is usually called every
2248 * scheduler tick (TICK_NSEC).
2250 static void update_cpu_load(struct rq
*this_rq
)
2252 unsigned long this_load
= this_rq
->load
.weight
;
2255 this_rq
->nr_load_updates
++;
2257 /* Update our load: */
2258 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2259 unsigned long old_load
, new_load
;
2261 /* scale is effectively 1 << i now, and >> i divides by scale */
2263 old_load
= this_rq
->cpu_load
[i
];
2264 new_load
= this_load
;
2266 * Round up the averaging division if load is increasing. This
2267 * prevents us from getting stuck on 9 if the load is 10, for
2270 if (new_load
> old_load
)
2271 new_load
+= scale
-1;
2272 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2279 * double_rq_lock - safely lock two runqueues
2281 * Note this does not disable interrupts like task_rq_lock,
2282 * you need to do so manually before calling.
2284 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2285 __acquires(rq1
->lock
)
2286 __acquires(rq2
->lock
)
2288 BUG_ON(!irqs_disabled());
2290 spin_lock(&rq1
->lock
);
2291 __acquire(rq2
->lock
); /* Fake it out ;) */
2294 spin_lock(&rq1
->lock
);
2295 spin_lock(&rq2
->lock
);
2297 spin_lock(&rq2
->lock
);
2298 spin_lock(&rq1
->lock
);
2301 update_rq_clock(rq1
);
2302 update_rq_clock(rq2
);
2306 * double_rq_unlock - safely unlock two runqueues
2308 * Note this does not restore interrupts like task_rq_unlock,
2309 * you need to do so manually after calling.
2311 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2312 __releases(rq1
->lock
)
2313 __releases(rq2
->lock
)
2315 spin_unlock(&rq1
->lock
);
2317 spin_unlock(&rq2
->lock
);
2319 __release(rq2
->lock
);
2323 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2325 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2326 __releases(this_rq
->lock
)
2327 __acquires(busiest
->lock
)
2328 __acquires(this_rq
->lock
)
2332 if (unlikely(!irqs_disabled())) {
2333 /* printk() doesn't work good under rq->lock */
2334 spin_unlock(&this_rq
->lock
);
2337 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2338 if (busiest
< this_rq
) {
2339 spin_unlock(&this_rq
->lock
);
2340 spin_lock(&busiest
->lock
);
2341 spin_lock(&this_rq
->lock
);
2344 spin_lock(&busiest
->lock
);
2350 * If dest_cpu is allowed for this process, migrate the task to it.
2351 * This is accomplished by forcing the cpu_allowed mask to only
2352 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2353 * the cpu_allowed mask is restored.
2355 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2357 struct migration_req req
;
2358 unsigned long flags
;
2361 rq
= task_rq_lock(p
, &flags
);
2362 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2363 || unlikely(cpu_is_offline(dest_cpu
)))
2366 /* force the process onto the specified CPU */
2367 if (migrate_task(p
, dest_cpu
, &req
)) {
2368 /* Need to wait for migration thread (might exit: take ref). */
2369 struct task_struct
*mt
= rq
->migration_thread
;
2371 get_task_struct(mt
);
2372 task_rq_unlock(rq
, &flags
);
2373 wake_up_process(mt
);
2374 put_task_struct(mt
);
2375 wait_for_completion(&req
.done
);
2380 task_rq_unlock(rq
, &flags
);
2384 * sched_exec - execve() is a valuable balancing opportunity, because at
2385 * this point the task has the smallest effective memory and cache footprint.
2387 void sched_exec(void)
2389 int new_cpu
, this_cpu
= get_cpu();
2390 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2392 if (new_cpu
!= this_cpu
)
2393 sched_migrate_task(current
, new_cpu
);
2397 * pull_task - move a task from a remote runqueue to the local runqueue.
2398 * Both runqueues must be locked.
2400 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2401 struct rq
*this_rq
, int this_cpu
)
2403 deactivate_task(src_rq
, p
, 0);
2404 set_task_cpu(p
, this_cpu
);
2405 activate_task(this_rq
, p
, 0);
2407 * Note that idle threads have a prio of MAX_PRIO, for this test
2408 * to be always true for them.
2410 check_preempt_curr(this_rq
, p
);
2414 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2417 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2418 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2422 * We do not migrate tasks that are:
2423 * 1) running (obviously), or
2424 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2425 * 3) are cache-hot on their current CPU.
2427 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2428 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2433 if (task_running(rq
, p
)) {
2434 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2439 * Aggressive migration if:
2440 * 1) task is cache cold, or
2441 * 2) too many balance attempts have failed.
2444 if (!task_hot(p
, rq
->clock
, sd
) ||
2445 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2446 #ifdef CONFIG_SCHEDSTATS
2447 if (task_hot(p
, rq
->clock
, sd
)) {
2448 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2449 schedstat_inc(p
, se
.nr_forced_migrations
);
2455 if (task_hot(p
, rq
->clock
, sd
)) {
2456 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2462 static unsigned long
2463 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2464 unsigned long max_load_move
, struct sched_domain
*sd
,
2465 enum cpu_idle_type idle
, int *all_pinned
,
2466 int *this_best_prio
, struct rq_iterator
*iterator
)
2468 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2469 struct task_struct
*p
;
2470 long rem_load_move
= max_load_move
;
2472 if (max_load_move
== 0)
2478 * Start the load-balancing iterator:
2480 p
= iterator
->start(iterator
->arg
);
2482 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2485 * To help distribute high priority tasks across CPUs we don't
2486 * skip a task if it will be the highest priority task (i.e. smallest
2487 * prio value) on its new queue regardless of its load weight
2489 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2490 SCHED_LOAD_SCALE_FUZZ
;
2491 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2492 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2493 p
= iterator
->next(iterator
->arg
);
2497 pull_task(busiest
, p
, this_rq
, this_cpu
);
2499 rem_load_move
-= p
->se
.load
.weight
;
2502 * We only want to steal up to the prescribed amount of weighted load.
2504 if (rem_load_move
> 0) {
2505 if (p
->prio
< *this_best_prio
)
2506 *this_best_prio
= p
->prio
;
2507 p
= iterator
->next(iterator
->arg
);
2512 * Right now, this is one of only two places pull_task() is called,
2513 * so we can safely collect pull_task() stats here rather than
2514 * inside pull_task().
2516 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2519 *all_pinned
= pinned
;
2521 return max_load_move
- rem_load_move
;
2525 * move_tasks tries to move up to max_load_move weighted load from busiest to
2526 * this_rq, as part of a balancing operation within domain "sd".
2527 * Returns 1 if successful and 0 otherwise.
2529 * Called with both runqueues locked.
2531 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2532 unsigned long max_load_move
,
2533 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2536 const struct sched_class
*class = sched_class_highest
;
2537 unsigned long total_load_moved
= 0;
2538 int this_best_prio
= this_rq
->curr
->prio
;
2542 class->load_balance(this_rq
, this_cpu
, busiest
,
2543 max_load_move
- total_load_moved
,
2544 sd
, idle
, all_pinned
, &this_best_prio
);
2545 class = class->next
;
2546 } while (class && max_load_move
> total_load_moved
);
2548 return total_load_moved
> 0;
2552 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2553 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2554 struct rq_iterator
*iterator
)
2556 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2560 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2561 pull_task(busiest
, p
, this_rq
, this_cpu
);
2563 * Right now, this is only the second place pull_task()
2564 * is called, so we can safely collect pull_task()
2565 * stats here rather than inside pull_task().
2567 schedstat_inc(sd
, lb_gained
[idle
]);
2571 p
= iterator
->next(iterator
->arg
);
2578 * move_one_task tries to move exactly one task from busiest to this_rq, as
2579 * part of active balancing operations within "domain".
2580 * Returns 1 if successful and 0 otherwise.
2582 * Called with both runqueues locked.
2584 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2585 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2587 const struct sched_class
*class;
2589 for (class = sched_class_highest
; class; class = class->next
)
2590 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2597 * find_busiest_group finds and returns the busiest CPU group within the
2598 * domain. It calculates and returns the amount of weighted load which
2599 * should be moved to restore balance via the imbalance parameter.
2601 static struct sched_group
*
2602 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2603 unsigned long *imbalance
, enum cpu_idle_type idle
,
2604 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2606 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2607 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2608 unsigned long max_pull
;
2609 unsigned long busiest_load_per_task
, busiest_nr_running
;
2610 unsigned long this_load_per_task
, this_nr_running
;
2611 int load_idx
, group_imb
= 0;
2612 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2613 int power_savings_balance
= 1;
2614 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2615 unsigned long min_nr_running
= ULONG_MAX
;
2616 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2619 max_load
= this_load
= total_load
= total_pwr
= 0;
2620 busiest_load_per_task
= busiest_nr_running
= 0;
2621 this_load_per_task
= this_nr_running
= 0;
2622 if (idle
== CPU_NOT_IDLE
)
2623 load_idx
= sd
->busy_idx
;
2624 else if (idle
== CPU_NEWLY_IDLE
)
2625 load_idx
= sd
->newidle_idx
;
2627 load_idx
= sd
->idle_idx
;
2630 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2633 int __group_imb
= 0;
2634 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2635 unsigned long sum_nr_running
, sum_weighted_load
;
2637 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2640 balance_cpu
= first_cpu(group
->cpumask
);
2642 /* Tally up the load of all CPUs in the group */
2643 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2645 min_cpu_load
= ~0UL;
2647 for_each_cpu_mask(i
, group
->cpumask
) {
2650 if (!cpu_isset(i
, *cpus
))
2655 if (*sd_idle
&& rq
->nr_running
)
2658 /* Bias balancing toward cpus of our domain */
2660 if (idle_cpu(i
) && !first_idle_cpu
) {
2665 load
= target_load(i
, load_idx
);
2667 load
= source_load(i
, load_idx
);
2668 if (load
> max_cpu_load
)
2669 max_cpu_load
= load
;
2670 if (min_cpu_load
> load
)
2671 min_cpu_load
= load
;
2675 sum_nr_running
+= rq
->nr_running
;
2676 sum_weighted_load
+= weighted_cpuload(i
);
2680 * First idle cpu or the first cpu(busiest) in this sched group
2681 * is eligible for doing load balancing at this and above
2682 * domains. In the newly idle case, we will allow all the cpu's
2683 * to do the newly idle load balance.
2685 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2686 balance_cpu
!= this_cpu
&& balance
) {
2691 total_load
+= avg_load
;
2692 total_pwr
+= group
->__cpu_power
;
2694 /* Adjust by relative CPU power of the group */
2695 avg_load
= sg_div_cpu_power(group
,
2696 avg_load
* SCHED_LOAD_SCALE
);
2698 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2701 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2704 this_load
= avg_load
;
2706 this_nr_running
= sum_nr_running
;
2707 this_load_per_task
= sum_weighted_load
;
2708 } else if (avg_load
> max_load
&&
2709 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2710 max_load
= avg_load
;
2712 busiest_nr_running
= sum_nr_running
;
2713 busiest_load_per_task
= sum_weighted_load
;
2714 group_imb
= __group_imb
;
2717 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2719 * Busy processors will not participate in power savings
2722 if (idle
== CPU_NOT_IDLE
||
2723 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2727 * If the local group is idle or completely loaded
2728 * no need to do power savings balance at this domain
2730 if (local_group
&& (this_nr_running
>= group_capacity
||
2732 power_savings_balance
= 0;
2735 * If a group is already running at full capacity or idle,
2736 * don't include that group in power savings calculations
2738 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2743 * Calculate the group which has the least non-idle load.
2744 * This is the group from where we need to pick up the load
2747 if ((sum_nr_running
< min_nr_running
) ||
2748 (sum_nr_running
== min_nr_running
&&
2749 first_cpu(group
->cpumask
) <
2750 first_cpu(group_min
->cpumask
))) {
2752 min_nr_running
= sum_nr_running
;
2753 min_load_per_task
= sum_weighted_load
/
2758 * Calculate the group which is almost near its
2759 * capacity but still has some space to pick up some load
2760 * from other group and save more power
2762 if (sum_nr_running
<= group_capacity
- 1) {
2763 if (sum_nr_running
> leader_nr_running
||
2764 (sum_nr_running
== leader_nr_running
&&
2765 first_cpu(group
->cpumask
) >
2766 first_cpu(group_leader
->cpumask
))) {
2767 group_leader
= group
;
2768 leader_nr_running
= sum_nr_running
;
2773 group
= group
->next
;
2774 } while (group
!= sd
->groups
);
2776 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2779 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2781 if (this_load
>= avg_load
||
2782 100*max_load
<= sd
->imbalance_pct
*this_load
)
2785 busiest_load_per_task
/= busiest_nr_running
;
2787 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2790 * We're trying to get all the cpus to the average_load, so we don't
2791 * want to push ourselves above the average load, nor do we wish to
2792 * reduce the max loaded cpu below the average load, as either of these
2793 * actions would just result in more rebalancing later, and ping-pong
2794 * tasks around. Thus we look for the minimum possible imbalance.
2795 * Negative imbalances (*we* are more loaded than anyone else) will
2796 * be counted as no imbalance for these purposes -- we can't fix that
2797 * by pulling tasks to us. Be careful of negative numbers as they'll
2798 * appear as very large values with unsigned longs.
2800 if (max_load
<= busiest_load_per_task
)
2804 * In the presence of smp nice balancing, certain scenarios can have
2805 * max load less than avg load(as we skip the groups at or below
2806 * its cpu_power, while calculating max_load..)
2808 if (max_load
< avg_load
) {
2810 goto small_imbalance
;
2813 /* Don't want to pull so many tasks that a group would go idle */
2814 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2816 /* How much load to actually move to equalise the imbalance */
2817 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2818 (avg_load
- this_load
) * this->__cpu_power
)
2822 * if *imbalance is less than the average load per runnable task
2823 * there is no gaurantee that any tasks will be moved so we'll have
2824 * a think about bumping its value to force at least one task to be
2827 if (*imbalance
< busiest_load_per_task
) {
2828 unsigned long tmp
, pwr_now
, pwr_move
;
2832 pwr_move
= pwr_now
= 0;
2834 if (this_nr_running
) {
2835 this_load_per_task
/= this_nr_running
;
2836 if (busiest_load_per_task
> this_load_per_task
)
2839 this_load_per_task
= SCHED_LOAD_SCALE
;
2841 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2842 busiest_load_per_task
* imbn
) {
2843 *imbalance
= busiest_load_per_task
;
2848 * OK, we don't have enough imbalance to justify moving tasks,
2849 * however we may be able to increase total CPU power used by
2853 pwr_now
+= busiest
->__cpu_power
*
2854 min(busiest_load_per_task
, max_load
);
2855 pwr_now
+= this->__cpu_power
*
2856 min(this_load_per_task
, this_load
);
2857 pwr_now
/= SCHED_LOAD_SCALE
;
2859 /* Amount of load we'd subtract */
2860 tmp
= sg_div_cpu_power(busiest
,
2861 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2863 pwr_move
+= busiest
->__cpu_power
*
2864 min(busiest_load_per_task
, max_load
- tmp
);
2866 /* Amount of load we'd add */
2867 if (max_load
* busiest
->__cpu_power
<
2868 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2869 tmp
= sg_div_cpu_power(this,
2870 max_load
* busiest
->__cpu_power
);
2872 tmp
= sg_div_cpu_power(this,
2873 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2874 pwr_move
+= this->__cpu_power
*
2875 min(this_load_per_task
, this_load
+ tmp
);
2876 pwr_move
/= SCHED_LOAD_SCALE
;
2878 /* Move if we gain throughput */
2879 if (pwr_move
> pwr_now
)
2880 *imbalance
= busiest_load_per_task
;
2886 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2887 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2890 if (this == group_leader
&& group_leader
!= group_min
) {
2891 *imbalance
= min_load_per_task
;
2901 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2904 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2905 unsigned long imbalance
, cpumask_t
*cpus
)
2907 struct rq
*busiest
= NULL
, *rq
;
2908 unsigned long max_load
= 0;
2911 for_each_cpu_mask(i
, group
->cpumask
) {
2914 if (!cpu_isset(i
, *cpus
))
2918 wl
= weighted_cpuload(i
);
2920 if (rq
->nr_running
== 1 && wl
> imbalance
)
2923 if (wl
> max_load
) {
2933 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2934 * so long as it is large enough.
2936 #define MAX_PINNED_INTERVAL 512
2939 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2940 * tasks if there is an imbalance.
2942 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2943 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2946 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2947 struct sched_group
*group
;
2948 unsigned long imbalance
;
2950 cpumask_t cpus
= CPU_MASK_ALL
;
2951 unsigned long flags
;
2954 * When power savings policy is enabled for the parent domain, idle
2955 * sibling can pick up load irrespective of busy siblings. In this case,
2956 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2957 * portraying it as CPU_NOT_IDLE.
2959 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2960 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2963 schedstat_inc(sd
, lb_count
[idle
]);
2966 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2973 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2977 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2979 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2983 BUG_ON(busiest
== this_rq
);
2985 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2988 if (busiest
->nr_running
> 1) {
2990 * Attempt to move tasks. If find_busiest_group has found
2991 * an imbalance but busiest->nr_running <= 1, the group is
2992 * still unbalanced. ld_moved simply stays zero, so it is
2993 * correctly treated as an imbalance.
2995 local_irq_save(flags
);
2996 double_rq_lock(this_rq
, busiest
);
2997 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2998 imbalance
, sd
, idle
, &all_pinned
);
2999 double_rq_unlock(this_rq
, busiest
);
3000 local_irq_restore(flags
);
3003 * some other cpu did the load balance for us.
3005 if (ld_moved
&& this_cpu
!= smp_processor_id())
3006 resched_cpu(this_cpu
);
3008 /* All tasks on this runqueue were pinned by CPU affinity */
3009 if (unlikely(all_pinned
)) {
3010 cpu_clear(cpu_of(busiest
), cpus
);
3011 if (!cpus_empty(cpus
))
3018 schedstat_inc(sd
, lb_failed
[idle
]);
3019 sd
->nr_balance_failed
++;
3021 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3023 spin_lock_irqsave(&busiest
->lock
, flags
);
3025 /* don't kick the migration_thread, if the curr
3026 * task on busiest cpu can't be moved to this_cpu
3028 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3029 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3031 goto out_one_pinned
;
3034 if (!busiest
->active_balance
) {
3035 busiest
->active_balance
= 1;
3036 busiest
->push_cpu
= this_cpu
;
3039 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3041 wake_up_process(busiest
->migration_thread
);
3044 * We've kicked active balancing, reset the failure
3047 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3050 sd
->nr_balance_failed
= 0;
3052 if (likely(!active_balance
)) {
3053 /* We were unbalanced, so reset the balancing interval */
3054 sd
->balance_interval
= sd
->min_interval
;
3057 * If we've begun active balancing, start to back off. This
3058 * case may not be covered by the all_pinned logic if there
3059 * is only 1 task on the busy runqueue (because we don't call
3062 if (sd
->balance_interval
< sd
->max_interval
)
3063 sd
->balance_interval
*= 2;
3066 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3067 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3072 schedstat_inc(sd
, lb_balanced
[idle
]);
3074 sd
->nr_balance_failed
= 0;
3077 /* tune up the balancing interval */
3078 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3079 (sd
->balance_interval
< sd
->max_interval
))
3080 sd
->balance_interval
*= 2;
3082 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3083 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3089 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3090 * tasks if there is an imbalance.
3092 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3093 * this_rq is locked.
3096 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
3098 struct sched_group
*group
;
3099 struct rq
*busiest
= NULL
;
3100 unsigned long imbalance
;
3104 cpumask_t cpus
= CPU_MASK_ALL
;
3107 * When power savings policy is enabled for the parent domain, idle
3108 * sibling can pick up load irrespective of busy siblings. In this case,
3109 * let the state of idle sibling percolate up as IDLE, instead of
3110 * portraying it as CPU_NOT_IDLE.
3112 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3113 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3116 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3118 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3119 &sd_idle
, &cpus
, NULL
);
3121 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3125 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
3128 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3132 BUG_ON(busiest
== this_rq
);
3134 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3137 if (busiest
->nr_running
> 1) {
3138 /* Attempt to move tasks */
3139 double_lock_balance(this_rq
, busiest
);
3140 /* this_rq->clock is already updated */
3141 update_rq_clock(busiest
);
3142 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3143 imbalance
, sd
, CPU_NEWLY_IDLE
,
3145 spin_unlock(&busiest
->lock
);
3147 if (unlikely(all_pinned
)) {
3148 cpu_clear(cpu_of(busiest
), cpus
);
3149 if (!cpus_empty(cpus
))
3155 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3156 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3157 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3160 sd
->nr_balance_failed
= 0;
3165 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3166 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3167 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3169 sd
->nr_balance_failed
= 0;
3175 * idle_balance is called by schedule() if this_cpu is about to become
3176 * idle. Attempts to pull tasks from other CPUs.
3178 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3180 struct sched_domain
*sd
;
3181 int pulled_task
= -1;
3182 unsigned long next_balance
= jiffies
+ HZ
;
3184 for_each_domain(this_cpu
, sd
) {
3185 unsigned long interval
;
3187 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3190 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3191 /* If we've pulled tasks over stop searching: */
3192 pulled_task
= load_balance_newidle(this_cpu
,
3195 interval
= msecs_to_jiffies(sd
->balance_interval
);
3196 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3197 next_balance
= sd
->last_balance
+ interval
;
3201 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3203 * We are going idle. next_balance may be set based on
3204 * a busy processor. So reset next_balance.
3206 this_rq
->next_balance
= next_balance
;
3211 * active_load_balance is run by migration threads. It pushes running tasks
3212 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3213 * running on each physical CPU where possible, and avoids physical /
3214 * logical imbalances.
3216 * Called with busiest_rq locked.
3218 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3220 int target_cpu
= busiest_rq
->push_cpu
;
3221 struct sched_domain
*sd
;
3222 struct rq
*target_rq
;
3224 /* Is there any task to move? */
3225 if (busiest_rq
->nr_running
<= 1)
3228 target_rq
= cpu_rq(target_cpu
);
3231 * This condition is "impossible", if it occurs
3232 * we need to fix it. Originally reported by
3233 * Bjorn Helgaas on a 128-cpu setup.
3235 BUG_ON(busiest_rq
== target_rq
);
3237 /* move a task from busiest_rq to target_rq */
3238 double_lock_balance(busiest_rq
, target_rq
);
3239 update_rq_clock(busiest_rq
);
3240 update_rq_clock(target_rq
);
3242 /* Search for an sd spanning us and the target CPU. */
3243 for_each_domain(target_cpu
, sd
) {
3244 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3245 cpu_isset(busiest_cpu
, sd
->span
))
3250 schedstat_inc(sd
, alb_count
);
3252 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3254 schedstat_inc(sd
, alb_pushed
);
3256 schedstat_inc(sd
, alb_failed
);
3258 spin_unlock(&target_rq
->lock
);
3263 atomic_t load_balancer
;
3265 } nohz ____cacheline_aligned
= {
3266 .load_balancer
= ATOMIC_INIT(-1),
3267 .cpu_mask
= CPU_MASK_NONE
,
3271 * This routine will try to nominate the ilb (idle load balancing)
3272 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3273 * load balancing on behalf of all those cpus. If all the cpus in the system
3274 * go into this tickless mode, then there will be no ilb owner (as there is
3275 * no need for one) and all the cpus will sleep till the next wakeup event
3278 * For the ilb owner, tick is not stopped. And this tick will be used
3279 * for idle load balancing. ilb owner will still be part of
3282 * While stopping the tick, this cpu will become the ilb owner if there
3283 * is no other owner. And will be the owner till that cpu becomes busy
3284 * or if all cpus in the system stop their ticks at which point
3285 * there is no need for ilb owner.
3287 * When the ilb owner becomes busy, it nominates another owner, during the
3288 * next busy scheduler_tick()
3290 int select_nohz_load_balancer(int stop_tick
)
3292 int cpu
= smp_processor_id();
3295 cpu_set(cpu
, nohz
.cpu_mask
);
3296 cpu_rq(cpu
)->in_nohz_recently
= 1;
3299 * If we are going offline and still the leader, give up!
3301 if (cpu_is_offline(cpu
) &&
3302 atomic_read(&nohz
.load_balancer
) == cpu
) {
3303 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3308 /* time for ilb owner also to sleep */
3309 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3310 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3311 atomic_set(&nohz
.load_balancer
, -1);
3315 if (atomic_read(&nohz
.load_balancer
) == -1) {
3316 /* make me the ilb owner */
3317 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3319 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3322 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3325 cpu_clear(cpu
, nohz
.cpu_mask
);
3327 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3328 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3335 static DEFINE_SPINLOCK(balancing
);
3338 * It checks each scheduling domain to see if it is due to be balanced,
3339 * and initiates a balancing operation if so.
3341 * Balancing parameters are set up in arch_init_sched_domains.
3343 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3346 struct rq
*rq
= cpu_rq(cpu
);
3347 unsigned long interval
;
3348 struct sched_domain
*sd
;
3349 /* Earliest time when we have to do rebalance again */
3350 unsigned long next_balance
= jiffies
+ 60*HZ
;
3351 int update_next_balance
= 0;
3353 for_each_domain(cpu
, sd
) {
3354 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3357 interval
= sd
->balance_interval
;
3358 if (idle
!= CPU_IDLE
)
3359 interval
*= sd
->busy_factor
;
3361 /* scale ms to jiffies */
3362 interval
= msecs_to_jiffies(interval
);
3363 if (unlikely(!interval
))
3365 if (interval
> HZ
*NR_CPUS
/10)
3366 interval
= HZ
*NR_CPUS
/10;
3369 if (sd
->flags
& SD_SERIALIZE
) {
3370 if (!spin_trylock(&balancing
))
3374 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3375 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3377 * We've pulled tasks over so either we're no
3378 * longer idle, or one of our SMT siblings is
3381 idle
= CPU_NOT_IDLE
;
3383 sd
->last_balance
= jiffies
;
3385 if (sd
->flags
& SD_SERIALIZE
)
3386 spin_unlock(&balancing
);
3388 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3389 next_balance
= sd
->last_balance
+ interval
;
3390 update_next_balance
= 1;
3394 * Stop the load balance at this level. There is another
3395 * CPU in our sched group which is doing load balancing more
3403 * next_balance will be updated only when there is a need.
3404 * When the cpu is attached to null domain for ex, it will not be
3407 if (likely(update_next_balance
))
3408 rq
->next_balance
= next_balance
;
3412 * run_rebalance_domains is triggered when needed from the scheduler tick.
3413 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3414 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3416 static void run_rebalance_domains(struct softirq_action
*h
)
3418 int this_cpu
= smp_processor_id();
3419 struct rq
*this_rq
= cpu_rq(this_cpu
);
3420 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3421 CPU_IDLE
: CPU_NOT_IDLE
;
3423 rebalance_domains(this_cpu
, idle
);
3427 * If this cpu is the owner for idle load balancing, then do the
3428 * balancing on behalf of the other idle cpus whose ticks are
3431 if (this_rq
->idle_at_tick
&&
3432 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3433 cpumask_t cpus
= nohz
.cpu_mask
;
3437 cpu_clear(this_cpu
, cpus
);
3438 for_each_cpu_mask(balance_cpu
, cpus
) {
3440 * If this cpu gets work to do, stop the load balancing
3441 * work being done for other cpus. Next load
3442 * balancing owner will pick it up.
3447 rebalance_domains(balance_cpu
, CPU_IDLE
);
3449 rq
= cpu_rq(balance_cpu
);
3450 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3451 this_rq
->next_balance
= rq
->next_balance
;
3458 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3460 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3461 * idle load balancing owner or decide to stop the periodic load balancing,
3462 * if the whole system is idle.
3464 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3468 * If we were in the nohz mode recently and busy at the current
3469 * scheduler tick, then check if we need to nominate new idle
3472 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3473 rq
->in_nohz_recently
= 0;
3475 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3476 cpu_clear(cpu
, nohz
.cpu_mask
);
3477 atomic_set(&nohz
.load_balancer
, -1);
3480 if (atomic_read(&nohz
.load_balancer
) == -1) {
3482 * simple selection for now: Nominate the
3483 * first cpu in the nohz list to be the next
3486 * TBD: Traverse the sched domains and nominate
3487 * the nearest cpu in the nohz.cpu_mask.
3489 int ilb
= first_cpu(nohz
.cpu_mask
);
3497 * If this cpu is idle and doing idle load balancing for all the
3498 * cpus with ticks stopped, is it time for that to stop?
3500 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3501 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3507 * If this cpu is idle and the idle load balancing is done by
3508 * someone else, then no need raise the SCHED_SOFTIRQ
3510 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3511 cpu_isset(cpu
, nohz
.cpu_mask
))
3514 if (time_after_eq(jiffies
, rq
->next_balance
))
3515 raise_softirq(SCHED_SOFTIRQ
);
3518 #else /* CONFIG_SMP */
3521 * on UP we do not need to balance between CPUs:
3523 static inline void idle_balance(int cpu
, struct rq
*rq
)
3529 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3531 EXPORT_PER_CPU_SYMBOL(kstat
);
3534 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3535 * that have not yet been banked in case the task is currently running.
3537 unsigned long long task_sched_runtime(struct task_struct
*p
)
3539 unsigned long flags
;
3543 rq
= task_rq_lock(p
, &flags
);
3544 ns
= p
->se
.sum_exec_runtime
;
3545 if (task_current(rq
, p
)) {
3546 update_rq_clock(rq
);
3547 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3548 if ((s64
)delta_exec
> 0)
3551 task_rq_unlock(rq
, &flags
);
3557 * Account user cpu time to a process.
3558 * @p: the process that the cpu time gets accounted to
3559 * @cputime: the cpu time spent in user space since the last update
3561 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3563 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3566 p
->utime
= cputime_add(p
->utime
, cputime
);
3568 /* Add user time to cpustat. */
3569 tmp
= cputime_to_cputime64(cputime
);
3570 if (TASK_NICE(p
) > 0)
3571 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3573 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3577 * Account guest cpu time to a process.
3578 * @p: the process that the cpu time gets accounted to
3579 * @cputime: the cpu time spent in virtual machine since the last update
3581 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3584 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3586 tmp
= cputime_to_cputime64(cputime
);
3588 p
->utime
= cputime_add(p
->utime
, cputime
);
3589 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3591 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3592 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3596 * Account scaled user cpu time to a process.
3597 * @p: the process that the cpu time gets accounted to
3598 * @cputime: the cpu time spent in user space since the last update
3600 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3602 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3606 * Account system cpu time to a process.
3607 * @p: the process that the cpu time gets accounted to
3608 * @hardirq_offset: the offset to subtract from hardirq_count()
3609 * @cputime: the cpu time spent in kernel space since the last update
3611 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3614 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3615 struct rq
*rq
= this_rq();
3618 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3619 return account_guest_time(p
, cputime
);
3621 p
->stime
= cputime_add(p
->stime
, cputime
);
3623 /* Add system time to cpustat. */
3624 tmp
= cputime_to_cputime64(cputime
);
3625 if (hardirq_count() - hardirq_offset
)
3626 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3627 else if (softirq_count())
3628 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3629 else if (p
!= rq
->idle
)
3630 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3631 else if (atomic_read(&rq
->nr_iowait
) > 0)
3632 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3634 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3635 /* Account for system time used */
3636 acct_update_integrals(p
);
3640 * Account scaled system cpu time to a process.
3641 * @p: the process that the cpu time gets accounted to
3642 * @hardirq_offset: the offset to subtract from hardirq_count()
3643 * @cputime: the cpu time spent in kernel space since the last update
3645 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3647 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3651 * Account for involuntary wait time.
3652 * @p: the process from which the cpu time has been stolen
3653 * @steal: the cpu time spent in involuntary wait
3655 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3657 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3658 cputime64_t tmp
= cputime_to_cputime64(steal
);
3659 struct rq
*rq
= this_rq();
3661 if (p
== rq
->idle
) {
3662 p
->stime
= cputime_add(p
->stime
, steal
);
3663 if (atomic_read(&rq
->nr_iowait
) > 0)
3664 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3666 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3668 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3672 * This function gets called by the timer code, with HZ frequency.
3673 * We call it with interrupts disabled.
3675 * It also gets called by the fork code, when changing the parent's
3678 void scheduler_tick(void)
3680 int cpu
= smp_processor_id();
3681 struct rq
*rq
= cpu_rq(cpu
);
3682 struct task_struct
*curr
= rq
->curr
;
3683 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3685 spin_lock(&rq
->lock
);
3686 __update_rq_clock(rq
);
3688 * Let rq->clock advance by at least TICK_NSEC:
3690 if (unlikely(rq
->clock
< next_tick
))
3691 rq
->clock
= next_tick
;
3692 rq
->tick_timestamp
= rq
->clock
;
3693 update_cpu_load(rq
);
3694 curr
->sched_class
->task_tick(rq
, curr
, 0);
3695 update_sched_rt_period(rq
);
3696 spin_unlock(&rq
->lock
);
3699 rq
->idle_at_tick
= idle_cpu(cpu
);
3700 trigger_load_balance(rq
, cpu
);
3704 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3706 void fastcall
add_preempt_count(int val
)
3711 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3713 preempt_count() += val
;
3715 * Spinlock count overflowing soon?
3717 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3720 EXPORT_SYMBOL(add_preempt_count
);
3722 void fastcall
sub_preempt_count(int val
)
3727 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3730 * Is the spinlock portion underflowing?
3732 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3733 !(preempt_count() & PREEMPT_MASK
)))
3736 preempt_count() -= val
;
3738 EXPORT_SYMBOL(sub_preempt_count
);
3743 * Print scheduling while atomic bug:
3745 static noinline
void __schedule_bug(struct task_struct
*prev
)
3747 struct pt_regs
*regs
= get_irq_regs();
3749 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3750 prev
->comm
, prev
->pid
, preempt_count());
3752 debug_show_held_locks(prev
);
3753 if (irqs_disabled())
3754 print_irqtrace_events(prev
);
3763 * Various schedule()-time debugging checks and statistics:
3765 static inline void schedule_debug(struct task_struct
*prev
)
3768 * Test if we are atomic. Since do_exit() needs to call into
3769 * schedule() atomically, we ignore that path for now.
3770 * Otherwise, whine if we are scheduling when we should not be.
3772 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3773 __schedule_bug(prev
);
3775 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3777 schedstat_inc(this_rq(), sched_count
);
3778 #ifdef CONFIG_SCHEDSTATS
3779 if (unlikely(prev
->lock_depth
>= 0)) {
3780 schedstat_inc(this_rq(), bkl_count
);
3781 schedstat_inc(prev
, sched_info
.bkl_count
);
3787 * Pick up the highest-prio task:
3789 static inline struct task_struct
*
3790 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3792 const struct sched_class
*class;
3793 struct task_struct
*p
;
3796 * Optimization: we know that if all tasks are in
3797 * the fair class we can call that function directly:
3799 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3800 p
= fair_sched_class
.pick_next_task(rq
);
3805 class = sched_class_highest
;
3807 p
= class->pick_next_task(rq
);
3811 * Will never be NULL as the idle class always
3812 * returns a non-NULL p:
3814 class = class->next
;
3819 * schedule() is the main scheduler function.
3821 asmlinkage
void __sched
schedule(void)
3823 struct task_struct
*prev
, *next
;
3830 cpu
= smp_processor_id();
3834 switch_count
= &prev
->nivcsw
;
3836 release_kernel_lock(prev
);
3837 need_resched_nonpreemptible
:
3839 schedule_debug(prev
);
3844 * Do the rq-clock update outside the rq lock:
3846 local_irq_disable();
3847 __update_rq_clock(rq
);
3848 spin_lock(&rq
->lock
);
3849 clear_tsk_need_resched(prev
);
3851 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3852 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3853 unlikely(signal_pending(prev
)))) {
3854 prev
->state
= TASK_RUNNING
;
3856 deactivate_task(rq
, prev
, 1);
3858 switch_count
= &prev
->nvcsw
;
3862 if (prev
->sched_class
->pre_schedule
)
3863 prev
->sched_class
->pre_schedule(rq
, prev
);
3866 if (unlikely(!rq
->nr_running
))
3867 idle_balance(cpu
, rq
);
3869 prev
->sched_class
->put_prev_task(rq
, prev
);
3870 next
= pick_next_task(rq
, prev
);
3872 sched_info_switch(prev
, next
);
3874 if (likely(prev
!= next
)) {
3879 context_switch(rq
, prev
, next
); /* unlocks the rq */
3881 * the context switch might have flipped the stack from under
3882 * us, hence refresh the local variables.
3884 cpu
= smp_processor_id();
3887 spin_unlock_irq(&rq
->lock
);
3891 if (unlikely(reacquire_kernel_lock(current
) < 0))
3892 goto need_resched_nonpreemptible
;
3894 preempt_enable_no_resched();
3895 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3898 EXPORT_SYMBOL(schedule
);
3900 #ifdef CONFIG_PREEMPT
3902 * this is the entry point to schedule() from in-kernel preemption
3903 * off of preempt_enable. Kernel preemptions off return from interrupt
3904 * occur there and call schedule directly.
3906 asmlinkage
void __sched
preempt_schedule(void)
3908 struct thread_info
*ti
= current_thread_info();
3909 #ifdef CONFIG_PREEMPT_BKL
3910 struct task_struct
*task
= current
;
3911 int saved_lock_depth
;
3914 * If there is a non-zero preempt_count or interrupts are disabled,
3915 * we do not want to preempt the current task. Just return..
3917 if (likely(ti
->preempt_count
|| irqs_disabled()))
3921 add_preempt_count(PREEMPT_ACTIVE
);
3924 * We keep the big kernel semaphore locked, but we
3925 * clear ->lock_depth so that schedule() doesnt
3926 * auto-release the semaphore:
3928 #ifdef CONFIG_PREEMPT_BKL
3929 saved_lock_depth
= task
->lock_depth
;
3930 task
->lock_depth
= -1;
3933 #ifdef CONFIG_PREEMPT_BKL
3934 task
->lock_depth
= saved_lock_depth
;
3936 sub_preempt_count(PREEMPT_ACTIVE
);
3939 * Check again in case we missed a preemption opportunity
3940 * between schedule and now.
3943 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3945 EXPORT_SYMBOL(preempt_schedule
);
3948 * this is the entry point to schedule() from kernel preemption
3949 * off of irq context.
3950 * Note, that this is called and return with irqs disabled. This will
3951 * protect us against recursive calling from irq.
3953 asmlinkage
void __sched
preempt_schedule_irq(void)
3955 struct thread_info
*ti
= current_thread_info();
3956 #ifdef CONFIG_PREEMPT_BKL
3957 struct task_struct
*task
= current
;
3958 int saved_lock_depth
;
3960 /* Catch callers which need to be fixed */
3961 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3964 add_preempt_count(PREEMPT_ACTIVE
);
3967 * We keep the big kernel semaphore locked, but we
3968 * clear ->lock_depth so that schedule() doesnt
3969 * auto-release the semaphore:
3971 #ifdef CONFIG_PREEMPT_BKL
3972 saved_lock_depth
= task
->lock_depth
;
3973 task
->lock_depth
= -1;
3977 local_irq_disable();
3978 #ifdef CONFIG_PREEMPT_BKL
3979 task
->lock_depth
= saved_lock_depth
;
3981 sub_preempt_count(PREEMPT_ACTIVE
);
3984 * Check again in case we missed a preemption opportunity
3985 * between schedule and now.
3988 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3991 #endif /* CONFIG_PREEMPT */
3993 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3996 return try_to_wake_up(curr
->private, mode
, sync
);
3998 EXPORT_SYMBOL(default_wake_function
);
4001 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4002 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4003 * number) then we wake all the non-exclusive tasks and one exclusive task.
4005 * There are circumstances in which we can try to wake a task which has already
4006 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4007 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4009 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4010 int nr_exclusive
, int sync
, void *key
)
4012 wait_queue_t
*curr
, *next
;
4014 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4015 unsigned flags
= curr
->flags
;
4017 if (curr
->func(curr
, mode
, sync
, key
) &&
4018 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4024 * __wake_up - wake up threads blocked on a waitqueue.
4026 * @mode: which threads
4027 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4028 * @key: is directly passed to the wakeup function
4030 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4031 int nr_exclusive
, void *key
)
4033 unsigned long flags
;
4035 spin_lock_irqsave(&q
->lock
, flags
);
4036 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4037 spin_unlock_irqrestore(&q
->lock
, flags
);
4039 EXPORT_SYMBOL(__wake_up
);
4042 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4044 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4046 __wake_up_common(q
, mode
, 1, 0, NULL
);
4050 * __wake_up_sync - wake up threads blocked on a waitqueue.
4052 * @mode: which threads
4053 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4055 * The sync wakeup differs that the waker knows that it will schedule
4056 * away soon, so while the target thread will be woken up, it will not
4057 * be migrated to another CPU - ie. the two threads are 'synchronized'
4058 * with each other. This can prevent needless bouncing between CPUs.
4060 * On UP it can prevent extra preemption.
4063 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4065 unsigned long flags
;
4071 if (unlikely(!nr_exclusive
))
4074 spin_lock_irqsave(&q
->lock
, flags
);
4075 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4076 spin_unlock_irqrestore(&q
->lock
, flags
);
4078 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4080 void complete(struct completion
*x
)
4082 unsigned long flags
;
4084 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4086 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
4088 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4090 EXPORT_SYMBOL(complete
);
4092 void complete_all(struct completion
*x
)
4094 unsigned long flags
;
4096 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4097 x
->done
+= UINT_MAX
/2;
4098 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
4100 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4102 EXPORT_SYMBOL(complete_all
);
4104 static inline long __sched
4105 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4108 DECLARE_WAITQUEUE(wait
, current
);
4110 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4111 __add_wait_queue_tail(&x
->wait
, &wait
);
4113 if (state
== TASK_INTERRUPTIBLE
&&
4114 signal_pending(current
)) {
4115 __remove_wait_queue(&x
->wait
, &wait
);
4116 return -ERESTARTSYS
;
4118 __set_current_state(state
);
4119 spin_unlock_irq(&x
->wait
.lock
);
4120 timeout
= schedule_timeout(timeout
);
4121 spin_lock_irq(&x
->wait
.lock
);
4123 __remove_wait_queue(&x
->wait
, &wait
);
4127 __remove_wait_queue(&x
->wait
, &wait
);
4134 wait_for_common(struct completion
*x
, long timeout
, int state
)
4138 spin_lock_irq(&x
->wait
.lock
);
4139 timeout
= do_wait_for_common(x
, timeout
, state
);
4140 spin_unlock_irq(&x
->wait
.lock
);
4144 void __sched
wait_for_completion(struct completion
*x
)
4146 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4148 EXPORT_SYMBOL(wait_for_completion
);
4150 unsigned long __sched
4151 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4153 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4155 EXPORT_SYMBOL(wait_for_completion_timeout
);
4157 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4159 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4160 if (t
== -ERESTARTSYS
)
4164 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4166 unsigned long __sched
4167 wait_for_completion_interruptible_timeout(struct completion
*x
,
4168 unsigned long timeout
)
4170 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4172 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4175 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4177 unsigned long flags
;
4180 init_waitqueue_entry(&wait
, current
);
4182 __set_current_state(state
);
4184 spin_lock_irqsave(&q
->lock
, flags
);
4185 __add_wait_queue(q
, &wait
);
4186 spin_unlock(&q
->lock
);
4187 timeout
= schedule_timeout(timeout
);
4188 spin_lock_irq(&q
->lock
);
4189 __remove_wait_queue(q
, &wait
);
4190 spin_unlock_irqrestore(&q
->lock
, flags
);
4195 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4197 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4199 EXPORT_SYMBOL(interruptible_sleep_on
);
4202 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4204 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4206 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4208 void __sched
sleep_on(wait_queue_head_t
*q
)
4210 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4212 EXPORT_SYMBOL(sleep_on
);
4214 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4216 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4218 EXPORT_SYMBOL(sleep_on_timeout
);
4220 #ifdef CONFIG_RT_MUTEXES
4223 * rt_mutex_setprio - set the current priority of a task
4225 * @prio: prio value (kernel-internal form)
4227 * This function changes the 'effective' priority of a task. It does
4228 * not touch ->normal_prio like __setscheduler().
4230 * Used by the rt_mutex code to implement priority inheritance logic.
4232 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4234 unsigned long flags
;
4235 int oldprio
, on_rq
, running
;
4237 const struct sched_class
*prev_class
= p
->sched_class
;
4239 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4241 rq
= task_rq_lock(p
, &flags
);
4242 update_rq_clock(rq
);
4245 on_rq
= p
->se
.on_rq
;
4246 running
= task_current(rq
, p
);
4248 dequeue_task(rq
, p
, 0);
4250 p
->sched_class
->put_prev_task(rq
, p
);
4254 p
->sched_class
= &rt_sched_class
;
4256 p
->sched_class
= &fair_sched_class
;
4262 p
->sched_class
->set_curr_task(rq
);
4264 enqueue_task(rq
, p
, 0);
4266 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4268 task_rq_unlock(rq
, &flags
);
4273 void set_user_nice(struct task_struct
*p
, long nice
)
4275 int old_prio
, delta
, on_rq
;
4276 unsigned long flags
;
4279 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4282 * We have to be careful, if called from sys_setpriority(),
4283 * the task might be in the middle of scheduling on another CPU.
4285 rq
= task_rq_lock(p
, &flags
);
4286 update_rq_clock(rq
);
4288 * The RT priorities are set via sched_setscheduler(), but we still
4289 * allow the 'normal' nice value to be set - but as expected
4290 * it wont have any effect on scheduling until the task is
4291 * SCHED_FIFO/SCHED_RR:
4293 if (task_has_rt_policy(p
)) {
4294 p
->static_prio
= NICE_TO_PRIO(nice
);
4297 on_rq
= p
->se
.on_rq
;
4299 dequeue_task(rq
, p
, 0);
4301 p
->static_prio
= NICE_TO_PRIO(nice
);
4304 p
->prio
= effective_prio(p
);
4305 delta
= p
->prio
- old_prio
;
4308 enqueue_task(rq
, p
, 0);
4310 * If the task increased its priority or is running and
4311 * lowered its priority, then reschedule its CPU:
4313 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4314 resched_task(rq
->curr
);
4317 task_rq_unlock(rq
, &flags
);
4319 EXPORT_SYMBOL(set_user_nice
);
4322 * can_nice - check if a task can reduce its nice value
4326 int can_nice(const struct task_struct
*p
, const int nice
)
4328 /* convert nice value [19,-20] to rlimit style value [1,40] */
4329 int nice_rlim
= 20 - nice
;
4331 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4332 capable(CAP_SYS_NICE
));
4335 #ifdef __ARCH_WANT_SYS_NICE
4338 * sys_nice - change the priority of the current process.
4339 * @increment: priority increment
4341 * sys_setpriority is a more generic, but much slower function that
4342 * does similar things.
4344 asmlinkage
long sys_nice(int increment
)
4349 * Setpriority might change our priority at the same moment.
4350 * We don't have to worry. Conceptually one call occurs first
4351 * and we have a single winner.
4353 if (increment
< -40)
4358 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4364 if (increment
< 0 && !can_nice(current
, nice
))
4367 retval
= security_task_setnice(current
, nice
);
4371 set_user_nice(current
, nice
);
4378 * task_prio - return the priority value of a given task.
4379 * @p: the task in question.
4381 * This is the priority value as seen by users in /proc.
4382 * RT tasks are offset by -200. Normal tasks are centered
4383 * around 0, value goes from -16 to +15.
4385 int task_prio(const struct task_struct
*p
)
4387 return p
->prio
- MAX_RT_PRIO
;
4391 * task_nice - return the nice value of a given task.
4392 * @p: the task in question.
4394 int task_nice(const struct task_struct
*p
)
4396 return TASK_NICE(p
);
4398 EXPORT_SYMBOL_GPL(task_nice
);
4401 * idle_cpu - is a given cpu idle currently?
4402 * @cpu: the processor in question.
4404 int idle_cpu(int cpu
)
4406 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4410 * idle_task - return the idle task for a given cpu.
4411 * @cpu: the processor in question.
4413 struct task_struct
*idle_task(int cpu
)
4415 return cpu_rq(cpu
)->idle
;
4419 * find_process_by_pid - find a process with a matching PID value.
4420 * @pid: the pid in question.
4422 static struct task_struct
*find_process_by_pid(pid_t pid
)
4424 return pid
? find_task_by_vpid(pid
) : current
;
4427 /* Actually do priority change: must hold rq lock. */
4429 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4431 BUG_ON(p
->se
.on_rq
);
4434 switch (p
->policy
) {
4438 p
->sched_class
= &fair_sched_class
;
4442 p
->sched_class
= &rt_sched_class
;
4446 p
->rt_priority
= prio
;
4447 p
->normal_prio
= normal_prio(p
);
4448 /* we are holding p->pi_lock already */
4449 p
->prio
= rt_mutex_getprio(p
);
4454 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4455 * @p: the task in question.
4456 * @policy: new policy.
4457 * @param: structure containing the new RT priority.
4459 * NOTE that the task may be already dead.
4461 int sched_setscheduler(struct task_struct
*p
, int policy
,
4462 struct sched_param
*param
)
4464 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4465 unsigned long flags
;
4466 const struct sched_class
*prev_class
= p
->sched_class
;
4469 /* may grab non-irq protected spin_locks */
4470 BUG_ON(in_interrupt());
4472 /* double check policy once rq lock held */
4474 policy
= oldpolicy
= p
->policy
;
4475 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4476 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4477 policy
!= SCHED_IDLE
)
4480 * Valid priorities for SCHED_FIFO and SCHED_RR are
4481 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4482 * SCHED_BATCH and SCHED_IDLE is 0.
4484 if (param
->sched_priority
< 0 ||
4485 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4486 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4488 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4492 * Allow unprivileged RT tasks to decrease priority:
4494 if (!capable(CAP_SYS_NICE
)) {
4495 if (rt_policy(policy
)) {
4496 unsigned long rlim_rtprio
;
4498 if (!lock_task_sighand(p
, &flags
))
4500 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4501 unlock_task_sighand(p
, &flags
);
4503 /* can't set/change the rt policy */
4504 if (policy
!= p
->policy
&& !rlim_rtprio
)
4507 /* can't increase priority */
4508 if (param
->sched_priority
> p
->rt_priority
&&
4509 param
->sched_priority
> rlim_rtprio
)
4513 * Like positive nice levels, dont allow tasks to
4514 * move out of SCHED_IDLE either:
4516 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4519 /* can't change other user's priorities */
4520 if ((current
->euid
!= p
->euid
) &&
4521 (current
->euid
!= p
->uid
))
4525 retval
= security_task_setscheduler(p
, policy
, param
);
4529 * make sure no PI-waiters arrive (or leave) while we are
4530 * changing the priority of the task:
4532 spin_lock_irqsave(&p
->pi_lock
, flags
);
4534 * To be able to change p->policy safely, the apropriate
4535 * runqueue lock must be held.
4537 rq
= __task_rq_lock(p
);
4538 /* recheck policy now with rq lock held */
4539 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4540 policy
= oldpolicy
= -1;
4541 __task_rq_unlock(rq
);
4542 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4545 update_rq_clock(rq
);
4546 on_rq
= p
->se
.on_rq
;
4547 running
= task_current(rq
, p
);
4549 deactivate_task(rq
, p
, 0);
4551 p
->sched_class
->put_prev_task(rq
, p
);
4555 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4559 p
->sched_class
->set_curr_task(rq
);
4561 activate_task(rq
, p
, 0);
4563 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4565 __task_rq_unlock(rq
);
4566 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4568 rt_mutex_adjust_pi(p
);
4572 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4575 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4577 struct sched_param lparam
;
4578 struct task_struct
*p
;
4581 if (!param
|| pid
< 0)
4583 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4588 p
= find_process_by_pid(pid
);
4590 retval
= sched_setscheduler(p
, policy
, &lparam
);
4597 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4598 * @pid: the pid in question.
4599 * @policy: new policy.
4600 * @param: structure containing the new RT priority.
4603 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4605 /* negative values for policy are not valid */
4609 return do_sched_setscheduler(pid
, policy
, param
);
4613 * sys_sched_setparam - set/change the RT priority of a thread
4614 * @pid: the pid in question.
4615 * @param: structure containing the new RT priority.
4617 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4619 return do_sched_setscheduler(pid
, -1, param
);
4623 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4624 * @pid: the pid in question.
4626 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4628 struct task_struct
*p
;
4635 read_lock(&tasklist_lock
);
4636 p
= find_process_by_pid(pid
);
4638 retval
= security_task_getscheduler(p
);
4642 read_unlock(&tasklist_lock
);
4647 * sys_sched_getscheduler - get the RT priority of a thread
4648 * @pid: the pid in question.
4649 * @param: structure containing the RT priority.
4651 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4653 struct sched_param lp
;
4654 struct task_struct
*p
;
4657 if (!param
|| pid
< 0)
4660 read_lock(&tasklist_lock
);
4661 p
= find_process_by_pid(pid
);
4666 retval
= security_task_getscheduler(p
);
4670 lp
.sched_priority
= p
->rt_priority
;
4671 read_unlock(&tasklist_lock
);
4674 * This one might sleep, we cannot do it with a spinlock held ...
4676 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4681 read_unlock(&tasklist_lock
);
4685 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4687 cpumask_t cpus_allowed
;
4688 struct task_struct
*p
;
4692 read_lock(&tasklist_lock
);
4694 p
= find_process_by_pid(pid
);
4696 read_unlock(&tasklist_lock
);
4702 * It is not safe to call set_cpus_allowed with the
4703 * tasklist_lock held. We will bump the task_struct's
4704 * usage count and then drop tasklist_lock.
4707 read_unlock(&tasklist_lock
);
4710 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4711 !capable(CAP_SYS_NICE
))
4714 retval
= security_task_setscheduler(p
, 0, NULL
);
4718 cpus_allowed
= cpuset_cpus_allowed(p
);
4719 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4721 retval
= set_cpus_allowed(p
, new_mask
);
4724 cpus_allowed
= cpuset_cpus_allowed(p
);
4725 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4727 * We must have raced with a concurrent cpuset
4728 * update. Just reset the cpus_allowed to the
4729 * cpuset's cpus_allowed
4731 new_mask
= cpus_allowed
;
4741 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4742 cpumask_t
*new_mask
)
4744 if (len
< sizeof(cpumask_t
)) {
4745 memset(new_mask
, 0, sizeof(cpumask_t
));
4746 } else if (len
> sizeof(cpumask_t
)) {
4747 len
= sizeof(cpumask_t
);
4749 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4753 * sys_sched_setaffinity - set the cpu affinity of a process
4754 * @pid: pid of the process
4755 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4756 * @user_mask_ptr: user-space pointer to the new cpu mask
4758 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4759 unsigned long __user
*user_mask_ptr
)
4764 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4768 return sched_setaffinity(pid
, new_mask
);
4772 * Represents all cpu's present in the system
4773 * In systems capable of hotplug, this map could dynamically grow
4774 * as new cpu's are detected in the system via any platform specific
4775 * method, such as ACPI for e.g.
4778 cpumask_t cpu_present_map __read_mostly
;
4779 EXPORT_SYMBOL(cpu_present_map
);
4782 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4783 EXPORT_SYMBOL(cpu_online_map
);
4785 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4786 EXPORT_SYMBOL(cpu_possible_map
);
4789 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4791 struct task_struct
*p
;
4795 read_lock(&tasklist_lock
);
4798 p
= find_process_by_pid(pid
);
4802 retval
= security_task_getscheduler(p
);
4806 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4809 read_unlock(&tasklist_lock
);
4816 * sys_sched_getaffinity - get the cpu affinity of a process
4817 * @pid: pid of the process
4818 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4819 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4821 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4822 unsigned long __user
*user_mask_ptr
)
4827 if (len
< sizeof(cpumask_t
))
4830 ret
= sched_getaffinity(pid
, &mask
);
4834 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4837 return sizeof(cpumask_t
);
4841 * sys_sched_yield - yield the current processor to other threads.
4843 * This function yields the current CPU to other tasks. If there are no
4844 * other threads running on this CPU then this function will return.
4846 asmlinkage
long sys_sched_yield(void)
4848 struct rq
*rq
= this_rq_lock();
4850 schedstat_inc(rq
, yld_count
);
4851 current
->sched_class
->yield_task(rq
);
4854 * Since we are going to call schedule() anyway, there's
4855 * no need to preempt or enable interrupts:
4857 __release(rq
->lock
);
4858 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4859 _raw_spin_unlock(&rq
->lock
);
4860 preempt_enable_no_resched();
4867 static void __cond_resched(void)
4869 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4870 __might_sleep(__FILE__
, __LINE__
);
4873 * The BKS might be reacquired before we have dropped
4874 * PREEMPT_ACTIVE, which could trigger a second
4875 * cond_resched() call.
4878 add_preempt_count(PREEMPT_ACTIVE
);
4880 sub_preempt_count(PREEMPT_ACTIVE
);
4881 } while (need_resched());
4884 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
4885 int __sched
_cond_resched(void)
4887 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4888 system_state
== SYSTEM_RUNNING
) {
4894 EXPORT_SYMBOL(_cond_resched
);
4898 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4899 * call schedule, and on return reacquire the lock.
4901 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4902 * operations here to prevent schedule() from being called twice (once via
4903 * spin_unlock(), once by hand).
4905 int cond_resched_lock(spinlock_t
*lock
)
4909 if (need_lockbreak(lock
)) {
4915 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4916 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4917 _raw_spin_unlock(lock
);
4918 preempt_enable_no_resched();
4925 EXPORT_SYMBOL(cond_resched_lock
);
4927 int __sched
cond_resched_softirq(void)
4929 BUG_ON(!in_softirq());
4931 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4939 EXPORT_SYMBOL(cond_resched_softirq
);
4942 * yield - yield the current processor to other threads.
4944 * This is a shortcut for kernel-space yielding - it marks the
4945 * thread runnable and calls sys_sched_yield().
4947 void __sched
yield(void)
4949 set_current_state(TASK_RUNNING
);
4952 EXPORT_SYMBOL(yield
);
4955 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4956 * that process accounting knows that this is a task in IO wait state.
4958 * But don't do that if it is a deliberate, throttling IO wait (this task
4959 * has set its backing_dev_info: the queue against which it should throttle)
4961 void __sched
io_schedule(void)
4963 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4965 delayacct_blkio_start();
4966 atomic_inc(&rq
->nr_iowait
);
4968 atomic_dec(&rq
->nr_iowait
);
4969 delayacct_blkio_end();
4971 EXPORT_SYMBOL(io_schedule
);
4973 long __sched
io_schedule_timeout(long timeout
)
4975 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4978 delayacct_blkio_start();
4979 atomic_inc(&rq
->nr_iowait
);
4980 ret
= schedule_timeout(timeout
);
4981 atomic_dec(&rq
->nr_iowait
);
4982 delayacct_blkio_end();
4987 * sys_sched_get_priority_max - return maximum RT priority.
4988 * @policy: scheduling class.
4990 * this syscall returns the maximum rt_priority that can be used
4991 * by a given scheduling class.
4993 asmlinkage
long sys_sched_get_priority_max(int policy
)
5000 ret
= MAX_USER_RT_PRIO
-1;
5012 * sys_sched_get_priority_min - return minimum RT priority.
5013 * @policy: scheduling class.
5015 * this syscall returns the minimum rt_priority that can be used
5016 * by a given scheduling class.
5018 asmlinkage
long sys_sched_get_priority_min(int policy
)
5036 * sys_sched_rr_get_interval - return the default timeslice of a process.
5037 * @pid: pid of the process.
5038 * @interval: userspace pointer to the timeslice value.
5040 * this syscall writes the default timeslice value of a given process
5041 * into the user-space timespec buffer. A value of '0' means infinity.
5044 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5046 struct task_struct
*p
;
5047 unsigned int time_slice
;
5055 read_lock(&tasklist_lock
);
5056 p
= find_process_by_pid(pid
);
5060 retval
= security_task_getscheduler(p
);
5065 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5066 * tasks that are on an otherwise idle runqueue:
5069 if (p
->policy
== SCHED_RR
) {
5070 time_slice
= DEF_TIMESLICE
;
5072 struct sched_entity
*se
= &p
->se
;
5073 unsigned long flags
;
5076 rq
= task_rq_lock(p
, &flags
);
5077 if (rq
->cfs
.load
.weight
)
5078 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5079 task_rq_unlock(rq
, &flags
);
5081 read_unlock(&tasklist_lock
);
5082 jiffies_to_timespec(time_slice
, &t
);
5083 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5087 read_unlock(&tasklist_lock
);
5091 static const char stat_nam
[] = "RSDTtZX";
5093 void sched_show_task(struct task_struct
*p
)
5095 unsigned long free
= 0;
5098 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5099 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5100 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5101 #if BITS_PER_LONG == 32
5102 if (state
== TASK_RUNNING
)
5103 printk(KERN_CONT
" running ");
5105 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5107 if (state
== TASK_RUNNING
)
5108 printk(KERN_CONT
" running task ");
5110 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5112 #ifdef CONFIG_DEBUG_STACK_USAGE
5114 unsigned long *n
= end_of_stack(p
);
5117 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5120 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5121 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5123 if (state
!= TASK_RUNNING
)
5124 show_stack(p
, NULL
);
5127 void show_state_filter(unsigned long state_filter
)
5129 struct task_struct
*g
, *p
;
5131 #if BITS_PER_LONG == 32
5133 " task PC stack pid father\n");
5136 " task PC stack pid father\n");
5138 read_lock(&tasklist_lock
);
5139 do_each_thread(g
, p
) {
5141 * reset the NMI-timeout, listing all files on a slow
5142 * console might take alot of time:
5144 touch_nmi_watchdog();
5145 if (!state_filter
|| (p
->state
& state_filter
))
5147 } while_each_thread(g
, p
);
5149 touch_all_softlockup_watchdogs();
5151 #ifdef CONFIG_SCHED_DEBUG
5152 sysrq_sched_debug_show();
5154 read_unlock(&tasklist_lock
);
5156 * Only show locks if all tasks are dumped:
5158 if (state_filter
== -1)
5159 debug_show_all_locks();
5162 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5164 idle
->sched_class
= &idle_sched_class
;
5168 * init_idle - set up an idle thread for a given CPU
5169 * @idle: task in question
5170 * @cpu: cpu the idle task belongs to
5172 * NOTE: this function does not set the idle thread's NEED_RESCHED
5173 * flag, to make booting more robust.
5175 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5177 struct rq
*rq
= cpu_rq(cpu
);
5178 unsigned long flags
;
5181 idle
->se
.exec_start
= sched_clock();
5183 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5184 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5185 __set_task_cpu(idle
, cpu
);
5187 spin_lock_irqsave(&rq
->lock
, flags
);
5188 rq
->curr
= rq
->idle
= idle
;
5189 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5192 spin_unlock_irqrestore(&rq
->lock
, flags
);
5194 /* Set the preempt count _outside_ the spinlocks! */
5195 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5196 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5198 task_thread_info(idle
)->preempt_count
= 0;
5201 * The idle tasks have their own, simple scheduling class:
5203 idle
->sched_class
= &idle_sched_class
;
5207 * In a system that switches off the HZ timer nohz_cpu_mask
5208 * indicates which cpus entered this state. This is used
5209 * in the rcu update to wait only for active cpus. For system
5210 * which do not switch off the HZ timer nohz_cpu_mask should
5211 * always be CPU_MASK_NONE.
5213 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5216 * Increase the granularity value when there are more CPUs,
5217 * because with more CPUs the 'effective latency' as visible
5218 * to users decreases. But the relationship is not linear,
5219 * so pick a second-best guess by going with the log2 of the
5222 * This idea comes from the SD scheduler of Con Kolivas:
5224 static inline void sched_init_granularity(void)
5226 unsigned int factor
= 1 + ilog2(num_online_cpus());
5227 const unsigned long limit
= 200000000;
5229 sysctl_sched_min_granularity
*= factor
;
5230 if (sysctl_sched_min_granularity
> limit
)
5231 sysctl_sched_min_granularity
= limit
;
5233 sysctl_sched_latency
*= factor
;
5234 if (sysctl_sched_latency
> limit
)
5235 sysctl_sched_latency
= limit
;
5237 sysctl_sched_wakeup_granularity
*= factor
;
5238 sysctl_sched_batch_wakeup_granularity
*= factor
;
5243 * This is how migration works:
5245 * 1) we queue a struct migration_req structure in the source CPU's
5246 * runqueue and wake up that CPU's migration thread.
5247 * 2) we down() the locked semaphore => thread blocks.
5248 * 3) migration thread wakes up (implicitly it forces the migrated
5249 * thread off the CPU)
5250 * 4) it gets the migration request and checks whether the migrated
5251 * task is still in the wrong runqueue.
5252 * 5) if it's in the wrong runqueue then the migration thread removes
5253 * it and puts it into the right queue.
5254 * 6) migration thread up()s the semaphore.
5255 * 7) we wake up and the migration is done.
5259 * Change a given task's CPU affinity. Migrate the thread to a
5260 * proper CPU and schedule it away if the CPU it's executing on
5261 * is removed from the allowed bitmask.
5263 * NOTE: the caller must have a valid reference to the task, the
5264 * task must not exit() & deallocate itself prematurely. The
5265 * call is not atomic; no spinlocks may be held.
5267 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5269 struct migration_req req
;
5270 unsigned long flags
;
5274 rq
= task_rq_lock(p
, &flags
);
5275 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5280 if (p
->sched_class
->set_cpus_allowed
)
5281 p
->sched_class
->set_cpus_allowed(p
, &new_mask
);
5283 p
->cpus_allowed
= new_mask
;
5284 p
->nr_cpus_allowed
= cpus_weight(new_mask
);
5287 /* Can the task run on the task's current CPU? If so, we're done */
5288 if (cpu_isset(task_cpu(p
), new_mask
))
5291 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5292 /* Need help from migration thread: drop lock and wait. */
5293 task_rq_unlock(rq
, &flags
);
5294 wake_up_process(rq
->migration_thread
);
5295 wait_for_completion(&req
.done
);
5296 tlb_migrate_finish(p
->mm
);
5300 task_rq_unlock(rq
, &flags
);
5304 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5307 * Move (not current) task off this cpu, onto dest cpu. We're doing
5308 * this because either it can't run here any more (set_cpus_allowed()
5309 * away from this CPU, or CPU going down), or because we're
5310 * attempting to rebalance this task on exec (sched_exec).
5312 * So we race with normal scheduler movements, but that's OK, as long
5313 * as the task is no longer on this CPU.
5315 * Returns non-zero if task was successfully migrated.
5317 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5319 struct rq
*rq_dest
, *rq_src
;
5322 if (unlikely(cpu_is_offline(dest_cpu
)))
5325 rq_src
= cpu_rq(src_cpu
);
5326 rq_dest
= cpu_rq(dest_cpu
);
5328 double_rq_lock(rq_src
, rq_dest
);
5329 /* Already moved. */
5330 if (task_cpu(p
) != src_cpu
)
5332 /* Affinity changed (again). */
5333 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5336 on_rq
= p
->se
.on_rq
;
5338 deactivate_task(rq_src
, p
, 0);
5340 set_task_cpu(p
, dest_cpu
);
5342 activate_task(rq_dest
, p
, 0);
5343 check_preempt_curr(rq_dest
, p
);
5347 double_rq_unlock(rq_src
, rq_dest
);
5352 * migration_thread - this is a highprio system thread that performs
5353 * thread migration by bumping thread off CPU then 'pushing' onto
5356 static int migration_thread(void *data
)
5358 int cpu
= (long)data
;
5362 BUG_ON(rq
->migration_thread
!= current
);
5364 set_current_state(TASK_INTERRUPTIBLE
);
5365 while (!kthread_should_stop()) {
5366 struct migration_req
*req
;
5367 struct list_head
*head
;
5369 spin_lock_irq(&rq
->lock
);
5371 if (cpu_is_offline(cpu
)) {
5372 spin_unlock_irq(&rq
->lock
);
5376 if (rq
->active_balance
) {
5377 active_load_balance(rq
, cpu
);
5378 rq
->active_balance
= 0;
5381 head
= &rq
->migration_queue
;
5383 if (list_empty(head
)) {
5384 spin_unlock_irq(&rq
->lock
);
5386 set_current_state(TASK_INTERRUPTIBLE
);
5389 req
= list_entry(head
->next
, struct migration_req
, list
);
5390 list_del_init(head
->next
);
5392 spin_unlock(&rq
->lock
);
5393 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5396 complete(&req
->done
);
5398 __set_current_state(TASK_RUNNING
);
5402 /* Wait for kthread_stop */
5403 set_current_state(TASK_INTERRUPTIBLE
);
5404 while (!kthread_should_stop()) {
5406 set_current_state(TASK_INTERRUPTIBLE
);
5408 __set_current_state(TASK_RUNNING
);
5412 #ifdef CONFIG_HOTPLUG_CPU
5414 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5418 local_irq_disable();
5419 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5425 * Figure out where task on dead CPU should go, use force if necessary.
5426 * NOTE: interrupts should be disabled by the caller
5428 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5430 unsigned long flags
;
5437 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5438 cpus_and(mask
, mask
, p
->cpus_allowed
);
5439 dest_cpu
= any_online_cpu(mask
);
5441 /* On any allowed CPU? */
5442 if (dest_cpu
== NR_CPUS
)
5443 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5445 /* No more Mr. Nice Guy. */
5446 if (dest_cpu
== NR_CPUS
) {
5447 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5449 * Try to stay on the same cpuset, where the
5450 * current cpuset may be a subset of all cpus.
5451 * The cpuset_cpus_allowed_locked() variant of
5452 * cpuset_cpus_allowed() will not block. It must be
5453 * called within calls to cpuset_lock/cpuset_unlock.
5455 rq
= task_rq_lock(p
, &flags
);
5456 p
->cpus_allowed
= cpus_allowed
;
5457 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5458 task_rq_unlock(rq
, &flags
);
5461 * Don't tell them about moving exiting tasks or
5462 * kernel threads (both mm NULL), since they never
5465 if (p
->mm
&& printk_ratelimit()) {
5466 printk(KERN_INFO
"process %d (%s) no "
5467 "longer affine to cpu%d\n",
5468 task_pid_nr(p
), p
->comm
, dead_cpu
);
5471 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5475 * While a dead CPU has no uninterruptible tasks queued at this point,
5476 * it might still have a nonzero ->nr_uninterruptible counter, because
5477 * for performance reasons the counter is not stricly tracking tasks to
5478 * their home CPUs. So we just add the counter to another CPU's counter,
5479 * to keep the global sum constant after CPU-down:
5481 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5483 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5484 unsigned long flags
;
5486 local_irq_save(flags
);
5487 double_rq_lock(rq_src
, rq_dest
);
5488 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5489 rq_src
->nr_uninterruptible
= 0;
5490 double_rq_unlock(rq_src
, rq_dest
);
5491 local_irq_restore(flags
);
5494 /* Run through task list and migrate tasks from the dead cpu. */
5495 static void migrate_live_tasks(int src_cpu
)
5497 struct task_struct
*p
, *t
;
5499 read_lock(&tasklist_lock
);
5501 do_each_thread(t
, p
) {
5505 if (task_cpu(p
) == src_cpu
)
5506 move_task_off_dead_cpu(src_cpu
, p
);
5507 } while_each_thread(t
, p
);
5509 read_unlock(&tasklist_lock
);
5513 * Schedules idle task to be the next runnable task on current CPU.
5514 * It does so by boosting its priority to highest possible.
5515 * Used by CPU offline code.
5517 void sched_idle_next(void)
5519 int this_cpu
= smp_processor_id();
5520 struct rq
*rq
= cpu_rq(this_cpu
);
5521 struct task_struct
*p
= rq
->idle
;
5522 unsigned long flags
;
5524 /* cpu has to be offline */
5525 BUG_ON(cpu_online(this_cpu
));
5528 * Strictly not necessary since rest of the CPUs are stopped by now
5529 * and interrupts disabled on the current cpu.
5531 spin_lock_irqsave(&rq
->lock
, flags
);
5533 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5535 update_rq_clock(rq
);
5536 activate_task(rq
, p
, 0);
5538 spin_unlock_irqrestore(&rq
->lock
, flags
);
5542 * Ensures that the idle task is using init_mm right before its cpu goes
5545 void idle_task_exit(void)
5547 struct mm_struct
*mm
= current
->active_mm
;
5549 BUG_ON(cpu_online(smp_processor_id()));
5552 switch_mm(mm
, &init_mm
, current
);
5556 /* called under rq->lock with disabled interrupts */
5557 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5559 struct rq
*rq
= cpu_rq(dead_cpu
);
5561 /* Must be exiting, otherwise would be on tasklist. */
5562 BUG_ON(!p
->exit_state
);
5564 /* Cannot have done final schedule yet: would have vanished. */
5565 BUG_ON(p
->state
== TASK_DEAD
);
5570 * Drop lock around migration; if someone else moves it,
5571 * that's OK. No task can be added to this CPU, so iteration is
5574 spin_unlock_irq(&rq
->lock
);
5575 move_task_off_dead_cpu(dead_cpu
, p
);
5576 spin_lock_irq(&rq
->lock
);
5581 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5582 static void migrate_dead_tasks(unsigned int dead_cpu
)
5584 struct rq
*rq
= cpu_rq(dead_cpu
);
5585 struct task_struct
*next
;
5588 if (!rq
->nr_running
)
5590 update_rq_clock(rq
);
5591 next
= pick_next_task(rq
, rq
->curr
);
5594 migrate_dead(dead_cpu
, next
);
5598 #endif /* CONFIG_HOTPLUG_CPU */
5600 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5602 static struct ctl_table sd_ctl_dir
[] = {
5604 .procname
= "sched_domain",
5610 static struct ctl_table sd_ctl_root
[] = {
5612 .ctl_name
= CTL_KERN
,
5613 .procname
= "kernel",
5615 .child
= sd_ctl_dir
,
5620 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5622 struct ctl_table
*entry
=
5623 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5628 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5630 struct ctl_table
*entry
;
5633 * In the intermediate directories, both the child directory and
5634 * procname are dynamically allocated and could fail but the mode
5635 * will always be set. In the lowest directory the names are
5636 * static strings and all have proc handlers.
5638 for (entry
= *tablep
; entry
->mode
; entry
++) {
5640 sd_free_ctl_entry(&entry
->child
);
5641 if (entry
->proc_handler
== NULL
)
5642 kfree(entry
->procname
);
5650 set_table_entry(struct ctl_table
*entry
,
5651 const char *procname
, void *data
, int maxlen
,
5652 mode_t mode
, proc_handler
*proc_handler
)
5654 entry
->procname
= procname
;
5656 entry
->maxlen
= maxlen
;
5658 entry
->proc_handler
= proc_handler
;
5661 static struct ctl_table
*
5662 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5664 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5669 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5670 sizeof(long), 0644, proc_doulongvec_minmax
);
5671 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5672 sizeof(long), 0644, proc_doulongvec_minmax
);
5673 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5674 sizeof(int), 0644, proc_dointvec_minmax
);
5675 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5676 sizeof(int), 0644, proc_dointvec_minmax
);
5677 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5678 sizeof(int), 0644, proc_dointvec_minmax
);
5679 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5680 sizeof(int), 0644, proc_dointvec_minmax
);
5681 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5682 sizeof(int), 0644, proc_dointvec_minmax
);
5683 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5684 sizeof(int), 0644, proc_dointvec_minmax
);
5685 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5686 sizeof(int), 0644, proc_dointvec_minmax
);
5687 set_table_entry(&table
[9], "cache_nice_tries",
5688 &sd
->cache_nice_tries
,
5689 sizeof(int), 0644, proc_dointvec_minmax
);
5690 set_table_entry(&table
[10], "flags", &sd
->flags
,
5691 sizeof(int), 0644, proc_dointvec_minmax
);
5692 /* &table[11] is terminator */
5697 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5699 struct ctl_table
*entry
, *table
;
5700 struct sched_domain
*sd
;
5701 int domain_num
= 0, i
;
5704 for_each_domain(cpu
, sd
)
5706 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5711 for_each_domain(cpu
, sd
) {
5712 snprintf(buf
, 32, "domain%d", i
);
5713 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5715 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5722 static struct ctl_table_header
*sd_sysctl_header
;
5723 static void register_sched_domain_sysctl(void)
5725 int i
, cpu_num
= num_online_cpus();
5726 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5729 WARN_ON(sd_ctl_dir
[0].child
);
5730 sd_ctl_dir
[0].child
= entry
;
5735 for_each_online_cpu(i
) {
5736 snprintf(buf
, 32, "cpu%d", i
);
5737 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5739 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5743 WARN_ON(sd_sysctl_header
);
5744 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5747 /* may be called multiple times per register */
5748 static void unregister_sched_domain_sysctl(void)
5750 if (sd_sysctl_header
)
5751 unregister_sysctl_table(sd_sysctl_header
);
5752 sd_sysctl_header
= NULL
;
5753 if (sd_ctl_dir
[0].child
)
5754 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5757 static void register_sched_domain_sysctl(void)
5760 static void unregister_sched_domain_sysctl(void)
5766 * migration_call - callback that gets triggered when a CPU is added.
5767 * Here we can start up the necessary migration thread for the new CPU.
5769 static int __cpuinit
5770 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5772 struct task_struct
*p
;
5773 int cpu
= (long)hcpu
;
5774 unsigned long flags
;
5779 case CPU_UP_PREPARE
:
5780 case CPU_UP_PREPARE_FROZEN
:
5781 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5784 kthread_bind(p
, cpu
);
5785 /* Must be high prio: stop_machine expects to yield to it. */
5786 rq
= task_rq_lock(p
, &flags
);
5787 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5788 task_rq_unlock(rq
, &flags
);
5789 cpu_rq(cpu
)->migration_thread
= p
;
5793 case CPU_ONLINE_FROZEN
:
5794 /* Strictly unnecessary, as first user will wake it. */
5795 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5797 /* Update our root-domain */
5799 spin_lock_irqsave(&rq
->lock
, flags
);
5801 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5802 cpu_set(cpu
, rq
->rd
->online
);
5804 spin_unlock_irqrestore(&rq
->lock
, flags
);
5807 #ifdef CONFIG_HOTPLUG_CPU
5808 case CPU_UP_CANCELED
:
5809 case CPU_UP_CANCELED_FROZEN
:
5810 if (!cpu_rq(cpu
)->migration_thread
)
5812 /* Unbind it from offline cpu so it can run. Fall thru. */
5813 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5814 any_online_cpu(cpu_online_map
));
5815 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5816 cpu_rq(cpu
)->migration_thread
= NULL
;
5820 case CPU_DEAD_FROZEN
:
5821 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5822 migrate_live_tasks(cpu
);
5824 kthread_stop(rq
->migration_thread
);
5825 rq
->migration_thread
= NULL
;
5826 /* Idle task back to normal (off runqueue, low prio) */
5827 spin_lock_irq(&rq
->lock
);
5828 update_rq_clock(rq
);
5829 deactivate_task(rq
, rq
->idle
, 0);
5830 rq
->idle
->static_prio
= MAX_PRIO
;
5831 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5832 rq
->idle
->sched_class
= &idle_sched_class
;
5833 migrate_dead_tasks(cpu
);
5834 spin_unlock_irq(&rq
->lock
);
5836 migrate_nr_uninterruptible(rq
);
5837 BUG_ON(rq
->nr_running
!= 0);
5840 * No need to migrate the tasks: it was best-effort if
5841 * they didn't take sched_hotcpu_mutex. Just wake up
5844 spin_lock_irq(&rq
->lock
);
5845 while (!list_empty(&rq
->migration_queue
)) {
5846 struct migration_req
*req
;
5848 req
= list_entry(rq
->migration_queue
.next
,
5849 struct migration_req
, list
);
5850 list_del_init(&req
->list
);
5851 complete(&req
->done
);
5853 spin_unlock_irq(&rq
->lock
);
5856 case CPU_DOWN_PREPARE
:
5857 /* Update our root-domain */
5859 spin_lock_irqsave(&rq
->lock
, flags
);
5861 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5862 cpu_clear(cpu
, rq
->rd
->online
);
5864 spin_unlock_irqrestore(&rq
->lock
, flags
);
5871 /* Register at highest priority so that task migration (migrate_all_tasks)
5872 * happens before everything else.
5874 static struct notifier_block __cpuinitdata migration_notifier
= {
5875 .notifier_call
= migration_call
,
5879 void __init
migration_init(void)
5881 void *cpu
= (void *)(long)smp_processor_id();
5884 /* Start one for the boot CPU: */
5885 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5886 BUG_ON(err
== NOTIFY_BAD
);
5887 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5888 register_cpu_notifier(&migration_notifier
);
5894 /* Number of possible processor ids */
5895 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5896 EXPORT_SYMBOL(nr_cpu_ids
);
5898 #ifdef CONFIG_SCHED_DEBUG
5900 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
5902 struct sched_group
*group
= sd
->groups
;
5903 cpumask_t groupmask
;
5906 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5907 cpus_clear(groupmask
);
5909 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5911 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5912 printk("does not load-balance\n");
5914 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5919 printk(KERN_CONT
"span %s\n", str
);
5921 if (!cpu_isset(cpu
, sd
->span
)) {
5922 printk(KERN_ERR
"ERROR: domain->span does not contain "
5925 if (!cpu_isset(cpu
, group
->cpumask
)) {
5926 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5930 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5934 printk(KERN_ERR
"ERROR: group is NULL\n");
5938 if (!group
->__cpu_power
) {
5939 printk(KERN_CONT
"\n");
5940 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5945 if (!cpus_weight(group
->cpumask
)) {
5946 printk(KERN_CONT
"\n");
5947 printk(KERN_ERR
"ERROR: empty group\n");
5951 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5952 printk(KERN_CONT
"\n");
5953 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5957 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5959 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5960 printk(KERN_CONT
" %s", str
);
5962 group
= group
->next
;
5963 } while (group
!= sd
->groups
);
5964 printk(KERN_CONT
"\n");
5966 if (!cpus_equal(sd
->span
, groupmask
))
5967 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5969 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
5970 printk(KERN_ERR
"ERROR: parent span is not a superset "
5971 "of domain->span\n");
5975 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5980 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5984 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5987 if (sched_domain_debug_one(sd
, cpu
, level
))
5996 # define sched_domain_debug(sd, cpu) do { } while (0)
5999 static int sd_degenerate(struct sched_domain
*sd
)
6001 if (cpus_weight(sd
->span
) == 1)
6004 /* Following flags need at least 2 groups */
6005 if (sd
->flags
& (SD_LOAD_BALANCE
|
6006 SD_BALANCE_NEWIDLE
|
6010 SD_SHARE_PKG_RESOURCES
)) {
6011 if (sd
->groups
!= sd
->groups
->next
)
6015 /* Following flags don't use groups */
6016 if (sd
->flags
& (SD_WAKE_IDLE
|
6025 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6027 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6029 if (sd_degenerate(parent
))
6032 if (!cpus_equal(sd
->span
, parent
->span
))
6035 /* Does parent contain flags not in child? */
6036 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6037 if (cflags
& SD_WAKE_AFFINE
)
6038 pflags
&= ~SD_WAKE_BALANCE
;
6039 /* Flags needing groups don't count if only 1 group in parent */
6040 if (parent
->groups
== parent
->groups
->next
) {
6041 pflags
&= ~(SD_LOAD_BALANCE
|
6042 SD_BALANCE_NEWIDLE
|
6046 SD_SHARE_PKG_RESOURCES
);
6048 if (~cflags
& pflags
)
6054 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6056 unsigned long flags
;
6057 const struct sched_class
*class;
6059 spin_lock_irqsave(&rq
->lock
, flags
);
6062 struct root_domain
*old_rd
= rq
->rd
;
6064 for (class = sched_class_highest
; class; class = class->next
) {
6065 if (class->leave_domain
)
6066 class->leave_domain(rq
);
6069 cpu_clear(rq
->cpu
, old_rd
->span
);
6070 cpu_clear(rq
->cpu
, old_rd
->online
);
6072 if (atomic_dec_and_test(&old_rd
->refcount
))
6076 atomic_inc(&rd
->refcount
);
6079 cpu_set(rq
->cpu
, rd
->span
);
6080 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6081 cpu_set(rq
->cpu
, rd
->online
);
6083 for (class = sched_class_highest
; class; class = class->next
) {
6084 if (class->join_domain
)
6085 class->join_domain(rq
);
6088 spin_unlock_irqrestore(&rq
->lock
, flags
);
6091 static void init_rootdomain(struct root_domain
*rd
)
6093 memset(rd
, 0, sizeof(*rd
));
6095 cpus_clear(rd
->span
);
6096 cpus_clear(rd
->online
);
6099 static void init_defrootdomain(void)
6101 init_rootdomain(&def_root_domain
);
6102 atomic_set(&def_root_domain
.refcount
, 1);
6105 static struct root_domain
*alloc_rootdomain(void)
6107 struct root_domain
*rd
;
6109 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6113 init_rootdomain(rd
);
6119 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6120 * hold the hotplug lock.
6123 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6125 struct rq
*rq
= cpu_rq(cpu
);
6126 struct sched_domain
*tmp
;
6128 /* Remove the sched domains which do not contribute to scheduling. */
6129 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6130 struct sched_domain
*parent
= tmp
->parent
;
6133 if (sd_parent_degenerate(tmp
, parent
)) {
6134 tmp
->parent
= parent
->parent
;
6136 parent
->parent
->child
= tmp
;
6140 if (sd
&& sd_degenerate(sd
)) {
6146 sched_domain_debug(sd
, cpu
);
6148 rq_attach_root(rq
, rd
);
6149 rcu_assign_pointer(rq
->sd
, sd
);
6152 /* cpus with isolated domains */
6153 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6155 /* Setup the mask of cpus configured for isolated domains */
6156 static int __init
isolated_cpu_setup(char *str
)
6158 int ints
[NR_CPUS
], i
;
6160 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6161 cpus_clear(cpu_isolated_map
);
6162 for (i
= 1; i
<= ints
[0]; i
++)
6163 if (ints
[i
] < NR_CPUS
)
6164 cpu_set(ints
[i
], cpu_isolated_map
);
6168 __setup("isolcpus=", isolated_cpu_setup
);
6171 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6172 * to a function which identifies what group(along with sched group) a CPU
6173 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6174 * (due to the fact that we keep track of groups covered with a cpumask_t).
6176 * init_sched_build_groups will build a circular linked list of the groups
6177 * covered by the given span, and will set each group's ->cpumask correctly,
6178 * and ->cpu_power to 0.
6181 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
6182 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6183 struct sched_group
**sg
))
6185 struct sched_group
*first
= NULL
, *last
= NULL
;
6186 cpumask_t covered
= CPU_MASK_NONE
;
6189 for_each_cpu_mask(i
, span
) {
6190 struct sched_group
*sg
;
6191 int group
= group_fn(i
, cpu_map
, &sg
);
6194 if (cpu_isset(i
, covered
))
6197 sg
->cpumask
= CPU_MASK_NONE
;
6198 sg
->__cpu_power
= 0;
6200 for_each_cpu_mask(j
, span
) {
6201 if (group_fn(j
, cpu_map
, NULL
) != group
)
6204 cpu_set(j
, covered
);
6205 cpu_set(j
, sg
->cpumask
);
6216 #define SD_NODES_PER_DOMAIN 16
6221 * find_next_best_node - find the next node to include in a sched_domain
6222 * @node: node whose sched_domain we're building
6223 * @used_nodes: nodes already in the sched_domain
6225 * Find the next node to include in a given scheduling domain. Simply
6226 * finds the closest node not already in the @used_nodes map.
6228 * Should use nodemask_t.
6230 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6232 int i
, n
, val
, min_val
, best_node
= 0;
6236 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6237 /* Start at @node */
6238 n
= (node
+ i
) % MAX_NUMNODES
;
6240 if (!nr_cpus_node(n
))
6243 /* Skip already used nodes */
6244 if (test_bit(n
, used_nodes
))
6247 /* Simple min distance search */
6248 val
= node_distance(node
, n
);
6250 if (val
< min_val
) {
6256 set_bit(best_node
, used_nodes
);
6261 * sched_domain_node_span - get a cpumask for a node's sched_domain
6262 * @node: node whose cpumask we're constructing
6263 * @size: number of nodes to include in this span
6265 * Given a node, construct a good cpumask for its sched_domain to span. It
6266 * should be one that prevents unnecessary balancing, but also spreads tasks
6269 static cpumask_t
sched_domain_node_span(int node
)
6271 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6272 cpumask_t span
, nodemask
;
6276 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6278 nodemask
= node_to_cpumask(node
);
6279 cpus_or(span
, span
, nodemask
);
6280 set_bit(node
, used_nodes
);
6282 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6283 int next_node
= find_next_best_node(node
, used_nodes
);
6285 nodemask
= node_to_cpumask(next_node
);
6286 cpus_or(span
, span
, nodemask
);
6293 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6296 * SMT sched-domains:
6298 #ifdef CONFIG_SCHED_SMT
6299 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6300 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6303 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6306 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6312 * multi-core sched-domains:
6314 #ifdef CONFIG_SCHED_MC
6315 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6316 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6319 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6321 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6324 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6325 cpus_and(mask
, mask
, *cpu_map
);
6326 group
= first_cpu(mask
);
6328 *sg
= &per_cpu(sched_group_core
, group
);
6331 #elif defined(CONFIG_SCHED_MC)
6333 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6336 *sg
= &per_cpu(sched_group_core
, cpu
);
6341 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6342 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6345 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6348 #ifdef CONFIG_SCHED_MC
6349 cpumask_t mask
= cpu_coregroup_map(cpu
);
6350 cpus_and(mask
, mask
, *cpu_map
);
6351 group
= first_cpu(mask
);
6352 #elif defined(CONFIG_SCHED_SMT)
6353 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6354 cpus_and(mask
, mask
, *cpu_map
);
6355 group
= first_cpu(mask
);
6360 *sg
= &per_cpu(sched_group_phys
, group
);
6366 * The init_sched_build_groups can't handle what we want to do with node
6367 * groups, so roll our own. Now each node has its own list of groups which
6368 * gets dynamically allocated.
6370 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6371 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6373 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6374 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6376 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6377 struct sched_group
**sg
)
6379 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6382 cpus_and(nodemask
, nodemask
, *cpu_map
);
6383 group
= first_cpu(nodemask
);
6386 *sg
= &per_cpu(sched_group_allnodes
, group
);
6390 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6392 struct sched_group
*sg
= group_head
;
6398 for_each_cpu_mask(j
, sg
->cpumask
) {
6399 struct sched_domain
*sd
;
6401 sd
= &per_cpu(phys_domains
, j
);
6402 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6404 * Only add "power" once for each
6410 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6413 } while (sg
!= group_head
);
6418 /* Free memory allocated for various sched_group structures */
6419 static void free_sched_groups(const cpumask_t
*cpu_map
)
6423 for_each_cpu_mask(cpu
, *cpu_map
) {
6424 struct sched_group
**sched_group_nodes
6425 = sched_group_nodes_bycpu
[cpu
];
6427 if (!sched_group_nodes
)
6430 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6431 cpumask_t nodemask
= node_to_cpumask(i
);
6432 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6434 cpus_and(nodemask
, nodemask
, *cpu_map
);
6435 if (cpus_empty(nodemask
))
6445 if (oldsg
!= sched_group_nodes
[i
])
6448 kfree(sched_group_nodes
);
6449 sched_group_nodes_bycpu
[cpu
] = NULL
;
6453 static void free_sched_groups(const cpumask_t
*cpu_map
)
6459 * Initialize sched groups cpu_power.
6461 * cpu_power indicates the capacity of sched group, which is used while
6462 * distributing the load between different sched groups in a sched domain.
6463 * Typically cpu_power for all the groups in a sched domain will be same unless
6464 * there are asymmetries in the topology. If there are asymmetries, group
6465 * having more cpu_power will pickup more load compared to the group having
6468 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6469 * the maximum number of tasks a group can handle in the presence of other idle
6470 * or lightly loaded groups in the same sched domain.
6472 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6474 struct sched_domain
*child
;
6475 struct sched_group
*group
;
6477 WARN_ON(!sd
|| !sd
->groups
);
6479 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6484 sd
->groups
->__cpu_power
= 0;
6487 * For perf policy, if the groups in child domain share resources
6488 * (for example cores sharing some portions of the cache hierarchy
6489 * or SMT), then set this domain groups cpu_power such that each group
6490 * can handle only one task, when there are other idle groups in the
6491 * same sched domain.
6493 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6495 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6496 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6501 * add cpu_power of each child group to this groups cpu_power
6503 group
= child
->groups
;
6505 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6506 group
= group
->next
;
6507 } while (group
!= child
->groups
);
6511 * Build sched domains for a given set of cpus and attach the sched domains
6512 * to the individual cpus
6514 static int build_sched_domains(const cpumask_t
*cpu_map
)
6517 struct root_domain
*rd
;
6519 struct sched_group
**sched_group_nodes
= NULL
;
6520 int sd_allnodes
= 0;
6523 * Allocate the per-node list of sched groups
6525 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6527 if (!sched_group_nodes
) {
6528 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6531 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6534 rd
= alloc_rootdomain();
6536 printk(KERN_WARNING
"Cannot alloc root domain\n");
6541 * Set up domains for cpus specified by the cpu_map.
6543 for_each_cpu_mask(i
, *cpu_map
) {
6544 struct sched_domain
*sd
= NULL
, *p
;
6545 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6547 cpus_and(nodemask
, nodemask
, *cpu_map
);
6550 if (cpus_weight(*cpu_map
) >
6551 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6552 sd
= &per_cpu(allnodes_domains
, i
);
6553 *sd
= SD_ALLNODES_INIT
;
6554 sd
->span
= *cpu_map
;
6555 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6561 sd
= &per_cpu(node_domains
, i
);
6563 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6567 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6571 sd
= &per_cpu(phys_domains
, i
);
6573 sd
->span
= nodemask
;
6577 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6579 #ifdef CONFIG_SCHED_MC
6581 sd
= &per_cpu(core_domains
, i
);
6583 sd
->span
= cpu_coregroup_map(i
);
6584 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6587 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6590 #ifdef CONFIG_SCHED_SMT
6592 sd
= &per_cpu(cpu_domains
, i
);
6593 *sd
= SD_SIBLING_INIT
;
6594 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6595 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6598 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6602 #ifdef CONFIG_SCHED_SMT
6603 /* Set up CPU (sibling) groups */
6604 for_each_cpu_mask(i
, *cpu_map
) {
6605 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6606 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6607 if (i
!= first_cpu(this_sibling_map
))
6610 init_sched_build_groups(this_sibling_map
, cpu_map
,
6615 #ifdef CONFIG_SCHED_MC
6616 /* Set up multi-core groups */
6617 for_each_cpu_mask(i
, *cpu_map
) {
6618 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6619 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6620 if (i
!= first_cpu(this_core_map
))
6622 init_sched_build_groups(this_core_map
, cpu_map
,
6623 &cpu_to_core_group
);
6627 /* Set up physical groups */
6628 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6629 cpumask_t nodemask
= node_to_cpumask(i
);
6631 cpus_and(nodemask
, nodemask
, *cpu_map
);
6632 if (cpus_empty(nodemask
))
6635 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6639 /* Set up node groups */
6641 init_sched_build_groups(*cpu_map
, cpu_map
,
6642 &cpu_to_allnodes_group
);
6644 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6645 /* Set up node groups */
6646 struct sched_group
*sg
, *prev
;
6647 cpumask_t nodemask
= node_to_cpumask(i
);
6648 cpumask_t domainspan
;
6649 cpumask_t covered
= CPU_MASK_NONE
;
6652 cpus_and(nodemask
, nodemask
, *cpu_map
);
6653 if (cpus_empty(nodemask
)) {
6654 sched_group_nodes
[i
] = NULL
;
6658 domainspan
= sched_domain_node_span(i
);
6659 cpus_and(domainspan
, domainspan
, *cpu_map
);
6661 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6663 printk(KERN_WARNING
"Can not alloc domain group for "
6667 sched_group_nodes
[i
] = sg
;
6668 for_each_cpu_mask(j
, nodemask
) {
6669 struct sched_domain
*sd
;
6671 sd
= &per_cpu(node_domains
, j
);
6674 sg
->__cpu_power
= 0;
6675 sg
->cpumask
= nodemask
;
6677 cpus_or(covered
, covered
, nodemask
);
6680 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6681 cpumask_t tmp
, notcovered
;
6682 int n
= (i
+ j
) % MAX_NUMNODES
;
6684 cpus_complement(notcovered
, covered
);
6685 cpus_and(tmp
, notcovered
, *cpu_map
);
6686 cpus_and(tmp
, tmp
, domainspan
);
6687 if (cpus_empty(tmp
))
6690 nodemask
= node_to_cpumask(n
);
6691 cpus_and(tmp
, tmp
, nodemask
);
6692 if (cpus_empty(tmp
))
6695 sg
= kmalloc_node(sizeof(struct sched_group
),
6699 "Can not alloc domain group for node %d\n", j
);
6702 sg
->__cpu_power
= 0;
6704 sg
->next
= prev
->next
;
6705 cpus_or(covered
, covered
, tmp
);
6712 /* Calculate CPU power for physical packages and nodes */
6713 #ifdef CONFIG_SCHED_SMT
6714 for_each_cpu_mask(i
, *cpu_map
) {
6715 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6717 init_sched_groups_power(i
, sd
);
6720 #ifdef CONFIG_SCHED_MC
6721 for_each_cpu_mask(i
, *cpu_map
) {
6722 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6724 init_sched_groups_power(i
, sd
);
6728 for_each_cpu_mask(i
, *cpu_map
) {
6729 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6731 init_sched_groups_power(i
, sd
);
6735 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6736 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6739 struct sched_group
*sg
;
6741 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6742 init_numa_sched_groups_power(sg
);
6746 /* Attach the domains */
6747 for_each_cpu_mask(i
, *cpu_map
) {
6748 struct sched_domain
*sd
;
6749 #ifdef CONFIG_SCHED_SMT
6750 sd
= &per_cpu(cpu_domains
, i
);
6751 #elif defined(CONFIG_SCHED_MC)
6752 sd
= &per_cpu(core_domains
, i
);
6754 sd
= &per_cpu(phys_domains
, i
);
6756 cpu_attach_domain(sd
, rd
, i
);
6763 free_sched_groups(cpu_map
);
6768 static cpumask_t
*doms_cur
; /* current sched domains */
6769 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6772 * Special case: If a kmalloc of a doms_cur partition (array of
6773 * cpumask_t) fails, then fallback to a single sched domain,
6774 * as determined by the single cpumask_t fallback_doms.
6776 static cpumask_t fallback_doms
;
6779 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6780 * For now this just excludes isolated cpus, but could be used to
6781 * exclude other special cases in the future.
6783 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6788 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6790 doms_cur
= &fallback_doms
;
6791 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6792 err
= build_sched_domains(doms_cur
);
6793 register_sched_domain_sysctl();
6798 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6800 free_sched_groups(cpu_map
);
6804 * Detach sched domains from a group of cpus specified in cpu_map
6805 * These cpus will now be attached to the NULL domain
6807 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6811 unregister_sched_domain_sysctl();
6813 for_each_cpu_mask(i
, *cpu_map
)
6814 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6815 synchronize_sched();
6816 arch_destroy_sched_domains(cpu_map
);
6820 * Partition sched domains as specified by the 'ndoms_new'
6821 * cpumasks in the array doms_new[] of cpumasks. This compares
6822 * doms_new[] to the current sched domain partitioning, doms_cur[].
6823 * It destroys each deleted domain and builds each new domain.
6825 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6826 * The masks don't intersect (don't overlap.) We should setup one
6827 * sched domain for each mask. CPUs not in any of the cpumasks will
6828 * not be load balanced. If the same cpumask appears both in the
6829 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6832 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6833 * ownership of it and will kfree it when done with it. If the caller
6834 * failed the kmalloc call, then it can pass in doms_new == NULL,
6835 * and partition_sched_domains() will fallback to the single partition
6838 * Call with hotplug lock held
6840 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6846 /* always unregister in case we don't destroy any domains */
6847 unregister_sched_domain_sysctl();
6849 if (doms_new
== NULL
) {
6851 doms_new
= &fallback_doms
;
6852 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6855 /* Destroy deleted domains */
6856 for (i
= 0; i
< ndoms_cur
; i
++) {
6857 for (j
= 0; j
< ndoms_new
; j
++) {
6858 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
6861 /* no match - a current sched domain not in new doms_new[] */
6862 detach_destroy_domains(doms_cur
+ i
);
6867 /* Build new domains */
6868 for (i
= 0; i
< ndoms_new
; i
++) {
6869 for (j
= 0; j
< ndoms_cur
; j
++) {
6870 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
6873 /* no match - add a new doms_new */
6874 build_sched_domains(doms_new
+ i
);
6879 /* Remember the new sched domains */
6880 if (doms_cur
!= &fallback_doms
)
6882 doms_cur
= doms_new
;
6883 ndoms_cur
= ndoms_new
;
6885 register_sched_domain_sysctl();
6890 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6891 static int arch_reinit_sched_domains(void)
6896 detach_destroy_domains(&cpu_online_map
);
6897 err
= arch_init_sched_domains(&cpu_online_map
);
6903 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6907 if (buf
[0] != '0' && buf
[0] != '1')
6911 sched_smt_power_savings
= (buf
[0] == '1');
6913 sched_mc_power_savings
= (buf
[0] == '1');
6915 ret
= arch_reinit_sched_domains();
6917 return ret
? ret
: count
;
6920 #ifdef CONFIG_SCHED_MC
6921 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6923 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6925 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6926 const char *buf
, size_t count
)
6928 return sched_power_savings_store(buf
, count
, 0);
6930 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6931 sched_mc_power_savings_store
);
6934 #ifdef CONFIG_SCHED_SMT
6935 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6937 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6939 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6940 const char *buf
, size_t count
)
6942 return sched_power_savings_store(buf
, count
, 1);
6944 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6945 sched_smt_power_savings_store
);
6948 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6952 #ifdef CONFIG_SCHED_SMT
6954 err
= sysfs_create_file(&cls
->kset
.kobj
,
6955 &attr_sched_smt_power_savings
.attr
);
6957 #ifdef CONFIG_SCHED_MC
6958 if (!err
&& mc_capable())
6959 err
= sysfs_create_file(&cls
->kset
.kobj
,
6960 &attr_sched_mc_power_savings
.attr
);
6967 * Force a reinitialization of the sched domains hierarchy. The domains
6968 * and groups cannot be updated in place without racing with the balancing
6969 * code, so we temporarily attach all running cpus to the NULL domain
6970 * which will prevent rebalancing while the sched domains are recalculated.
6972 static int update_sched_domains(struct notifier_block
*nfb
,
6973 unsigned long action
, void *hcpu
)
6976 case CPU_UP_PREPARE
:
6977 case CPU_UP_PREPARE_FROZEN
:
6978 case CPU_DOWN_PREPARE
:
6979 case CPU_DOWN_PREPARE_FROZEN
:
6980 detach_destroy_domains(&cpu_online_map
);
6983 case CPU_UP_CANCELED
:
6984 case CPU_UP_CANCELED_FROZEN
:
6985 case CPU_DOWN_FAILED
:
6986 case CPU_DOWN_FAILED_FROZEN
:
6988 case CPU_ONLINE_FROZEN
:
6990 case CPU_DEAD_FROZEN
:
6992 * Fall through and re-initialise the domains.
6999 /* The hotplug lock is already held by cpu_up/cpu_down */
7000 arch_init_sched_domains(&cpu_online_map
);
7005 void __init
sched_init_smp(void)
7007 cpumask_t non_isolated_cpus
;
7010 arch_init_sched_domains(&cpu_online_map
);
7011 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7012 if (cpus_empty(non_isolated_cpus
))
7013 cpu_set(smp_processor_id(), non_isolated_cpus
);
7015 /* XXX: Theoretical race here - CPU may be hotplugged now */
7016 hotcpu_notifier(update_sched_domains
, 0);
7018 /* Move init over to a non-isolated CPU */
7019 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
7021 sched_init_granularity();
7023 #ifdef CONFIG_FAIR_GROUP_SCHED
7024 if (nr_cpu_ids
== 1)
7027 lb_monitor_task
= kthread_create(load_balance_monitor
, NULL
,
7029 if (!IS_ERR(lb_monitor_task
)) {
7030 lb_monitor_task
->flags
|= PF_NOFREEZE
;
7031 wake_up_process(lb_monitor_task
);
7033 printk(KERN_ERR
"Could not create load balance monitor thread"
7034 "(error = %ld) \n", PTR_ERR(lb_monitor_task
));
7039 void __init
sched_init_smp(void)
7041 sched_init_granularity();
7043 #endif /* CONFIG_SMP */
7045 int in_sched_functions(unsigned long addr
)
7047 return in_lock_functions(addr
) ||
7048 (addr
>= (unsigned long)__sched_text_start
7049 && addr
< (unsigned long)__sched_text_end
);
7052 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7054 cfs_rq
->tasks_timeline
= RB_ROOT
;
7055 #ifdef CONFIG_FAIR_GROUP_SCHED
7058 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7061 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7063 struct rt_prio_array
*array
;
7066 array
= &rt_rq
->active
;
7067 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7068 INIT_LIST_HEAD(array
->queue
+ i
);
7069 __clear_bit(i
, array
->bitmap
);
7071 /* delimiter for bitsearch: */
7072 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7075 rt_rq
->rt_nr_migratory
= 0;
7076 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7077 rt_rq
->overloaded
= 0;
7081 rt_rq
->rt_throttled
= 0;
7084 void __init
sched_init(void)
7086 int highest_cpu
= 0;
7090 init_defrootdomain();
7093 for_each_possible_cpu(i
) {
7097 spin_lock_init(&rq
->lock
);
7098 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7101 init_cfs_rq(&rq
->cfs
, rq
);
7102 #ifdef CONFIG_FAIR_GROUP_SCHED
7103 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7105 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
7106 struct sched_entity
*se
=
7107 &per_cpu(init_sched_entity
, i
);
7109 init_cfs_rq_p
[i
] = cfs_rq
;
7110 init_cfs_rq(cfs_rq
, rq
);
7111 cfs_rq
->tg
= &init_task_group
;
7112 list_add(&cfs_rq
->leaf_cfs_rq_list
,
7113 &rq
->leaf_cfs_rq_list
);
7115 init_sched_entity_p
[i
] = se
;
7116 se
->cfs_rq
= &rq
->cfs
;
7118 se
->load
.weight
= init_task_group_load
;
7119 se
->load
.inv_weight
=
7120 div64_64(1ULL<<32, init_task_group_load
);
7123 init_task_group
.shares
= init_task_group_load
;
7125 init_rt_rq(&rq
->rt
, rq
);
7126 rq
->rt_period_expire
= 0;
7128 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7129 rq
->cpu_load
[j
] = 0;
7133 rq
->active_balance
= 0;
7134 rq
->next_balance
= jiffies
;
7137 rq
->migration_thread
= NULL
;
7138 INIT_LIST_HEAD(&rq
->migration_queue
);
7139 rq_attach_root(rq
, &def_root_domain
);
7142 atomic_set(&rq
->nr_iowait
, 0);
7146 set_load_weight(&init_task
);
7148 #ifdef CONFIG_PREEMPT_NOTIFIERS
7149 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7153 nr_cpu_ids
= highest_cpu
+ 1;
7154 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7157 #ifdef CONFIG_RT_MUTEXES
7158 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7162 * The boot idle thread does lazy MMU switching as well:
7164 atomic_inc(&init_mm
.mm_count
);
7165 enter_lazy_tlb(&init_mm
, current
);
7168 * Make us the idle thread. Technically, schedule() should not be
7169 * called from this thread, however somewhere below it might be,
7170 * but because we are the idle thread, we just pick up running again
7171 * when this runqueue becomes "idle".
7173 init_idle(current
, smp_processor_id());
7175 * During early bootup we pretend to be a normal task:
7177 current
->sched_class
= &fair_sched_class
;
7180 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7181 void __might_sleep(char *file
, int line
)
7184 static unsigned long prev_jiffy
; /* ratelimiting */
7186 if ((in_atomic() || irqs_disabled()) &&
7187 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
7188 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7190 prev_jiffy
= jiffies
;
7191 printk(KERN_ERR
"BUG: sleeping function called from invalid"
7192 " context at %s:%d\n", file
, line
);
7193 printk("in_atomic():%d, irqs_disabled():%d\n",
7194 in_atomic(), irqs_disabled());
7195 debug_show_held_locks(current
);
7196 if (irqs_disabled())
7197 print_irqtrace_events(current
);
7202 EXPORT_SYMBOL(__might_sleep
);
7205 #ifdef CONFIG_MAGIC_SYSRQ
7206 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7209 update_rq_clock(rq
);
7210 on_rq
= p
->se
.on_rq
;
7212 deactivate_task(rq
, p
, 0);
7213 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7215 activate_task(rq
, p
, 0);
7216 resched_task(rq
->curr
);
7220 void normalize_rt_tasks(void)
7222 struct task_struct
*g
, *p
;
7223 unsigned long flags
;
7226 read_lock_irq(&tasklist_lock
);
7227 do_each_thread(g
, p
) {
7229 * Only normalize user tasks:
7234 p
->se
.exec_start
= 0;
7235 #ifdef CONFIG_SCHEDSTATS
7236 p
->se
.wait_start
= 0;
7237 p
->se
.sleep_start
= 0;
7238 p
->se
.block_start
= 0;
7240 task_rq(p
)->clock
= 0;
7244 * Renice negative nice level userspace
7247 if (TASK_NICE(p
) < 0 && p
->mm
)
7248 set_user_nice(p
, 0);
7252 spin_lock_irqsave(&p
->pi_lock
, flags
);
7253 rq
= __task_rq_lock(p
);
7255 normalize_task(rq
, p
);
7257 __task_rq_unlock(rq
);
7258 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
7259 } while_each_thread(g
, p
);
7261 read_unlock_irq(&tasklist_lock
);
7264 #endif /* CONFIG_MAGIC_SYSRQ */
7268 * These functions are only useful for the IA64 MCA handling.
7270 * They can only be called when the whole system has been
7271 * stopped - every CPU needs to be quiescent, and no scheduling
7272 * activity can take place. Using them for anything else would
7273 * be a serious bug, and as a result, they aren't even visible
7274 * under any other configuration.
7278 * curr_task - return the current task for a given cpu.
7279 * @cpu: the processor in question.
7281 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7283 struct task_struct
*curr_task(int cpu
)
7285 return cpu_curr(cpu
);
7289 * set_curr_task - set the current task for a given cpu.
7290 * @cpu: the processor in question.
7291 * @p: the task pointer to set.
7293 * Description: This function must only be used when non-maskable interrupts
7294 * are serviced on a separate stack. It allows the architecture to switch the
7295 * notion of the current task on a cpu in a non-blocking manner. This function
7296 * must be called with all CPU's synchronized, and interrupts disabled, the
7297 * and caller must save the original value of the current task (see
7298 * curr_task() above) and restore that value before reenabling interrupts and
7299 * re-starting the system.
7301 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7303 void set_curr_task(int cpu
, struct task_struct
*p
)
7310 #ifdef CONFIG_FAIR_GROUP_SCHED
7314 * distribute shares of all task groups among their schedulable entities,
7315 * to reflect load distribution across cpus.
7317 static int rebalance_shares(struct sched_domain
*sd
, int this_cpu
)
7319 struct cfs_rq
*cfs_rq
;
7320 struct rq
*rq
= cpu_rq(this_cpu
);
7321 cpumask_t sdspan
= sd
->span
;
7324 /* Walk thr' all the task groups that we have */
7325 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
7327 unsigned long total_load
= 0, total_shares
;
7328 struct task_group
*tg
= cfs_rq
->tg
;
7330 /* Gather total task load of this group across cpus */
7331 for_each_cpu_mask(i
, sdspan
)
7332 total_load
+= tg
->cfs_rq
[i
]->load
.weight
;
7334 /* Nothing to do if this group has no load */
7339 * tg->shares represents the number of cpu shares the task group
7340 * is eligible to hold on a single cpu. On N cpus, it is
7341 * eligible to hold (N * tg->shares) number of cpu shares.
7343 total_shares
= tg
->shares
* cpus_weight(sdspan
);
7346 * redistribute total_shares across cpus as per the task load
7349 for_each_cpu_mask(i
, sdspan
) {
7350 unsigned long local_load
, local_shares
;
7352 local_load
= tg
->cfs_rq
[i
]->load
.weight
;
7353 local_shares
= (local_load
* total_shares
) / total_load
;
7355 local_shares
= MIN_GROUP_SHARES
;
7356 if (local_shares
== tg
->se
[i
]->load
.weight
)
7359 spin_lock_irq(&cpu_rq(i
)->lock
);
7360 set_se_shares(tg
->se
[i
], local_shares
);
7361 spin_unlock_irq(&cpu_rq(i
)->lock
);
7370 * How frequently should we rebalance_shares() across cpus?
7372 * The more frequently we rebalance shares, the more accurate is the fairness
7373 * of cpu bandwidth distribution between task groups. However higher frequency
7374 * also implies increased scheduling overhead.
7376 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7377 * consecutive calls to rebalance_shares() in the same sched domain.
7379 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7380 * consecutive calls to rebalance_shares() in the same sched domain.
7382 * These settings allows for the appropriate trade-off between accuracy of
7383 * fairness and the associated overhead.
7387 /* default: 8ms, units: milliseconds */
7388 const_debug
unsigned int sysctl_sched_min_bal_int_shares
= 8;
7390 /* default: 128ms, units: milliseconds */
7391 const_debug
unsigned int sysctl_sched_max_bal_int_shares
= 128;
7393 /* kernel thread that runs rebalance_shares() periodically */
7394 static int load_balance_monitor(void *unused
)
7396 unsigned int timeout
= sysctl_sched_min_bal_int_shares
;
7397 struct sched_param schedparm
;
7401 * We don't want this thread's execution to be limited by the shares
7402 * assigned to default group (init_task_group). Hence make it run
7403 * as a SCHED_RR RT task at the lowest priority.
7405 schedparm
.sched_priority
= 1;
7406 ret
= sched_setscheduler(current
, SCHED_RR
, &schedparm
);
7408 printk(KERN_ERR
"Couldn't set SCHED_RR policy for load balance"
7409 " monitor thread (error = %d) \n", ret
);
7411 while (!kthread_should_stop()) {
7412 int i
, cpu
, balanced
= 1;
7414 /* Prevent cpus going down or coming up */
7416 /* lockout changes to doms_cur[] array */
7419 * Enter a rcu read-side critical section to safely walk rq->sd
7420 * chain on various cpus and to walk task group list
7421 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7425 for (i
= 0; i
< ndoms_cur
; i
++) {
7426 cpumask_t cpumap
= doms_cur
[i
];
7427 struct sched_domain
*sd
= NULL
, *sd_prev
= NULL
;
7429 cpu
= first_cpu(cpumap
);
7431 /* Find the highest domain at which to balance shares */
7432 for_each_domain(cpu
, sd
) {
7433 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7439 /* sd == NULL? No load balance reqd in this domain */
7443 balanced
&= rebalance_shares(sd
, cpu
);
7452 timeout
= sysctl_sched_min_bal_int_shares
;
7453 else if (timeout
< sysctl_sched_max_bal_int_shares
)
7456 msleep_interruptible(timeout
);
7461 #endif /* CONFIG_SMP */
7463 /* allocate runqueue etc for a new task group */
7464 struct task_group
*sched_create_group(void)
7466 struct task_group
*tg
;
7467 struct cfs_rq
*cfs_rq
;
7468 struct sched_entity
*se
;
7472 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7474 return ERR_PTR(-ENOMEM
);
7476 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
7479 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
7483 for_each_possible_cpu(i
) {
7486 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
7491 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
7496 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
7497 memset(se
, 0, sizeof(struct sched_entity
));
7499 tg
->cfs_rq
[i
] = cfs_rq
;
7500 init_cfs_rq(cfs_rq
, rq
);
7504 se
->cfs_rq
= &rq
->cfs
;
7506 se
->load
.weight
= NICE_0_LOAD
;
7507 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
7511 tg
->shares
= NICE_0_LOAD
;
7513 lock_task_group_list();
7514 for_each_possible_cpu(i
) {
7516 cfs_rq
= tg
->cfs_rq
[i
];
7517 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7519 unlock_task_group_list();
7524 for_each_possible_cpu(i
) {
7526 kfree(tg
->cfs_rq
[i
]);
7534 return ERR_PTR(-ENOMEM
);
7537 /* rcu callback to free various structures associated with a task group */
7538 static void free_sched_group(struct rcu_head
*rhp
)
7540 struct task_group
*tg
= container_of(rhp
, struct task_group
, rcu
);
7541 struct cfs_rq
*cfs_rq
;
7542 struct sched_entity
*se
;
7545 /* now it should be safe to free those cfs_rqs */
7546 for_each_possible_cpu(i
) {
7547 cfs_rq
= tg
->cfs_rq
[i
];
7559 /* Destroy runqueue etc associated with a task group */
7560 void sched_destroy_group(struct task_group
*tg
)
7562 struct cfs_rq
*cfs_rq
= NULL
;
7565 lock_task_group_list();
7566 for_each_possible_cpu(i
) {
7567 cfs_rq
= tg
->cfs_rq
[i
];
7568 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7570 unlock_task_group_list();
7574 /* wait for possible concurrent references to cfs_rqs complete */
7575 call_rcu(&tg
->rcu
, free_sched_group
);
7578 /* change task's runqueue when it moves between groups.
7579 * The caller of this function should have put the task in its new group
7580 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7581 * reflect its new group.
7583 void sched_move_task(struct task_struct
*tsk
)
7586 unsigned long flags
;
7589 rq
= task_rq_lock(tsk
, &flags
);
7591 if (tsk
->sched_class
!= &fair_sched_class
) {
7592 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7596 update_rq_clock(rq
);
7598 running
= task_current(rq
, tsk
);
7599 on_rq
= tsk
->se
.on_rq
;
7602 dequeue_task(rq
, tsk
, 0);
7603 if (unlikely(running
))
7604 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7607 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7610 if (unlikely(running
))
7611 tsk
->sched_class
->set_curr_task(rq
);
7612 enqueue_task(rq
, tsk
, 0);
7616 task_rq_unlock(rq
, &flags
);
7619 /* rq->lock to be locked by caller */
7620 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7622 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7623 struct rq
*rq
= cfs_rq
->rq
;
7627 shares
= MIN_GROUP_SHARES
;
7631 dequeue_entity(cfs_rq
, se
, 0);
7632 dec_cpu_load(rq
, se
->load
.weight
);
7635 se
->load
.weight
= shares
;
7636 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7639 enqueue_entity(cfs_rq
, se
, 0);
7640 inc_cpu_load(rq
, se
->load
.weight
);
7644 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7647 struct cfs_rq
*cfs_rq
;
7650 lock_task_group_list();
7651 if (tg
->shares
== shares
)
7654 if (shares
< MIN_GROUP_SHARES
)
7655 shares
= MIN_GROUP_SHARES
;
7658 * Prevent any load balance activity (rebalance_shares,
7659 * load_balance_fair) from referring to this group first,
7660 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7662 for_each_possible_cpu(i
) {
7663 cfs_rq
= tg
->cfs_rq
[i
];
7664 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7667 /* wait for any ongoing reference to this group to finish */
7668 synchronize_sched();
7671 * Now we are free to modify the group's share on each cpu
7672 * w/o tripping rebalance_share or load_balance_fair.
7674 tg
->shares
= shares
;
7675 for_each_possible_cpu(i
) {
7676 spin_lock_irq(&cpu_rq(i
)->lock
);
7677 set_se_shares(tg
->se
[i
], shares
);
7678 spin_unlock_irq(&cpu_rq(i
)->lock
);
7682 * Enable load balance activity on this group, by inserting it back on
7683 * each cpu's rq->leaf_cfs_rq_list.
7685 for_each_possible_cpu(i
) {
7687 cfs_rq
= tg
->cfs_rq
[i
];
7688 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7691 unlock_task_group_list();
7695 unsigned long sched_group_shares(struct task_group
*tg
)
7700 #endif /* CONFIG_FAIR_GROUP_SCHED */
7702 #ifdef CONFIG_FAIR_CGROUP_SCHED
7704 /* return corresponding task_group object of a cgroup */
7705 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7707 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7708 struct task_group
, css
);
7711 static struct cgroup_subsys_state
*
7712 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7714 struct task_group
*tg
;
7716 if (!cgrp
->parent
) {
7717 /* This is early initialization for the top cgroup */
7718 init_task_group
.css
.cgroup
= cgrp
;
7719 return &init_task_group
.css
;
7722 /* we support only 1-level deep hierarchical scheduler atm */
7723 if (cgrp
->parent
->parent
)
7724 return ERR_PTR(-EINVAL
);
7726 tg
= sched_create_group();
7728 return ERR_PTR(-ENOMEM
);
7730 /* Bind the cgroup to task_group object we just created */
7731 tg
->css
.cgroup
= cgrp
;
7737 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7739 struct task_group
*tg
= cgroup_tg(cgrp
);
7741 sched_destroy_group(tg
);
7745 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7746 struct task_struct
*tsk
)
7748 /* We don't support RT-tasks being in separate groups */
7749 if (tsk
->sched_class
!= &fair_sched_class
)
7756 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7757 struct cgroup
*old_cont
, struct task_struct
*tsk
)
7759 sched_move_task(tsk
);
7762 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7765 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7768 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7770 struct task_group
*tg
= cgroup_tg(cgrp
);
7772 return (u64
) tg
->shares
;
7775 static struct cftype cpu_files
[] = {
7778 .read_uint
= cpu_shares_read_uint
,
7779 .write_uint
= cpu_shares_write_uint
,
7783 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7785 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7788 struct cgroup_subsys cpu_cgroup_subsys
= {
7790 .create
= cpu_cgroup_create
,
7791 .destroy
= cpu_cgroup_destroy
,
7792 .can_attach
= cpu_cgroup_can_attach
,
7793 .attach
= cpu_cgroup_attach
,
7794 .populate
= cpu_cgroup_populate
,
7795 .subsys_id
= cpu_cgroup_subsys_id
,
7799 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7801 #ifdef CONFIG_CGROUP_CPUACCT
7804 * CPU accounting code for task groups.
7806 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7807 * (balbir@in.ibm.com).
7810 /* track cpu usage of a group of tasks */
7812 struct cgroup_subsys_state css
;
7813 /* cpuusage holds pointer to a u64-type object on every cpu */
7817 struct cgroup_subsys cpuacct_subsys
;
7819 /* return cpu accounting group corresponding to this container */
7820 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
7822 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
7823 struct cpuacct
, css
);
7826 /* return cpu accounting group to which this task belongs */
7827 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
7829 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
7830 struct cpuacct
, css
);
7833 /* create a new cpu accounting group */
7834 static struct cgroup_subsys_state
*cpuacct_create(
7835 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7837 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7840 return ERR_PTR(-ENOMEM
);
7842 ca
->cpuusage
= alloc_percpu(u64
);
7843 if (!ca
->cpuusage
) {
7845 return ERR_PTR(-ENOMEM
);
7851 /* destroy an existing cpu accounting group */
7853 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7855 struct cpuacct
*ca
= cgroup_ca(cont
);
7857 free_percpu(ca
->cpuusage
);
7861 /* return total cpu usage (in nanoseconds) of a group */
7862 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
7864 struct cpuacct
*ca
= cgroup_ca(cont
);
7865 u64 totalcpuusage
= 0;
7868 for_each_possible_cpu(i
) {
7869 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
7872 * Take rq->lock to make 64-bit addition safe on 32-bit
7875 spin_lock_irq(&cpu_rq(i
)->lock
);
7876 totalcpuusage
+= *cpuusage
;
7877 spin_unlock_irq(&cpu_rq(i
)->lock
);
7880 return totalcpuusage
;
7883 static struct cftype files
[] = {
7886 .read_uint
= cpuusage_read
,
7890 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7892 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
7896 * charge this task's execution time to its accounting group.
7898 * called with rq->lock held.
7900 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
7904 if (!cpuacct_subsys
.active
)
7909 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
7911 *cpuusage
+= cputime
;
7915 struct cgroup_subsys cpuacct_subsys
= {
7917 .create
= cpuacct_create
,
7918 .destroy
= cpuacct_destroy
,
7919 .populate
= cpuacct_populate
,
7920 .subsys_id
= cpuacct_subsys_id
,
7922 #endif /* CONFIG_CGROUP_CPUACCT */