4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/pid_namespace.h>
48 #include <linux/smp.h>
49 #include <linux/threads.h>
50 #include <linux/timer.h>
51 #include <linux/rcupdate.h>
52 #include <linux/cpu.h>
53 #include <linux/cpuset.h>
54 #include <linux/percpu.h>
55 #include <linux/kthread.h>
56 #include <linux/seq_file.h>
57 #include <linux/sysctl.h>
58 #include <linux/syscalls.h>
59 #include <linux/times.h>
60 #include <linux/tsacct_kern.h>
61 #include <linux/kprobes.h>
62 #include <linux/delayacct.h>
63 #include <linux/reciprocal_div.h>
64 #include <linux/unistd.h>
65 #include <linux/pagemap.h>
68 #include <asm/irq_regs.h>
71 * Scheduler clock - returns current time in nanosec units.
72 * This is default implementation.
73 * Architectures and sub-architectures can override this.
75 unsigned long long __attribute__((weak
)) sched_clock(void)
77 return (unsigned long long)jiffies
* (NSEC_PER_SEC
/ HZ
);
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Some helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / HZ))
104 #define NICE_0_LOAD SCHED_LOAD_SCALE
105 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
108 * These are the 'tuning knobs' of the scheduler:
110 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
111 * Timeslices get refilled after they expire.
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
122 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
131 sg
->__cpu_power
+= val
;
132 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
136 static inline int rt_policy(int policy
)
138 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
143 static inline int task_has_rt_policy(struct task_struct
*p
)
145 return rt_policy(p
->policy
);
149 * This is the priority-queue data structure of the RT scheduling class:
151 struct rt_prio_array
{
152 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
153 struct list_head queue
[MAX_RT_PRIO
];
156 #ifdef CONFIG_FAIR_GROUP_SCHED
158 #include <linux/cgroup.h>
162 /* task group related information */
164 #ifdef CONFIG_FAIR_CGROUP_SCHED
165 struct cgroup_subsys_state css
;
167 /* schedulable entities of this group on each cpu */
168 struct sched_entity
**se
;
169 /* runqueue "owned" by this group on each cpu */
170 struct cfs_rq
**cfs_rq
;
171 unsigned long shares
;
172 /* spinlock to serialize modification to shares */
177 /* Default task group's sched entity on each cpu */
178 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
179 /* Default task group's cfs_rq on each cpu */
180 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
182 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
183 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
185 /* Default task group.
186 * Every task in system belong to this group at bootup.
188 struct task_group init_task_group
= {
189 .se
= init_sched_entity_p
,
190 .cfs_rq
= init_cfs_rq_p
,
193 #ifdef CONFIG_FAIR_USER_SCHED
194 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
196 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
199 static int init_task_group_load
= INIT_TASK_GRP_LOAD
;
201 /* return group to which a task belongs */
202 static inline struct task_group
*task_group(struct task_struct
*p
)
204 struct task_group
*tg
;
206 #ifdef CONFIG_FAIR_USER_SCHED
208 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
209 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
210 struct task_group
, css
);
212 tg
= &init_task_group
;
217 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
218 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
)
220 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
221 p
->se
.parent
= task_group(p
)->se
[cpu
];
226 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
) { }
228 #endif /* CONFIG_FAIR_GROUP_SCHED */
230 /* CFS-related fields in a runqueue */
232 struct load_weight load
;
233 unsigned long nr_running
;
238 struct rb_root tasks_timeline
;
239 struct rb_node
*rb_leftmost
;
240 struct rb_node
*rb_load_balance_curr
;
241 /* 'curr' points to currently running entity on this cfs_rq.
242 * It is set to NULL otherwise (i.e when none are currently running).
244 struct sched_entity
*curr
;
246 unsigned long nr_spread_over
;
248 #ifdef CONFIG_FAIR_GROUP_SCHED
249 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
252 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
253 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
254 * (like users, containers etc.)
256 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
257 * list is used during load balance.
259 struct list_head leaf_cfs_rq_list
;
260 struct task_group
*tg
; /* group that "owns" this runqueue */
264 /* Real-Time classes' related field in a runqueue: */
266 struct rt_prio_array active
;
267 int rt_load_balance_idx
;
268 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
272 * This is the main, per-CPU runqueue data structure.
274 * Locking rule: those places that want to lock multiple runqueues
275 * (such as the load balancing or the thread migration code), lock
276 * acquire operations must be ordered by ascending &runqueue.
283 * nr_running and cpu_load should be in the same cacheline because
284 * remote CPUs use both these fields when doing load calculation.
286 unsigned long nr_running
;
287 #define CPU_LOAD_IDX_MAX 5
288 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
289 unsigned char idle_at_tick
;
291 unsigned char in_nohz_recently
;
293 /* capture load from *all* tasks on this cpu: */
294 struct load_weight load
;
295 unsigned long nr_load_updates
;
299 #ifdef CONFIG_FAIR_GROUP_SCHED
300 /* list of leaf cfs_rq on this cpu: */
301 struct list_head leaf_cfs_rq_list
;
306 * This is part of a global counter where only the total sum
307 * over all CPUs matters. A task can increase this counter on
308 * one CPU and if it got migrated afterwards it may decrease
309 * it on another CPU. Always updated under the runqueue lock:
311 unsigned long nr_uninterruptible
;
313 struct task_struct
*curr
, *idle
;
314 unsigned long next_balance
;
315 struct mm_struct
*prev_mm
;
317 u64 clock
, prev_clock_raw
;
320 unsigned int clock_warps
, clock_overflows
;
322 unsigned int clock_deep_idle_events
;
328 struct sched_domain
*sd
;
330 /* For active balancing */
333 /* cpu of this runqueue: */
336 struct task_struct
*migration_thread
;
337 struct list_head migration_queue
;
340 #ifdef CONFIG_SCHEDSTATS
342 struct sched_info rq_sched_info
;
344 /* sys_sched_yield() stats */
345 unsigned int yld_exp_empty
;
346 unsigned int yld_act_empty
;
347 unsigned int yld_both_empty
;
348 unsigned int yld_count
;
350 /* schedule() stats */
351 unsigned int sched_switch
;
352 unsigned int sched_count
;
353 unsigned int sched_goidle
;
355 /* try_to_wake_up() stats */
356 unsigned int ttwu_count
;
357 unsigned int ttwu_local
;
360 unsigned int bkl_count
;
362 struct lock_class_key rq_lock_key
;
365 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
366 static DEFINE_MUTEX(sched_hotcpu_mutex
);
368 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
370 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
373 static inline int cpu_of(struct rq
*rq
)
383 * Update the per-runqueue clock, as finegrained as the platform can give
384 * us, but without assuming monotonicity, etc.:
386 static void __update_rq_clock(struct rq
*rq
)
388 u64 prev_raw
= rq
->prev_clock_raw
;
389 u64 now
= sched_clock();
390 s64 delta
= now
- prev_raw
;
391 u64 clock
= rq
->clock
;
393 #ifdef CONFIG_SCHED_DEBUG
394 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
397 * Protect against sched_clock() occasionally going backwards:
399 if (unlikely(delta
< 0)) {
404 * Catch too large forward jumps too:
406 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
407 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
408 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
411 rq
->clock_overflows
++;
413 if (unlikely(delta
> rq
->clock_max_delta
))
414 rq
->clock_max_delta
= delta
;
419 rq
->prev_clock_raw
= now
;
423 static void update_rq_clock(struct rq
*rq
)
425 if (likely(smp_processor_id() == cpu_of(rq
)))
426 __update_rq_clock(rq
);
430 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
431 * See detach_destroy_domains: synchronize_sched for details.
433 * The domain tree of any CPU may only be accessed from within
434 * preempt-disabled sections.
436 #define for_each_domain(cpu, __sd) \
437 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
439 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
440 #define this_rq() (&__get_cpu_var(runqueues))
441 #define task_rq(p) cpu_rq(task_cpu(p))
442 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
445 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
447 #ifdef CONFIG_SCHED_DEBUG
448 # define const_debug __read_mostly
450 # define const_debug static const
454 * Debugging: various feature bits
457 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
458 SCHED_FEAT_WAKEUP_PREEMPT
= 2,
459 SCHED_FEAT_START_DEBIT
= 4,
460 SCHED_FEAT_TREE_AVG
= 8,
461 SCHED_FEAT_APPROX_AVG
= 16,
464 const_debug
unsigned int sysctl_sched_features
=
465 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
466 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
467 SCHED_FEAT_START_DEBIT
* 1 |
468 SCHED_FEAT_TREE_AVG
* 0 |
469 SCHED_FEAT_APPROX_AVG
* 0;
471 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
474 * Number of tasks to iterate in a single balance run.
475 * Limited because this is done with IRQs disabled.
477 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
480 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
481 * clock constructed from sched_clock():
483 unsigned long long cpu_clock(int cpu
)
485 unsigned long long now
;
489 local_irq_save(flags
);
493 local_irq_restore(flags
);
497 EXPORT_SYMBOL_GPL(cpu_clock
);
499 #ifndef prepare_arch_switch
500 # define prepare_arch_switch(next) do { } while (0)
502 #ifndef finish_arch_switch
503 # define finish_arch_switch(prev) do { } while (0)
506 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
507 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
509 return rq
->curr
== p
;
512 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
516 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
518 #ifdef CONFIG_DEBUG_SPINLOCK
519 /* this is a valid case when another task releases the spinlock */
520 rq
->lock
.owner
= current
;
523 * If we are tracking spinlock dependencies then we have to
524 * fix up the runqueue lock - which gets 'carried over' from
527 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
529 spin_unlock_irq(&rq
->lock
);
532 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
533 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
538 return rq
->curr
== p
;
542 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
546 * We can optimise this out completely for !SMP, because the
547 * SMP rebalancing from interrupt is the only thing that cares
552 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
553 spin_unlock_irq(&rq
->lock
);
555 spin_unlock(&rq
->lock
);
559 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
563 * After ->oncpu is cleared, the task can be moved to a different CPU.
564 * We must ensure this doesn't happen until the switch is completely
570 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
574 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
577 * __task_rq_lock - lock the runqueue a given task resides on.
578 * Must be called interrupts disabled.
580 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
584 struct rq
*rq
= task_rq(p
);
585 spin_lock(&rq
->lock
);
586 if (likely(rq
== task_rq(p
)))
588 spin_unlock(&rq
->lock
);
593 * task_rq_lock - lock the runqueue a given task resides on and disable
594 * interrupts. Note the ordering: we can safely lookup the task_rq without
595 * explicitly disabling preemption.
597 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
603 local_irq_save(*flags
);
605 spin_lock(&rq
->lock
);
606 if (likely(rq
== task_rq(p
)))
608 spin_unlock_irqrestore(&rq
->lock
, *flags
);
612 static void __task_rq_unlock(struct rq
*rq
)
615 spin_unlock(&rq
->lock
);
618 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
621 spin_unlock_irqrestore(&rq
->lock
, *flags
);
625 * this_rq_lock - lock this runqueue and disable interrupts.
627 static struct rq
*this_rq_lock(void)
634 spin_lock(&rq
->lock
);
640 * We are going deep-idle (irqs are disabled):
642 void sched_clock_idle_sleep_event(void)
644 struct rq
*rq
= cpu_rq(smp_processor_id());
646 spin_lock(&rq
->lock
);
647 __update_rq_clock(rq
);
648 spin_unlock(&rq
->lock
);
649 rq
->clock_deep_idle_events
++;
651 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
654 * We just idled delta nanoseconds (called with irqs disabled):
656 void sched_clock_idle_wakeup_event(u64 delta_ns
)
658 struct rq
*rq
= cpu_rq(smp_processor_id());
659 u64 now
= sched_clock();
661 rq
->idle_clock
+= delta_ns
;
663 * Override the previous timestamp and ignore all
664 * sched_clock() deltas that occured while we idled,
665 * and use the PM-provided delta_ns to advance the
668 spin_lock(&rq
->lock
);
669 rq
->prev_clock_raw
= now
;
670 rq
->clock
+= delta_ns
;
671 spin_unlock(&rq
->lock
);
673 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
676 * resched_task - mark a task 'to be rescheduled now'.
678 * On UP this means the setting of the need_resched flag, on SMP it
679 * might also involve a cross-CPU call to trigger the scheduler on
684 #ifndef tsk_is_polling
685 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
688 static void resched_task(struct task_struct
*p
)
692 assert_spin_locked(&task_rq(p
)->lock
);
694 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
697 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
700 if (cpu
== smp_processor_id())
703 /* NEED_RESCHED must be visible before we test polling */
705 if (!tsk_is_polling(p
))
706 smp_send_reschedule(cpu
);
709 static void resched_cpu(int cpu
)
711 struct rq
*rq
= cpu_rq(cpu
);
714 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
716 resched_task(cpu_curr(cpu
));
717 spin_unlock_irqrestore(&rq
->lock
, flags
);
720 static inline void resched_task(struct task_struct
*p
)
722 assert_spin_locked(&task_rq(p
)->lock
);
723 set_tsk_need_resched(p
);
727 #if BITS_PER_LONG == 32
728 # define WMULT_CONST (~0UL)
730 # define WMULT_CONST (1UL << 32)
733 #define WMULT_SHIFT 32
736 * Shift right and round:
738 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
741 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
742 struct load_weight
*lw
)
746 if (unlikely(!lw
->inv_weight
))
747 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
749 tmp
= (u64
)delta_exec
* weight
;
751 * Check whether we'd overflow the 64-bit multiplication:
753 if (unlikely(tmp
> WMULT_CONST
))
754 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
757 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
759 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
762 static inline unsigned long
763 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
765 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
768 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
773 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
779 * To aid in avoiding the subversion of "niceness" due to uneven distribution
780 * of tasks with abnormal "nice" values across CPUs the contribution that
781 * each task makes to its run queue's load is weighted according to its
782 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
783 * scaled version of the new time slice allocation that they receive on time
787 #define WEIGHT_IDLEPRIO 2
788 #define WMULT_IDLEPRIO (1 << 31)
791 * Nice levels are multiplicative, with a gentle 10% change for every
792 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
793 * nice 1, it will get ~10% less CPU time than another CPU-bound task
794 * that remained on nice 0.
796 * The "10% effect" is relative and cumulative: from _any_ nice level,
797 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
798 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
799 * If a task goes up by ~10% and another task goes down by ~10% then
800 * the relative distance between them is ~25%.)
802 static const int prio_to_weight
[40] = {
803 /* -20 */ 88761, 71755, 56483, 46273, 36291,
804 /* -15 */ 29154, 23254, 18705, 14949, 11916,
805 /* -10 */ 9548, 7620, 6100, 4904, 3906,
806 /* -5 */ 3121, 2501, 1991, 1586, 1277,
807 /* 0 */ 1024, 820, 655, 526, 423,
808 /* 5 */ 335, 272, 215, 172, 137,
809 /* 10 */ 110, 87, 70, 56, 45,
810 /* 15 */ 36, 29, 23, 18, 15,
814 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
816 * In cases where the weight does not change often, we can use the
817 * precalculated inverse to speed up arithmetics by turning divisions
818 * into multiplications:
820 static const u32 prio_to_wmult
[40] = {
821 /* -20 */ 48388, 59856, 76040, 92818, 118348,
822 /* -15 */ 147320, 184698, 229616, 287308, 360437,
823 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
824 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
825 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
826 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
827 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
828 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
831 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
834 * runqueue iterator, to support SMP load-balancing between different
835 * scheduling classes, without having to expose their internal data
836 * structures to the load-balancing proper:
840 struct task_struct
*(*start
)(void *);
841 struct task_struct
*(*next
)(void *);
846 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
847 unsigned long max_load_move
, struct sched_domain
*sd
,
848 enum cpu_idle_type idle
, int *all_pinned
,
849 int *this_best_prio
, struct rq_iterator
*iterator
);
852 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
853 struct sched_domain
*sd
, enum cpu_idle_type idle
,
854 struct rq_iterator
*iterator
);
857 #ifdef CONFIG_CGROUP_CPUACCT
858 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
860 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
863 #include "sched_stats.h"
864 #include "sched_idletask.c"
865 #include "sched_fair.c"
866 #include "sched_rt.c"
867 #ifdef CONFIG_SCHED_DEBUG
868 # include "sched_debug.c"
871 #define sched_class_highest (&rt_sched_class)
874 * Update delta_exec, delta_fair fields for rq.
876 * delta_fair clock advances at a rate inversely proportional to
877 * total load (rq->load.weight) on the runqueue, while
878 * delta_exec advances at the same rate as wall-clock (provided
881 * delta_exec / delta_fair is a measure of the (smoothened) load on this
882 * runqueue over any given interval. This (smoothened) load is used
883 * during load balance.
885 * This function is called /before/ updating rq->load
886 * and when switching tasks.
888 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
890 update_load_add(&rq
->load
, p
->se
.load
.weight
);
893 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
895 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
898 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
904 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
910 static void set_load_weight(struct task_struct
*p
)
912 if (task_has_rt_policy(p
)) {
913 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
914 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
919 * SCHED_IDLE tasks get minimal weight:
921 if (p
->policy
== SCHED_IDLE
) {
922 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
923 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
927 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
928 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
931 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
933 sched_info_queued(p
);
934 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
938 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
940 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
945 * __normal_prio - return the priority that is based on the static prio
947 static inline int __normal_prio(struct task_struct
*p
)
949 return p
->static_prio
;
953 * Calculate the expected normal priority: i.e. priority
954 * without taking RT-inheritance into account. Might be
955 * boosted by interactivity modifiers. Changes upon fork,
956 * setprio syscalls, and whenever the interactivity
957 * estimator recalculates.
959 static inline int normal_prio(struct task_struct
*p
)
963 if (task_has_rt_policy(p
))
964 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
966 prio
= __normal_prio(p
);
971 * Calculate the current priority, i.e. the priority
972 * taken into account by the scheduler. This value might
973 * be boosted by RT tasks, or might be boosted by
974 * interactivity modifiers. Will be RT if the task got
975 * RT-boosted. If not then it returns p->normal_prio.
977 static int effective_prio(struct task_struct
*p
)
979 p
->normal_prio
= normal_prio(p
);
981 * If we are RT tasks or we were boosted to RT priority,
982 * keep the priority unchanged. Otherwise, update priority
983 * to the normal priority:
985 if (!rt_prio(p
->prio
))
986 return p
->normal_prio
;
991 * activate_task - move a task to the runqueue.
993 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
995 if (task_contributes_to_load(p
))
996 rq
->nr_uninterruptible
--;
998 enqueue_task(rq
, p
, wakeup
);
999 inc_nr_running(p
, rq
);
1003 * deactivate_task - remove a task from the runqueue.
1005 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1007 if (task_contributes_to_load(p
))
1008 rq
->nr_uninterruptible
++;
1010 dequeue_task(rq
, p
, sleep
);
1011 dec_nr_running(p
, rq
);
1015 * task_curr - is this task currently executing on a CPU?
1016 * @p: the task in question.
1018 inline int task_curr(const struct task_struct
*p
)
1020 return cpu_curr(task_cpu(p
)) == p
;
1023 /* Used instead of source_load when we know the type == 0 */
1024 unsigned long weighted_cpuload(const int cpu
)
1026 return cpu_rq(cpu
)->load
.weight
;
1029 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1031 set_task_cfs_rq(p
, cpu
);
1034 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1035 * successfuly executed on another CPU. We must ensure that updates of
1036 * per-task data have been completed by this moment.
1039 task_thread_info(p
)->cpu
= cpu
;
1046 * Is this task likely cache-hot:
1049 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1053 if (p
->sched_class
!= &fair_sched_class
)
1056 if (sysctl_sched_migration_cost
== -1)
1058 if (sysctl_sched_migration_cost
== 0)
1061 delta
= now
- p
->se
.exec_start
;
1063 return delta
< (s64
)sysctl_sched_migration_cost
;
1067 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1069 int old_cpu
= task_cpu(p
);
1070 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1071 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1072 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1075 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1077 #ifdef CONFIG_SCHEDSTATS
1078 if (p
->se
.wait_start
)
1079 p
->se
.wait_start
-= clock_offset
;
1080 if (p
->se
.sleep_start
)
1081 p
->se
.sleep_start
-= clock_offset
;
1082 if (p
->se
.block_start
)
1083 p
->se
.block_start
-= clock_offset
;
1084 if (old_cpu
!= new_cpu
) {
1085 schedstat_inc(p
, se
.nr_migrations
);
1086 if (task_hot(p
, old_rq
->clock
, NULL
))
1087 schedstat_inc(p
, se
.nr_forced2_migrations
);
1090 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1091 new_cfsrq
->min_vruntime
;
1093 __set_task_cpu(p
, new_cpu
);
1096 struct migration_req
{
1097 struct list_head list
;
1099 struct task_struct
*task
;
1102 struct completion done
;
1106 * The task's runqueue lock must be held.
1107 * Returns true if you have to wait for migration thread.
1110 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1112 struct rq
*rq
= task_rq(p
);
1115 * If the task is not on a runqueue (and not running), then
1116 * it is sufficient to simply update the task's cpu field.
1118 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1119 set_task_cpu(p
, dest_cpu
);
1123 init_completion(&req
->done
);
1125 req
->dest_cpu
= dest_cpu
;
1126 list_add(&req
->list
, &rq
->migration_queue
);
1132 * wait_task_inactive - wait for a thread to unschedule.
1134 * The caller must ensure that the task *will* unschedule sometime soon,
1135 * else this function might spin for a *long* time. This function can't
1136 * be called with interrupts off, or it may introduce deadlock with
1137 * smp_call_function() if an IPI is sent by the same process we are
1138 * waiting to become inactive.
1140 void wait_task_inactive(struct task_struct
*p
)
1142 unsigned long flags
;
1148 * We do the initial early heuristics without holding
1149 * any task-queue locks at all. We'll only try to get
1150 * the runqueue lock when things look like they will
1156 * If the task is actively running on another CPU
1157 * still, just relax and busy-wait without holding
1160 * NOTE! Since we don't hold any locks, it's not
1161 * even sure that "rq" stays as the right runqueue!
1162 * But we don't care, since "task_running()" will
1163 * return false if the runqueue has changed and p
1164 * is actually now running somewhere else!
1166 while (task_running(rq
, p
))
1170 * Ok, time to look more closely! We need the rq
1171 * lock now, to be *sure*. If we're wrong, we'll
1172 * just go back and repeat.
1174 rq
= task_rq_lock(p
, &flags
);
1175 running
= task_running(rq
, p
);
1176 on_rq
= p
->se
.on_rq
;
1177 task_rq_unlock(rq
, &flags
);
1180 * Was it really running after all now that we
1181 * checked with the proper locks actually held?
1183 * Oops. Go back and try again..
1185 if (unlikely(running
)) {
1191 * It's not enough that it's not actively running,
1192 * it must be off the runqueue _entirely_, and not
1195 * So if it wa still runnable (but just not actively
1196 * running right now), it's preempted, and we should
1197 * yield - it could be a while.
1199 if (unlikely(on_rq
)) {
1200 schedule_timeout_uninterruptible(1);
1205 * Ahh, all good. It wasn't running, and it wasn't
1206 * runnable, which means that it will never become
1207 * running in the future either. We're all done!
1214 * kick_process - kick a running thread to enter/exit the kernel
1215 * @p: the to-be-kicked thread
1217 * Cause a process which is running on another CPU to enter
1218 * kernel-mode, without any delay. (to get signals handled.)
1220 * NOTE: this function doesnt have to take the runqueue lock,
1221 * because all it wants to ensure is that the remote task enters
1222 * the kernel. If the IPI races and the task has been migrated
1223 * to another CPU then no harm is done and the purpose has been
1226 void kick_process(struct task_struct
*p
)
1232 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1233 smp_send_reschedule(cpu
);
1238 * Return a low guess at the load of a migration-source cpu weighted
1239 * according to the scheduling class and "nice" value.
1241 * We want to under-estimate the load of migration sources, to
1242 * balance conservatively.
1244 static unsigned long source_load(int cpu
, int type
)
1246 struct rq
*rq
= cpu_rq(cpu
);
1247 unsigned long total
= weighted_cpuload(cpu
);
1252 return min(rq
->cpu_load
[type
-1], total
);
1256 * Return a high guess at the load of a migration-target cpu weighted
1257 * according to the scheduling class and "nice" value.
1259 static unsigned long target_load(int cpu
, int type
)
1261 struct rq
*rq
= cpu_rq(cpu
);
1262 unsigned long total
= weighted_cpuload(cpu
);
1267 return max(rq
->cpu_load
[type
-1], total
);
1271 * Return the average load per task on the cpu's run queue
1273 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1275 struct rq
*rq
= cpu_rq(cpu
);
1276 unsigned long total
= weighted_cpuload(cpu
);
1277 unsigned long n
= rq
->nr_running
;
1279 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1283 * find_idlest_group finds and returns the least busy CPU group within the
1286 static struct sched_group
*
1287 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1289 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1290 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1291 int load_idx
= sd
->forkexec_idx
;
1292 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1295 unsigned long load
, avg_load
;
1299 /* Skip over this group if it has no CPUs allowed */
1300 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1303 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1305 /* Tally up the load of all CPUs in the group */
1308 for_each_cpu_mask(i
, group
->cpumask
) {
1309 /* Bias balancing toward cpus of our domain */
1311 load
= source_load(i
, load_idx
);
1313 load
= target_load(i
, load_idx
);
1318 /* Adjust by relative CPU power of the group */
1319 avg_load
= sg_div_cpu_power(group
,
1320 avg_load
* SCHED_LOAD_SCALE
);
1323 this_load
= avg_load
;
1325 } else if (avg_load
< min_load
) {
1326 min_load
= avg_load
;
1329 } while (group
= group
->next
, group
!= sd
->groups
);
1331 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1337 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1340 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1343 unsigned long load
, min_load
= ULONG_MAX
;
1347 /* Traverse only the allowed CPUs */
1348 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1350 for_each_cpu_mask(i
, tmp
) {
1351 load
= weighted_cpuload(i
);
1353 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1363 * sched_balance_self: balance the current task (running on cpu) in domains
1364 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1367 * Balance, ie. select the least loaded group.
1369 * Returns the target CPU number, or the same CPU if no balancing is needed.
1371 * preempt must be disabled.
1373 static int sched_balance_self(int cpu
, int flag
)
1375 struct task_struct
*t
= current
;
1376 struct sched_domain
*tmp
, *sd
= NULL
;
1378 for_each_domain(cpu
, tmp
) {
1380 * If power savings logic is enabled for a domain, stop there.
1382 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1384 if (tmp
->flags
& flag
)
1390 struct sched_group
*group
;
1391 int new_cpu
, weight
;
1393 if (!(sd
->flags
& flag
)) {
1399 group
= find_idlest_group(sd
, t
, cpu
);
1405 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1406 if (new_cpu
== -1 || new_cpu
== cpu
) {
1407 /* Now try balancing at a lower domain level of cpu */
1412 /* Now try balancing at a lower domain level of new_cpu */
1415 weight
= cpus_weight(span
);
1416 for_each_domain(cpu
, tmp
) {
1417 if (weight
<= cpus_weight(tmp
->span
))
1419 if (tmp
->flags
& flag
)
1422 /* while loop will break here if sd == NULL */
1428 #endif /* CONFIG_SMP */
1431 * wake_idle() will wake a task on an idle cpu if task->cpu is
1432 * not idle and an idle cpu is available. The span of cpus to
1433 * search starts with cpus closest then further out as needed,
1434 * so we always favor a closer, idle cpu.
1436 * Returns the CPU we should wake onto.
1438 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1439 static int wake_idle(int cpu
, struct task_struct
*p
)
1442 struct sched_domain
*sd
;
1446 * If it is idle, then it is the best cpu to run this task.
1448 * This cpu is also the best, if it has more than one task already.
1449 * Siblings must be also busy(in most cases) as they didn't already
1450 * pickup the extra load from this cpu and hence we need not check
1451 * sibling runqueue info. This will avoid the checks and cache miss
1452 * penalities associated with that.
1454 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1457 for_each_domain(cpu
, sd
) {
1458 if (sd
->flags
& SD_WAKE_IDLE
) {
1459 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1460 for_each_cpu_mask(i
, tmp
) {
1462 if (i
!= task_cpu(p
)) {
1464 se
.nr_wakeups_idle
);
1476 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1483 * try_to_wake_up - wake up a thread
1484 * @p: the to-be-woken-up thread
1485 * @state: the mask of task states that can be woken
1486 * @sync: do a synchronous wakeup?
1488 * Put it on the run-queue if it's not already there. The "current"
1489 * thread is always on the run-queue (except when the actual
1490 * re-schedule is in progress), and as such you're allowed to do
1491 * the simpler "current->state = TASK_RUNNING" to mark yourself
1492 * runnable without the overhead of this.
1494 * returns failure only if the task is already active.
1496 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1498 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1499 unsigned long flags
;
1503 struct sched_domain
*sd
, *this_sd
= NULL
;
1504 unsigned long load
, this_load
;
1508 rq
= task_rq_lock(p
, &flags
);
1509 old_state
= p
->state
;
1510 if (!(old_state
& state
))
1518 this_cpu
= smp_processor_id();
1521 if (unlikely(task_running(rq
, p
)))
1526 schedstat_inc(rq
, ttwu_count
);
1527 if (cpu
== this_cpu
) {
1528 schedstat_inc(rq
, ttwu_local
);
1532 for_each_domain(this_cpu
, sd
) {
1533 if (cpu_isset(cpu
, sd
->span
)) {
1534 schedstat_inc(sd
, ttwu_wake_remote
);
1540 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1544 * Check for affine wakeup and passive balancing possibilities.
1547 int idx
= this_sd
->wake_idx
;
1548 unsigned int imbalance
;
1550 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1552 load
= source_load(cpu
, idx
);
1553 this_load
= target_load(this_cpu
, idx
);
1555 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1557 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1558 unsigned long tl
= this_load
;
1559 unsigned long tl_per_task
;
1562 * Attract cache-cold tasks on sync wakeups:
1564 if (sync
&& !task_hot(p
, rq
->clock
, this_sd
))
1567 schedstat_inc(p
, se
.nr_wakeups_affine_attempts
);
1568 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1571 * If sync wakeup then subtract the (maximum possible)
1572 * effect of the currently running task from the load
1573 * of the current CPU:
1576 tl
-= current
->se
.load
.weight
;
1579 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1580 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1582 * This domain has SD_WAKE_AFFINE and
1583 * p is cache cold in this domain, and
1584 * there is no bad imbalance.
1586 schedstat_inc(this_sd
, ttwu_move_affine
);
1587 schedstat_inc(p
, se
.nr_wakeups_affine
);
1593 * Start passive balancing when half the imbalance_pct
1596 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1597 if (imbalance
*this_load
<= 100*load
) {
1598 schedstat_inc(this_sd
, ttwu_move_balance
);
1599 schedstat_inc(p
, se
.nr_wakeups_passive
);
1605 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1607 new_cpu
= wake_idle(new_cpu
, p
);
1608 if (new_cpu
!= cpu
) {
1609 set_task_cpu(p
, new_cpu
);
1610 task_rq_unlock(rq
, &flags
);
1611 /* might preempt at this point */
1612 rq
= task_rq_lock(p
, &flags
);
1613 old_state
= p
->state
;
1614 if (!(old_state
& state
))
1619 this_cpu
= smp_processor_id();
1624 #endif /* CONFIG_SMP */
1625 schedstat_inc(p
, se
.nr_wakeups
);
1627 schedstat_inc(p
, se
.nr_wakeups_sync
);
1628 if (orig_cpu
!= cpu
)
1629 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1630 if (cpu
== this_cpu
)
1631 schedstat_inc(p
, se
.nr_wakeups_local
);
1633 schedstat_inc(p
, se
.nr_wakeups_remote
);
1634 update_rq_clock(rq
);
1635 activate_task(rq
, p
, 1);
1636 check_preempt_curr(rq
, p
);
1640 p
->state
= TASK_RUNNING
;
1642 task_rq_unlock(rq
, &flags
);
1647 int fastcall
wake_up_process(struct task_struct
*p
)
1649 return try_to_wake_up(p
, TASK_ALL
, 0);
1651 EXPORT_SYMBOL(wake_up_process
);
1653 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1655 return try_to_wake_up(p
, state
, 0);
1659 * Perform scheduler related setup for a newly forked process p.
1660 * p is forked by current.
1662 * __sched_fork() is basic setup used by init_idle() too:
1664 static void __sched_fork(struct task_struct
*p
)
1666 p
->se
.exec_start
= 0;
1667 p
->se
.sum_exec_runtime
= 0;
1668 p
->se
.prev_sum_exec_runtime
= 0;
1670 #ifdef CONFIG_SCHEDSTATS
1671 p
->se
.wait_start
= 0;
1672 p
->se
.sum_sleep_runtime
= 0;
1673 p
->se
.sleep_start
= 0;
1674 p
->se
.block_start
= 0;
1675 p
->se
.sleep_max
= 0;
1676 p
->se
.block_max
= 0;
1678 p
->se
.slice_max
= 0;
1682 INIT_LIST_HEAD(&p
->run_list
);
1685 #ifdef CONFIG_PREEMPT_NOTIFIERS
1686 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1690 * We mark the process as running here, but have not actually
1691 * inserted it onto the runqueue yet. This guarantees that
1692 * nobody will actually run it, and a signal or other external
1693 * event cannot wake it up and insert it on the runqueue either.
1695 p
->state
= TASK_RUNNING
;
1699 * fork()/clone()-time setup:
1701 void sched_fork(struct task_struct
*p
, int clone_flags
)
1703 int cpu
= get_cpu();
1708 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1710 set_task_cpu(p
, cpu
);
1713 * Make sure we do not leak PI boosting priority to the child:
1715 p
->prio
= current
->normal_prio
;
1716 if (!rt_prio(p
->prio
))
1717 p
->sched_class
= &fair_sched_class
;
1719 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1720 if (likely(sched_info_on()))
1721 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1723 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1726 #ifdef CONFIG_PREEMPT
1727 /* Want to start with kernel preemption disabled. */
1728 task_thread_info(p
)->preempt_count
= 1;
1734 * wake_up_new_task - wake up a newly created task for the first time.
1736 * This function will do some initial scheduler statistics housekeeping
1737 * that must be done for every newly created context, then puts the task
1738 * on the runqueue and wakes it.
1740 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1742 unsigned long flags
;
1745 rq
= task_rq_lock(p
, &flags
);
1746 BUG_ON(p
->state
!= TASK_RUNNING
);
1747 update_rq_clock(rq
);
1749 p
->prio
= effective_prio(p
);
1751 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
1752 activate_task(rq
, p
, 0);
1755 * Let the scheduling class do new task startup
1756 * management (if any):
1758 p
->sched_class
->task_new(rq
, p
);
1759 inc_nr_running(p
, rq
);
1761 check_preempt_curr(rq
, p
);
1762 task_rq_unlock(rq
, &flags
);
1765 #ifdef CONFIG_PREEMPT_NOTIFIERS
1768 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1769 * @notifier: notifier struct to register
1771 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1773 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1775 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1778 * preempt_notifier_unregister - no longer interested in preemption notifications
1779 * @notifier: notifier struct to unregister
1781 * This is safe to call from within a preemption notifier.
1783 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1785 hlist_del(¬ifier
->link
);
1787 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1789 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1791 struct preempt_notifier
*notifier
;
1792 struct hlist_node
*node
;
1794 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1795 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1799 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1800 struct task_struct
*next
)
1802 struct preempt_notifier
*notifier
;
1803 struct hlist_node
*node
;
1805 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1806 notifier
->ops
->sched_out(notifier
, next
);
1811 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1816 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1817 struct task_struct
*next
)
1824 * prepare_task_switch - prepare to switch tasks
1825 * @rq: the runqueue preparing to switch
1826 * @prev: the current task that is being switched out
1827 * @next: the task we are going to switch to.
1829 * This is called with the rq lock held and interrupts off. It must
1830 * be paired with a subsequent finish_task_switch after the context
1833 * prepare_task_switch sets up locking and calls architecture specific
1837 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1838 struct task_struct
*next
)
1840 fire_sched_out_preempt_notifiers(prev
, next
);
1841 prepare_lock_switch(rq
, next
);
1842 prepare_arch_switch(next
);
1846 * finish_task_switch - clean up after a task-switch
1847 * @rq: runqueue associated with task-switch
1848 * @prev: the thread we just switched away from.
1850 * finish_task_switch must be called after the context switch, paired
1851 * with a prepare_task_switch call before the context switch.
1852 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1853 * and do any other architecture-specific cleanup actions.
1855 * Note that we may have delayed dropping an mm in context_switch(). If
1856 * so, we finish that here outside of the runqueue lock. (Doing it
1857 * with the lock held can cause deadlocks; see schedule() for
1860 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1861 __releases(rq
->lock
)
1863 struct mm_struct
*mm
= rq
->prev_mm
;
1869 * A task struct has one reference for the use as "current".
1870 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1871 * schedule one last time. The schedule call will never return, and
1872 * the scheduled task must drop that reference.
1873 * The test for TASK_DEAD must occur while the runqueue locks are
1874 * still held, otherwise prev could be scheduled on another cpu, die
1875 * there before we look at prev->state, and then the reference would
1877 * Manfred Spraul <manfred@colorfullife.com>
1879 prev_state
= prev
->state
;
1880 finish_arch_switch(prev
);
1881 finish_lock_switch(rq
, prev
);
1882 fire_sched_in_preempt_notifiers(current
);
1885 if (unlikely(prev_state
== TASK_DEAD
)) {
1887 * Remove function-return probe instances associated with this
1888 * task and put them back on the free list.
1890 kprobe_flush_task(prev
);
1891 put_task_struct(prev
);
1896 * schedule_tail - first thing a freshly forked thread must call.
1897 * @prev: the thread we just switched away from.
1899 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1900 __releases(rq
->lock
)
1902 struct rq
*rq
= this_rq();
1904 finish_task_switch(rq
, prev
);
1905 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1906 /* In this case, finish_task_switch does not reenable preemption */
1909 if (current
->set_child_tid
)
1910 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1914 * context_switch - switch to the new MM and the new
1915 * thread's register state.
1918 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1919 struct task_struct
*next
)
1921 struct mm_struct
*mm
, *oldmm
;
1923 prepare_task_switch(rq
, prev
, next
);
1925 oldmm
= prev
->active_mm
;
1927 * For paravirt, this is coupled with an exit in switch_to to
1928 * combine the page table reload and the switch backend into
1931 arch_enter_lazy_cpu_mode();
1933 if (unlikely(!mm
)) {
1934 next
->active_mm
= oldmm
;
1935 atomic_inc(&oldmm
->mm_count
);
1936 enter_lazy_tlb(oldmm
, next
);
1938 switch_mm(oldmm
, mm
, next
);
1940 if (unlikely(!prev
->mm
)) {
1941 prev
->active_mm
= NULL
;
1942 rq
->prev_mm
= oldmm
;
1945 * Since the runqueue lock will be released by the next
1946 * task (which is an invalid locking op but in the case
1947 * of the scheduler it's an obvious special-case), so we
1948 * do an early lockdep release here:
1950 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1951 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1954 /* Here we just switch the register state and the stack. */
1955 switch_to(prev
, next
, prev
);
1959 * this_rq must be evaluated again because prev may have moved
1960 * CPUs since it called schedule(), thus the 'rq' on its stack
1961 * frame will be invalid.
1963 finish_task_switch(this_rq(), prev
);
1967 * nr_running, nr_uninterruptible and nr_context_switches:
1969 * externally visible scheduler statistics: current number of runnable
1970 * threads, current number of uninterruptible-sleeping threads, total
1971 * number of context switches performed since bootup.
1973 unsigned long nr_running(void)
1975 unsigned long i
, sum
= 0;
1977 for_each_online_cpu(i
)
1978 sum
+= cpu_rq(i
)->nr_running
;
1983 unsigned long nr_uninterruptible(void)
1985 unsigned long i
, sum
= 0;
1987 for_each_possible_cpu(i
)
1988 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1991 * Since we read the counters lockless, it might be slightly
1992 * inaccurate. Do not allow it to go below zero though:
1994 if (unlikely((long)sum
< 0))
2000 unsigned long long nr_context_switches(void)
2003 unsigned long long sum
= 0;
2005 for_each_possible_cpu(i
)
2006 sum
+= cpu_rq(i
)->nr_switches
;
2011 unsigned long nr_iowait(void)
2013 unsigned long i
, sum
= 0;
2015 for_each_possible_cpu(i
)
2016 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2021 unsigned long nr_active(void)
2023 unsigned long i
, running
= 0, uninterruptible
= 0;
2025 for_each_online_cpu(i
) {
2026 running
+= cpu_rq(i
)->nr_running
;
2027 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2030 if (unlikely((long)uninterruptible
< 0))
2031 uninterruptible
= 0;
2033 return running
+ uninterruptible
;
2037 * Update rq->cpu_load[] statistics. This function is usually called every
2038 * scheduler tick (TICK_NSEC).
2040 static void update_cpu_load(struct rq
*this_rq
)
2042 unsigned long this_load
= this_rq
->load
.weight
;
2045 this_rq
->nr_load_updates
++;
2047 /* Update our load: */
2048 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2049 unsigned long old_load
, new_load
;
2051 /* scale is effectively 1 << i now, and >> i divides by scale */
2053 old_load
= this_rq
->cpu_load
[i
];
2054 new_load
= this_load
;
2056 * Round up the averaging division if load is increasing. This
2057 * prevents us from getting stuck on 9 if the load is 10, for
2060 if (new_load
> old_load
)
2061 new_load
+= scale
-1;
2062 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2069 * double_rq_lock - safely lock two runqueues
2071 * Note this does not disable interrupts like task_rq_lock,
2072 * you need to do so manually before calling.
2074 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2075 __acquires(rq1
->lock
)
2076 __acquires(rq2
->lock
)
2078 BUG_ON(!irqs_disabled());
2080 spin_lock(&rq1
->lock
);
2081 __acquire(rq2
->lock
); /* Fake it out ;) */
2084 spin_lock(&rq1
->lock
);
2085 spin_lock(&rq2
->lock
);
2087 spin_lock(&rq2
->lock
);
2088 spin_lock(&rq1
->lock
);
2091 update_rq_clock(rq1
);
2092 update_rq_clock(rq2
);
2096 * double_rq_unlock - safely unlock two runqueues
2098 * Note this does not restore interrupts like task_rq_unlock,
2099 * you need to do so manually after calling.
2101 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2102 __releases(rq1
->lock
)
2103 __releases(rq2
->lock
)
2105 spin_unlock(&rq1
->lock
);
2107 spin_unlock(&rq2
->lock
);
2109 __release(rq2
->lock
);
2113 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2115 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2116 __releases(this_rq
->lock
)
2117 __acquires(busiest
->lock
)
2118 __acquires(this_rq
->lock
)
2120 if (unlikely(!irqs_disabled())) {
2121 /* printk() doesn't work good under rq->lock */
2122 spin_unlock(&this_rq
->lock
);
2125 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2126 if (busiest
< this_rq
) {
2127 spin_unlock(&this_rq
->lock
);
2128 spin_lock(&busiest
->lock
);
2129 spin_lock(&this_rq
->lock
);
2131 spin_lock(&busiest
->lock
);
2136 * If dest_cpu is allowed for this process, migrate the task to it.
2137 * This is accomplished by forcing the cpu_allowed mask to only
2138 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2139 * the cpu_allowed mask is restored.
2141 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2143 struct migration_req req
;
2144 unsigned long flags
;
2147 rq
= task_rq_lock(p
, &flags
);
2148 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2149 || unlikely(cpu_is_offline(dest_cpu
)))
2152 /* force the process onto the specified CPU */
2153 if (migrate_task(p
, dest_cpu
, &req
)) {
2154 /* Need to wait for migration thread (might exit: take ref). */
2155 struct task_struct
*mt
= rq
->migration_thread
;
2157 get_task_struct(mt
);
2158 task_rq_unlock(rq
, &flags
);
2159 wake_up_process(mt
);
2160 put_task_struct(mt
);
2161 wait_for_completion(&req
.done
);
2166 task_rq_unlock(rq
, &flags
);
2170 * sched_exec - execve() is a valuable balancing opportunity, because at
2171 * this point the task has the smallest effective memory and cache footprint.
2173 void sched_exec(void)
2175 int new_cpu
, this_cpu
= get_cpu();
2176 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2178 if (new_cpu
!= this_cpu
)
2179 sched_migrate_task(current
, new_cpu
);
2183 * pull_task - move a task from a remote runqueue to the local runqueue.
2184 * Both runqueues must be locked.
2186 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2187 struct rq
*this_rq
, int this_cpu
)
2189 deactivate_task(src_rq
, p
, 0);
2190 set_task_cpu(p
, this_cpu
);
2191 activate_task(this_rq
, p
, 0);
2193 * Note that idle threads have a prio of MAX_PRIO, for this test
2194 * to be always true for them.
2196 check_preempt_curr(this_rq
, p
);
2200 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2203 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2204 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2208 * We do not migrate tasks that are:
2209 * 1) running (obviously), or
2210 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2211 * 3) are cache-hot on their current CPU.
2213 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2214 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2219 if (task_running(rq
, p
)) {
2220 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2225 * Aggressive migration if:
2226 * 1) task is cache cold, or
2227 * 2) too many balance attempts have failed.
2230 if (!task_hot(p
, rq
->clock
, sd
) ||
2231 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2232 #ifdef CONFIG_SCHEDSTATS
2233 if (task_hot(p
, rq
->clock
, sd
)) {
2234 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2235 schedstat_inc(p
, se
.nr_forced_migrations
);
2241 if (task_hot(p
, rq
->clock
, sd
)) {
2242 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2248 static unsigned long
2249 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2250 unsigned long max_load_move
, struct sched_domain
*sd
,
2251 enum cpu_idle_type idle
, int *all_pinned
,
2252 int *this_best_prio
, struct rq_iterator
*iterator
)
2254 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2255 struct task_struct
*p
;
2256 long rem_load_move
= max_load_move
;
2258 if (max_load_move
== 0)
2264 * Start the load-balancing iterator:
2266 p
= iterator
->start(iterator
->arg
);
2268 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2271 * To help distribute high priority tasks across CPUs we don't
2272 * skip a task if it will be the highest priority task (i.e. smallest
2273 * prio value) on its new queue regardless of its load weight
2275 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2276 SCHED_LOAD_SCALE_FUZZ
;
2277 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2278 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2279 p
= iterator
->next(iterator
->arg
);
2283 pull_task(busiest
, p
, this_rq
, this_cpu
);
2285 rem_load_move
-= p
->se
.load
.weight
;
2288 * We only want to steal up to the prescribed amount of weighted load.
2290 if (rem_load_move
> 0) {
2291 if (p
->prio
< *this_best_prio
)
2292 *this_best_prio
= p
->prio
;
2293 p
= iterator
->next(iterator
->arg
);
2298 * Right now, this is one of only two places pull_task() is called,
2299 * so we can safely collect pull_task() stats here rather than
2300 * inside pull_task().
2302 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2305 *all_pinned
= pinned
;
2307 return max_load_move
- rem_load_move
;
2311 * move_tasks tries to move up to max_load_move weighted load from busiest to
2312 * this_rq, as part of a balancing operation within domain "sd".
2313 * Returns 1 if successful and 0 otherwise.
2315 * Called with both runqueues locked.
2317 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2318 unsigned long max_load_move
,
2319 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2322 const struct sched_class
*class = sched_class_highest
;
2323 unsigned long total_load_moved
= 0;
2324 int this_best_prio
= this_rq
->curr
->prio
;
2328 class->load_balance(this_rq
, this_cpu
, busiest
,
2329 max_load_move
- total_load_moved
,
2330 sd
, idle
, all_pinned
, &this_best_prio
);
2331 class = class->next
;
2332 } while (class && max_load_move
> total_load_moved
);
2334 return total_load_moved
> 0;
2338 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2339 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2340 struct rq_iterator
*iterator
)
2342 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2346 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2347 pull_task(busiest
, p
, this_rq
, this_cpu
);
2349 * Right now, this is only the second place pull_task()
2350 * is called, so we can safely collect pull_task()
2351 * stats here rather than inside pull_task().
2353 schedstat_inc(sd
, lb_gained
[idle
]);
2357 p
= iterator
->next(iterator
->arg
);
2364 * move_one_task tries to move exactly one task from busiest to this_rq, as
2365 * part of active balancing operations within "domain".
2366 * Returns 1 if successful and 0 otherwise.
2368 * Called with both runqueues locked.
2370 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2371 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2373 const struct sched_class
*class;
2375 for (class = sched_class_highest
; class; class = class->next
)
2376 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2383 * find_busiest_group finds and returns the busiest CPU group within the
2384 * domain. It calculates and returns the amount of weighted load which
2385 * should be moved to restore balance via the imbalance parameter.
2387 static struct sched_group
*
2388 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2389 unsigned long *imbalance
, enum cpu_idle_type idle
,
2390 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2392 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2393 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2394 unsigned long max_pull
;
2395 unsigned long busiest_load_per_task
, busiest_nr_running
;
2396 unsigned long this_load_per_task
, this_nr_running
;
2397 int load_idx
, group_imb
= 0;
2398 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2399 int power_savings_balance
= 1;
2400 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2401 unsigned long min_nr_running
= ULONG_MAX
;
2402 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2405 max_load
= this_load
= total_load
= total_pwr
= 0;
2406 busiest_load_per_task
= busiest_nr_running
= 0;
2407 this_load_per_task
= this_nr_running
= 0;
2408 if (idle
== CPU_NOT_IDLE
)
2409 load_idx
= sd
->busy_idx
;
2410 else if (idle
== CPU_NEWLY_IDLE
)
2411 load_idx
= sd
->newidle_idx
;
2413 load_idx
= sd
->idle_idx
;
2416 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2419 int __group_imb
= 0;
2420 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2421 unsigned long sum_nr_running
, sum_weighted_load
;
2423 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2426 balance_cpu
= first_cpu(group
->cpumask
);
2428 /* Tally up the load of all CPUs in the group */
2429 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2431 min_cpu_load
= ~0UL;
2433 for_each_cpu_mask(i
, group
->cpumask
) {
2436 if (!cpu_isset(i
, *cpus
))
2441 if (*sd_idle
&& rq
->nr_running
)
2444 /* Bias balancing toward cpus of our domain */
2446 if (idle_cpu(i
) && !first_idle_cpu
) {
2451 load
= target_load(i
, load_idx
);
2453 load
= source_load(i
, load_idx
);
2454 if (load
> max_cpu_load
)
2455 max_cpu_load
= load
;
2456 if (min_cpu_load
> load
)
2457 min_cpu_load
= load
;
2461 sum_nr_running
+= rq
->nr_running
;
2462 sum_weighted_load
+= weighted_cpuload(i
);
2466 * First idle cpu or the first cpu(busiest) in this sched group
2467 * is eligible for doing load balancing at this and above
2468 * domains. In the newly idle case, we will allow all the cpu's
2469 * to do the newly idle load balance.
2471 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2472 balance_cpu
!= this_cpu
&& balance
) {
2477 total_load
+= avg_load
;
2478 total_pwr
+= group
->__cpu_power
;
2480 /* Adjust by relative CPU power of the group */
2481 avg_load
= sg_div_cpu_power(group
,
2482 avg_load
* SCHED_LOAD_SCALE
);
2484 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2487 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2490 this_load
= avg_load
;
2492 this_nr_running
= sum_nr_running
;
2493 this_load_per_task
= sum_weighted_load
;
2494 } else if (avg_load
> max_load
&&
2495 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2496 max_load
= avg_load
;
2498 busiest_nr_running
= sum_nr_running
;
2499 busiest_load_per_task
= sum_weighted_load
;
2500 group_imb
= __group_imb
;
2503 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2505 * Busy processors will not participate in power savings
2508 if (idle
== CPU_NOT_IDLE
||
2509 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2513 * If the local group is idle or completely loaded
2514 * no need to do power savings balance at this domain
2516 if (local_group
&& (this_nr_running
>= group_capacity
||
2518 power_savings_balance
= 0;
2521 * If a group is already running at full capacity or idle,
2522 * don't include that group in power savings calculations
2524 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2529 * Calculate the group which has the least non-idle load.
2530 * This is the group from where we need to pick up the load
2533 if ((sum_nr_running
< min_nr_running
) ||
2534 (sum_nr_running
== min_nr_running
&&
2535 first_cpu(group
->cpumask
) <
2536 first_cpu(group_min
->cpumask
))) {
2538 min_nr_running
= sum_nr_running
;
2539 min_load_per_task
= sum_weighted_load
/
2544 * Calculate the group which is almost near its
2545 * capacity but still has some space to pick up some load
2546 * from other group and save more power
2548 if (sum_nr_running
<= group_capacity
- 1) {
2549 if (sum_nr_running
> leader_nr_running
||
2550 (sum_nr_running
== leader_nr_running
&&
2551 first_cpu(group
->cpumask
) >
2552 first_cpu(group_leader
->cpumask
))) {
2553 group_leader
= group
;
2554 leader_nr_running
= sum_nr_running
;
2559 group
= group
->next
;
2560 } while (group
!= sd
->groups
);
2562 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2565 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2567 if (this_load
>= avg_load
||
2568 100*max_load
<= sd
->imbalance_pct
*this_load
)
2571 busiest_load_per_task
/= busiest_nr_running
;
2573 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2576 * We're trying to get all the cpus to the average_load, so we don't
2577 * want to push ourselves above the average load, nor do we wish to
2578 * reduce the max loaded cpu below the average load, as either of these
2579 * actions would just result in more rebalancing later, and ping-pong
2580 * tasks around. Thus we look for the minimum possible imbalance.
2581 * Negative imbalances (*we* are more loaded than anyone else) will
2582 * be counted as no imbalance for these purposes -- we can't fix that
2583 * by pulling tasks to us. Be careful of negative numbers as they'll
2584 * appear as very large values with unsigned longs.
2586 if (max_load
<= busiest_load_per_task
)
2590 * In the presence of smp nice balancing, certain scenarios can have
2591 * max load less than avg load(as we skip the groups at or below
2592 * its cpu_power, while calculating max_load..)
2594 if (max_load
< avg_load
) {
2596 goto small_imbalance
;
2599 /* Don't want to pull so many tasks that a group would go idle */
2600 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2602 /* How much load to actually move to equalise the imbalance */
2603 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2604 (avg_load
- this_load
) * this->__cpu_power
)
2608 * if *imbalance is less than the average load per runnable task
2609 * there is no gaurantee that any tasks will be moved so we'll have
2610 * a think about bumping its value to force at least one task to be
2613 if (*imbalance
< busiest_load_per_task
) {
2614 unsigned long tmp
, pwr_now
, pwr_move
;
2618 pwr_move
= pwr_now
= 0;
2620 if (this_nr_running
) {
2621 this_load_per_task
/= this_nr_running
;
2622 if (busiest_load_per_task
> this_load_per_task
)
2625 this_load_per_task
= SCHED_LOAD_SCALE
;
2627 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2628 busiest_load_per_task
* imbn
) {
2629 *imbalance
= busiest_load_per_task
;
2634 * OK, we don't have enough imbalance to justify moving tasks,
2635 * however we may be able to increase total CPU power used by
2639 pwr_now
+= busiest
->__cpu_power
*
2640 min(busiest_load_per_task
, max_load
);
2641 pwr_now
+= this->__cpu_power
*
2642 min(this_load_per_task
, this_load
);
2643 pwr_now
/= SCHED_LOAD_SCALE
;
2645 /* Amount of load we'd subtract */
2646 tmp
= sg_div_cpu_power(busiest
,
2647 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2649 pwr_move
+= busiest
->__cpu_power
*
2650 min(busiest_load_per_task
, max_load
- tmp
);
2652 /* Amount of load we'd add */
2653 if (max_load
* busiest
->__cpu_power
<
2654 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2655 tmp
= sg_div_cpu_power(this,
2656 max_load
* busiest
->__cpu_power
);
2658 tmp
= sg_div_cpu_power(this,
2659 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2660 pwr_move
+= this->__cpu_power
*
2661 min(this_load_per_task
, this_load
+ tmp
);
2662 pwr_move
/= SCHED_LOAD_SCALE
;
2664 /* Move if we gain throughput */
2665 if (pwr_move
> pwr_now
)
2666 *imbalance
= busiest_load_per_task
;
2672 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2673 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2676 if (this == group_leader
&& group_leader
!= group_min
) {
2677 *imbalance
= min_load_per_task
;
2687 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2690 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2691 unsigned long imbalance
, cpumask_t
*cpus
)
2693 struct rq
*busiest
= NULL
, *rq
;
2694 unsigned long max_load
= 0;
2697 for_each_cpu_mask(i
, group
->cpumask
) {
2700 if (!cpu_isset(i
, *cpus
))
2704 wl
= weighted_cpuload(i
);
2706 if (rq
->nr_running
== 1 && wl
> imbalance
)
2709 if (wl
> max_load
) {
2719 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2720 * so long as it is large enough.
2722 #define MAX_PINNED_INTERVAL 512
2725 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2726 * tasks if there is an imbalance.
2728 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2729 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2732 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2733 struct sched_group
*group
;
2734 unsigned long imbalance
;
2736 cpumask_t cpus
= CPU_MASK_ALL
;
2737 unsigned long flags
;
2740 * When power savings policy is enabled for the parent domain, idle
2741 * sibling can pick up load irrespective of busy siblings. In this case,
2742 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2743 * portraying it as CPU_NOT_IDLE.
2745 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2746 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2749 schedstat_inc(sd
, lb_count
[idle
]);
2752 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2759 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2763 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2765 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2769 BUG_ON(busiest
== this_rq
);
2771 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2774 if (busiest
->nr_running
> 1) {
2776 * Attempt to move tasks. If find_busiest_group has found
2777 * an imbalance but busiest->nr_running <= 1, the group is
2778 * still unbalanced. ld_moved simply stays zero, so it is
2779 * correctly treated as an imbalance.
2781 local_irq_save(flags
);
2782 double_rq_lock(this_rq
, busiest
);
2783 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2784 imbalance
, sd
, idle
, &all_pinned
);
2785 double_rq_unlock(this_rq
, busiest
);
2786 local_irq_restore(flags
);
2789 * some other cpu did the load balance for us.
2791 if (ld_moved
&& this_cpu
!= smp_processor_id())
2792 resched_cpu(this_cpu
);
2794 /* All tasks on this runqueue were pinned by CPU affinity */
2795 if (unlikely(all_pinned
)) {
2796 cpu_clear(cpu_of(busiest
), cpus
);
2797 if (!cpus_empty(cpus
))
2804 schedstat_inc(sd
, lb_failed
[idle
]);
2805 sd
->nr_balance_failed
++;
2807 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2809 spin_lock_irqsave(&busiest
->lock
, flags
);
2811 /* don't kick the migration_thread, if the curr
2812 * task on busiest cpu can't be moved to this_cpu
2814 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2815 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2817 goto out_one_pinned
;
2820 if (!busiest
->active_balance
) {
2821 busiest
->active_balance
= 1;
2822 busiest
->push_cpu
= this_cpu
;
2825 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2827 wake_up_process(busiest
->migration_thread
);
2830 * We've kicked active balancing, reset the failure
2833 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2836 sd
->nr_balance_failed
= 0;
2838 if (likely(!active_balance
)) {
2839 /* We were unbalanced, so reset the balancing interval */
2840 sd
->balance_interval
= sd
->min_interval
;
2843 * If we've begun active balancing, start to back off. This
2844 * case may not be covered by the all_pinned logic if there
2845 * is only 1 task on the busy runqueue (because we don't call
2848 if (sd
->balance_interval
< sd
->max_interval
)
2849 sd
->balance_interval
*= 2;
2852 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2853 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2858 schedstat_inc(sd
, lb_balanced
[idle
]);
2860 sd
->nr_balance_failed
= 0;
2863 /* tune up the balancing interval */
2864 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2865 (sd
->balance_interval
< sd
->max_interval
))
2866 sd
->balance_interval
*= 2;
2868 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2869 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2875 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2876 * tasks if there is an imbalance.
2878 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2879 * this_rq is locked.
2882 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2884 struct sched_group
*group
;
2885 struct rq
*busiest
= NULL
;
2886 unsigned long imbalance
;
2890 cpumask_t cpus
= CPU_MASK_ALL
;
2893 * When power savings policy is enabled for the parent domain, idle
2894 * sibling can pick up load irrespective of busy siblings. In this case,
2895 * let the state of idle sibling percolate up as IDLE, instead of
2896 * portraying it as CPU_NOT_IDLE.
2898 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2899 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2902 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
2904 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2905 &sd_idle
, &cpus
, NULL
);
2907 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2911 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2914 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2918 BUG_ON(busiest
== this_rq
);
2920 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2923 if (busiest
->nr_running
> 1) {
2924 /* Attempt to move tasks */
2925 double_lock_balance(this_rq
, busiest
);
2926 /* this_rq->clock is already updated */
2927 update_rq_clock(busiest
);
2928 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2929 imbalance
, sd
, CPU_NEWLY_IDLE
,
2931 spin_unlock(&busiest
->lock
);
2933 if (unlikely(all_pinned
)) {
2934 cpu_clear(cpu_of(busiest
), cpus
);
2935 if (!cpus_empty(cpus
))
2941 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2942 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2943 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2946 sd
->nr_balance_failed
= 0;
2951 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2952 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2953 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2955 sd
->nr_balance_failed
= 0;
2961 * idle_balance is called by schedule() if this_cpu is about to become
2962 * idle. Attempts to pull tasks from other CPUs.
2964 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2966 struct sched_domain
*sd
;
2967 int pulled_task
= -1;
2968 unsigned long next_balance
= jiffies
+ HZ
;
2970 for_each_domain(this_cpu
, sd
) {
2971 unsigned long interval
;
2973 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2976 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2977 /* If we've pulled tasks over stop searching: */
2978 pulled_task
= load_balance_newidle(this_cpu
,
2981 interval
= msecs_to_jiffies(sd
->balance_interval
);
2982 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2983 next_balance
= sd
->last_balance
+ interval
;
2987 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2989 * We are going idle. next_balance may be set based on
2990 * a busy processor. So reset next_balance.
2992 this_rq
->next_balance
= next_balance
;
2997 * active_load_balance is run by migration threads. It pushes running tasks
2998 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2999 * running on each physical CPU where possible, and avoids physical /
3000 * logical imbalances.
3002 * Called with busiest_rq locked.
3004 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3006 int target_cpu
= busiest_rq
->push_cpu
;
3007 struct sched_domain
*sd
;
3008 struct rq
*target_rq
;
3010 /* Is there any task to move? */
3011 if (busiest_rq
->nr_running
<= 1)
3014 target_rq
= cpu_rq(target_cpu
);
3017 * This condition is "impossible", if it occurs
3018 * we need to fix it. Originally reported by
3019 * Bjorn Helgaas on a 128-cpu setup.
3021 BUG_ON(busiest_rq
== target_rq
);
3023 /* move a task from busiest_rq to target_rq */
3024 double_lock_balance(busiest_rq
, target_rq
);
3025 update_rq_clock(busiest_rq
);
3026 update_rq_clock(target_rq
);
3028 /* Search for an sd spanning us and the target CPU. */
3029 for_each_domain(target_cpu
, sd
) {
3030 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3031 cpu_isset(busiest_cpu
, sd
->span
))
3036 schedstat_inc(sd
, alb_count
);
3038 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3040 schedstat_inc(sd
, alb_pushed
);
3042 schedstat_inc(sd
, alb_failed
);
3044 spin_unlock(&target_rq
->lock
);
3049 atomic_t load_balancer
;
3051 } nohz ____cacheline_aligned
= {
3052 .load_balancer
= ATOMIC_INIT(-1),
3053 .cpu_mask
= CPU_MASK_NONE
,
3057 * This routine will try to nominate the ilb (idle load balancing)
3058 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3059 * load balancing on behalf of all those cpus. If all the cpus in the system
3060 * go into this tickless mode, then there will be no ilb owner (as there is
3061 * no need for one) and all the cpus will sleep till the next wakeup event
3064 * For the ilb owner, tick is not stopped. And this tick will be used
3065 * for idle load balancing. ilb owner will still be part of
3068 * While stopping the tick, this cpu will become the ilb owner if there
3069 * is no other owner. And will be the owner till that cpu becomes busy
3070 * or if all cpus in the system stop their ticks at which point
3071 * there is no need for ilb owner.
3073 * When the ilb owner becomes busy, it nominates another owner, during the
3074 * next busy scheduler_tick()
3076 int select_nohz_load_balancer(int stop_tick
)
3078 int cpu
= smp_processor_id();
3081 cpu_set(cpu
, nohz
.cpu_mask
);
3082 cpu_rq(cpu
)->in_nohz_recently
= 1;
3085 * If we are going offline and still the leader, give up!
3087 if (cpu_is_offline(cpu
) &&
3088 atomic_read(&nohz
.load_balancer
) == cpu
) {
3089 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3094 /* time for ilb owner also to sleep */
3095 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3096 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3097 atomic_set(&nohz
.load_balancer
, -1);
3101 if (atomic_read(&nohz
.load_balancer
) == -1) {
3102 /* make me the ilb owner */
3103 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3105 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3108 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3111 cpu_clear(cpu
, nohz
.cpu_mask
);
3113 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3114 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3121 static DEFINE_SPINLOCK(balancing
);
3124 * It checks each scheduling domain to see if it is due to be balanced,
3125 * and initiates a balancing operation if so.
3127 * Balancing parameters are set up in arch_init_sched_domains.
3129 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3132 struct rq
*rq
= cpu_rq(cpu
);
3133 unsigned long interval
;
3134 struct sched_domain
*sd
;
3135 /* Earliest time when we have to do rebalance again */
3136 unsigned long next_balance
= jiffies
+ 60*HZ
;
3137 int update_next_balance
= 0;
3139 for_each_domain(cpu
, sd
) {
3140 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3143 interval
= sd
->balance_interval
;
3144 if (idle
!= CPU_IDLE
)
3145 interval
*= sd
->busy_factor
;
3147 /* scale ms to jiffies */
3148 interval
= msecs_to_jiffies(interval
);
3149 if (unlikely(!interval
))
3151 if (interval
> HZ
*NR_CPUS
/10)
3152 interval
= HZ
*NR_CPUS
/10;
3155 if (sd
->flags
& SD_SERIALIZE
) {
3156 if (!spin_trylock(&balancing
))
3160 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3161 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3163 * We've pulled tasks over so either we're no
3164 * longer idle, or one of our SMT siblings is
3167 idle
= CPU_NOT_IDLE
;
3169 sd
->last_balance
= jiffies
;
3171 if (sd
->flags
& SD_SERIALIZE
)
3172 spin_unlock(&balancing
);
3174 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3175 next_balance
= sd
->last_balance
+ interval
;
3176 update_next_balance
= 1;
3180 * Stop the load balance at this level. There is another
3181 * CPU in our sched group which is doing load balancing more
3189 * next_balance will be updated only when there is a need.
3190 * When the cpu is attached to null domain for ex, it will not be
3193 if (likely(update_next_balance
))
3194 rq
->next_balance
= next_balance
;
3198 * run_rebalance_domains is triggered when needed from the scheduler tick.
3199 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3200 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3202 static void run_rebalance_domains(struct softirq_action
*h
)
3204 int this_cpu
= smp_processor_id();
3205 struct rq
*this_rq
= cpu_rq(this_cpu
);
3206 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3207 CPU_IDLE
: CPU_NOT_IDLE
;
3209 rebalance_domains(this_cpu
, idle
);
3213 * If this cpu is the owner for idle load balancing, then do the
3214 * balancing on behalf of the other idle cpus whose ticks are
3217 if (this_rq
->idle_at_tick
&&
3218 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3219 cpumask_t cpus
= nohz
.cpu_mask
;
3223 cpu_clear(this_cpu
, cpus
);
3224 for_each_cpu_mask(balance_cpu
, cpus
) {
3226 * If this cpu gets work to do, stop the load balancing
3227 * work being done for other cpus. Next load
3228 * balancing owner will pick it up.
3233 rebalance_domains(balance_cpu
, CPU_IDLE
);
3235 rq
= cpu_rq(balance_cpu
);
3236 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3237 this_rq
->next_balance
= rq
->next_balance
;
3244 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3246 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3247 * idle load balancing owner or decide to stop the periodic load balancing,
3248 * if the whole system is idle.
3250 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3254 * If we were in the nohz mode recently and busy at the current
3255 * scheduler tick, then check if we need to nominate new idle
3258 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3259 rq
->in_nohz_recently
= 0;
3261 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3262 cpu_clear(cpu
, nohz
.cpu_mask
);
3263 atomic_set(&nohz
.load_balancer
, -1);
3266 if (atomic_read(&nohz
.load_balancer
) == -1) {
3268 * simple selection for now: Nominate the
3269 * first cpu in the nohz list to be the next
3272 * TBD: Traverse the sched domains and nominate
3273 * the nearest cpu in the nohz.cpu_mask.
3275 int ilb
= first_cpu(nohz
.cpu_mask
);
3283 * If this cpu is idle and doing idle load balancing for all the
3284 * cpus with ticks stopped, is it time for that to stop?
3286 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3287 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3293 * If this cpu is idle and the idle load balancing is done by
3294 * someone else, then no need raise the SCHED_SOFTIRQ
3296 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3297 cpu_isset(cpu
, nohz
.cpu_mask
))
3300 if (time_after_eq(jiffies
, rq
->next_balance
))
3301 raise_softirq(SCHED_SOFTIRQ
);
3304 #else /* CONFIG_SMP */
3307 * on UP we do not need to balance between CPUs:
3309 static inline void idle_balance(int cpu
, struct rq
*rq
)
3315 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3317 EXPORT_PER_CPU_SYMBOL(kstat
);
3320 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3321 * that have not yet been banked in case the task is currently running.
3323 unsigned long long task_sched_runtime(struct task_struct
*p
)
3325 unsigned long flags
;
3329 rq
= task_rq_lock(p
, &flags
);
3330 ns
= p
->se
.sum_exec_runtime
;
3331 if (rq
->curr
== p
) {
3332 update_rq_clock(rq
);
3333 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3334 if ((s64
)delta_exec
> 0)
3337 task_rq_unlock(rq
, &flags
);
3343 * Account user cpu time to a process.
3344 * @p: the process that the cpu time gets accounted to
3345 * @cputime: the cpu time spent in user space since the last update
3347 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3349 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3352 p
->utime
= cputime_add(p
->utime
, cputime
);
3354 /* Add user time to cpustat. */
3355 tmp
= cputime_to_cputime64(cputime
);
3356 if (TASK_NICE(p
) > 0)
3357 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3359 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3363 * Account guest cpu time to a process.
3364 * @p: the process that the cpu time gets accounted to
3365 * @cputime: the cpu time spent in virtual machine since the last update
3367 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3370 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3372 tmp
= cputime_to_cputime64(cputime
);
3374 p
->utime
= cputime_add(p
->utime
, cputime
);
3375 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3377 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3378 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3382 * Account scaled user cpu time to a process.
3383 * @p: the process that the cpu time gets accounted to
3384 * @cputime: the cpu time spent in user space since the last update
3386 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3388 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3392 * Account system cpu time to a process.
3393 * @p: the process that the cpu time gets accounted to
3394 * @hardirq_offset: the offset to subtract from hardirq_count()
3395 * @cputime: the cpu time spent in kernel space since the last update
3397 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3400 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3401 struct rq
*rq
= this_rq();
3404 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3405 return account_guest_time(p
, cputime
);
3407 p
->stime
= cputime_add(p
->stime
, cputime
);
3409 /* Add system time to cpustat. */
3410 tmp
= cputime_to_cputime64(cputime
);
3411 if (hardirq_count() - hardirq_offset
)
3412 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3413 else if (softirq_count())
3414 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3415 else if (p
!= rq
->idle
)
3416 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3417 else if (atomic_read(&rq
->nr_iowait
) > 0)
3418 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3420 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3421 /* Account for system time used */
3422 acct_update_integrals(p
);
3426 * Account scaled system cpu time to a process.
3427 * @p: the process that the cpu time gets accounted to
3428 * @hardirq_offset: the offset to subtract from hardirq_count()
3429 * @cputime: the cpu time spent in kernel space since the last update
3431 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3433 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3437 * Account for involuntary wait time.
3438 * @p: the process from which the cpu time has been stolen
3439 * @steal: the cpu time spent in involuntary wait
3441 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3443 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3444 cputime64_t tmp
= cputime_to_cputime64(steal
);
3445 struct rq
*rq
= this_rq();
3447 if (p
== rq
->idle
) {
3448 p
->stime
= cputime_add(p
->stime
, steal
);
3449 if (atomic_read(&rq
->nr_iowait
) > 0)
3450 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3452 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3454 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3458 * This function gets called by the timer code, with HZ frequency.
3459 * We call it with interrupts disabled.
3461 * It also gets called by the fork code, when changing the parent's
3464 void scheduler_tick(void)
3466 int cpu
= smp_processor_id();
3467 struct rq
*rq
= cpu_rq(cpu
);
3468 struct task_struct
*curr
= rq
->curr
;
3469 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3471 spin_lock(&rq
->lock
);
3472 __update_rq_clock(rq
);
3474 * Let rq->clock advance by at least TICK_NSEC:
3476 if (unlikely(rq
->clock
< next_tick
))
3477 rq
->clock
= next_tick
;
3478 rq
->tick_timestamp
= rq
->clock
;
3479 update_cpu_load(rq
);
3480 if (curr
!= rq
->idle
) /* FIXME: needed? */
3481 curr
->sched_class
->task_tick(rq
, curr
);
3482 spin_unlock(&rq
->lock
);
3485 rq
->idle_at_tick
= idle_cpu(cpu
);
3486 trigger_load_balance(rq
, cpu
);
3490 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3492 void fastcall
add_preempt_count(int val
)
3497 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3499 preempt_count() += val
;
3501 * Spinlock count overflowing soon?
3503 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3506 EXPORT_SYMBOL(add_preempt_count
);
3508 void fastcall
sub_preempt_count(int val
)
3513 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3516 * Is the spinlock portion underflowing?
3518 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3519 !(preempt_count() & PREEMPT_MASK
)))
3522 preempt_count() -= val
;
3524 EXPORT_SYMBOL(sub_preempt_count
);
3529 * Print scheduling while atomic bug:
3531 static noinline
void __schedule_bug(struct task_struct
*prev
)
3533 struct pt_regs
*regs
= get_irq_regs();
3535 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3536 prev
->comm
, prev
->pid
, preempt_count());
3538 debug_show_held_locks(prev
);
3539 if (irqs_disabled())
3540 print_irqtrace_events(prev
);
3549 * Various schedule()-time debugging checks and statistics:
3551 static inline void schedule_debug(struct task_struct
*prev
)
3554 * Test if we are atomic. Since do_exit() needs to call into
3555 * schedule() atomically, we ignore that path for now.
3556 * Otherwise, whine if we are scheduling when we should not be.
3558 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3559 __schedule_bug(prev
);
3561 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3563 schedstat_inc(this_rq(), sched_count
);
3564 #ifdef CONFIG_SCHEDSTATS
3565 if (unlikely(prev
->lock_depth
>= 0)) {
3566 schedstat_inc(this_rq(), bkl_count
);
3567 schedstat_inc(prev
, sched_info
.bkl_count
);
3573 * Pick up the highest-prio task:
3575 static inline struct task_struct
*
3576 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3578 const struct sched_class
*class;
3579 struct task_struct
*p
;
3582 * Optimization: we know that if all tasks are in
3583 * the fair class we can call that function directly:
3585 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3586 p
= fair_sched_class
.pick_next_task(rq
);
3591 class = sched_class_highest
;
3593 p
= class->pick_next_task(rq
);
3597 * Will never be NULL as the idle class always
3598 * returns a non-NULL p:
3600 class = class->next
;
3605 * schedule() is the main scheduler function.
3607 asmlinkage
void __sched
schedule(void)
3609 struct task_struct
*prev
, *next
;
3616 cpu
= smp_processor_id();
3620 switch_count
= &prev
->nivcsw
;
3622 release_kernel_lock(prev
);
3623 need_resched_nonpreemptible
:
3625 schedule_debug(prev
);
3628 * Do the rq-clock update outside the rq lock:
3630 local_irq_disable();
3631 __update_rq_clock(rq
);
3632 spin_lock(&rq
->lock
);
3633 clear_tsk_need_resched(prev
);
3635 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3636 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3637 unlikely(signal_pending(prev
)))) {
3638 prev
->state
= TASK_RUNNING
;
3640 deactivate_task(rq
, prev
, 1);
3642 switch_count
= &prev
->nvcsw
;
3645 if (unlikely(!rq
->nr_running
))
3646 idle_balance(cpu
, rq
);
3648 prev
->sched_class
->put_prev_task(rq
, prev
);
3649 next
= pick_next_task(rq
, prev
);
3651 sched_info_switch(prev
, next
);
3653 if (likely(prev
!= next
)) {
3658 context_switch(rq
, prev
, next
); /* unlocks the rq */
3660 spin_unlock_irq(&rq
->lock
);
3662 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3663 cpu
= smp_processor_id();
3665 goto need_resched_nonpreemptible
;
3667 preempt_enable_no_resched();
3668 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3671 EXPORT_SYMBOL(schedule
);
3673 #ifdef CONFIG_PREEMPT
3675 * this is the entry point to schedule() from in-kernel preemption
3676 * off of preempt_enable. Kernel preemptions off return from interrupt
3677 * occur there and call schedule directly.
3679 asmlinkage
void __sched
preempt_schedule(void)
3681 struct thread_info
*ti
= current_thread_info();
3682 #ifdef CONFIG_PREEMPT_BKL
3683 struct task_struct
*task
= current
;
3684 int saved_lock_depth
;
3687 * If there is a non-zero preempt_count or interrupts are disabled,
3688 * we do not want to preempt the current task. Just return..
3690 if (likely(ti
->preempt_count
|| irqs_disabled()))
3694 add_preempt_count(PREEMPT_ACTIVE
);
3697 * We keep the big kernel semaphore locked, but we
3698 * clear ->lock_depth so that schedule() doesnt
3699 * auto-release the semaphore:
3701 #ifdef CONFIG_PREEMPT_BKL
3702 saved_lock_depth
= task
->lock_depth
;
3703 task
->lock_depth
= -1;
3706 #ifdef CONFIG_PREEMPT_BKL
3707 task
->lock_depth
= saved_lock_depth
;
3709 sub_preempt_count(PREEMPT_ACTIVE
);
3712 * Check again in case we missed a preemption opportunity
3713 * between schedule and now.
3716 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3718 EXPORT_SYMBOL(preempt_schedule
);
3721 * this is the entry point to schedule() from kernel preemption
3722 * off of irq context.
3723 * Note, that this is called and return with irqs disabled. This will
3724 * protect us against recursive calling from irq.
3726 asmlinkage
void __sched
preempt_schedule_irq(void)
3728 struct thread_info
*ti
= current_thread_info();
3729 #ifdef CONFIG_PREEMPT_BKL
3730 struct task_struct
*task
= current
;
3731 int saved_lock_depth
;
3733 /* Catch callers which need to be fixed */
3734 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3737 add_preempt_count(PREEMPT_ACTIVE
);
3740 * We keep the big kernel semaphore locked, but we
3741 * clear ->lock_depth so that schedule() doesnt
3742 * auto-release the semaphore:
3744 #ifdef CONFIG_PREEMPT_BKL
3745 saved_lock_depth
= task
->lock_depth
;
3746 task
->lock_depth
= -1;
3750 local_irq_disable();
3751 #ifdef CONFIG_PREEMPT_BKL
3752 task
->lock_depth
= saved_lock_depth
;
3754 sub_preempt_count(PREEMPT_ACTIVE
);
3757 * Check again in case we missed a preemption opportunity
3758 * between schedule and now.
3761 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3764 #endif /* CONFIG_PREEMPT */
3766 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3769 return try_to_wake_up(curr
->private, mode
, sync
);
3771 EXPORT_SYMBOL(default_wake_function
);
3774 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3775 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3776 * number) then we wake all the non-exclusive tasks and one exclusive task.
3778 * There are circumstances in which we can try to wake a task which has already
3779 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3780 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3782 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3783 int nr_exclusive
, int sync
, void *key
)
3785 wait_queue_t
*curr
, *next
;
3787 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3788 unsigned flags
= curr
->flags
;
3790 if (curr
->func(curr
, mode
, sync
, key
) &&
3791 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3797 * __wake_up - wake up threads blocked on a waitqueue.
3799 * @mode: which threads
3800 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3801 * @key: is directly passed to the wakeup function
3803 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3804 int nr_exclusive
, void *key
)
3806 unsigned long flags
;
3808 spin_lock_irqsave(&q
->lock
, flags
);
3809 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3810 spin_unlock_irqrestore(&q
->lock
, flags
);
3812 EXPORT_SYMBOL(__wake_up
);
3815 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3817 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3819 __wake_up_common(q
, mode
, 1, 0, NULL
);
3823 * __wake_up_sync - wake up threads blocked on a waitqueue.
3825 * @mode: which threads
3826 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3828 * The sync wakeup differs that the waker knows that it will schedule
3829 * away soon, so while the target thread will be woken up, it will not
3830 * be migrated to another CPU - ie. the two threads are 'synchronized'
3831 * with each other. This can prevent needless bouncing between CPUs.
3833 * On UP it can prevent extra preemption.
3836 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3838 unsigned long flags
;
3844 if (unlikely(!nr_exclusive
))
3847 spin_lock_irqsave(&q
->lock
, flags
);
3848 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3849 spin_unlock_irqrestore(&q
->lock
, flags
);
3851 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3853 void complete(struct completion
*x
)
3855 unsigned long flags
;
3857 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3859 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3860 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3862 EXPORT_SYMBOL(complete
);
3864 void complete_all(struct completion
*x
)
3866 unsigned long flags
;
3868 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3869 x
->done
+= UINT_MAX
/2;
3870 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3871 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3873 EXPORT_SYMBOL(complete_all
);
3875 static inline long __sched
3876 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3879 DECLARE_WAITQUEUE(wait
, current
);
3881 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3882 __add_wait_queue_tail(&x
->wait
, &wait
);
3884 if ((state
== TASK_INTERRUPTIBLE
&&
3885 signal_pending(current
)) ||
3886 (state
== TASK_KILLABLE
&&
3887 fatal_signal_pending(current
))) {
3888 __remove_wait_queue(&x
->wait
, &wait
);
3889 return -ERESTARTSYS
;
3891 __set_current_state(state
);
3892 spin_unlock_irq(&x
->wait
.lock
);
3893 timeout
= schedule_timeout(timeout
);
3894 spin_lock_irq(&x
->wait
.lock
);
3896 __remove_wait_queue(&x
->wait
, &wait
);
3900 __remove_wait_queue(&x
->wait
, &wait
);
3907 wait_for_common(struct completion
*x
, long timeout
, int state
)
3911 spin_lock_irq(&x
->wait
.lock
);
3912 timeout
= do_wait_for_common(x
, timeout
, state
);
3913 spin_unlock_irq(&x
->wait
.lock
);
3917 void __sched
wait_for_completion(struct completion
*x
)
3919 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3921 EXPORT_SYMBOL(wait_for_completion
);
3923 unsigned long __sched
3924 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3926 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3928 EXPORT_SYMBOL(wait_for_completion_timeout
);
3930 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3932 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3933 if (t
== -ERESTARTSYS
)
3937 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3939 unsigned long __sched
3940 wait_for_completion_interruptible_timeout(struct completion
*x
,
3941 unsigned long timeout
)
3943 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3945 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3947 int __sched
wait_for_completion_killable(struct completion
*x
)
3949 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3950 if (t
== -ERESTARTSYS
)
3954 EXPORT_SYMBOL(wait_for_completion_killable
);
3957 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3959 unsigned long flags
;
3962 init_waitqueue_entry(&wait
, current
);
3964 __set_current_state(state
);
3966 spin_lock_irqsave(&q
->lock
, flags
);
3967 __add_wait_queue(q
, &wait
);
3968 spin_unlock(&q
->lock
);
3969 timeout
= schedule_timeout(timeout
);
3970 spin_lock_irq(&q
->lock
);
3971 __remove_wait_queue(q
, &wait
);
3972 spin_unlock_irqrestore(&q
->lock
, flags
);
3977 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3979 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3981 EXPORT_SYMBOL(interruptible_sleep_on
);
3984 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3986 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3988 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3990 void __sched
sleep_on(wait_queue_head_t
*q
)
3992 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3994 EXPORT_SYMBOL(sleep_on
);
3996 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3998 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4000 EXPORT_SYMBOL(sleep_on_timeout
);
4002 #ifdef CONFIG_RT_MUTEXES
4005 * rt_mutex_setprio - set the current priority of a task
4007 * @prio: prio value (kernel-internal form)
4009 * This function changes the 'effective' priority of a task. It does
4010 * not touch ->normal_prio like __setscheduler().
4012 * Used by the rt_mutex code to implement priority inheritance logic.
4014 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4016 unsigned long flags
;
4017 int oldprio
, on_rq
, running
;
4020 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4022 rq
= task_rq_lock(p
, &flags
);
4023 update_rq_clock(rq
);
4026 on_rq
= p
->se
.on_rq
;
4027 running
= task_running(rq
, p
);
4029 dequeue_task(rq
, p
, 0);
4031 p
->sched_class
->put_prev_task(rq
, p
);
4035 p
->sched_class
= &rt_sched_class
;
4037 p
->sched_class
= &fair_sched_class
;
4043 p
->sched_class
->set_curr_task(rq
);
4044 enqueue_task(rq
, p
, 0);
4046 * Reschedule if we are currently running on this runqueue and
4047 * our priority decreased, or if we are not currently running on
4048 * this runqueue and our priority is higher than the current's
4051 if (p
->prio
> oldprio
)
4052 resched_task(rq
->curr
);
4054 check_preempt_curr(rq
, p
);
4057 task_rq_unlock(rq
, &flags
);
4062 void set_user_nice(struct task_struct
*p
, long nice
)
4064 int old_prio
, delta
, on_rq
;
4065 unsigned long flags
;
4068 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4071 * We have to be careful, if called from sys_setpriority(),
4072 * the task might be in the middle of scheduling on another CPU.
4074 rq
= task_rq_lock(p
, &flags
);
4075 update_rq_clock(rq
);
4077 * The RT priorities are set via sched_setscheduler(), but we still
4078 * allow the 'normal' nice value to be set - but as expected
4079 * it wont have any effect on scheduling until the task is
4080 * SCHED_FIFO/SCHED_RR:
4082 if (task_has_rt_policy(p
)) {
4083 p
->static_prio
= NICE_TO_PRIO(nice
);
4086 on_rq
= p
->se
.on_rq
;
4088 dequeue_task(rq
, p
, 0);
4092 p
->static_prio
= NICE_TO_PRIO(nice
);
4095 p
->prio
= effective_prio(p
);
4096 delta
= p
->prio
- old_prio
;
4099 enqueue_task(rq
, p
, 0);
4102 * If the task increased its priority or is running and
4103 * lowered its priority, then reschedule its CPU:
4105 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4106 resched_task(rq
->curr
);
4109 task_rq_unlock(rq
, &flags
);
4111 EXPORT_SYMBOL(set_user_nice
);
4114 * can_nice - check if a task can reduce its nice value
4118 int can_nice(const struct task_struct
*p
, const int nice
)
4120 /* convert nice value [19,-20] to rlimit style value [1,40] */
4121 int nice_rlim
= 20 - nice
;
4123 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4124 capable(CAP_SYS_NICE
));
4127 #ifdef __ARCH_WANT_SYS_NICE
4130 * sys_nice - change the priority of the current process.
4131 * @increment: priority increment
4133 * sys_setpriority is a more generic, but much slower function that
4134 * does similar things.
4136 asmlinkage
long sys_nice(int increment
)
4141 * Setpriority might change our priority at the same moment.
4142 * We don't have to worry. Conceptually one call occurs first
4143 * and we have a single winner.
4145 if (increment
< -40)
4150 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4156 if (increment
< 0 && !can_nice(current
, nice
))
4159 retval
= security_task_setnice(current
, nice
);
4163 set_user_nice(current
, nice
);
4170 * task_prio - return the priority value of a given task.
4171 * @p: the task in question.
4173 * This is the priority value as seen by users in /proc.
4174 * RT tasks are offset by -200. Normal tasks are centered
4175 * around 0, value goes from -16 to +15.
4177 int task_prio(const struct task_struct
*p
)
4179 return p
->prio
- MAX_RT_PRIO
;
4183 * task_nice - return the nice value of a given task.
4184 * @p: the task in question.
4186 int task_nice(const struct task_struct
*p
)
4188 return TASK_NICE(p
);
4190 EXPORT_SYMBOL_GPL(task_nice
);
4193 * idle_cpu - is a given cpu idle currently?
4194 * @cpu: the processor in question.
4196 int idle_cpu(int cpu
)
4198 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4202 * idle_task - return the idle task for a given cpu.
4203 * @cpu: the processor in question.
4205 struct task_struct
*idle_task(int cpu
)
4207 return cpu_rq(cpu
)->idle
;
4211 * find_process_by_pid - find a process with a matching PID value.
4212 * @pid: the pid in question.
4214 static struct task_struct
*find_process_by_pid(pid_t pid
)
4216 return pid
? find_task_by_vpid(pid
) : current
;
4219 /* Actually do priority change: must hold rq lock. */
4221 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4223 BUG_ON(p
->se
.on_rq
);
4226 switch (p
->policy
) {
4230 p
->sched_class
= &fair_sched_class
;
4234 p
->sched_class
= &rt_sched_class
;
4238 p
->rt_priority
= prio
;
4239 p
->normal_prio
= normal_prio(p
);
4240 /* we are holding p->pi_lock already */
4241 p
->prio
= rt_mutex_getprio(p
);
4246 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4247 * @p: the task in question.
4248 * @policy: new policy.
4249 * @param: structure containing the new RT priority.
4251 * NOTE that the task may be already dead.
4253 int sched_setscheduler(struct task_struct
*p
, int policy
,
4254 struct sched_param
*param
)
4256 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4257 unsigned long flags
;
4260 /* may grab non-irq protected spin_locks */
4261 BUG_ON(in_interrupt());
4263 /* double check policy once rq lock held */
4265 policy
= oldpolicy
= p
->policy
;
4266 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4267 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4268 policy
!= SCHED_IDLE
)
4271 * Valid priorities for SCHED_FIFO and SCHED_RR are
4272 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4273 * SCHED_BATCH and SCHED_IDLE is 0.
4275 if (param
->sched_priority
< 0 ||
4276 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4277 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4279 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4283 * Allow unprivileged RT tasks to decrease priority:
4285 if (!capable(CAP_SYS_NICE
)) {
4286 if (rt_policy(policy
)) {
4287 unsigned long rlim_rtprio
;
4289 if (!lock_task_sighand(p
, &flags
))
4291 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4292 unlock_task_sighand(p
, &flags
);
4294 /* can't set/change the rt policy */
4295 if (policy
!= p
->policy
&& !rlim_rtprio
)
4298 /* can't increase priority */
4299 if (param
->sched_priority
> p
->rt_priority
&&
4300 param
->sched_priority
> rlim_rtprio
)
4304 * Like positive nice levels, dont allow tasks to
4305 * move out of SCHED_IDLE either:
4307 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4310 /* can't change other user's priorities */
4311 if ((current
->euid
!= p
->euid
) &&
4312 (current
->euid
!= p
->uid
))
4316 retval
= security_task_setscheduler(p
, policy
, param
);
4320 * make sure no PI-waiters arrive (or leave) while we are
4321 * changing the priority of the task:
4323 spin_lock_irqsave(&p
->pi_lock
, flags
);
4325 * To be able to change p->policy safely, the apropriate
4326 * runqueue lock must be held.
4328 rq
= __task_rq_lock(p
);
4329 /* recheck policy now with rq lock held */
4330 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4331 policy
= oldpolicy
= -1;
4332 __task_rq_unlock(rq
);
4333 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4336 update_rq_clock(rq
);
4337 on_rq
= p
->se
.on_rq
;
4338 running
= task_running(rq
, p
);
4340 deactivate_task(rq
, p
, 0);
4342 p
->sched_class
->put_prev_task(rq
, p
);
4346 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4350 p
->sched_class
->set_curr_task(rq
);
4351 activate_task(rq
, p
, 0);
4353 * Reschedule if we are currently running on this runqueue and
4354 * our priority decreased, or if we are not currently running on
4355 * this runqueue and our priority is higher than the current's
4358 if (p
->prio
> oldprio
)
4359 resched_task(rq
->curr
);
4361 check_preempt_curr(rq
, p
);
4364 __task_rq_unlock(rq
);
4365 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4367 rt_mutex_adjust_pi(p
);
4371 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4374 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4376 struct sched_param lparam
;
4377 struct task_struct
*p
;
4380 if (!param
|| pid
< 0)
4382 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4387 p
= find_process_by_pid(pid
);
4389 retval
= sched_setscheduler(p
, policy
, &lparam
);
4396 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4397 * @pid: the pid in question.
4398 * @policy: new policy.
4399 * @param: structure containing the new RT priority.
4402 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4404 /* negative values for policy are not valid */
4408 return do_sched_setscheduler(pid
, policy
, param
);
4412 * sys_sched_setparam - set/change the RT priority of a thread
4413 * @pid: the pid in question.
4414 * @param: structure containing the new RT priority.
4416 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4418 return do_sched_setscheduler(pid
, -1, param
);
4422 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4423 * @pid: the pid in question.
4425 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4427 struct task_struct
*p
;
4434 read_lock(&tasklist_lock
);
4435 p
= find_process_by_pid(pid
);
4437 retval
= security_task_getscheduler(p
);
4441 read_unlock(&tasklist_lock
);
4446 * sys_sched_getscheduler - get the RT priority of a thread
4447 * @pid: the pid in question.
4448 * @param: structure containing the RT priority.
4450 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4452 struct sched_param lp
;
4453 struct task_struct
*p
;
4456 if (!param
|| pid
< 0)
4459 read_lock(&tasklist_lock
);
4460 p
= find_process_by_pid(pid
);
4465 retval
= security_task_getscheduler(p
);
4469 lp
.sched_priority
= p
->rt_priority
;
4470 read_unlock(&tasklist_lock
);
4473 * This one might sleep, we cannot do it with a spinlock held ...
4475 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4480 read_unlock(&tasklist_lock
);
4484 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4486 cpumask_t cpus_allowed
;
4487 struct task_struct
*p
;
4490 mutex_lock(&sched_hotcpu_mutex
);
4491 read_lock(&tasklist_lock
);
4493 p
= find_process_by_pid(pid
);
4495 read_unlock(&tasklist_lock
);
4496 mutex_unlock(&sched_hotcpu_mutex
);
4501 * It is not safe to call set_cpus_allowed with the
4502 * tasklist_lock held. We will bump the task_struct's
4503 * usage count and then drop tasklist_lock.
4506 read_unlock(&tasklist_lock
);
4509 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4510 !capable(CAP_SYS_NICE
))
4513 retval
= security_task_setscheduler(p
, 0, NULL
);
4517 cpus_allowed
= cpuset_cpus_allowed(p
);
4518 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4520 retval
= set_cpus_allowed(p
, new_mask
);
4523 cpus_allowed
= cpuset_cpus_allowed(p
);
4524 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4526 * We must have raced with a concurrent cpuset
4527 * update. Just reset the cpus_allowed to the
4528 * cpuset's cpus_allowed
4530 new_mask
= cpus_allowed
;
4536 mutex_unlock(&sched_hotcpu_mutex
);
4540 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4541 cpumask_t
*new_mask
)
4543 if (len
< sizeof(cpumask_t
)) {
4544 memset(new_mask
, 0, sizeof(cpumask_t
));
4545 } else if (len
> sizeof(cpumask_t
)) {
4546 len
= sizeof(cpumask_t
);
4548 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4552 * sys_sched_setaffinity - set the cpu affinity of a process
4553 * @pid: pid of the process
4554 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4555 * @user_mask_ptr: user-space pointer to the new cpu mask
4557 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4558 unsigned long __user
*user_mask_ptr
)
4563 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4567 return sched_setaffinity(pid
, new_mask
);
4571 * Represents all cpu's present in the system
4572 * In systems capable of hotplug, this map could dynamically grow
4573 * as new cpu's are detected in the system via any platform specific
4574 * method, such as ACPI for e.g.
4577 cpumask_t cpu_present_map __read_mostly
;
4578 EXPORT_SYMBOL(cpu_present_map
);
4581 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4582 EXPORT_SYMBOL(cpu_online_map
);
4584 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4585 EXPORT_SYMBOL(cpu_possible_map
);
4588 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4590 struct task_struct
*p
;
4593 mutex_lock(&sched_hotcpu_mutex
);
4594 read_lock(&tasklist_lock
);
4597 p
= find_process_by_pid(pid
);
4601 retval
= security_task_getscheduler(p
);
4605 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4608 read_unlock(&tasklist_lock
);
4609 mutex_unlock(&sched_hotcpu_mutex
);
4615 * sys_sched_getaffinity - get the cpu affinity of a process
4616 * @pid: pid of the process
4617 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4618 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4620 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4621 unsigned long __user
*user_mask_ptr
)
4626 if (len
< sizeof(cpumask_t
))
4629 ret
= sched_getaffinity(pid
, &mask
);
4633 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4636 return sizeof(cpumask_t
);
4640 * sys_sched_yield - yield the current processor to other threads.
4642 * This function yields the current CPU to other tasks. If there are no
4643 * other threads running on this CPU then this function will return.
4645 asmlinkage
long sys_sched_yield(void)
4647 struct rq
*rq
= this_rq_lock();
4649 schedstat_inc(rq
, yld_count
);
4650 current
->sched_class
->yield_task(rq
);
4653 * Since we are going to call schedule() anyway, there's
4654 * no need to preempt or enable interrupts:
4656 __release(rq
->lock
);
4657 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4658 _raw_spin_unlock(&rq
->lock
);
4659 preempt_enable_no_resched();
4666 static void __cond_resched(void)
4668 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4669 __might_sleep(__FILE__
, __LINE__
);
4672 * The BKS might be reacquired before we have dropped
4673 * PREEMPT_ACTIVE, which could trigger a second
4674 * cond_resched() call.
4677 add_preempt_count(PREEMPT_ACTIVE
);
4679 sub_preempt_count(PREEMPT_ACTIVE
);
4680 } while (need_resched());
4683 int __sched
cond_resched(void)
4685 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4686 system_state
== SYSTEM_RUNNING
) {
4692 EXPORT_SYMBOL(cond_resched
);
4695 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4696 * call schedule, and on return reacquire the lock.
4698 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4699 * operations here to prevent schedule() from being called twice (once via
4700 * spin_unlock(), once by hand).
4702 int cond_resched_lock(spinlock_t
*lock
)
4706 if (need_lockbreak(lock
)) {
4712 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4713 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4714 _raw_spin_unlock(lock
);
4715 preempt_enable_no_resched();
4722 EXPORT_SYMBOL(cond_resched_lock
);
4724 int __sched
cond_resched_softirq(void)
4726 BUG_ON(!in_softirq());
4728 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4736 EXPORT_SYMBOL(cond_resched_softirq
);
4739 * yield - yield the current processor to other threads.
4741 * This is a shortcut for kernel-space yielding - it marks the
4742 * thread runnable and calls sys_sched_yield().
4744 void __sched
yield(void)
4746 set_current_state(TASK_RUNNING
);
4749 EXPORT_SYMBOL(yield
);
4752 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4753 * that process accounting knows that this is a task in IO wait state.
4755 * But don't do that if it is a deliberate, throttling IO wait (this task
4756 * has set its backing_dev_info: the queue against which it should throttle)
4758 void __sched
io_schedule(void)
4760 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4762 delayacct_blkio_start();
4763 atomic_inc(&rq
->nr_iowait
);
4765 atomic_dec(&rq
->nr_iowait
);
4766 delayacct_blkio_end();
4768 EXPORT_SYMBOL(io_schedule
);
4770 long __sched
io_schedule_timeout(long timeout
)
4772 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4775 delayacct_blkio_start();
4776 atomic_inc(&rq
->nr_iowait
);
4777 ret
= schedule_timeout(timeout
);
4778 atomic_dec(&rq
->nr_iowait
);
4779 delayacct_blkio_end();
4784 * sys_sched_get_priority_max - return maximum RT priority.
4785 * @policy: scheduling class.
4787 * this syscall returns the maximum rt_priority that can be used
4788 * by a given scheduling class.
4790 asmlinkage
long sys_sched_get_priority_max(int policy
)
4797 ret
= MAX_USER_RT_PRIO
-1;
4809 * sys_sched_get_priority_min - return minimum RT priority.
4810 * @policy: scheduling class.
4812 * this syscall returns the minimum rt_priority that can be used
4813 * by a given scheduling class.
4815 asmlinkage
long sys_sched_get_priority_min(int policy
)
4833 * sys_sched_rr_get_interval - return the default timeslice of a process.
4834 * @pid: pid of the process.
4835 * @interval: userspace pointer to the timeslice value.
4837 * this syscall writes the default timeslice value of a given process
4838 * into the user-space timespec buffer. A value of '0' means infinity.
4841 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4843 struct task_struct
*p
;
4844 unsigned int time_slice
;
4852 read_lock(&tasklist_lock
);
4853 p
= find_process_by_pid(pid
);
4857 retval
= security_task_getscheduler(p
);
4862 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4863 * tasks that are on an otherwise idle runqueue:
4866 if (p
->policy
== SCHED_RR
) {
4867 time_slice
= DEF_TIMESLICE
;
4869 struct sched_entity
*se
= &p
->se
;
4870 unsigned long flags
;
4873 rq
= task_rq_lock(p
, &flags
);
4874 if (rq
->cfs
.load
.weight
)
4875 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
4876 task_rq_unlock(rq
, &flags
);
4878 read_unlock(&tasklist_lock
);
4879 jiffies_to_timespec(time_slice
, &t
);
4880 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4884 read_unlock(&tasklist_lock
);
4888 static const char stat_nam
[] = "RSDTtZX";
4890 static void show_task(struct task_struct
*p
)
4892 unsigned long free
= 0;
4895 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4896 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
4897 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4898 #if BITS_PER_LONG == 32
4899 if (state
== TASK_RUNNING
)
4900 printk(KERN_CONT
" running ");
4902 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4904 if (state
== TASK_RUNNING
)
4905 printk(KERN_CONT
" running task ");
4907 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4909 #ifdef CONFIG_DEBUG_STACK_USAGE
4911 unsigned long *n
= end_of_stack(p
);
4914 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4917 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
4918 task_pid_nr(p
), task_pid_nr(p
->parent
));
4920 if (state
!= TASK_RUNNING
)
4921 show_stack(p
, NULL
);
4924 void show_state_filter(unsigned long state_filter
)
4926 struct task_struct
*g
, *p
;
4928 #if BITS_PER_LONG == 32
4930 " task PC stack pid father\n");
4933 " task PC stack pid father\n");
4935 read_lock(&tasklist_lock
);
4936 do_each_thread(g
, p
) {
4938 * reset the NMI-timeout, listing all files on a slow
4939 * console might take alot of time:
4941 touch_nmi_watchdog();
4942 if (!state_filter
|| (p
->state
& state_filter
))
4944 } while_each_thread(g
, p
);
4946 touch_all_softlockup_watchdogs();
4948 #ifdef CONFIG_SCHED_DEBUG
4949 sysrq_sched_debug_show();
4951 read_unlock(&tasklist_lock
);
4953 * Only show locks if all tasks are dumped:
4955 if (state_filter
== -1)
4956 debug_show_all_locks();
4959 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4961 idle
->sched_class
= &idle_sched_class
;
4965 * init_idle - set up an idle thread for a given CPU
4966 * @idle: task in question
4967 * @cpu: cpu the idle task belongs to
4969 * NOTE: this function does not set the idle thread's NEED_RESCHED
4970 * flag, to make booting more robust.
4972 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4974 struct rq
*rq
= cpu_rq(cpu
);
4975 unsigned long flags
;
4978 idle
->se
.exec_start
= sched_clock();
4980 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4981 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4982 __set_task_cpu(idle
, cpu
);
4984 spin_lock_irqsave(&rq
->lock
, flags
);
4985 rq
->curr
= rq
->idle
= idle
;
4986 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4989 spin_unlock_irqrestore(&rq
->lock
, flags
);
4991 /* Set the preempt count _outside_ the spinlocks! */
4992 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4993 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4995 task_thread_info(idle
)->preempt_count
= 0;
4998 * The idle tasks have their own, simple scheduling class:
5000 idle
->sched_class
= &idle_sched_class
;
5004 * In a system that switches off the HZ timer nohz_cpu_mask
5005 * indicates which cpus entered this state. This is used
5006 * in the rcu update to wait only for active cpus. For system
5007 * which do not switch off the HZ timer nohz_cpu_mask should
5008 * always be CPU_MASK_NONE.
5010 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5013 * Increase the granularity value when there are more CPUs,
5014 * because with more CPUs the 'effective latency' as visible
5015 * to users decreases. But the relationship is not linear,
5016 * so pick a second-best guess by going with the log2 of the
5019 * This idea comes from the SD scheduler of Con Kolivas:
5021 static inline void sched_init_granularity(void)
5023 unsigned int factor
= 1 + ilog2(num_online_cpus());
5024 const unsigned long limit
= 200000000;
5026 sysctl_sched_min_granularity
*= factor
;
5027 if (sysctl_sched_min_granularity
> limit
)
5028 sysctl_sched_min_granularity
= limit
;
5030 sysctl_sched_latency
*= factor
;
5031 if (sysctl_sched_latency
> limit
)
5032 sysctl_sched_latency
= limit
;
5034 sysctl_sched_wakeup_granularity
*= factor
;
5035 sysctl_sched_batch_wakeup_granularity
*= factor
;
5040 * This is how migration works:
5042 * 1) we queue a struct migration_req structure in the source CPU's
5043 * runqueue and wake up that CPU's migration thread.
5044 * 2) we down() the locked semaphore => thread blocks.
5045 * 3) migration thread wakes up (implicitly it forces the migrated
5046 * thread off the CPU)
5047 * 4) it gets the migration request and checks whether the migrated
5048 * task is still in the wrong runqueue.
5049 * 5) if it's in the wrong runqueue then the migration thread removes
5050 * it and puts it into the right queue.
5051 * 6) migration thread up()s the semaphore.
5052 * 7) we wake up and the migration is done.
5056 * Change a given task's CPU affinity. Migrate the thread to a
5057 * proper CPU and schedule it away if the CPU it's executing on
5058 * is removed from the allowed bitmask.
5060 * NOTE: the caller must have a valid reference to the task, the
5061 * task must not exit() & deallocate itself prematurely. The
5062 * call is not atomic; no spinlocks may be held.
5064 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5066 struct migration_req req
;
5067 unsigned long flags
;
5071 rq
= task_rq_lock(p
, &flags
);
5072 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5077 p
->cpus_allowed
= new_mask
;
5078 /* Can the task run on the task's current CPU? If so, we're done */
5079 if (cpu_isset(task_cpu(p
), new_mask
))
5082 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5083 /* Need help from migration thread: drop lock and wait. */
5084 task_rq_unlock(rq
, &flags
);
5085 wake_up_process(rq
->migration_thread
);
5086 wait_for_completion(&req
.done
);
5087 tlb_migrate_finish(p
->mm
);
5091 task_rq_unlock(rq
, &flags
);
5095 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5098 * Move (not current) task off this cpu, onto dest cpu. We're doing
5099 * this because either it can't run here any more (set_cpus_allowed()
5100 * away from this CPU, or CPU going down), or because we're
5101 * attempting to rebalance this task on exec (sched_exec).
5103 * So we race with normal scheduler movements, but that's OK, as long
5104 * as the task is no longer on this CPU.
5106 * Returns non-zero if task was successfully migrated.
5108 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5110 struct rq
*rq_dest
, *rq_src
;
5113 if (unlikely(cpu_is_offline(dest_cpu
)))
5116 rq_src
= cpu_rq(src_cpu
);
5117 rq_dest
= cpu_rq(dest_cpu
);
5119 double_rq_lock(rq_src
, rq_dest
);
5120 /* Already moved. */
5121 if (task_cpu(p
) != src_cpu
)
5123 /* Affinity changed (again). */
5124 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5127 on_rq
= p
->se
.on_rq
;
5129 deactivate_task(rq_src
, p
, 0);
5131 set_task_cpu(p
, dest_cpu
);
5133 activate_task(rq_dest
, p
, 0);
5134 check_preempt_curr(rq_dest
, p
);
5138 double_rq_unlock(rq_src
, rq_dest
);
5143 * migration_thread - this is a highprio system thread that performs
5144 * thread migration by bumping thread off CPU then 'pushing' onto
5147 static int migration_thread(void *data
)
5149 int cpu
= (long)data
;
5153 BUG_ON(rq
->migration_thread
!= current
);
5155 set_current_state(TASK_INTERRUPTIBLE
);
5156 while (!kthread_should_stop()) {
5157 struct migration_req
*req
;
5158 struct list_head
*head
;
5160 spin_lock_irq(&rq
->lock
);
5162 if (cpu_is_offline(cpu
)) {
5163 spin_unlock_irq(&rq
->lock
);
5167 if (rq
->active_balance
) {
5168 active_load_balance(rq
, cpu
);
5169 rq
->active_balance
= 0;
5172 head
= &rq
->migration_queue
;
5174 if (list_empty(head
)) {
5175 spin_unlock_irq(&rq
->lock
);
5177 set_current_state(TASK_INTERRUPTIBLE
);
5180 req
= list_entry(head
->next
, struct migration_req
, list
);
5181 list_del_init(head
->next
);
5183 spin_unlock(&rq
->lock
);
5184 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5187 complete(&req
->done
);
5189 __set_current_state(TASK_RUNNING
);
5193 /* Wait for kthread_stop */
5194 set_current_state(TASK_INTERRUPTIBLE
);
5195 while (!kthread_should_stop()) {
5197 set_current_state(TASK_INTERRUPTIBLE
);
5199 __set_current_state(TASK_RUNNING
);
5203 #ifdef CONFIG_HOTPLUG_CPU
5205 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5209 local_irq_disable();
5210 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5216 * Figure out where task on dead CPU should go, use force if necessary.
5217 * NOTE: interrupts should be disabled by the caller
5219 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5221 unsigned long flags
;
5228 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5229 cpus_and(mask
, mask
, p
->cpus_allowed
);
5230 dest_cpu
= any_online_cpu(mask
);
5232 /* On any allowed CPU? */
5233 if (dest_cpu
== NR_CPUS
)
5234 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5236 /* No more Mr. Nice Guy. */
5237 if (dest_cpu
== NR_CPUS
) {
5238 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5240 * Try to stay on the same cpuset, where the
5241 * current cpuset may be a subset of all cpus.
5242 * The cpuset_cpus_allowed_locked() variant of
5243 * cpuset_cpus_allowed() will not block. It must be
5244 * called within calls to cpuset_lock/cpuset_unlock.
5246 rq
= task_rq_lock(p
, &flags
);
5247 p
->cpus_allowed
= cpus_allowed
;
5248 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5249 task_rq_unlock(rq
, &flags
);
5252 * Don't tell them about moving exiting tasks or
5253 * kernel threads (both mm NULL), since they never
5256 if (p
->mm
&& printk_ratelimit()) {
5257 printk(KERN_INFO
"process %d (%s) no "
5258 "longer affine to cpu%d\n",
5259 task_pid_nr(p
), p
->comm
, dead_cpu
);
5262 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5266 * While a dead CPU has no uninterruptible tasks queued at this point,
5267 * it might still have a nonzero ->nr_uninterruptible counter, because
5268 * for performance reasons the counter is not stricly tracking tasks to
5269 * their home CPUs. So we just add the counter to another CPU's counter,
5270 * to keep the global sum constant after CPU-down:
5272 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5274 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5275 unsigned long flags
;
5277 local_irq_save(flags
);
5278 double_rq_lock(rq_src
, rq_dest
);
5279 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5280 rq_src
->nr_uninterruptible
= 0;
5281 double_rq_unlock(rq_src
, rq_dest
);
5282 local_irq_restore(flags
);
5285 /* Run through task list and migrate tasks from the dead cpu. */
5286 static void migrate_live_tasks(int src_cpu
)
5288 struct task_struct
*p
, *t
;
5290 read_lock(&tasklist_lock
);
5292 do_each_thread(t
, p
) {
5296 if (task_cpu(p
) == src_cpu
)
5297 move_task_off_dead_cpu(src_cpu
, p
);
5298 } while_each_thread(t
, p
);
5300 read_unlock(&tasklist_lock
);
5304 * Schedules idle task to be the next runnable task on current CPU.
5305 * It does so by boosting its priority to highest possible.
5306 * Used by CPU offline code.
5308 void sched_idle_next(void)
5310 int this_cpu
= smp_processor_id();
5311 struct rq
*rq
= cpu_rq(this_cpu
);
5312 struct task_struct
*p
= rq
->idle
;
5313 unsigned long flags
;
5315 /* cpu has to be offline */
5316 BUG_ON(cpu_online(this_cpu
));
5319 * Strictly not necessary since rest of the CPUs are stopped by now
5320 * and interrupts disabled on the current cpu.
5322 spin_lock_irqsave(&rq
->lock
, flags
);
5324 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5326 update_rq_clock(rq
);
5327 activate_task(rq
, p
, 0);
5329 spin_unlock_irqrestore(&rq
->lock
, flags
);
5333 * Ensures that the idle task is using init_mm right before its cpu goes
5336 void idle_task_exit(void)
5338 struct mm_struct
*mm
= current
->active_mm
;
5340 BUG_ON(cpu_online(smp_processor_id()));
5343 switch_mm(mm
, &init_mm
, current
);
5347 /* called under rq->lock with disabled interrupts */
5348 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5350 struct rq
*rq
= cpu_rq(dead_cpu
);
5352 /* Must be exiting, otherwise would be on tasklist. */
5353 BUG_ON(!p
->exit_state
);
5355 /* Cannot have done final schedule yet: would have vanished. */
5356 BUG_ON(p
->state
== TASK_DEAD
);
5361 * Drop lock around migration; if someone else moves it,
5362 * that's OK. No task can be added to this CPU, so iteration is
5365 spin_unlock_irq(&rq
->lock
);
5366 move_task_off_dead_cpu(dead_cpu
, p
);
5367 spin_lock_irq(&rq
->lock
);
5372 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5373 static void migrate_dead_tasks(unsigned int dead_cpu
)
5375 struct rq
*rq
= cpu_rq(dead_cpu
);
5376 struct task_struct
*next
;
5379 if (!rq
->nr_running
)
5381 update_rq_clock(rq
);
5382 next
= pick_next_task(rq
, rq
->curr
);
5385 migrate_dead(dead_cpu
, next
);
5389 #endif /* CONFIG_HOTPLUG_CPU */
5391 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5393 static struct ctl_table sd_ctl_dir
[] = {
5395 .procname
= "sched_domain",
5401 static struct ctl_table sd_ctl_root
[] = {
5403 .ctl_name
= CTL_KERN
,
5404 .procname
= "kernel",
5406 .child
= sd_ctl_dir
,
5411 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5413 struct ctl_table
*entry
=
5414 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5419 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5421 struct ctl_table
*entry
;
5424 * In the intermediate directories, both the child directory and
5425 * procname are dynamically allocated and could fail but the mode
5426 * will always be set. In the lowest directory the names are
5427 * static strings and all have proc handlers.
5429 for (entry
= *tablep
; entry
->mode
; entry
++) {
5431 sd_free_ctl_entry(&entry
->child
);
5432 if (entry
->proc_handler
== NULL
)
5433 kfree(entry
->procname
);
5441 set_table_entry(struct ctl_table
*entry
,
5442 const char *procname
, void *data
, int maxlen
,
5443 mode_t mode
, proc_handler
*proc_handler
)
5445 entry
->procname
= procname
;
5447 entry
->maxlen
= maxlen
;
5449 entry
->proc_handler
= proc_handler
;
5452 static struct ctl_table
*
5453 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5455 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5460 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5461 sizeof(long), 0644, proc_doulongvec_minmax
);
5462 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5463 sizeof(long), 0644, proc_doulongvec_minmax
);
5464 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5465 sizeof(int), 0644, proc_dointvec_minmax
);
5466 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5467 sizeof(int), 0644, proc_dointvec_minmax
);
5468 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5469 sizeof(int), 0644, proc_dointvec_minmax
);
5470 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5471 sizeof(int), 0644, proc_dointvec_minmax
);
5472 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5473 sizeof(int), 0644, proc_dointvec_minmax
);
5474 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5475 sizeof(int), 0644, proc_dointvec_minmax
);
5476 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5477 sizeof(int), 0644, proc_dointvec_minmax
);
5478 set_table_entry(&table
[9], "cache_nice_tries",
5479 &sd
->cache_nice_tries
,
5480 sizeof(int), 0644, proc_dointvec_minmax
);
5481 set_table_entry(&table
[10], "flags", &sd
->flags
,
5482 sizeof(int), 0644, proc_dointvec_minmax
);
5483 /* &table[11] is terminator */
5488 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5490 struct ctl_table
*entry
, *table
;
5491 struct sched_domain
*sd
;
5492 int domain_num
= 0, i
;
5495 for_each_domain(cpu
, sd
)
5497 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5502 for_each_domain(cpu
, sd
) {
5503 snprintf(buf
, 32, "domain%d", i
);
5504 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5506 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5513 static struct ctl_table_header
*sd_sysctl_header
;
5514 static void register_sched_domain_sysctl(void)
5516 int i
, cpu_num
= num_online_cpus();
5517 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5520 WARN_ON(sd_ctl_dir
[0].child
);
5521 sd_ctl_dir
[0].child
= entry
;
5526 for_each_online_cpu(i
) {
5527 snprintf(buf
, 32, "cpu%d", i
);
5528 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5530 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5534 WARN_ON(sd_sysctl_header
);
5535 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5538 /* may be called multiple times per register */
5539 static void unregister_sched_domain_sysctl(void)
5541 if (sd_sysctl_header
)
5542 unregister_sysctl_table(sd_sysctl_header
);
5543 sd_sysctl_header
= NULL
;
5544 if (sd_ctl_dir
[0].child
)
5545 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5548 static void register_sched_domain_sysctl(void)
5551 static void unregister_sched_domain_sysctl(void)
5557 * migration_call - callback that gets triggered when a CPU is added.
5558 * Here we can start up the necessary migration thread for the new CPU.
5560 static int __cpuinit
5561 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5563 struct task_struct
*p
;
5564 int cpu
= (long)hcpu
;
5565 unsigned long flags
;
5569 case CPU_LOCK_ACQUIRE
:
5570 mutex_lock(&sched_hotcpu_mutex
);
5573 case CPU_UP_PREPARE
:
5574 case CPU_UP_PREPARE_FROZEN
:
5575 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5578 kthread_bind(p
, cpu
);
5579 /* Must be high prio: stop_machine expects to yield to it. */
5580 rq
= task_rq_lock(p
, &flags
);
5581 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5582 task_rq_unlock(rq
, &flags
);
5583 cpu_rq(cpu
)->migration_thread
= p
;
5587 case CPU_ONLINE_FROZEN
:
5588 /* Strictly unnecessary, as first user will wake it. */
5589 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5592 #ifdef CONFIG_HOTPLUG_CPU
5593 case CPU_UP_CANCELED
:
5594 case CPU_UP_CANCELED_FROZEN
:
5595 if (!cpu_rq(cpu
)->migration_thread
)
5597 /* Unbind it from offline cpu so it can run. Fall thru. */
5598 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5599 any_online_cpu(cpu_online_map
));
5600 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5601 cpu_rq(cpu
)->migration_thread
= NULL
;
5605 case CPU_DEAD_FROZEN
:
5606 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5607 migrate_live_tasks(cpu
);
5609 kthread_stop(rq
->migration_thread
);
5610 rq
->migration_thread
= NULL
;
5611 /* Idle task back to normal (off runqueue, low prio) */
5612 spin_lock_irq(&rq
->lock
);
5613 update_rq_clock(rq
);
5614 deactivate_task(rq
, rq
->idle
, 0);
5615 rq
->idle
->static_prio
= MAX_PRIO
;
5616 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5617 rq
->idle
->sched_class
= &idle_sched_class
;
5618 migrate_dead_tasks(cpu
);
5619 spin_unlock_irq(&rq
->lock
);
5621 migrate_nr_uninterruptible(rq
);
5622 BUG_ON(rq
->nr_running
!= 0);
5625 * No need to migrate the tasks: it was best-effort if
5626 * they didn't take sched_hotcpu_mutex. Just wake up
5629 spin_lock_irq(&rq
->lock
);
5630 while (!list_empty(&rq
->migration_queue
)) {
5631 struct migration_req
*req
;
5633 req
= list_entry(rq
->migration_queue
.next
,
5634 struct migration_req
, list
);
5635 list_del_init(&req
->list
);
5636 complete(&req
->done
);
5638 spin_unlock_irq(&rq
->lock
);
5641 case CPU_LOCK_RELEASE
:
5642 mutex_unlock(&sched_hotcpu_mutex
);
5648 /* Register at highest priority so that task migration (migrate_all_tasks)
5649 * happens before everything else.
5651 static struct notifier_block __cpuinitdata migration_notifier
= {
5652 .notifier_call
= migration_call
,
5656 void __init
migration_init(void)
5658 void *cpu
= (void *)(long)smp_processor_id();
5661 /* Start one for the boot CPU: */
5662 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5663 BUG_ON(err
== NOTIFY_BAD
);
5664 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5665 register_cpu_notifier(&migration_notifier
);
5671 /* Number of possible processor ids */
5672 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5673 EXPORT_SYMBOL(nr_cpu_ids
);
5675 #ifdef CONFIG_SCHED_DEBUG
5677 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
5679 struct sched_group
*group
= sd
->groups
;
5680 cpumask_t groupmask
;
5683 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5684 cpus_clear(groupmask
);
5686 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5688 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5689 printk("does not load-balance\n");
5691 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5696 printk(KERN_CONT
"span %s\n", str
);
5698 if (!cpu_isset(cpu
, sd
->span
)) {
5699 printk(KERN_ERR
"ERROR: domain->span does not contain "
5702 if (!cpu_isset(cpu
, group
->cpumask
)) {
5703 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5707 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5711 printk(KERN_ERR
"ERROR: group is NULL\n");
5715 if (!group
->__cpu_power
) {
5716 printk(KERN_CONT
"\n");
5717 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5722 if (!cpus_weight(group
->cpumask
)) {
5723 printk(KERN_CONT
"\n");
5724 printk(KERN_ERR
"ERROR: empty group\n");
5728 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5729 printk(KERN_CONT
"\n");
5730 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5734 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5736 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5737 printk(KERN_CONT
" %s", str
);
5739 group
= group
->next
;
5740 } while (group
!= sd
->groups
);
5741 printk(KERN_CONT
"\n");
5743 if (!cpus_equal(sd
->span
, groupmask
))
5744 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5746 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
5747 printk(KERN_ERR
"ERROR: parent span is not a superset "
5748 "of domain->span\n");
5752 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5757 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5761 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5764 if (sched_domain_debug_one(sd
, cpu
, level
))
5773 # define sched_domain_debug(sd, cpu) do { } while (0)
5776 static int sd_degenerate(struct sched_domain
*sd
)
5778 if (cpus_weight(sd
->span
) == 1)
5781 /* Following flags need at least 2 groups */
5782 if (sd
->flags
& (SD_LOAD_BALANCE
|
5783 SD_BALANCE_NEWIDLE
|
5787 SD_SHARE_PKG_RESOURCES
)) {
5788 if (sd
->groups
!= sd
->groups
->next
)
5792 /* Following flags don't use groups */
5793 if (sd
->flags
& (SD_WAKE_IDLE
|
5802 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5804 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5806 if (sd_degenerate(parent
))
5809 if (!cpus_equal(sd
->span
, parent
->span
))
5812 /* Does parent contain flags not in child? */
5813 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5814 if (cflags
& SD_WAKE_AFFINE
)
5815 pflags
&= ~SD_WAKE_BALANCE
;
5816 /* Flags needing groups don't count if only 1 group in parent */
5817 if (parent
->groups
== parent
->groups
->next
) {
5818 pflags
&= ~(SD_LOAD_BALANCE
|
5819 SD_BALANCE_NEWIDLE
|
5823 SD_SHARE_PKG_RESOURCES
);
5825 if (~cflags
& pflags
)
5832 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5833 * hold the hotplug lock.
5835 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5837 struct rq
*rq
= cpu_rq(cpu
);
5838 struct sched_domain
*tmp
;
5840 /* Remove the sched domains which do not contribute to scheduling. */
5841 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5842 struct sched_domain
*parent
= tmp
->parent
;
5845 if (sd_parent_degenerate(tmp
, parent
)) {
5846 tmp
->parent
= parent
->parent
;
5848 parent
->parent
->child
= tmp
;
5852 if (sd
&& sd_degenerate(sd
)) {
5858 sched_domain_debug(sd
, cpu
);
5860 rcu_assign_pointer(rq
->sd
, sd
);
5863 /* cpus with isolated domains */
5864 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5866 /* Setup the mask of cpus configured for isolated domains */
5867 static int __init
isolated_cpu_setup(char *str
)
5869 int ints
[NR_CPUS
], i
;
5871 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5872 cpus_clear(cpu_isolated_map
);
5873 for (i
= 1; i
<= ints
[0]; i
++)
5874 if (ints
[i
] < NR_CPUS
)
5875 cpu_set(ints
[i
], cpu_isolated_map
);
5879 __setup("isolcpus=", isolated_cpu_setup
);
5882 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5883 * to a function which identifies what group(along with sched group) a CPU
5884 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5885 * (due to the fact that we keep track of groups covered with a cpumask_t).
5887 * init_sched_build_groups will build a circular linked list of the groups
5888 * covered by the given span, and will set each group's ->cpumask correctly,
5889 * and ->cpu_power to 0.
5892 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5893 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5894 struct sched_group
**sg
))
5896 struct sched_group
*first
= NULL
, *last
= NULL
;
5897 cpumask_t covered
= CPU_MASK_NONE
;
5900 for_each_cpu_mask(i
, span
) {
5901 struct sched_group
*sg
;
5902 int group
= group_fn(i
, cpu_map
, &sg
);
5905 if (cpu_isset(i
, covered
))
5908 sg
->cpumask
= CPU_MASK_NONE
;
5909 sg
->__cpu_power
= 0;
5911 for_each_cpu_mask(j
, span
) {
5912 if (group_fn(j
, cpu_map
, NULL
) != group
)
5915 cpu_set(j
, covered
);
5916 cpu_set(j
, sg
->cpumask
);
5927 #define SD_NODES_PER_DOMAIN 16
5932 * find_next_best_node - find the next node to include in a sched_domain
5933 * @node: node whose sched_domain we're building
5934 * @used_nodes: nodes already in the sched_domain
5936 * Find the next node to include in a given scheduling domain. Simply
5937 * finds the closest node not already in the @used_nodes map.
5939 * Should use nodemask_t.
5941 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5943 int i
, n
, val
, min_val
, best_node
= 0;
5947 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5948 /* Start at @node */
5949 n
= (node
+ i
) % MAX_NUMNODES
;
5951 if (!nr_cpus_node(n
))
5954 /* Skip already used nodes */
5955 if (test_bit(n
, used_nodes
))
5958 /* Simple min distance search */
5959 val
= node_distance(node
, n
);
5961 if (val
< min_val
) {
5967 set_bit(best_node
, used_nodes
);
5972 * sched_domain_node_span - get a cpumask for a node's sched_domain
5973 * @node: node whose cpumask we're constructing
5974 * @size: number of nodes to include in this span
5976 * Given a node, construct a good cpumask for its sched_domain to span. It
5977 * should be one that prevents unnecessary balancing, but also spreads tasks
5980 static cpumask_t
sched_domain_node_span(int node
)
5982 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5983 cpumask_t span
, nodemask
;
5987 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5989 nodemask
= node_to_cpumask(node
);
5990 cpus_or(span
, span
, nodemask
);
5991 set_bit(node
, used_nodes
);
5993 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5994 int next_node
= find_next_best_node(node
, used_nodes
);
5996 nodemask
= node_to_cpumask(next_node
);
5997 cpus_or(span
, span
, nodemask
);
6004 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6007 * SMT sched-domains:
6009 #ifdef CONFIG_SCHED_SMT
6010 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6011 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6014 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6017 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6023 * multi-core sched-domains:
6025 #ifdef CONFIG_SCHED_MC
6026 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6027 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6030 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6032 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6035 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6036 cpus_and(mask
, mask
, *cpu_map
);
6037 group
= first_cpu(mask
);
6039 *sg
= &per_cpu(sched_group_core
, group
);
6042 #elif defined(CONFIG_SCHED_MC)
6044 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6047 *sg
= &per_cpu(sched_group_core
, cpu
);
6052 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6053 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6056 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6059 #ifdef CONFIG_SCHED_MC
6060 cpumask_t mask
= cpu_coregroup_map(cpu
);
6061 cpus_and(mask
, mask
, *cpu_map
);
6062 group
= first_cpu(mask
);
6063 #elif defined(CONFIG_SCHED_SMT)
6064 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6065 cpus_and(mask
, mask
, *cpu_map
);
6066 group
= first_cpu(mask
);
6071 *sg
= &per_cpu(sched_group_phys
, group
);
6077 * The init_sched_build_groups can't handle what we want to do with node
6078 * groups, so roll our own. Now each node has its own list of groups which
6079 * gets dynamically allocated.
6081 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6082 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6084 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6085 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6087 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6088 struct sched_group
**sg
)
6090 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6093 cpus_and(nodemask
, nodemask
, *cpu_map
);
6094 group
= first_cpu(nodemask
);
6097 *sg
= &per_cpu(sched_group_allnodes
, group
);
6101 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6103 struct sched_group
*sg
= group_head
;
6109 for_each_cpu_mask(j
, sg
->cpumask
) {
6110 struct sched_domain
*sd
;
6112 sd
= &per_cpu(phys_domains
, j
);
6113 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6115 * Only add "power" once for each
6121 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6124 } while (sg
!= group_head
);
6129 /* Free memory allocated for various sched_group structures */
6130 static void free_sched_groups(const cpumask_t
*cpu_map
)
6134 for_each_cpu_mask(cpu
, *cpu_map
) {
6135 struct sched_group
**sched_group_nodes
6136 = sched_group_nodes_bycpu
[cpu
];
6138 if (!sched_group_nodes
)
6141 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6142 cpumask_t nodemask
= node_to_cpumask(i
);
6143 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6145 cpus_and(nodemask
, nodemask
, *cpu_map
);
6146 if (cpus_empty(nodemask
))
6156 if (oldsg
!= sched_group_nodes
[i
])
6159 kfree(sched_group_nodes
);
6160 sched_group_nodes_bycpu
[cpu
] = NULL
;
6164 static void free_sched_groups(const cpumask_t
*cpu_map
)
6170 * Initialize sched groups cpu_power.
6172 * cpu_power indicates the capacity of sched group, which is used while
6173 * distributing the load between different sched groups in a sched domain.
6174 * Typically cpu_power for all the groups in a sched domain will be same unless
6175 * there are asymmetries in the topology. If there are asymmetries, group
6176 * having more cpu_power will pickup more load compared to the group having
6179 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6180 * the maximum number of tasks a group can handle in the presence of other idle
6181 * or lightly loaded groups in the same sched domain.
6183 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6185 struct sched_domain
*child
;
6186 struct sched_group
*group
;
6188 WARN_ON(!sd
|| !sd
->groups
);
6190 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6195 sd
->groups
->__cpu_power
= 0;
6198 * For perf policy, if the groups in child domain share resources
6199 * (for example cores sharing some portions of the cache hierarchy
6200 * or SMT), then set this domain groups cpu_power such that each group
6201 * can handle only one task, when there are other idle groups in the
6202 * same sched domain.
6204 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6206 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6207 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6212 * add cpu_power of each child group to this groups cpu_power
6214 group
= child
->groups
;
6216 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6217 group
= group
->next
;
6218 } while (group
!= child
->groups
);
6222 * Build sched domains for a given set of cpus and attach the sched domains
6223 * to the individual cpus
6225 static int build_sched_domains(const cpumask_t
*cpu_map
)
6229 struct sched_group
**sched_group_nodes
= NULL
;
6230 int sd_allnodes
= 0;
6233 * Allocate the per-node list of sched groups
6235 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6237 if (!sched_group_nodes
) {
6238 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6241 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6245 * Set up domains for cpus specified by the cpu_map.
6247 for_each_cpu_mask(i
, *cpu_map
) {
6248 struct sched_domain
*sd
= NULL
, *p
;
6249 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6251 cpus_and(nodemask
, nodemask
, *cpu_map
);
6254 if (cpus_weight(*cpu_map
) >
6255 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6256 sd
= &per_cpu(allnodes_domains
, i
);
6257 *sd
= SD_ALLNODES_INIT
;
6258 sd
->span
= *cpu_map
;
6259 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6265 sd
= &per_cpu(node_domains
, i
);
6267 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6271 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6275 sd
= &per_cpu(phys_domains
, i
);
6277 sd
->span
= nodemask
;
6281 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6283 #ifdef CONFIG_SCHED_MC
6285 sd
= &per_cpu(core_domains
, i
);
6287 sd
->span
= cpu_coregroup_map(i
);
6288 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6291 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6294 #ifdef CONFIG_SCHED_SMT
6296 sd
= &per_cpu(cpu_domains
, i
);
6297 *sd
= SD_SIBLING_INIT
;
6298 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6299 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6302 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6306 #ifdef CONFIG_SCHED_SMT
6307 /* Set up CPU (sibling) groups */
6308 for_each_cpu_mask(i
, *cpu_map
) {
6309 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6310 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6311 if (i
!= first_cpu(this_sibling_map
))
6314 init_sched_build_groups(this_sibling_map
, cpu_map
,
6319 #ifdef CONFIG_SCHED_MC
6320 /* Set up multi-core groups */
6321 for_each_cpu_mask(i
, *cpu_map
) {
6322 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6323 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6324 if (i
!= first_cpu(this_core_map
))
6326 init_sched_build_groups(this_core_map
, cpu_map
,
6327 &cpu_to_core_group
);
6331 /* Set up physical groups */
6332 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6333 cpumask_t nodemask
= node_to_cpumask(i
);
6335 cpus_and(nodemask
, nodemask
, *cpu_map
);
6336 if (cpus_empty(nodemask
))
6339 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6343 /* Set up node groups */
6345 init_sched_build_groups(*cpu_map
, cpu_map
,
6346 &cpu_to_allnodes_group
);
6348 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6349 /* Set up node groups */
6350 struct sched_group
*sg
, *prev
;
6351 cpumask_t nodemask
= node_to_cpumask(i
);
6352 cpumask_t domainspan
;
6353 cpumask_t covered
= CPU_MASK_NONE
;
6356 cpus_and(nodemask
, nodemask
, *cpu_map
);
6357 if (cpus_empty(nodemask
)) {
6358 sched_group_nodes
[i
] = NULL
;
6362 domainspan
= sched_domain_node_span(i
);
6363 cpus_and(domainspan
, domainspan
, *cpu_map
);
6365 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6367 printk(KERN_WARNING
"Can not alloc domain group for "
6371 sched_group_nodes
[i
] = sg
;
6372 for_each_cpu_mask(j
, nodemask
) {
6373 struct sched_domain
*sd
;
6375 sd
= &per_cpu(node_domains
, j
);
6378 sg
->__cpu_power
= 0;
6379 sg
->cpumask
= nodemask
;
6381 cpus_or(covered
, covered
, nodemask
);
6384 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6385 cpumask_t tmp
, notcovered
;
6386 int n
= (i
+ j
) % MAX_NUMNODES
;
6388 cpus_complement(notcovered
, covered
);
6389 cpus_and(tmp
, notcovered
, *cpu_map
);
6390 cpus_and(tmp
, tmp
, domainspan
);
6391 if (cpus_empty(tmp
))
6394 nodemask
= node_to_cpumask(n
);
6395 cpus_and(tmp
, tmp
, nodemask
);
6396 if (cpus_empty(tmp
))
6399 sg
= kmalloc_node(sizeof(struct sched_group
),
6403 "Can not alloc domain group for node %d\n", j
);
6406 sg
->__cpu_power
= 0;
6408 sg
->next
= prev
->next
;
6409 cpus_or(covered
, covered
, tmp
);
6416 /* Calculate CPU power for physical packages and nodes */
6417 #ifdef CONFIG_SCHED_SMT
6418 for_each_cpu_mask(i
, *cpu_map
) {
6419 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6421 init_sched_groups_power(i
, sd
);
6424 #ifdef CONFIG_SCHED_MC
6425 for_each_cpu_mask(i
, *cpu_map
) {
6426 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6428 init_sched_groups_power(i
, sd
);
6432 for_each_cpu_mask(i
, *cpu_map
) {
6433 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6435 init_sched_groups_power(i
, sd
);
6439 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6440 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6443 struct sched_group
*sg
;
6445 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6446 init_numa_sched_groups_power(sg
);
6450 /* Attach the domains */
6451 for_each_cpu_mask(i
, *cpu_map
) {
6452 struct sched_domain
*sd
;
6453 #ifdef CONFIG_SCHED_SMT
6454 sd
= &per_cpu(cpu_domains
, i
);
6455 #elif defined(CONFIG_SCHED_MC)
6456 sd
= &per_cpu(core_domains
, i
);
6458 sd
= &per_cpu(phys_domains
, i
);
6460 cpu_attach_domain(sd
, i
);
6467 free_sched_groups(cpu_map
);
6472 static cpumask_t
*doms_cur
; /* current sched domains */
6473 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6476 * Special case: If a kmalloc of a doms_cur partition (array of
6477 * cpumask_t) fails, then fallback to a single sched domain,
6478 * as determined by the single cpumask_t fallback_doms.
6480 static cpumask_t fallback_doms
;
6483 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6484 * For now this just excludes isolated cpus, but could be used to
6485 * exclude other special cases in the future.
6487 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6492 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6494 doms_cur
= &fallback_doms
;
6495 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6496 err
= build_sched_domains(doms_cur
);
6497 register_sched_domain_sysctl();
6502 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6504 free_sched_groups(cpu_map
);
6508 * Detach sched domains from a group of cpus specified in cpu_map
6509 * These cpus will now be attached to the NULL domain
6511 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6515 unregister_sched_domain_sysctl();
6517 for_each_cpu_mask(i
, *cpu_map
)
6518 cpu_attach_domain(NULL
, i
);
6519 synchronize_sched();
6520 arch_destroy_sched_domains(cpu_map
);
6524 * Partition sched domains as specified by the 'ndoms_new'
6525 * cpumasks in the array doms_new[] of cpumasks. This compares
6526 * doms_new[] to the current sched domain partitioning, doms_cur[].
6527 * It destroys each deleted domain and builds each new domain.
6529 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6530 * The masks don't intersect (don't overlap.) We should setup one
6531 * sched domain for each mask. CPUs not in any of the cpumasks will
6532 * not be load balanced. If the same cpumask appears both in the
6533 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6536 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6537 * ownership of it and will kfree it when done with it. If the caller
6538 * failed the kmalloc call, then it can pass in doms_new == NULL,
6539 * and partition_sched_domains() will fallback to the single partition
6542 * Call with hotplug lock held
6544 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6548 /* always unregister in case we don't destroy any domains */
6549 unregister_sched_domain_sysctl();
6551 if (doms_new
== NULL
) {
6553 doms_new
= &fallback_doms
;
6554 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6557 /* Destroy deleted domains */
6558 for (i
= 0; i
< ndoms_cur
; i
++) {
6559 for (j
= 0; j
< ndoms_new
; j
++) {
6560 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
6563 /* no match - a current sched domain not in new doms_new[] */
6564 detach_destroy_domains(doms_cur
+ i
);
6569 /* Build new domains */
6570 for (i
= 0; i
< ndoms_new
; i
++) {
6571 for (j
= 0; j
< ndoms_cur
; j
++) {
6572 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
6575 /* no match - add a new doms_new */
6576 build_sched_domains(doms_new
+ i
);
6581 /* Remember the new sched domains */
6582 if (doms_cur
!= &fallback_doms
)
6584 doms_cur
= doms_new
;
6585 ndoms_cur
= ndoms_new
;
6587 register_sched_domain_sysctl();
6590 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6591 static int arch_reinit_sched_domains(void)
6595 mutex_lock(&sched_hotcpu_mutex
);
6596 detach_destroy_domains(&cpu_online_map
);
6597 err
= arch_init_sched_domains(&cpu_online_map
);
6598 mutex_unlock(&sched_hotcpu_mutex
);
6603 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6607 if (buf
[0] != '0' && buf
[0] != '1')
6611 sched_smt_power_savings
= (buf
[0] == '1');
6613 sched_mc_power_savings
= (buf
[0] == '1');
6615 ret
= arch_reinit_sched_domains();
6617 return ret
? ret
: count
;
6620 #ifdef CONFIG_SCHED_MC
6621 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6623 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6625 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6626 const char *buf
, size_t count
)
6628 return sched_power_savings_store(buf
, count
, 0);
6630 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6631 sched_mc_power_savings_store
);
6634 #ifdef CONFIG_SCHED_SMT
6635 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6637 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6639 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6640 const char *buf
, size_t count
)
6642 return sched_power_savings_store(buf
, count
, 1);
6644 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6645 sched_smt_power_savings_store
);
6648 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6652 #ifdef CONFIG_SCHED_SMT
6654 err
= sysfs_create_file(&cls
->kset
.kobj
,
6655 &attr_sched_smt_power_savings
.attr
);
6657 #ifdef CONFIG_SCHED_MC
6658 if (!err
&& mc_capable())
6659 err
= sysfs_create_file(&cls
->kset
.kobj
,
6660 &attr_sched_mc_power_savings
.attr
);
6667 * Force a reinitialization of the sched domains hierarchy. The domains
6668 * and groups cannot be updated in place without racing with the balancing
6669 * code, so we temporarily attach all running cpus to the NULL domain
6670 * which will prevent rebalancing while the sched domains are recalculated.
6672 static int update_sched_domains(struct notifier_block
*nfb
,
6673 unsigned long action
, void *hcpu
)
6676 case CPU_UP_PREPARE
:
6677 case CPU_UP_PREPARE_FROZEN
:
6678 case CPU_DOWN_PREPARE
:
6679 case CPU_DOWN_PREPARE_FROZEN
:
6680 detach_destroy_domains(&cpu_online_map
);
6683 case CPU_UP_CANCELED
:
6684 case CPU_UP_CANCELED_FROZEN
:
6685 case CPU_DOWN_FAILED
:
6686 case CPU_DOWN_FAILED_FROZEN
:
6688 case CPU_ONLINE_FROZEN
:
6690 case CPU_DEAD_FROZEN
:
6692 * Fall through and re-initialise the domains.
6699 /* The hotplug lock is already held by cpu_up/cpu_down */
6700 arch_init_sched_domains(&cpu_online_map
);
6705 void __init
sched_init_smp(void)
6707 cpumask_t non_isolated_cpus
;
6709 mutex_lock(&sched_hotcpu_mutex
);
6710 arch_init_sched_domains(&cpu_online_map
);
6711 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6712 if (cpus_empty(non_isolated_cpus
))
6713 cpu_set(smp_processor_id(), non_isolated_cpus
);
6714 mutex_unlock(&sched_hotcpu_mutex
);
6715 /* XXX: Theoretical race here - CPU may be hotplugged now */
6716 hotcpu_notifier(update_sched_domains
, 0);
6718 /* Move init over to a non-isolated CPU */
6719 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6721 sched_init_granularity();
6724 void __init
sched_init_smp(void)
6726 sched_init_granularity();
6728 #endif /* CONFIG_SMP */
6730 int in_sched_functions(unsigned long addr
)
6732 return in_lock_functions(addr
) ||
6733 (addr
>= (unsigned long)__sched_text_start
6734 && addr
< (unsigned long)__sched_text_end
);
6737 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6739 cfs_rq
->tasks_timeline
= RB_ROOT
;
6740 #ifdef CONFIG_FAIR_GROUP_SCHED
6743 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6746 void __init
sched_init(void)
6748 int highest_cpu
= 0;
6751 for_each_possible_cpu(i
) {
6752 struct rt_prio_array
*array
;
6756 spin_lock_init(&rq
->lock
);
6757 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6760 init_cfs_rq(&rq
->cfs
, rq
);
6761 #ifdef CONFIG_FAIR_GROUP_SCHED
6762 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6764 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6765 struct sched_entity
*se
=
6766 &per_cpu(init_sched_entity
, i
);
6768 init_cfs_rq_p
[i
] = cfs_rq
;
6769 init_cfs_rq(cfs_rq
, rq
);
6770 cfs_rq
->tg
= &init_task_group
;
6771 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6772 &rq
->leaf_cfs_rq_list
);
6774 init_sched_entity_p
[i
] = se
;
6775 se
->cfs_rq
= &rq
->cfs
;
6777 se
->load
.weight
= init_task_group_load
;
6778 se
->load
.inv_weight
=
6779 div64_64(1ULL<<32, init_task_group_load
);
6782 init_task_group
.shares
= init_task_group_load
;
6783 spin_lock_init(&init_task_group
.lock
);
6786 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6787 rq
->cpu_load
[j
] = 0;
6790 rq
->active_balance
= 0;
6791 rq
->next_balance
= jiffies
;
6794 rq
->migration_thread
= NULL
;
6795 INIT_LIST_HEAD(&rq
->migration_queue
);
6797 atomic_set(&rq
->nr_iowait
, 0);
6799 array
= &rq
->rt
.active
;
6800 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6801 INIT_LIST_HEAD(array
->queue
+ j
);
6802 __clear_bit(j
, array
->bitmap
);
6805 /* delimiter for bitsearch: */
6806 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6809 set_load_weight(&init_task
);
6811 #ifdef CONFIG_PREEMPT_NOTIFIERS
6812 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6816 nr_cpu_ids
= highest_cpu
+ 1;
6817 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6820 #ifdef CONFIG_RT_MUTEXES
6821 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6825 * The boot idle thread does lazy MMU switching as well:
6827 atomic_inc(&init_mm
.mm_count
);
6828 enter_lazy_tlb(&init_mm
, current
);
6831 * Make us the idle thread. Technically, schedule() should not be
6832 * called from this thread, however somewhere below it might be,
6833 * but because we are the idle thread, we just pick up running again
6834 * when this runqueue becomes "idle".
6836 init_idle(current
, smp_processor_id());
6838 * During early bootup we pretend to be a normal task:
6840 current
->sched_class
= &fair_sched_class
;
6843 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6844 void __might_sleep(char *file
, int line
)
6847 static unsigned long prev_jiffy
; /* ratelimiting */
6849 if ((in_atomic() || irqs_disabled()) &&
6850 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6851 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6853 prev_jiffy
= jiffies
;
6854 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6855 " context at %s:%d\n", file
, line
);
6856 printk("in_atomic():%d, irqs_disabled():%d\n",
6857 in_atomic(), irqs_disabled());
6858 debug_show_held_locks(current
);
6859 if (irqs_disabled())
6860 print_irqtrace_events(current
);
6865 EXPORT_SYMBOL(__might_sleep
);
6868 #ifdef CONFIG_MAGIC_SYSRQ
6869 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6872 update_rq_clock(rq
);
6873 on_rq
= p
->se
.on_rq
;
6875 deactivate_task(rq
, p
, 0);
6876 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6878 activate_task(rq
, p
, 0);
6879 resched_task(rq
->curr
);
6883 void normalize_rt_tasks(void)
6885 struct task_struct
*g
, *p
;
6886 unsigned long flags
;
6889 read_lock_irq(&tasklist_lock
);
6890 do_each_thread(g
, p
) {
6892 * Only normalize user tasks:
6897 p
->se
.exec_start
= 0;
6898 #ifdef CONFIG_SCHEDSTATS
6899 p
->se
.wait_start
= 0;
6900 p
->se
.sleep_start
= 0;
6901 p
->se
.block_start
= 0;
6903 task_rq(p
)->clock
= 0;
6907 * Renice negative nice level userspace
6910 if (TASK_NICE(p
) < 0 && p
->mm
)
6911 set_user_nice(p
, 0);
6915 spin_lock_irqsave(&p
->pi_lock
, flags
);
6916 rq
= __task_rq_lock(p
);
6918 normalize_task(rq
, p
);
6920 __task_rq_unlock(rq
);
6921 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6922 } while_each_thread(g
, p
);
6924 read_unlock_irq(&tasklist_lock
);
6927 #endif /* CONFIG_MAGIC_SYSRQ */
6931 * These functions are only useful for the IA64 MCA handling.
6933 * They can only be called when the whole system has been
6934 * stopped - every CPU needs to be quiescent, and no scheduling
6935 * activity can take place. Using them for anything else would
6936 * be a serious bug, and as a result, they aren't even visible
6937 * under any other configuration.
6941 * curr_task - return the current task for a given cpu.
6942 * @cpu: the processor in question.
6944 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6946 struct task_struct
*curr_task(int cpu
)
6948 return cpu_curr(cpu
);
6952 * set_curr_task - set the current task for a given cpu.
6953 * @cpu: the processor in question.
6954 * @p: the task pointer to set.
6956 * Description: This function must only be used when non-maskable interrupts
6957 * are serviced on a separate stack. It allows the architecture to switch the
6958 * notion of the current task on a cpu in a non-blocking manner. This function
6959 * must be called with all CPU's synchronized, and interrupts disabled, the
6960 * and caller must save the original value of the current task (see
6961 * curr_task() above) and restore that value before reenabling interrupts and
6962 * re-starting the system.
6964 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6966 void set_curr_task(int cpu
, struct task_struct
*p
)
6973 #ifdef CONFIG_FAIR_GROUP_SCHED
6975 /* allocate runqueue etc for a new task group */
6976 struct task_group
*sched_create_group(void)
6978 struct task_group
*tg
;
6979 struct cfs_rq
*cfs_rq
;
6980 struct sched_entity
*se
;
6984 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6986 return ERR_PTR(-ENOMEM
);
6988 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
6991 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
6995 for_each_possible_cpu(i
) {
6998 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
7003 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
7008 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
7009 memset(se
, 0, sizeof(struct sched_entity
));
7011 tg
->cfs_rq
[i
] = cfs_rq
;
7012 init_cfs_rq(cfs_rq
, rq
);
7016 se
->cfs_rq
= &rq
->cfs
;
7018 se
->load
.weight
= NICE_0_LOAD
;
7019 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
7023 for_each_possible_cpu(i
) {
7025 cfs_rq
= tg
->cfs_rq
[i
];
7026 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7029 tg
->shares
= NICE_0_LOAD
;
7030 spin_lock_init(&tg
->lock
);
7035 for_each_possible_cpu(i
) {
7037 kfree(tg
->cfs_rq
[i
]);
7045 return ERR_PTR(-ENOMEM
);
7048 /* rcu callback to free various structures associated with a task group */
7049 static void free_sched_group(struct rcu_head
*rhp
)
7051 struct task_group
*tg
= container_of(rhp
, struct task_group
, rcu
);
7052 struct cfs_rq
*cfs_rq
;
7053 struct sched_entity
*se
;
7056 /* now it should be safe to free those cfs_rqs */
7057 for_each_possible_cpu(i
) {
7058 cfs_rq
= tg
->cfs_rq
[i
];
7070 /* Destroy runqueue etc associated with a task group */
7071 void sched_destroy_group(struct task_group
*tg
)
7073 struct cfs_rq
*cfs_rq
= NULL
;
7076 for_each_possible_cpu(i
) {
7077 cfs_rq
= tg
->cfs_rq
[i
];
7078 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7083 /* wait for possible concurrent references to cfs_rqs complete */
7084 call_rcu(&tg
->rcu
, free_sched_group
);
7087 /* change task's runqueue when it moves between groups.
7088 * The caller of this function should have put the task in its new group
7089 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7090 * reflect its new group.
7092 void sched_move_task(struct task_struct
*tsk
)
7095 unsigned long flags
;
7098 rq
= task_rq_lock(tsk
, &flags
);
7100 if (tsk
->sched_class
!= &fair_sched_class
) {
7101 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7105 update_rq_clock(rq
);
7107 running
= task_running(rq
, tsk
);
7108 on_rq
= tsk
->se
.on_rq
;
7111 dequeue_task(rq
, tsk
, 0);
7112 if (unlikely(running
))
7113 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7116 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7119 if (unlikely(running
))
7120 tsk
->sched_class
->set_curr_task(rq
);
7121 enqueue_task(rq
, tsk
, 0);
7125 task_rq_unlock(rq
, &flags
);
7128 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7130 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7131 struct rq
*rq
= cfs_rq
->rq
;
7134 spin_lock_irq(&rq
->lock
);
7138 dequeue_entity(cfs_rq
, se
, 0);
7140 se
->load
.weight
= shares
;
7141 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7144 enqueue_entity(cfs_rq
, se
, 0);
7146 spin_unlock_irq(&rq
->lock
);
7149 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7153 spin_lock(&tg
->lock
);
7154 if (tg
->shares
== shares
)
7157 tg
->shares
= shares
;
7158 for_each_possible_cpu(i
)
7159 set_se_shares(tg
->se
[i
], shares
);
7162 spin_unlock(&tg
->lock
);
7166 unsigned long sched_group_shares(struct task_group
*tg
)
7171 #endif /* CONFIG_FAIR_GROUP_SCHED */
7173 #ifdef CONFIG_FAIR_CGROUP_SCHED
7175 /* return corresponding task_group object of a cgroup */
7176 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7178 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7179 struct task_group
, css
);
7182 static struct cgroup_subsys_state
*
7183 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7185 struct task_group
*tg
;
7187 if (!cgrp
->parent
) {
7188 /* This is early initialization for the top cgroup */
7189 init_task_group
.css
.cgroup
= cgrp
;
7190 return &init_task_group
.css
;
7193 /* we support only 1-level deep hierarchical scheduler atm */
7194 if (cgrp
->parent
->parent
)
7195 return ERR_PTR(-EINVAL
);
7197 tg
= sched_create_group();
7199 return ERR_PTR(-ENOMEM
);
7201 /* Bind the cgroup to task_group object we just created */
7202 tg
->css
.cgroup
= cgrp
;
7208 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7210 struct task_group
*tg
= cgroup_tg(cgrp
);
7212 sched_destroy_group(tg
);
7216 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7217 struct task_struct
*tsk
)
7219 /* We don't support RT-tasks being in separate groups */
7220 if (tsk
->sched_class
!= &fair_sched_class
)
7227 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7228 struct cgroup
*old_cont
, struct task_struct
*tsk
)
7230 sched_move_task(tsk
);
7233 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7236 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7239 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7241 struct task_group
*tg
= cgroup_tg(cgrp
);
7243 return (u64
) tg
->shares
;
7246 static struct cftype cpu_files
[] = {
7249 .read_uint
= cpu_shares_read_uint
,
7250 .write_uint
= cpu_shares_write_uint
,
7254 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7256 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7259 struct cgroup_subsys cpu_cgroup_subsys
= {
7261 .create
= cpu_cgroup_create
,
7262 .destroy
= cpu_cgroup_destroy
,
7263 .can_attach
= cpu_cgroup_can_attach
,
7264 .attach
= cpu_cgroup_attach
,
7265 .populate
= cpu_cgroup_populate
,
7266 .subsys_id
= cpu_cgroup_subsys_id
,
7270 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7272 #ifdef CONFIG_CGROUP_CPUACCT
7275 * CPU accounting code for task groups.
7277 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7278 * (balbir@in.ibm.com).
7281 /* track cpu usage of a group of tasks */
7283 struct cgroup_subsys_state css
;
7284 /* cpuusage holds pointer to a u64-type object on every cpu */
7288 struct cgroup_subsys cpuacct_subsys
;
7290 /* return cpu accounting group corresponding to this container */
7291 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
7293 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
7294 struct cpuacct
, css
);
7297 /* return cpu accounting group to which this task belongs */
7298 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
7300 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
7301 struct cpuacct
, css
);
7304 /* create a new cpu accounting group */
7305 static struct cgroup_subsys_state
*cpuacct_create(
7306 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7308 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7311 return ERR_PTR(-ENOMEM
);
7313 ca
->cpuusage
= alloc_percpu(u64
);
7314 if (!ca
->cpuusage
) {
7316 return ERR_PTR(-ENOMEM
);
7322 /* destroy an existing cpu accounting group */
7324 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7326 struct cpuacct
*ca
= cgroup_ca(cont
);
7328 free_percpu(ca
->cpuusage
);
7332 /* return total cpu usage (in nanoseconds) of a group */
7333 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
7335 struct cpuacct
*ca
= cgroup_ca(cont
);
7336 u64 totalcpuusage
= 0;
7339 for_each_possible_cpu(i
) {
7340 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
7343 * Take rq->lock to make 64-bit addition safe on 32-bit
7346 spin_lock_irq(&cpu_rq(i
)->lock
);
7347 totalcpuusage
+= *cpuusage
;
7348 spin_unlock_irq(&cpu_rq(i
)->lock
);
7351 return totalcpuusage
;
7354 static struct cftype files
[] = {
7357 .read_uint
= cpuusage_read
,
7361 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7363 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
7367 * charge this task's execution time to its accounting group.
7369 * called with rq->lock held.
7371 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
7375 if (!cpuacct_subsys
.active
)
7380 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
7382 *cpuusage
+= cputime
;
7386 struct cgroup_subsys cpuacct_subsys
= {
7388 .create
= cpuacct_create
,
7389 .destroy
= cpuacct_destroy
,
7390 .populate
= cpuacct_populate
,
7391 .subsys_id
= cpuacct_subsys_id
,
7393 #endif /* CONFIG_CGROUP_CPUACCT */