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/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
64 #include <linux/pagemap.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __attribute__((weak
)) sched_clock(void)
75 return (unsigned long long)jiffies
* (1000000000 / HZ
);
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Some helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
100 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
109 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
110 * Timeslices get refilled after they expire.
112 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
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 #define SCALE_PRIO(x, prio) \
137 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
140 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
141 * to time slice values: [800ms ... 100ms ... 5ms]
143 static unsigned int static_prio_timeslice(int static_prio
)
145 if (static_prio
== NICE_TO_PRIO(19))
148 if (static_prio
< NICE_TO_PRIO(0))
149 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
151 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
154 static inline int rt_policy(int policy
)
156 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
161 static inline int task_has_rt_policy(struct task_struct
*p
)
163 return rt_policy(p
->policy
);
167 * This is the priority-queue data structure of the RT scheduling class:
169 struct rt_prio_array
{
170 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
171 struct list_head queue
[MAX_RT_PRIO
];
174 /* CFS-related fields in a runqueue */
176 struct load_weight load
;
177 unsigned long nr_running
;
182 struct rb_root tasks_timeline
;
183 struct rb_node
*rb_leftmost
;
184 struct rb_node
*rb_load_balance_curr
;
185 /* 'curr' points to currently running entity on this cfs_rq.
186 * It is set to NULL otherwise (i.e when none are currently running).
188 struct sched_entity
*curr
;
189 #ifdef CONFIG_FAIR_GROUP_SCHED
190 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
192 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
193 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
194 * (like users, containers etc.)
196 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
197 * list is used during load balance.
199 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
203 /* Real-Time classes' related field in a runqueue: */
205 struct rt_prio_array active
;
206 int rt_load_balance_idx
;
207 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
211 * This is the main, per-CPU runqueue data structure.
213 * Locking rule: those places that want to lock multiple runqueues
214 * (such as the load balancing or the thread migration code), lock
215 * acquire operations must be ordered by ascending &runqueue.
218 spinlock_t lock
; /* runqueue lock */
221 * nr_running and cpu_load should be in the same cacheline because
222 * remote CPUs use both these fields when doing load calculation.
224 unsigned long nr_running
;
225 #define CPU_LOAD_IDX_MAX 5
226 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
227 unsigned char idle_at_tick
;
229 unsigned char in_nohz_recently
;
231 struct load_weight load
; /* capture load from *all* tasks on this cpu */
232 unsigned long nr_load_updates
;
236 #ifdef CONFIG_FAIR_GROUP_SCHED
237 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
242 * This is part of a global counter where only the total sum
243 * over all CPUs matters. A task can increase this counter on
244 * one CPU and if it got migrated afterwards it may decrease
245 * it on another CPU. Always updated under the runqueue lock:
247 unsigned long nr_uninterruptible
;
249 struct task_struct
*curr
, *idle
;
250 unsigned long next_balance
;
251 struct mm_struct
*prev_mm
;
253 u64 clock
, prev_clock_raw
;
256 unsigned int clock_warps
, clock_overflows
;
258 unsigned int clock_deep_idle_events
;
264 struct sched_domain
*sd
;
266 /* For active balancing */
269 int cpu
; /* cpu of this runqueue */
271 struct task_struct
*migration_thread
;
272 struct list_head migration_queue
;
275 #ifdef CONFIG_SCHEDSTATS
277 struct sched_info rq_sched_info
;
279 /* sys_sched_yield() stats */
280 unsigned long yld_exp_empty
;
281 unsigned long yld_act_empty
;
282 unsigned long yld_both_empty
;
283 unsigned long yld_cnt
;
285 /* schedule() stats */
286 unsigned long sched_switch
;
287 unsigned long sched_cnt
;
288 unsigned long sched_goidle
;
290 /* try_to_wake_up() stats */
291 unsigned long ttwu_cnt
;
292 unsigned long ttwu_local
;
294 struct lock_class_key rq_lock_key
;
297 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
298 static DEFINE_MUTEX(sched_hotcpu_mutex
);
300 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
302 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
305 static inline int cpu_of(struct rq
*rq
)
315 * Update the per-runqueue clock, as finegrained as the platform can give
316 * us, but without assuming monotonicity, etc.:
318 static void __update_rq_clock(struct rq
*rq
)
320 u64 prev_raw
= rq
->prev_clock_raw
;
321 u64 now
= sched_clock();
322 s64 delta
= now
- prev_raw
;
323 u64 clock
= rq
->clock
;
325 #ifdef CONFIG_SCHED_DEBUG
326 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
329 * Protect against sched_clock() occasionally going backwards:
331 if (unlikely(delta
< 0)) {
336 * Catch too large forward jumps too:
338 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
339 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
340 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
343 rq
->clock_overflows
++;
345 if (unlikely(delta
> rq
->clock_max_delta
))
346 rq
->clock_max_delta
= delta
;
351 rq
->prev_clock_raw
= now
;
355 static void update_rq_clock(struct rq
*rq
)
357 if (likely(smp_processor_id() == cpu_of(rq
)))
358 __update_rq_clock(rq
);
362 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
363 * See detach_destroy_domains: synchronize_sched for details.
365 * The domain tree of any CPU may only be accessed from within
366 * preempt-disabled sections.
368 #define for_each_domain(cpu, __sd) \
369 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
371 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
372 #define this_rq() (&__get_cpu_var(runqueues))
373 #define task_rq(p) cpu_rq(task_cpu(p))
374 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
377 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
379 #ifdef CONFIG_SCHED_DEBUG
380 # define const_debug __read_mostly
382 # define const_debug static const
386 * Debugging: various feature bits
389 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
390 SCHED_FEAT_START_DEBIT
= 2,
391 SCHED_FEAT_USE_TREE_AVG
= 4,
392 SCHED_FEAT_APPROX_AVG
= 8,
395 const_debug
unsigned int sysctl_sched_features
=
396 SCHED_FEAT_NEW_FAIR_SLEEPERS
*1 |
397 SCHED_FEAT_START_DEBIT
*1 |
398 SCHED_FEAT_USE_TREE_AVG
*0 |
399 SCHED_FEAT_APPROX_AVG
*0;
401 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
404 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
405 * clock constructed from sched_clock():
407 unsigned long long cpu_clock(int cpu
)
409 unsigned long long now
;
413 local_irq_save(flags
);
417 local_irq_restore(flags
);
422 #ifdef CONFIG_FAIR_GROUP_SCHED
423 /* Change a task's ->cfs_rq if it moves across CPUs */
424 static inline void set_task_cfs_rq(struct task_struct
*p
)
426 p
->se
.cfs_rq
= &task_rq(p
)->cfs
;
429 static inline void set_task_cfs_rq(struct task_struct
*p
)
434 #ifndef prepare_arch_switch
435 # define prepare_arch_switch(next) do { } while (0)
437 #ifndef finish_arch_switch
438 # define finish_arch_switch(prev) do { } while (0)
441 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
442 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
444 return rq
->curr
== p
;
447 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
451 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
453 #ifdef CONFIG_DEBUG_SPINLOCK
454 /* this is a valid case when another task releases the spinlock */
455 rq
->lock
.owner
= current
;
458 * If we are tracking spinlock dependencies then we have to
459 * fix up the runqueue lock - which gets 'carried over' from
462 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
464 spin_unlock_irq(&rq
->lock
);
467 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
468 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
473 return rq
->curr
== p
;
477 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
481 * We can optimise this out completely for !SMP, because the
482 * SMP rebalancing from interrupt is the only thing that cares
487 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
488 spin_unlock_irq(&rq
->lock
);
490 spin_unlock(&rq
->lock
);
494 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
498 * After ->oncpu is cleared, the task can be moved to a different CPU.
499 * We must ensure this doesn't happen until the switch is completely
505 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
509 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
512 * __task_rq_lock - lock the runqueue a given task resides on.
513 * Must be called interrupts disabled.
515 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
522 spin_lock(&rq
->lock
);
523 if (unlikely(rq
!= task_rq(p
))) {
524 spin_unlock(&rq
->lock
);
525 goto repeat_lock_task
;
531 * task_rq_lock - lock the runqueue a given task resides on and disable
532 * interrupts. Note the ordering: we can safely lookup the task_rq without
533 * explicitly disabling preemption.
535 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
541 local_irq_save(*flags
);
543 spin_lock(&rq
->lock
);
544 if (unlikely(rq
!= task_rq(p
))) {
545 spin_unlock_irqrestore(&rq
->lock
, *flags
);
546 goto repeat_lock_task
;
551 static inline void __task_rq_unlock(struct rq
*rq
)
554 spin_unlock(&rq
->lock
);
557 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
560 spin_unlock_irqrestore(&rq
->lock
, *flags
);
564 * this_rq_lock - lock this runqueue and disable interrupts.
566 static inline struct rq
*this_rq_lock(void)
573 spin_lock(&rq
->lock
);
579 * We are going deep-idle (irqs are disabled):
581 void sched_clock_idle_sleep_event(void)
583 struct rq
*rq
= cpu_rq(smp_processor_id());
585 spin_lock(&rq
->lock
);
586 __update_rq_clock(rq
);
587 spin_unlock(&rq
->lock
);
588 rq
->clock_deep_idle_events
++;
590 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
593 * We just idled delta nanoseconds (called with irqs disabled):
595 void sched_clock_idle_wakeup_event(u64 delta_ns
)
597 struct rq
*rq
= cpu_rq(smp_processor_id());
598 u64 now
= sched_clock();
600 rq
->idle_clock
+= delta_ns
;
602 * Override the previous timestamp and ignore all
603 * sched_clock() deltas that occured while we idled,
604 * and use the PM-provided delta_ns to advance the
607 spin_lock(&rq
->lock
);
608 rq
->prev_clock_raw
= now
;
609 rq
->clock
+= delta_ns
;
610 spin_unlock(&rq
->lock
);
612 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
615 * resched_task - mark a task 'to be rescheduled now'.
617 * On UP this means the setting of the need_resched flag, on SMP it
618 * might also involve a cross-CPU call to trigger the scheduler on
623 #ifndef tsk_is_polling
624 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
627 static void resched_task(struct task_struct
*p
)
631 assert_spin_locked(&task_rq(p
)->lock
);
633 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
636 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
639 if (cpu
== smp_processor_id())
642 /* NEED_RESCHED must be visible before we test polling */
644 if (!tsk_is_polling(p
))
645 smp_send_reschedule(cpu
);
648 static void resched_cpu(int cpu
)
650 struct rq
*rq
= cpu_rq(cpu
);
653 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
655 resched_task(cpu_curr(cpu
));
656 spin_unlock_irqrestore(&rq
->lock
, flags
);
659 static inline void resched_task(struct task_struct
*p
)
661 assert_spin_locked(&task_rq(p
)->lock
);
662 set_tsk_need_resched(p
);
666 #if BITS_PER_LONG == 32
667 # define WMULT_CONST (~0UL)
669 # define WMULT_CONST (1UL << 32)
672 #define WMULT_SHIFT 32
675 * Shift right and round:
677 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
680 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
681 struct load_weight
*lw
)
685 if (unlikely(!lw
->inv_weight
))
686 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
688 tmp
= (u64
)delta_exec
* weight
;
690 * Check whether we'd overflow the 64-bit multiplication:
692 if (unlikely(tmp
> WMULT_CONST
))
693 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
696 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
698 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
701 static inline unsigned long
702 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
704 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
707 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
712 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
718 * To aid in avoiding the subversion of "niceness" due to uneven distribution
719 * of tasks with abnormal "nice" values across CPUs the contribution that
720 * each task makes to its run queue's load is weighted according to its
721 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
722 * scaled version of the new time slice allocation that they receive on time
726 #define WEIGHT_IDLEPRIO 2
727 #define WMULT_IDLEPRIO (1 << 31)
730 * Nice levels are multiplicative, with a gentle 10% change for every
731 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
732 * nice 1, it will get ~10% less CPU time than another CPU-bound task
733 * that remained on nice 0.
735 * The "10% effect" is relative and cumulative: from _any_ nice level,
736 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
737 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
738 * If a task goes up by ~10% and another task goes down by ~10% then
739 * the relative distance between them is ~25%.)
741 static const int prio_to_weight
[40] = {
742 /* -20 */ 88761, 71755, 56483, 46273, 36291,
743 /* -15 */ 29154, 23254, 18705, 14949, 11916,
744 /* -10 */ 9548, 7620, 6100, 4904, 3906,
745 /* -5 */ 3121, 2501, 1991, 1586, 1277,
746 /* 0 */ 1024, 820, 655, 526, 423,
747 /* 5 */ 335, 272, 215, 172, 137,
748 /* 10 */ 110, 87, 70, 56, 45,
749 /* 15 */ 36, 29, 23, 18, 15,
753 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
755 * In cases where the weight does not change often, we can use the
756 * precalculated inverse to speed up arithmetics by turning divisions
757 * into multiplications:
759 static const u32 prio_to_wmult
[40] = {
760 /* -20 */ 48388, 59856, 76040, 92818, 118348,
761 /* -15 */ 147320, 184698, 229616, 287308, 360437,
762 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
763 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
764 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
765 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
766 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
767 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
770 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
773 * runqueue iterator, to support SMP load-balancing between different
774 * scheduling classes, without having to expose their internal data
775 * structures to the load-balancing proper:
779 struct task_struct
*(*start
)(void *);
780 struct task_struct
*(*next
)(void *);
783 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
784 unsigned long max_nr_move
, unsigned long max_load_move
,
785 struct sched_domain
*sd
, enum cpu_idle_type idle
,
786 int *all_pinned
, unsigned long *load_moved
,
787 int *this_best_prio
, struct rq_iterator
*iterator
);
789 #include "sched_stats.h"
790 #include "sched_rt.c"
791 #include "sched_fair.c"
792 #include "sched_idletask.c"
793 #ifdef CONFIG_SCHED_DEBUG
794 # include "sched_debug.c"
797 #define sched_class_highest (&rt_sched_class)
800 * Update delta_exec, delta_fair fields for rq.
802 * delta_fair clock advances at a rate inversely proportional to
803 * total load (rq->load.weight) on the runqueue, while
804 * delta_exec advances at the same rate as wall-clock (provided
807 * delta_exec / delta_fair is a measure of the (smoothened) load on this
808 * runqueue over any given interval. This (smoothened) load is used
809 * during load balance.
811 * This function is called /before/ updating rq->load
812 * and when switching tasks.
814 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
816 update_load_add(&rq
->load
, p
->se
.load
.weight
);
819 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
821 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
824 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
830 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
836 static void set_load_weight(struct task_struct
*p
)
838 if (task_has_rt_policy(p
)) {
839 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
840 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
845 * SCHED_IDLE tasks get minimal weight:
847 if (p
->policy
== SCHED_IDLE
) {
848 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
849 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
853 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
854 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
857 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
859 sched_info_queued(p
);
860 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
864 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
866 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
871 * __normal_prio - return the priority that is based on the static prio
873 static inline int __normal_prio(struct task_struct
*p
)
875 return p
->static_prio
;
879 * Calculate the expected normal priority: i.e. priority
880 * without taking RT-inheritance into account. Might be
881 * boosted by interactivity modifiers. Changes upon fork,
882 * setprio syscalls, and whenever the interactivity
883 * estimator recalculates.
885 static inline int normal_prio(struct task_struct
*p
)
889 if (task_has_rt_policy(p
))
890 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
892 prio
= __normal_prio(p
);
897 * Calculate the current priority, i.e. the priority
898 * taken into account by the scheduler. This value might
899 * be boosted by RT tasks, or might be boosted by
900 * interactivity modifiers. Will be RT if the task got
901 * RT-boosted. If not then it returns p->normal_prio.
903 static int effective_prio(struct task_struct
*p
)
905 p
->normal_prio
= normal_prio(p
);
907 * If we are RT tasks or we were boosted to RT priority,
908 * keep the priority unchanged. Otherwise, update priority
909 * to the normal priority:
911 if (!rt_prio(p
->prio
))
912 return p
->normal_prio
;
917 * activate_task - move a task to the runqueue.
919 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
921 if (p
->state
== TASK_UNINTERRUPTIBLE
)
922 rq
->nr_uninterruptible
--;
924 enqueue_task(rq
, p
, wakeup
);
925 inc_nr_running(p
, rq
);
929 * activate_idle_task - move idle task to the _front_ of runqueue.
931 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
935 if (p
->state
== TASK_UNINTERRUPTIBLE
)
936 rq
->nr_uninterruptible
--;
938 enqueue_task(rq
, p
, 0);
939 inc_nr_running(p
, rq
);
943 * deactivate_task - remove a task from the runqueue.
945 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
947 if (p
->state
== TASK_UNINTERRUPTIBLE
)
948 rq
->nr_uninterruptible
++;
950 dequeue_task(rq
, p
, sleep
);
951 dec_nr_running(p
, rq
);
955 * task_curr - is this task currently executing on a CPU?
956 * @p: the task in question.
958 inline int task_curr(const struct task_struct
*p
)
960 return cpu_curr(task_cpu(p
)) == p
;
963 /* Used instead of source_load when we know the type == 0 */
964 unsigned long weighted_cpuload(const int cpu
)
966 return cpu_rq(cpu
)->load
.weight
;
969 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
972 task_thread_info(p
)->cpu
= cpu
;
979 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
981 int old_cpu
= task_cpu(p
);
982 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
985 clock_offset
= old_rq
->clock
- new_rq
->clock
;
987 #ifdef CONFIG_SCHEDSTATS
988 if (p
->se
.wait_start
)
989 p
->se
.wait_start
-= clock_offset
;
990 if (p
->se
.sleep_start
)
991 p
->se
.sleep_start
-= clock_offset
;
992 if (p
->se
.block_start
)
993 p
->se
.block_start
-= clock_offset
;
995 if (likely(new_rq
->cfs
.min_vruntime
))
996 p
->se
.vruntime
-= old_rq
->cfs
.min_vruntime
-
997 new_rq
->cfs
.min_vruntime
;
999 __set_task_cpu(p
, new_cpu
);
1002 struct migration_req
{
1003 struct list_head list
;
1005 struct task_struct
*task
;
1008 struct completion done
;
1012 * The task's runqueue lock must be held.
1013 * Returns true if you have to wait for migration thread.
1016 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1018 struct rq
*rq
= task_rq(p
);
1021 * If the task is not on a runqueue (and not running), then
1022 * it is sufficient to simply update the task's cpu field.
1024 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1025 set_task_cpu(p
, dest_cpu
);
1029 init_completion(&req
->done
);
1031 req
->dest_cpu
= dest_cpu
;
1032 list_add(&req
->list
, &rq
->migration_queue
);
1038 * wait_task_inactive - wait for a thread to unschedule.
1040 * The caller must ensure that the task *will* unschedule sometime soon,
1041 * else this function might spin for a *long* time. This function can't
1042 * be called with interrupts off, or it may introduce deadlock with
1043 * smp_call_function() if an IPI is sent by the same process we are
1044 * waiting to become inactive.
1046 void wait_task_inactive(struct task_struct
*p
)
1048 unsigned long flags
;
1054 * We do the initial early heuristics without holding
1055 * any task-queue locks at all. We'll only try to get
1056 * the runqueue lock when things look like they will
1062 * If the task is actively running on another CPU
1063 * still, just relax and busy-wait without holding
1066 * NOTE! Since we don't hold any locks, it's not
1067 * even sure that "rq" stays as the right runqueue!
1068 * But we don't care, since "task_running()" will
1069 * return false if the runqueue has changed and p
1070 * is actually now running somewhere else!
1072 while (task_running(rq
, p
))
1076 * Ok, time to look more closely! We need the rq
1077 * lock now, to be *sure*. If we're wrong, we'll
1078 * just go back and repeat.
1080 rq
= task_rq_lock(p
, &flags
);
1081 running
= task_running(rq
, p
);
1082 on_rq
= p
->se
.on_rq
;
1083 task_rq_unlock(rq
, &flags
);
1086 * Was it really running after all now that we
1087 * checked with the proper locks actually held?
1089 * Oops. Go back and try again..
1091 if (unlikely(running
)) {
1097 * It's not enough that it's not actively running,
1098 * it must be off the runqueue _entirely_, and not
1101 * So if it wa still runnable (but just not actively
1102 * running right now), it's preempted, and we should
1103 * yield - it could be a while.
1105 if (unlikely(on_rq
)) {
1111 * Ahh, all good. It wasn't running, and it wasn't
1112 * runnable, which means that it will never become
1113 * running in the future either. We're all done!
1118 * kick_process - kick a running thread to enter/exit the kernel
1119 * @p: the to-be-kicked thread
1121 * Cause a process which is running on another CPU to enter
1122 * kernel-mode, without any delay. (to get signals handled.)
1124 * NOTE: this function doesnt have to take the runqueue lock,
1125 * because all it wants to ensure is that the remote task enters
1126 * the kernel. If the IPI races and the task has been migrated
1127 * to another CPU then no harm is done and the purpose has been
1130 void kick_process(struct task_struct
*p
)
1136 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1137 smp_send_reschedule(cpu
);
1142 * Return a low guess at the load of a migration-source cpu weighted
1143 * according to the scheduling class and "nice" value.
1145 * We want to under-estimate the load of migration sources, to
1146 * balance conservatively.
1148 static inline unsigned long source_load(int cpu
, int type
)
1150 struct rq
*rq
= cpu_rq(cpu
);
1151 unsigned long total
= weighted_cpuload(cpu
);
1156 return min(rq
->cpu_load
[type
-1], total
);
1160 * Return a high guess at the load of a migration-target cpu weighted
1161 * according to the scheduling class and "nice" value.
1163 static inline unsigned long target_load(int cpu
, int type
)
1165 struct rq
*rq
= cpu_rq(cpu
);
1166 unsigned long total
= weighted_cpuload(cpu
);
1171 return max(rq
->cpu_load
[type
-1], total
);
1175 * Return the average load per task on the cpu's run queue
1177 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1179 struct rq
*rq
= cpu_rq(cpu
);
1180 unsigned long total
= weighted_cpuload(cpu
);
1181 unsigned long n
= rq
->nr_running
;
1183 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1187 * find_idlest_group finds and returns the least busy CPU group within the
1190 static struct sched_group
*
1191 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1193 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1194 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1195 int load_idx
= sd
->forkexec_idx
;
1196 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1199 unsigned long load
, avg_load
;
1203 /* Skip over this group if it has no CPUs allowed */
1204 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1207 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1209 /* Tally up the load of all CPUs in the group */
1212 for_each_cpu_mask(i
, group
->cpumask
) {
1213 /* Bias balancing toward cpus of our domain */
1215 load
= source_load(i
, load_idx
);
1217 load
= target_load(i
, load_idx
);
1222 /* Adjust by relative CPU power of the group */
1223 avg_load
= sg_div_cpu_power(group
,
1224 avg_load
* SCHED_LOAD_SCALE
);
1227 this_load
= avg_load
;
1229 } else if (avg_load
< min_load
) {
1230 min_load
= avg_load
;
1234 group
= group
->next
;
1235 } while (group
!= sd
->groups
);
1237 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1243 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1246 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1249 unsigned long load
, min_load
= ULONG_MAX
;
1253 /* Traverse only the allowed CPUs */
1254 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1256 for_each_cpu_mask(i
, tmp
) {
1257 load
= weighted_cpuload(i
);
1259 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1269 * sched_balance_self: balance the current task (running on cpu) in domains
1270 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1273 * Balance, ie. select the least loaded group.
1275 * Returns the target CPU number, or the same CPU if no balancing is needed.
1277 * preempt must be disabled.
1279 static int sched_balance_self(int cpu
, int flag
)
1281 struct task_struct
*t
= current
;
1282 struct sched_domain
*tmp
, *sd
= NULL
;
1284 for_each_domain(cpu
, tmp
) {
1286 * If power savings logic is enabled for a domain, stop there.
1288 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1290 if (tmp
->flags
& flag
)
1296 struct sched_group
*group
;
1297 int new_cpu
, weight
;
1299 if (!(sd
->flags
& flag
)) {
1305 group
= find_idlest_group(sd
, t
, cpu
);
1311 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1312 if (new_cpu
== -1 || new_cpu
== cpu
) {
1313 /* Now try balancing at a lower domain level of cpu */
1318 /* Now try balancing at a lower domain level of new_cpu */
1321 weight
= cpus_weight(span
);
1322 for_each_domain(cpu
, tmp
) {
1323 if (weight
<= cpus_weight(tmp
->span
))
1325 if (tmp
->flags
& flag
)
1328 /* while loop will break here if sd == NULL */
1334 #endif /* CONFIG_SMP */
1337 * wake_idle() will wake a task on an idle cpu if task->cpu is
1338 * not idle and an idle cpu is available. The span of cpus to
1339 * search starts with cpus closest then further out as needed,
1340 * so we always favor a closer, idle cpu.
1342 * Returns the CPU we should wake onto.
1344 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1345 static int wake_idle(int cpu
, struct task_struct
*p
)
1348 struct sched_domain
*sd
;
1352 * If it is idle, then it is the best cpu to run this task.
1354 * This cpu is also the best, if it has more than one task already.
1355 * Siblings must be also busy(in most cases) as they didn't already
1356 * pickup the extra load from this cpu and hence we need not check
1357 * sibling runqueue info. This will avoid the checks and cache miss
1358 * penalities associated with that.
1360 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1363 for_each_domain(cpu
, sd
) {
1364 if (sd
->flags
& SD_WAKE_IDLE
) {
1365 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1366 for_each_cpu_mask(i
, tmp
) {
1377 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1384 * try_to_wake_up - wake up a thread
1385 * @p: the to-be-woken-up thread
1386 * @state: the mask of task states that can be woken
1387 * @sync: do a synchronous wakeup?
1389 * Put it on the run-queue if it's not already there. The "current"
1390 * thread is always on the run-queue (except when the actual
1391 * re-schedule is in progress), and as such you're allowed to do
1392 * the simpler "current->state = TASK_RUNNING" to mark yourself
1393 * runnable without the overhead of this.
1395 * returns failure only if the task is already active.
1397 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1399 int cpu
, this_cpu
, success
= 0;
1400 unsigned long flags
;
1404 struct sched_domain
*sd
, *this_sd
= NULL
;
1405 unsigned long load
, this_load
;
1409 rq
= task_rq_lock(p
, &flags
);
1410 old_state
= p
->state
;
1411 if (!(old_state
& state
))
1418 this_cpu
= smp_processor_id();
1421 if (unlikely(task_running(rq
, p
)))
1426 schedstat_inc(rq
, ttwu_cnt
);
1427 if (cpu
== this_cpu
) {
1428 schedstat_inc(rq
, ttwu_local
);
1432 for_each_domain(this_cpu
, sd
) {
1433 if (cpu_isset(cpu
, sd
->span
)) {
1434 schedstat_inc(sd
, ttwu_wake_remote
);
1440 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1444 * Check for affine wakeup and passive balancing possibilities.
1447 int idx
= this_sd
->wake_idx
;
1448 unsigned int imbalance
;
1450 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1452 load
= source_load(cpu
, idx
);
1453 this_load
= target_load(this_cpu
, idx
);
1455 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1457 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1458 unsigned long tl
= this_load
;
1459 unsigned long tl_per_task
;
1461 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1464 * If sync wakeup then subtract the (maximum possible)
1465 * effect of the currently running task from the load
1466 * of the current CPU:
1469 tl
-= current
->se
.load
.weight
;
1472 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1473 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1475 * This domain has SD_WAKE_AFFINE and
1476 * p is cache cold in this domain, and
1477 * there is no bad imbalance.
1479 schedstat_inc(this_sd
, ttwu_move_affine
);
1485 * Start passive balancing when half the imbalance_pct
1488 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1489 if (imbalance
*this_load
<= 100*load
) {
1490 schedstat_inc(this_sd
, ttwu_move_balance
);
1496 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1498 new_cpu
= wake_idle(new_cpu
, p
);
1499 if (new_cpu
!= cpu
) {
1500 set_task_cpu(p
, new_cpu
);
1501 task_rq_unlock(rq
, &flags
);
1502 /* might preempt at this point */
1503 rq
= task_rq_lock(p
, &flags
);
1504 old_state
= p
->state
;
1505 if (!(old_state
& state
))
1510 this_cpu
= smp_processor_id();
1515 #endif /* CONFIG_SMP */
1516 update_rq_clock(rq
);
1517 activate_task(rq
, p
, 1);
1519 * Sync wakeups (i.e. those types of wakeups where the waker
1520 * has indicated that it will leave the CPU in short order)
1521 * don't trigger a preemption, if the woken up task will run on
1522 * this cpu. (in this case the 'I will reschedule' promise of
1523 * the waker guarantees that the freshly woken up task is going
1524 * to be considered on this CPU.)
1526 if (!sync
|| cpu
!= this_cpu
)
1527 check_preempt_curr(rq
, p
);
1531 p
->state
= TASK_RUNNING
;
1533 task_rq_unlock(rq
, &flags
);
1538 int fastcall
wake_up_process(struct task_struct
*p
)
1540 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1541 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1543 EXPORT_SYMBOL(wake_up_process
);
1545 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1547 return try_to_wake_up(p
, state
, 0);
1551 * Perform scheduler related setup for a newly forked process p.
1552 * p is forked by current.
1554 * __sched_fork() is basic setup used by init_idle() too:
1556 static void __sched_fork(struct task_struct
*p
)
1558 p
->se
.exec_start
= 0;
1559 p
->se
.sum_exec_runtime
= 0;
1560 p
->se
.prev_sum_exec_runtime
= 0;
1562 #ifdef CONFIG_SCHEDSTATS
1563 p
->se
.wait_start
= 0;
1564 p
->se
.sum_sleep_runtime
= 0;
1565 p
->se
.sleep_start
= 0;
1566 p
->se
.block_start
= 0;
1567 p
->se
.sleep_max
= 0;
1568 p
->se
.block_max
= 0;
1570 p
->se
.slice_max
= 0;
1574 INIT_LIST_HEAD(&p
->run_list
);
1577 #ifdef CONFIG_PREEMPT_NOTIFIERS
1578 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1582 * We mark the process as running here, but have not actually
1583 * inserted it onto the runqueue yet. This guarantees that
1584 * nobody will actually run it, and a signal or other external
1585 * event cannot wake it up and insert it on the runqueue either.
1587 p
->state
= TASK_RUNNING
;
1591 * fork()/clone()-time setup:
1593 void sched_fork(struct task_struct
*p
, int clone_flags
)
1595 int cpu
= get_cpu();
1600 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1602 __set_task_cpu(p
, cpu
);
1605 * Make sure we do not leak PI boosting priority to the child:
1607 p
->prio
= current
->normal_prio
;
1609 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1610 if (likely(sched_info_on()))
1611 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1613 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1616 #ifdef CONFIG_PREEMPT
1617 /* Want to start with kernel preemption disabled. */
1618 task_thread_info(p
)->preempt_count
= 1;
1624 * wake_up_new_task - wake up a newly created task for the first time.
1626 * This function will do some initial scheduler statistics housekeeping
1627 * that must be done for every newly created context, then puts the task
1628 * on the runqueue and wakes it.
1630 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1632 unsigned long flags
;
1636 rq
= task_rq_lock(p
, &flags
);
1637 BUG_ON(p
->state
!= TASK_RUNNING
);
1638 this_cpu
= smp_processor_id(); /* parent's CPU */
1639 update_rq_clock(rq
);
1641 p
->prio
= effective_prio(p
);
1643 if (rt_prio(p
->prio
))
1644 p
->sched_class
= &rt_sched_class
;
1646 p
->sched_class
= &fair_sched_class
;
1648 if (task_cpu(p
) != this_cpu
|| !p
->sched_class
->task_new
||
1649 !current
->se
.on_rq
) {
1650 activate_task(rq
, p
, 0);
1653 * Let the scheduling class do new task startup
1654 * management (if any):
1656 p
->sched_class
->task_new(rq
, p
);
1657 inc_nr_running(p
, rq
);
1659 check_preempt_curr(rq
, p
);
1660 task_rq_unlock(rq
, &flags
);
1663 #ifdef CONFIG_PREEMPT_NOTIFIERS
1666 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1667 * @notifier: notifier struct to register
1669 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1671 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1673 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1676 * preempt_notifier_unregister - no longer interested in preemption notifications
1677 * @notifier: notifier struct to unregister
1679 * This is safe to call from within a preemption notifier.
1681 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1683 hlist_del(¬ifier
->link
);
1685 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1687 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1689 struct preempt_notifier
*notifier
;
1690 struct hlist_node
*node
;
1692 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1693 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1697 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1698 struct task_struct
*next
)
1700 struct preempt_notifier
*notifier
;
1701 struct hlist_node
*node
;
1703 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1704 notifier
->ops
->sched_out(notifier
, next
);
1709 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1714 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1715 struct task_struct
*next
)
1722 * prepare_task_switch - prepare to switch tasks
1723 * @rq: the runqueue preparing to switch
1724 * @prev: the current task that is being switched out
1725 * @next: the task we are going to switch to.
1727 * This is called with the rq lock held and interrupts off. It must
1728 * be paired with a subsequent finish_task_switch after the context
1731 * prepare_task_switch sets up locking and calls architecture specific
1735 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1736 struct task_struct
*next
)
1738 fire_sched_out_preempt_notifiers(prev
, next
);
1739 prepare_lock_switch(rq
, next
);
1740 prepare_arch_switch(next
);
1744 * finish_task_switch - clean up after a task-switch
1745 * @rq: runqueue associated with task-switch
1746 * @prev: the thread we just switched away from.
1748 * finish_task_switch must be called after the context switch, paired
1749 * with a prepare_task_switch call before the context switch.
1750 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1751 * and do any other architecture-specific cleanup actions.
1753 * Note that we may have delayed dropping an mm in context_switch(). If
1754 * so, we finish that here outside of the runqueue lock. (Doing it
1755 * with the lock held can cause deadlocks; see schedule() for
1758 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1759 __releases(rq
->lock
)
1761 struct mm_struct
*mm
= rq
->prev_mm
;
1767 * A task struct has one reference for the use as "current".
1768 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1769 * schedule one last time. The schedule call will never return, and
1770 * the scheduled task must drop that reference.
1771 * The test for TASK_DEAD must occur while the runqueue locks are
1772 * still held, otherwise prev could be scheduled on another cpu, die
1773 * there before we look at prev->state, and then the reference would
1775 * Manfred Spraul <manfred@colorfullife.com>
1777 prev_state
= prev
->state
;
1778 finish_arch_switch(prev
);
1779 finish_lock_switch(rq
, prev
);
1780 fire_sched_in_preempt_notifiers(current
);
1783 if (unlikely(prev_state
== TASK_DEAD
)) {
1785 * Remove function-return probe instances associated with this
1786 * task and put them back on the free list.
1788 kprobe_flush_task(prev
);
1789 put_task_struct(prev
);
1794 * schedule_tail - first thing a freshly forked thread must call.
1795 * @prev: the thread we just switched away from.
1797 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1798 __releases(rq
->lock
)
1800 struct rq
*rq
= this_rq();
1802 finish_task_switch(rq
, prev
);
1803 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1804 /* In this case, finish_task_switch does not reenable preemption */
1807 if (current
->set_child_tid
)
1808 put_user(current
->pid
, current
->set_child_tid
);
1812 * context_switch - switch to the new MM and the new
1813 * thread's register state.
1816 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1817 struct task_struct
*next
)
1819 struct mm_struct
*mm
, *oldmm
;
1821 prepare_task_switch(rq
, prev
, next
);
1823 oldmm
= prev
->active_mm
;
1825 * For paravirt, this is coupled with an exit in switch_to to
1826 * combine the page table reload and the switch backend into
1829 arch_enter_lazy_cpu_mode();
1831 if (unlikely(!mm
)) {
1832 next
->active_mm
= oldmm
;
1833 atomic_inc(&oldmm
->mm_count
);
1834 enter_lazy_tlb(oldmm
, next
);
1836 switch_mm(oldmm
, mm
, next
);
1838 if (unlikely(!prev
->mm
)) {
1839 prev
->active_mm
= NULL
;
1840 rq
->prev_mm
= oldmm
;
1843 * Since the runqueue lock will be released by the next
1844 * task (which is an invalid locking op but in the case
1845 * of the scheduler it's an obvious special-case), so we
1846 * do an early lockdep release here:
1848 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1849 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1852 /* Here we just switch the register state and the stack. */
1853 switch_to(prev
, next
, prev
);
1857 * this_rq must be evaluated again because prev may have moved
1858 * CPUs since it called schedule(), thus the 'rq' on its stack
1859 * frame will be invalid.
1861 finish_task_switch(this_rq(), prev
);
1865 * nr_running, nr_uninterruptible and nr_context_switches:
1867 * externally visible scheduler statistics: current number of runnable
1868 * threads, current number of uninterruptible-sleeping threads, total
1869 * number of context switches performed since bootup.
1871 unsigned long nr_running(void)
1873 unsigned long i
, sum
= 0;
1875 for_each_online_cpu(i
)
1876 sum
+= cpu_rq(i
)->nr_running
;
1881 unsigned long nr_uninterruptible(void)
1883 unsigned long i
, sum
= 0;
1885 for_each_possible_cpu(i
)
1886 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1889 * Since we read the counters lockless, it might be slightly
1890 * inaccurate. Do not allow it to go below zero though:
1892 if (unlikely((long)sum
< 0))
1898 unsigned long long nr_context_switches(void)
1901 unsigned long long sum
= 0;
1903 for_each_possible_cpu(i
)
1904 sum
+= cpu_rq(i
)->nr_switches
;
1909 unsigned long nr_iowait(void)
1911 unsigned long i
, sum
= 0;
1913 for_each_possible_cpu(i
)
1914 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1919 unsigned long nr_active(void)
1921 unsigned long i
, running
= 0, uninterruptible
= 0;
1923 for_each_online_cpu(i
) {
1924 running
+= cpu_rq(i
)->nr_running
;
1925 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1928 if (unlikely((long)uninterruptible
< 0))
1929 uninterruptible
= 0;
1931 return running
+ uninterruptible
;
1935 * Update rq->cpu_load[] statistics. This function is usually called every
1936 * scheduler tick (TICK_NSEC).
1938 static void update_cpu_load(struct rq
*this_rq
)
1940 unsigned long this_load
= this_rq
->load
.weight
;
1943 this_rq
->nr_load_updates
++;
1945 /* Update our load: */
1946 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
1947 unsigned long old_load
, new_load
;
1949 /* scale is effectively 1 << i now, and >> i divides by scale */
1951 old_load
= this_rq
->cpu_load
[i
];
1952 new_load
= this_load
;
1954 * Round up the averaging division if load is increasing. This
1955 * prevents us from getting stuck on 9 if the load is 10, for
1958 if (new_load
> old_load
)
1959 new_load
+= scale
-1;
1960 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
1967 * double_rq_lock - safely lock two runqueues
1969 * Note this does not disable interrupts like task_rq_lock,
1970 * you need to do so manually before calling.
1972 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1973 __acquires(rq1
->lock
)
1974 __acquires(rq2
->lock
)
1976 BUG_ON(!irqs_disabled());
1978 spin_lock(&rq1
->lock
);
1979 __acquire(rq2
->lock
); /* Fake it out ;) */
1982 spin_lock(&rq1
->lock
);
1983 spin_lock(&rq2
->lock
);
1985 spin_lock(&rq2
->lock
);
1986 spin_lock(&rq1
->lock
);
1989 update_rq_clock(rq1
);
1990 update_rq_clock(rq2
);
1994 * double_rq_unlock - safely unlock two runqueues
1996 * Note this does not restore interrupts like task_rq_unlock,
1997 * you need to do so manually after calling.
1999 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2000 __releases(rq1
->lock
)
2001 __releases(rq2
->lock
)
2003 spin_unlock(&rq1
->lock
);
2005 spin_unlock(&rq2
->lock
);
2007 __release(rq2
->lock
);
2011 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2013 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2014 __releases(this_rq
->lock
)
2015 __acquires(busiest
->lock
)
2016 __acquires(this_rq
->lock
)
2018 if (unlikely(!irqs_disabled())) {
2019 /* printk() doesn't work good under rq->lock */
2020 spin_unlock(&this_rq
->lock
);
2023 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2024 if (busiest
< this_rq
) {
2025 spin_unlock(&this_rq
->lock
);
2026 spin_lock(&busiest
->lock
);
2027 spin_lock(&this_rq
->lock
);
2029 spin_lock(&busiest
->lock
);
2034 * If dest_cpu is allowed for this process, migrate the task to it.
2035 * This is accomplished by forcing the cpu_allowed mask to only
2036 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2037 * the cpu_allowed mask is restored.
2039 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2041 struct migration_req req
;
2042 unsigned long flags
;
2045 rq
= task_rq_lock(p
, &flags
);
2046 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2047 || unlikely(cpu_is_offline(dest_cpu
)))
2050 /* force the process onto the specified CPU */
2051 if (migrate_task(p
, dest_cpu
, &req
)) {
2052 /* Need to wait for migration thread (might exit: take ref). */
2053 struct task_struct
*mt
= rq
->migration_thread
;
2055 get_task_struct(mt
);
2056 task_rq_unlock(rq
, &flags
);
2057 wake_up_process(mt
);
2058 put_task_struct(mt
);
2059 wait_for_completion(&req
.done
);
2064 task_rq_unlock(rq
, &flags
);
2068 * sched_exec - execve() is a valuable balancing opportunity, because at
2069 * this point the task has the smallest effective memory and cache footprint.
2071 void sched_exec(void)
2073 int new_cpu
, this_cpu
= get_cpu();
2074 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2076 if (new_cpu
!= this_cpu
)
2077 sched_migrate_task(current
, new_cpu
);
2081 * pull_task - move a task from a remote runqueue to the local runqueue.
2082 * Both runqueues must be locked.
2084 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2085 struct rq
*this_rq
, int this_cpu
)
2087 deactivate_task(src_rq
, p
, 0);
2088 set_task_cpu(p
, this_cpu
);
2089 activate_task(this_rq
, p
, 0);
2091 * Note that idle threads have a prio of MAX_PRIO, for this test
2092 * to be always true for them.
2094 check_preempt_curr(this_rq
, p
);
2098 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2101 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2102 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2106 * We do not migrate tasks that are:
2107 * 1) running (obviously), or
2108 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2109 * 3) are cache-hot on their current CPU.
2111 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2115 if (task_running(rq
, p
))
2121 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2122 unsigned long max_nr_move
, unsigned long max_load_move
,
2123 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2124 int *all_pinned
, unsigned long *load_moved
,
2125 int *this_best_prio
, struct rq_iterator
*iterator
)
2127 int pulled
= 0, pinned
= 0, skip_for_load
;
2128 struct task_struct
*p
;
2129 long rem_load_move
= max_load_move
;
2131 if (max_nr_move
== 0 || max_load_move
== 0)
2137 * Start the load-balancing iterator:
2139 p
= iterator
->start(iterator
->arg
);
2144 * To help distribute high priority tasks accross CPUs we don't
2145 * skip a task if it will be the highest priority task (i.e. smallest
2146 * prio value) on its new queue regardless of its load weight
2148 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2149 SCHED_LOAD_SCALE_FUZZ
;
2150 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2151 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2152 p
= iterator
->next(iterator
->arg
);
2156 pull_task(busiest
, p
, this_rq
, this_cpu
);
2158 rem_load_move
-= p
->se
.load
.weight
;
2161 * We only want to steal up to the prescribed number of tasks
2162 * and the prescribed amount of weighted load.
2164 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2165 if (p
->prio
< *this_best_prio
)
2166 *this_best_prio
= p
->prio
;
2167 p
= iterator
->next(iterator
->arg
);
2172 * Right now, this is the only place pull_task() is called,
2173 * so we can safely collect pull_task() stats here rather than
2174 * inside pull_task().
2176 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2179 *all_pinned
= pinned
;
2180 *load_moved
= max_load_move
- rem_load_move
;
2185 * move_tasks tries to move up to max_load_move weighted load from busiest to
2186 * this_rq, as part of a balancing operation within domain "sd".
2187 * Returns 1 if successful and 0 otherwise.
2189 * Called with both runqueues locked.
2191 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2192 unsigned long max_load_move
,
2193 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2196 struct sched_class
*class = sched_class_highest
;
2197 unsigned long total_load_moved
= 0;
2198 int this_best_prio
= this_rq
->curr
->prio
;
2202 class->load_balance(this_rq
, this_cpu
, busiest
,
2203 ULONG_MAX
, max_load_move
- total_load_moved
,
2204 sd
, idle
, all_pinned
, &this_best_prio
);
2205 class = class->next
;
2206 } while (class && max_load_move
> total_load_moved
);
2208 return total_load_moved
> 0;
2212 * move_one_task tries to move exactly one task from busiest to this_rq, as
2213 * part of active balancing operations within "domain".
2214 * Returns 1 if successful and 0 otherwise.
2216 * Called with both runqueues locked.
2218 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2219 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2221 struct sched_class
*class;
2222 int this_best_prio
= MAX_PRIO
;
2224 for (class = sched_class_highest
; class; class = class->next
)
2225 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2226 1, ULONG_MAX
, sd
, idle
, NULL
,
2234 * find_busiest_group finds and returns the busiest CPU group within the
2235 * domain. It calculates and returns the amount of weighted load which
2236 * should be moved to restore balance via the imbalance parameter.
2238 static struct sched_group
*
2239 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2240 unsigned long *imbalance
, enum cpu_idle_type idle
,
2241 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2243 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2244 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2245 unsigned long max_pull
;
2246 unsigned long busiest_load_per_task
, busiest_nr_running
;
2247 unsigned long this_load_per_task
, this_nr_running
;
2249 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2250 int power_savings_balance
= 1;
2251 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2252 unsigned long min_nr_running
= ULONG_MAX
;
2253 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2256 max_load
= this_load
= total_load
= total_pwr
= 0;
2257 busiest_load_per_task
= busiest_nr_running
= 0;
2258 this_load_per_task
= this_nr_running
= 0;
2259 if (idle
== CPU_NOT_IDLE
)
2260 load_idx
= sd
->busy_idx
;
2261 else if (idle
== CPU_NEWLY_IDLE
)
2262 load_idx
= sd
->newidle_idx
;
2264 load_idx
= sd
->idle_idx
;
2267 unsigned long load
, group_capacity
;
2270 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2271 unsigned long sum_nr_running
, sum_weighted_load
;
2273 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2276 balance_cpu
= first_cpu(group
->cpumask
);
2278 /* Tally up the load of all CPUs in the group */
2279 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2281 for_each_cpu_mask(i
, group
->cpumask
) {
2284 if (!cpu_isset(i
, *cpus
))
2289 if (*sd_idle
&& rq
->nr_running
)
2292 /* Bias balancing toward cpus of our domain */
2294 if (idle_cpu(i
) && !first_idle_cpu
) {
2299 load
= target_load(i
, load_idx
);
2301 load
= source_load(i
, load_idx
);
2304 sum_nr_running
+= rq
->nr_running
;
2305 sum_weighted_load
+= weighted_cpuload(i
);
2309 * First idle cpu or the first cpu(busiest) in this sched group
2310 * is eligible for doing load balancing at this and above
2311 * domains. In the newly idle case, we will allow all the cpu's
2312 * to do the newly idle load balance.
2314 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2315 balance_cpu
!= this_cpu
&& balance
) {
2320 total_load
+= avg_load
;
2321 total_pwr
+= group
->__cpu_power
;
2323 /* Adjust by relative CPU power of the group */
2324 avg_load
= sg_div_cpu_power(group
,
2325 avg_load
* SCHED_LOAD_SCALE
);
2327 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2330 this_load
= avg_load
;
2332 this_nr_running
= sum_nr_running
;
2333 this_load_per_task
= sum_weighted_load
;
2334 } else if (avg_load
> max_load
&&
2335 sum_nr_running
> group_capacity
) {
2336 max_load
= avg_load
;
2338 busiest_nr_running
= sum_nr_running
;
2339 busiest_load_per_task
= sum_weighted_load
;
2342 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2344 * Busy processors will not participate in power savings
2347 if (idle
== CPU_NOT_IDLE
||
2348 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2352 * If the local group is idle or completely loaded
2353 * no need to do power savings balance at this domain
2355 if (local_group
&& (this_nr_running
>= group_capacity
||
2357 power_savings_balance
= 0;
2360 * If a group is already running at full capacity or idle,
2361 * don't include that group in power savings calculations
2363 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2368 * Calculate the group which has the least non-idle load.
2369 * This is the group from where we need to pick up the load
2372 if ((sum_nr_running
< min_nr_running
) ||
2373 (sum_nr_running
== min_nr_running
&&
2374 first_cpu(group
->cpumask
) <
2375 first_cpu(group_min
->cpumask
))) {
2377 min_nr_running
= sum_nr_running
;
2378 min_load_per_task
= sum_weighted_load
/
2383 * Calculate the group which is almost near its
2384 * capacity but still has some space to pick up some load
2385 * from other group and save more power
2387 if (sum_nr_running
<= group_capacity
- 1) {
2388 if (sum_nr_running
> leader_nr_running
||
2389 (sum_nr_running
== leader_nr_running
&&
2390 first_cpu(group
->cpumask
) >
2391 first_cpu(group_leader
->cpumask
))) {
2392 group_leader
= group
;
2393 leader_nr_running
= sum_nr_running
;
2398 group
= group
->next
;
2399 } while (group
!= sd
->groups
);
2401 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2404 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2406 if (this_load
>= avg_load
||
2407 100*max_load
<= sd
->imbalance_pct
*this_load
)
2410 busiest_load_per_task
/= busiest_nr_running
;
2412 * We're trying to get all the cpus to the average_load, so we don't
2413 * want to push ourselves above the average load, nor do we wish to
2414 * reduce the max loaded cpu below the average load, as either of these
2415 * actions would just result in more rebalancing later, and ping-pong
2416 * tasks around. Thus we look for the minimum possible imbalance.
2417 * Negative imbalances (*we* are more loaded than anyone else) will
2418 * be counted as no imbalance for these purposes -- we can't fix that
2419 * by pulling tasks to us. Be careful of negative numbers as they'll
2420 * appear as very large values with unsigned longs.
2422 if (max_load
<= busiest_load_per_task
)
2426 * In the presence of smp nice balancing, certain scenarios can have
2427 * max load less than avg load(as we skip the groups at or below
2428 * its cpu_power, while calculating max_load..)
2430 if (max_load
< avg_load
) {
2432 goto small_imbalance
;
2435 /* Don't want to pull so many tasks that a group would go idle */
2436 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2438 /* How much load to actually move to equalise the imbalance */
2439 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2440 (avg_load
- this_load
) * this->__cpu_power
)
2444 * if *imbalance is less than the average load per runnable task
2445 * there is no gaurantee that any tasks will be moved so we'll have
2446 * a think about bumping its value to force at least one task to be
2449 if (*imbalance
< busiest_load_per_task
) {
2450 unsigned long tmp
, pwr_now
, pwr_move
;
2454 pwr_move
= pwr_now
= 0;
2456 if (this_nr_running
) {
2457 this_load_per_task
/= this_nr_running
;
2458 if (busiest_load_per_task
> this_load_per_task
)
2461 this_load_per_task
= SCHED_LOAD_SCALE
;
2463 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2464 busiest_load_per_task
* imbn
) {
2465 *imbalance
= busiest_load_per_task
;
2470 * OK, we don't have enough imbalance to justify moving tasks,
2471 * however we may be able to increase total CPU power used by
2475 pwr_now
+= busiest
->__cpu_power
*
2476 min(busiest_load_per_task
, max_load
);
2477 pwr_now
+= this->__cpu_power
*
2478 min(this_load_per_task
, this_load
);
2479 pwr_now
/= SCHED_LOAD_SCALE
;
2481 /* Amount of load we'd subtract */
2482 tmp
= sg_div_cpu_power(busiest
,
2483 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2485 pwr_move
+= busiest
->__cpu_power
*
2486 min(busiest_load_per_task
, max_load
- tmp
);
2488 /* Amount of load we'd add */
2489 if (max_load
* busiest
->__cpu_power
<
2490 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2491 tmp
= sg_div_cpu_power(this,
2492 max_load
* busiest
->__cpu_power
);
2494 tmp
= sg_div_cpu_power(this,
2495 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2496 pwr_move
+= this->__cpu_power
*
2497 min(this_load_per_task
, this_load
+ tmp
);
2498 pwr_move
/= SCHED_LOAD_SCALE
;
2500 /* Move if we gain throughput */
2501 if (pwr_move
> pwr_now
)
2502 *imbalance
= busiest_load_per_task
;
2508 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2509 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2512 if (this == group_leader
&& group_leader
!= group_min
) {
2513 *imbalance
= min_load_per_task
;
2523 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2526 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2527 unsigned long imbalance
, cpumask_t
*cpus
)
2529 struct rq
*busiest
= NULL
, *rq
;
2530 unsigned long max_load
= 0;
2533 for_each_cpu_mask(i
, group
->cpumask
) {
2536 if (!cpu_isset(i
, *cpus
))
2540 wl
= weighted_cpuload(i
);
2542 if (rq
->nr_running
== 1 && wl
> imbalance
)
2545 if (wl
> max_load
) {
2555 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2556 * so long as it is large enough.
2558 #define MAX_PINNED_INTERVAL 512
2561 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2562 * tasks if there is an imbalance.
2564 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2565 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2568 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2569 struct sched_group
*group
;
2570 unsigned long imbalance
;
2572 cpumask_t cpus
= CPU_MASK_ALL
;
2573 unsigned long flags
;
2576 * When power savings policy is enabled for the parent domain, idle
2577 * sibling can pick up load irrespective of busy siblings. In this case,
2578 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2579 * portraying it as CPU_NOT_IDLE.
2581 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2582 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2585 schedstat_inc(sd
, lb_cnt
[idle
]);
2588 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2595 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2599 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2601 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2605 BUG_ON(busiest
== this_rq
);
2607 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2610 if (busiest
->nr_running
> 1) {
2612 * Attempt to move tasks. If find_busiest_group has found
2613 * an imbalance but busiest->nr_running <= 1, the group is
2614 * still unbalanced. ld_moved simply stays zero, so it is
2615 * correctly treated as an imbalance.
2617 local_irq_save(flags
);
2618 double_rq_lock(this_rq
, busiest
);
2619 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2620 imbalance
, sd
, idle
, &all_pinned
);
2621 double_rq_unlock(this_rq
, busiest
);
2622 local_irq_restore(flags
);
2625 * some other cpu did the load balance for us.
2627 if (ld_moved
&& this_cpu
!= smp_processor_id())
2628 resched_cpu(this_cpu
);
2630 /* All tasks on this runqueue were pinned by CPU affinity */
2631 if (unlikely(all_pinned
)) {
2632 cpu_clear(cpu_of(busiest
), cpus
);
2633 if (!cpus_empty(cpus
))
2640 schedstat_inc(sd
, lb_failed
[idle
]);
2641 sd
->nr_balance_failed
++;
2643 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2645 spin_lock_irqsave(&busiest
->lock
, flags
);
2647 /* don't kick the migration_thread, if the curr
2648 * task on busiest cpu can't be moved to this_cpu
2650 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2651 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2653 goto out_one_pinned
;
2656 if (!busiest
->active_balance
) {
2657 busiest
->active_balance
= 1;
2658 busiest
->push_cpu
= this_cpu
;
2661 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2663 wake_up_process(busiest
->migration_thread
);
2666 * We've kicked active balancing, reset the failure
2669 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2672 sd
->nr_balance_failed
= 0;
2674 if (likely(!active_balance
)) {
2675 /* We were unbalanced, so reset the balancing interval */
2676 sd
->balance_interval
= sd
->min_interval
;
2679 * If we've begun active balancing, start to back off. This
2680 * case may not be covered by the all_pinned logic if there
2681 * is only 1 task on the busy runqueue (because we don't call
2684 if (sd
->balance_interval
< sd
->max_interval
)
2685 sd
->balance_interval
*= 2;
2688 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2689 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2694 schedstat_inc(sd
, lb_balanced
[idle
]);
2696 sd
->nr_balance_failed
= 0;
2699 /* tune up the balancing interval */
2700 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2701 (sd
->balance_interval
< sd
->max_interval
))
2702 sd
->balance_interval
*= 2;
2704 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2705 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2711 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2712 * tasks if there is an imbalance.
2714 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2715 * this_rq is locked.
2718 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2720 struct sched_group
*group
;
2721 struct rq
*busiest
= NULL
;
2722 unsigned long imbalance
;
2726 cpumask_t cpus
= CPU_MASK_ALL
;
2729 * When power savings policy is enabled for the parent domain, idle
2730 * sibling can pick up load irrespective of busy siblings. In this case,
2731 * let the state of idle sibling percolate up as IDLE, instead of
2732 * portraying it as CPU_NOT_IDLE.
2734 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2735 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2738 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2740 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2741 &sd_idle
, &cpus
, NULL
);
2743 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2747 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2750 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2754 BUG_ON(busiest
== this_rq
);
2756 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2759 if (busiest
->nr_running
> 1) {
2760 /* Attempt to move tasks */
2761 double_lock_balance(this_rq
, busiest
);
2762 /* this_rq->clock is already updated */
2763 update_rq_clock(busiest
);
2764 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2765 imbalance
, sd
, CPU_NEWLY_IDLE
,
2767 spin_unlock(&busiest
->lock
);
2769 if (unlikely(all_pinned
)) {
2770 cpu_clear(cpu_of(busiest
), cpus
);
2771 if (!cpus_empty(cpus
))
2777 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2778 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2779 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2782 sd
->nr_balance_failed
= 0;
2787 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2788 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2789 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2791 sd
->nr_balance_failed
= 0;
2797 * idle_balance is called by schedule() if this_cpu is about to become
2798 * idle. Attempts to pull tasks from other CPUs.
2800 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2802 struct sched_domain
*sd
;
2803 int pulled_task
= -1;
2804 unsigned long next_balance
= jiffies
+ HZ
;
2806 for_each_domain(this_cpu
, sd
) {
2807 unsigned long interval
;
2809 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2812 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2813 /* If we've pulled tasks over stop searching: */
2814 pulled_task
= load_balance_newidle(this_cpu
,
2817 interval
= msecs_to_jiffies(sd
->balance_interval
);
2818 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2819 next_balance
= sd
->last_balance
+ interval
;
2823 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2825 * We are going idle. next_balance may be set based on
2826 * a busy processor. So reset next_balance.
2828 this_rq
->next_balance
= next_balance
;
2833 * active_load_balance is run by migration threads. It pushes running tasks
2834 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2835 * running on each physical CPU where possible, and avoids physical /
2836 * logical imbalances.
2838 * Called with busiest_rq locked.
2840 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2842 int target_cpu
= busiest_rq
->push_cpu
;
2843 struct sched_domain
*sd
;
2844 struct rq
*target_rq
;
2846 /* Is there any task to move? */
2847 if (busiest_rq
->nr_running
<= 1)
2850 target_rq
= cpu_rq(target_cpu
);
2853 * This condition is "impossible", if it occurs
2854 * we need to fix it. Originally reported by
2855 * Bjorn Helgaas on a 128-cpu setup.
2857 BUG_ON(busiest_rq
== target_rq
);
2859 /* move a task from busiest_rq to target_rq */
2860 double_lock_balance(busiest_rq
, target_rq
);
2861 update_rq_clock(busiest_rq
);
2862 update_rq_clock(target_rq
);
2864 /* Search for an sd spanning us and the target CPU. */
2865 for_each_domain(target_cpu
, sd
) {
2866 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2867 cpu_isset(busiest_cpu
, sd
->span
))
2872 schedstat_inc(sd
, alb_cnt
);
2874 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2876 schedstat_inc(sd
, alb_pushed
);
2878 schedstat_inc(sd
, alb_failed
);
2880 spin_unlock(&target_rq
->lock
);
2885 atomic_t load_balancer
;
2887 } nohz ____cacheline_aligned
= {
2888 .load_balancer
= ATOMIC_INIT(-1),
2889 .cpu_mask
= CPU_MASK_NONE
,
2893 * This routine will try to nominate the ilb (idle load balancing)
2894 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2895 * load balancing on behalf of all those cpus. If all the cpus in the system
2896 * go into this tickless mode, then there will be no ilb owner (as there is
2897 * no need for one) and all the cpus will sleep till the next wakeup event
2900 * For the ilb owner, tick is not stopped. And this tick will be used
2901 * for idle load balancing. ilb owner will still be part of
2904 * While stopping the tick, this cpu will become the ilb owner if there
2905 * is no other owner. And will be the owner till that cpu becomes busy
2906 * or if all cpus in the system stop their ticks at which point
2907 * there is no need for ilb owner.
2909 * When the ilb owner becomes busy, it nominates another owner, during the
2910 * next busy scheduler_tick()
2912 int select_nohz_load_balancer(int stop_tick
)
2914 int cpu
= smp_processor_id();
2917 cpu_set(cpu
, nohz
.cpu_mask
);
2918 cpu_rq(cpu
)->in_nohz_recently
= 1;
2921 * If we are going offline and still the leader, give up!
2923 if (cpu_is_offline(cpu
) &&
2924 atomic_read(&nohz
.load_balancer
) == cpu
) {
2925 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2930 /* time for ilb owner also to sleep */
2931 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2932 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2933 atomic_set(&nohz
.load_balancer
, -1);
2937 if (atomic_read(&nohz
.load_balancer
) == -1) {
2938 /* make me the ilb owner */
2939 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2941 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
2944 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
2947 cpu_clear(cpu
, nohz
.cpu_mask
);
2949 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2950 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2957 static DEFINE_SPINLOCK(balancing
);
2960 * It checks each scheduling domain to see if it is due to be balanced,
2961 * and initiates a balancing operation if so.
2963 * Balancing parameters are set up in arch_init_sched_domains.
2965 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
2968 struct rq
*rq
= cpu_rq(cpu
);
2969 unsigned long interval
;
2970 struct sched_domain
*sd
;
2971 /* Earliest time when we have to do rebalance again */
2972 unsigned long next_balance
= jiffies
+ 60*HZ
;
2973 int update_next_balance
= 0;
2975 for_each_domain(cpu
, sd
) {
2976 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2979 interval
= sd
->balance_interval
;
2980 if (idle
!= CPU_IDLE
)
2981 interval
*= sd
->busy_factor
;
2983 /* scale ms to jiffies */
2984 interval
= msecs_to_jiffies(interval
);
2985 if (unlikely(!interval
))
2987 if (interval
> HZ
*NR_CPUS
/10)
2988 interval
= HZ
*NR_CPUS
/10;
2991 if (sd
->flags
& SD_SERIALIZE
) {
2992 if (!spin_trylock(&balancing
))
2996 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
2997 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
2999 * We've pulled tasks over so either we're no
3000 * longer idle, or one of our SMT siblings is
3003 idle
= CPU_NOT_IDLE
;
3005 sd
->last_balance
= jiffies
;
3007 if (sd
->flags
& SD_SERIALIZE
)
3008 spin_unlock(&balancing
);
3010 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3011 next_balance
= sd
->last_balance
+ interval
;
3012 update_next_balance
= 1;
3016 * Stop the load balance at this level. There is another
3017 * CPU in our sched group which is doing load balancing more
3025 * next_balance will be updated only when there is a need.
3026 * When the cpu is attached to null domain for ex, it will not be
3029 if (likely(update_next_balance
))
3030 rq
->next_balance
= next_balance
;
3034 * run_rebalance_domains is triggered when needed from the scheduler tick.
3035 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3036 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3038 static void run_rebalance_domains(struct softirq_action
*h
)
3040 int this_cpu
= smp_processor_id();
3041 struct rq
*this_rq
= cpu_rq(this_cpu
);
3042 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3043 CPU_IDLE
: CPU_NOT_IDLE
;
3045 rebalance_domains(this_cpu
, idle
);
3049 * If this cpu is the owner for idle load balancing, then do the
3050 * balancing on behalf of the other idle cpus whose ticks are
3053 if (this_rq
->idle_at_tick
&&
3054 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3055 cpumask_t cpus
= nohz
.cpu_mask
;
3059 cpu_clear(this_cpu
, cpus
);
3060 for_each_cpu_mask(balance_cpu
, cpus
) {
3062 * If this cpu gets work to do, stop the load balancing
3063 * work being done for other cpus. Next load
3064 * balancing owner will pick it up.
3069 rebalance_domains(balance_cpu
, CPU_IDLE
);
3071 rq
= cpu_rq(balance_cpu
);
3072 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3073 this_rq
->next_balance
= rq
->next_balance
;
3080 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3082 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3083 * idle load balancing owner or decide to stop the periodic load balancing,
3084 * if the whole system is idle.
3086 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3090 * If we were in the nohz mode recently and busy at the current
3091 * scheduler tick, then check if we need to nominate new idle
3094 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3095 rq
->in_nohz_recently
= 0;
3097 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3098 cpu_clear(cpu
, nohz
.cpu_mask
);
3099 atomic_set(&nohz
.load_balancer
, -1);
3102 if (atomic_read(&nohz
.load_balancer
) == -1) {
3104 * simple selection for now: Nominate the
3105 * first cpu in the nohz list to be the next
3108 * TBD: Traverse the sched domains and nominate
3109 * the nearest cpu in the nohz.cpu_mask.
3111 int ilb
= first_cpu(nohz
.cpu_mask
);
3119 * If this cpu is idle and doing idle load balancing for all the
3120 * cpus with ticks stopped, is it time for that to stop?
3122 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3123 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3129 * If this cpu is idle and the idle load balancing is done by
3130 * someone else, then no need raise the SCHED_SOFTIRQ
3132 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3133 cpu_isset(cpu
, nohz
.cpu_mask
))
3136 if (time_after_eq(jiffies
, rq
->next_balance
))
3137 raise_softirq(SCHED_SOFTIRQ
);
3140 #else /* CONFIG_SMP */
3143 * on UP we do not need to balance between CPUs:
3145 static inline void idle_balance(int cpu
, struct rq
*rq
)
3149 /* Avoid "used but not defined" warning on UP */
3150 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3151 unsigned long max_nr_move
, unsigned long max_load_move
,
3152 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3153 int *all_pinned
, unsigned long *load_moved
,
3154 int *this_best_prio
, struct rq_iterator
*iterator
)
3163 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3165 EXPORT_PER_CPU_SYMBOL(kstat
);
3168 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3169 * that have not yet been banked in case the task is currently running.
3171 unsigned long long task_sched_runtime(struct task_struct
*p
)
3173 unsigned long flags
;
3177 rq
= task_rq_lock(p
, &flags
);
3178 ns
= p
->se
.sum_exec_runtime
;
3179 if (rq
->curr
== p
) {
3180 update_rq_clock(rq
);
3181 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3182 if ((s64
)delta_exec
> 0)
3185 task_rq_unlock(rq
, &flags
);
3191 * Account user cpu time to a process.
3192 * @p: the process that the cpu time gets accounted to
3193 * @hardirq_offset: the offset to subtract from hardirq_count()
3194 * @cputime: the cpu time spent in user space since the last update
3196 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3198 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3201 p
->utime
= cputime_add(p
->utime
, cputime
);
3203 /* Add user time to cpustat. */
3204 tmp
= cputime_to_cputime64(cputime
);
3205 if (TASK_NICE(p
) > 0)
3206 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3208 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3212 * Account system cpu time to a process.
3213 * @p: the process that the cpu time gets accounted to
3214 * @hardirq_offset: the offset to subtract from hardirq_count()
3215 * @cputime: the cpu time spent in kernel space since the last update
3217 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3220 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3221 struct rq
*rq
= this_rq();
3224 p
->stime
= cputime_add(p
->stime
, cputime
);
3226 /* Add system time to cpustat. */
3227 tmp
= cputime_to_cputime64(cputime
);
3228 if (hardirq_count() - hardirq_offset
)
3229 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3230 else if (softirq_count())
3231 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3232 else if (p
!= rq
->idle
)
3233 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3234 else if (atomic_read(&rq
->nr_iowait
) > 0)
3235 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3237 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3238 /* Account for system time used */
3239 acct_update_integrals(p
);
3243 * Account for involuntary wait time.
3244 * @p: the process from which the cpu time has been stolen
3245 * @steal: the cpu time spent in involuntary wait
3247 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3249 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3250 cputime64_t tmp
= cputime_to_cputime64(steal
);
3251 struct rq
*rq
= this_rq();
3253 if (p
== rq
->idle
) {
3254 p
->stime
= cputime_add(p
->stime
, steal
);
3255 if (atomic_read(&rq
->nr_iowait
) > 0)
3256 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3258 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3260 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3264 * This function gets called by the timer code, with HZ frequency.
3265 * We call it with interrupts disabled.
3267 * It also gets called by the fork code, when changing the parent's
3270 void scheduler_tick(void)
3272 int cpu
= smp_processor_id();
3273 struct rq
*rq
= cpu_rq(cpu
);
3274 struct task_struct
*curr
= rq
->curr
;
3275 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3277 spin_lock(&rq
->lock
);
3278 __update_rq_clock(rq
);
3280 * Let rq->clock advance by at least TICK_NSEC:
3282 if (unlikely(rq
->clock
< next_tick
))
3283 rq
->clock
= next_tick
;
3284 rq
->tick_timestamp
= rq
->clock
;
3285 update_cpu_load(rq
);
3286 if (curr
!= rq
->idle
) /* FIXME: needed? */
3287 curr
->sched_class
->task_tick(rq
, curr
);
3288 spin_unlock(&rq
->lock
);
3291 rq
->idle_at_tick
= idle_cpu(cpu
);
3292 trigger_load_balance(rq
, cpu
);
3296 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3298 void fastcall
add_preempt_count(int val
)
3303 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3305 preempt_count() += val
;
3307 * Spinlock count overflowing soon?
3309 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3312 EXPORT_SYMBOL(add_preempt_count
);
3314 void fastcall
sub_preempt_count(int val
)
3319 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3322 * Is the spinlock portion underflowing?
3324 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3325 !(preempt_count() & PREEMPT_MASK
)))
3328 preempt_count() -= val
;
3330 EXPORT_SYMBOL(sub_preempt_count
);
3335 * Print scheduling while atomic bug:
3337 static noinline
void __schedule_bug(struct task_struct
*prev
)
3339 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3340 prev
->comm
, preempt_count(), prev
->pid
);
3341 debug_show_held_locks(prev
);
3342 if (irqs_disabled())
3343 print_irqtrace_events(prev
);
3348 * Various schedule()-time debugging checks and statistics:
3350 static inline void schedule_debug(struct task_struct
*prev
)
3353 * Test if we are atomic. Since do_exit() needs to call into
3354 * schedule() atomically, we ignore that path for now.
3355 * Otherwise, whine if we are scheduling when we should not be.
3357 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3358 __schedule_bug(prev
);
3360 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3362 schedstat_inc(this_rq(), sched_cnt
);
3366 * Pick up the highest-prio task:
3368 static inline struct task_struct
*
3369 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3371 struct sched_class
*class;
3372 struct task_struct
*p
;
3375 * Optimization: we know that if all tasks are in
3376 * the fair class we can call that function directly:
3378 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3379 p
= fair_sched_class
.pick_next_task(rq
);
3384 class = sched_class_highest
;
3386 p
= class->pick_next_task(rq
);
3390 * Will never be NULL as the idle class always
3391 * returns a non-NULL p:
3393 class = class->next
;
3398 * schedule() is the main scheduler function.
3400 asmlinkage
void __sched
schedule(void)
3402 struct task_struct
*prev
, *next
;
3409 cpu
= smp_processor_id();
3413 switch_count
= &prev
->nivcsw
;
3415 release_kernel_lock(prev
);
3416 need_resched_nonpreemptible
:
3418 schedule_debug(prev
);
3420 spin_lock_irq(&rq
->lock
);
3421 clear_tsk_need_resched(prev
);
3422 __update_rq_clock(rq
);
3424 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3425 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3426 unlikely(signal_pending(prev
)))) {
3427 prev
->state
= TASK_RUNNING
;
3429 deactivate_task(rq
, prev
, 1);
3431 switch_count
= &prev
->nvcsw
;
3434 if (unlikely(!rq
->nr_running
))
3435 idle_balance(cpu
, rq
);
3437 prev
->sched_class
->put_prev_task(rq
, prev
);
3438 next
= pick_next_task(rq
, prev
);
3440 sched_info_switch(prev
, next
);
3442 if (likely(prev
!= next
)) {
3447 context_switch(rq
, prev
, next
); /* unlocks the rq */
3449 spin_unlock_irq(&rq
->lock
);
3451 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3452 cpu
= smp_processor_id();
3454 goto need_resched_nonpreemptible
;
3456 preempt_enable_no_resched();
3457 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3460 EXPORT_SYMBOL(schedule
);
3462 #ifdef CONFIG_PREEMPT
3464 * this is the entry point to schedule() from in-kernel preemption
3465 * off of preempt_enable. Kernel preemptions off return from interrupt
3466 * occur there and call schedule directly.
3468 asmlinkage
void __sched
preempt_schedule(void)
3470 struct thread_info
*ti
= current_thread_info();
3471 #ifdef CONFIG_PREEMPT_BKL
3472 struct task_struct
*task
= current
;
3473 int saved_lock_depth
;
3476 * If there is a non-zero preempt_count or interrupts are disabled,
3477 * we do not want to preempt the current task. Just return..
3479 if (likely(ti
->preempt_count
|| irqs_disabled()))
3483 add_preempt_count(PREEMPT_ACTIVE
);
3485 * We keep the big kernel semaphore locked, but we
3486 * clear ->lock_depth so that schedule() doesnt
3487 * auto-release the semaphore:
3489 #ifdef CONFIG_PREEMPT_BKL
3490 saved_lock_depth
= task
->lock_depth
;
3491 task
->lock_depth
= -1;
3494 #ifdef CONFIG_PREEMPT_BKL
3495 task
->lock_depth
= saved_lock_depth
;
3497 sub_preempt_count(PREEMPT_ACTIVE
);
3499 /* we could miss a preemption opportunity between schedule and now */
3501 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3504 EXPORT_SYMBOL(preempt_schedule
);
3507 * this is the entry point to schedule() from kernel preemption
3508 * off of irq context.
3509 * Note, that this is called and return with irqs disabled. This will
3510 * protect us against recursive calling from irq.
3512 asmlinkage
void __sched
preempt_schedule_irq(void)
3514 struct thread_info
*ti
= current_thread_info();
3515 #ifdef CONFIG_PREEMPT_BKL
3516 struct task_struct
*task
= current
;
3517 int saved_lock_depth
;
3519 /* Catch callers which need to be fixed */
3520 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3523 add_preempt_count(PREEMPT_ACTIVE
);
3525 * We keep the big kernel semaphore locked, but we
3526 * clear ->lock_depth so that schedule() doesnt
3527 * auto-release the semaphore:
3529 #ifdef CONFIG_PREEMPT_BKL
3530 saved_lock_depth
= task
->lock_depth
;
3531 task
->lock_depth
= -1;
3535 local_irq_disable();
3536 #ifdef CONFIG_PREEMPT_BKL
3537 task
->lock_depth
= saved_lock_depth
;
3539 sub_preempt_count(PREEMPT_ACTIVE
);
3541 /* we could miss a preemption opportunity between schedule and now */
3543 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3547 #endif /* CONFIG_PREEMPT */
3549 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3552 return try_to_wake_up(curr
->private, mode
, sync
);
3554 EXPORT_SYMBOL(default_wake_function
);
3557 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3558 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3559 * number) then we wake all the non-exclusive tasks and one exclusive task.
3561 * There are circumstances in which we can try to wake a task which has already
3562 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3563 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3565 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3566 int nr_exclusive
, int sync
, void *key
)
3568 wait_queue_t
*curr
, *next
;
3570 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3571 unsigned flags
= curr
->flags
;
3573 if (curr
->func(curr
, mode
, sync
, key
) &&
3574 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3580 * __wake_up - wake up threads blocked on a waitqueue.
3582 * @mode: which threads
3583 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3584 * @key: is directly passed to the wakeup function
3586 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3587 int nr_exclusive
, void *key
)
3589 unsigned long flags
;
3591 spin_lock_irqsave(&q
->lock
, flags
);
3592 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3593 spin_unlock_irqrestore(&q
->lock
, flags
);
3595 EXPORT_SYMBOL(__wake_up
);
3598 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3600 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3602 __wake_up_common(q
, mode
, 1, 0, NULL
);
3606 * __wake_up_sync - wake up threads blocked on a waitqueue.
3608 * @mode: which threads
3609 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3611 * The sync wakeup differs that the waker knows that it will schedule
3612 * away soon, so while the target thread will be woken up, it will not
3613 * be migrated to another CPU - ie. the two threads are 'synchronized'
3614 * with each other. This can prevent needless bouncing between CPUs.
3616 * On UP it can prevent extra preemption.
3619 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3621 unsigned long flags
;
3627 if (unlikely(!nr_exclusive
))
3630 spin_lock_irqsave(&q
->lock
, flags
);
3631 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3632 spin_unlock_irqrestore(&q
->lock
, flags
);
3634 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3636 void fastcall
complete(struct completion
*x
)
3638 unsigned long flags
;
3640 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3642 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3644 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3646 EXPORT_SYMBOL(complete
);
3648 void fastcall
complete_all(struct completion
*x
)
3650 unsigned long flags
;
3652 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3653 x
->done
+= UINT_MAX
/2;
3654 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3656 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3658 EXPORT_SYMBOL(complete_all
);
3660 void fastcall __sched
wait_for_completion(struct completion
*x
)
3664 spin_lock_irq(&x
->wait
.lock
);
3666 DECLARE_WAITQUEUE(wait
, current
);
3668 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3669 __add_wait_queue_tail(&x
->wait
, &wait
);
3671 __set_current_state(TASK_UNINTERRUPTIBLE
);
3672 spin_unlock_irq(&x
->wait
.lock
);
3674 spin_lock_irq(&x
->wait
.lock
);
3676 __remove_wait_queue(&x
->wait
, &wait
);
3679 spin_unlock_irq(&x
->wait
.lock
);
3681 EXPORT_SYMBOL(wait_for_completion
);
3683 unsigned long fastcall __sched
3684 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3688 spin_lock_irq(&x
->wait
.lock
);
3690 DECLARE_WAITQUEUE(wait
, current
);
3692 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3693 __add_wait_queue_tail(&x
->wait
, &wait
);
3695 __set_current_state(TASK_UNINTERRUPTIBLE
);
3696 spin_unlock_irq(&x
->wait
.lock
);
3697 timeout
= schedule_timeout(timeout
);
3698 spin_lock_irq(&x
->wait
.lock
);
3700 __remove_wait_queue(&x
->wait
, &wait
);
3704 __remove_wait_queue(&x
->wait
, &wait
);
3708 spin_unlock_irq(&x
->wait
.lock
);
3711 EXPORT_SYMBOL(wait_for_completion_timeout
);
3713 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3719 spin_lock_irq(&x
->wait
.lock
);
3721 DECLARE_WAITQUEUE(wait
, current
);
3723 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3724 __add_wait_queue_tail(&x
->wait
, &wait
);
3726 if (signal_pending(current
)) {
3728 __remove_wait_queue(&x
->wait
, &wait
);
3731 __set_current_state(TASK_INTERRUPTIBLE
);
3732 spin_unlock_irq(&x
->wait
.lock
);
3734 spin_lock_irq(&x
->wait
.lock
);
3736 __remove_wait_queue(&x
->wait
, &wait
);
3740 spin_unlock_irq(&x
->wait
.lock
);
3744 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3746 unsigned long fastcall __sched
3747 wait_for_completion_interruptible_timeout(struct completion
*x
,
3748 unsigned long timeout
)
3752 spin_lock_irq(&x
->wait
.lock
);
3754 DECLARE_WAITQUEUE(wait
, current
);
3756 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3757 __add_wait_queue_tail(&x
->wait
, &wait
);
3759 if (signal_pending(current
)) {
3760 timeout
= -ERESTARTSYS
;
3761 __remove_wait_queue(&x
->wait
, &wait
);
3764 __set_current_state(TASK_INTERRUPTIBLE
);
3765 spin_unlock_irq(&x
->wait
.lock
);
3766 timeout
= schedule_timeout(timeout
);
3767 spin_lock_irq(&x
->wait
.lock
);
3769 __remove_wait_queue(&x
->wait
, &wait
);
3773 __remove_wait_queue(&x
->wait
, &wait
);
3777 spin_unlock_irq(&x
->wait
.lock
);
3780 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3783 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3785 spin_lock_irqsave(&q
->lock
, *flags
);
3786 __add_wait_queue(q
, wait
);
3787 spin_unlock(&q
->lock
);
3791 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3793 spin_lock_irq(&q
->lock
);
3794 __remove_wait_queue(q
, wait
);
3795 spin_unlock_irqrestore(&q
->lock
, *flags
);
3798 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3800 unsigned long flags
;
3803 init_waitqueue_entry(&wait
, current
);
3805 current
->state
= TASK_INTERRUPTIBLE
;
3807 sleep_on_head(q
, &wait
, &flags
);
3809 sleep_on_tail(q
, &wait
, &flags
);
3811 EXPORT_SYMBOL(interruptible_sleep_on
);
3814 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3816 unsigned long flags
;
3819 init_waitqueue_entry(&wait
, current
);
3821 current
->state
= TASK_INTERRUPTIBLE
;
3823 sleep_on_head(q
, &wait
, &flags
);
3824 timeout
= schedule_timeout(timeout
);
3825 sleep_on_tail(q
, &wait
, &flags
);
3829 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3831 void __sched
sleep_on(wait_queue_head_t
*q
)
3833 unsigned long flags
;
3836 init_waitqueue_entry(&wait
, current
);
3838 current
->state
= TASK_UNINTERRUPTIBLE
;
3840 sleep_on_head(q
, &wait
, &flags
);
3842 sleep_on_tail(q
, &wait
, &flags
);
3844 EXPORT_SYMBOL(sleep_on
);
3846 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3848 unsigned long flags
;
3851 init_waitqueue_entry(&wait
, current
);
3853 current
->state
= TASK_UNINTERRUPTIBLE
;
3855 sleep_on_head(q
, &wait
, &flags
);
3856 timeout
= schedule_timeout(timeout
);
3857 sleep_on_tail(q
, &wait
, &flags
);
3861 EXPORT_SYMBOL(sleep_on_timeout
);
3863 #ifdef CONFIG_RT_MUTEXES
3866 * rt_mutex_setprio - set the current priority of a task
3868 * @prio: prio value (kernel-internal form)
3870 * This function changes the 'effective' priority of a task. It does
3871 * not touch ->normal_prio like __setscheduler().
3873 * Used by the rt_mutex code to implement priority inheritance logic.
3875 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3877 unsigned long flags
;
3881 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3883 rq
= task_rq_lock(p
, &flags
);
3884 update_rq_clock(rq
);
3887 on_rq
= p
->se
.on_rq
;
3889 dequeue_task(rq
, p
, 0);
3892 p
->sched_class
= &rt_sched_class
;
3894 p
->sched_class
= &fair_sched_class
;
3899 enqueue_task(rq
, p
, 0);
3901 * Reschedule if we are currently running on this runqueue and
3902 * our priority decreased, or if we are not currently running on
3903 * this runqueue and our priority is higher than the current's
3905 if (task_running(rq
, p
)) {
3906 if (p
->prio
> oldprio
)
3907 resched_task(rq
->curr
);
3909 check_preempt_curr(rq
, p
);
3912 task_rq_unlock(rq
, &flags
);
3917 void set_user_nice(struct task_struct
*p
, long nice
)
3919 int old_prio
, delta
, on_rq
;
3920 unsigned long flags
;
3923 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3926 * We have to be careful, if called from sys_setpriority(),
3927 * the task might be in the middle of scheduling on another CPU.
3929 rq
= task_rq_lock(p
, &flags
);
3930 update_rq_clock(rq
);
3932 * The RT priorities are set via sched_setscheduler(), but we still
3933 * allow the 'normal' nice value to be set - but as expected
3934 * it wont have any effect on scheduling until the task is
3935 * SCHED_FIFO/SCHED_RR:
3937 if (task_has_rt_policy(p
)) {
3938 p
->static_prio
= NICE_TO_PRIO(nice
);
3941 on_rq
= p
->se
.on_rq
;
3943 dequeue_task(rq
, p
, 0);
3947 p
->static_prio
= NICE_TO_PRIO(nice
);
3950 p
->prio
= effective_prio(p
);
3951 delta
= p
->prio
- old_prio
;
3954 enqueue_task(rq
, p
, 0);
3957 * If the task increased its priority or is running and
3958 * lowered its priority, then reschedule its CPU:
3960 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3961 resched_task(rq
->curr
);
3964 task_rq_unlock(rq
, &flags
);
3966 EXPORT_SYMBOL(set_user_nice
);
3969 * can_nice - check if a task can reduce its nice value
3973 int can_nice(const struct task_struct
*p
, const int nice
)
3975 /* convert nice value [19,-20] to rlimit style value [1,40] */
3976 int nice_rlim
= 20 - nice
;
3978 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3979 capable(CAP_SYS_NICE
));
3982 #ifdef __ARCH_WANT_SYS_NICE
3985 * sys_nice - change the priority of the current process.
3986 * @increment: priority increment
3988 * sys_setpriority is a more generic, but much slower function that
3989 * does similar things.
3991 asmlinkage
long sys_nice(int increment
)
3996 * Setpriority might change our priority at the same moment.
3997 * We don't have to worry. Conceptually one call occurs first
3998 * and we have a single winner.
4000 if (increment
< -40)
4005 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4011 if (increment
< 0 && !can_nice(current
, nice
))
4014 retval
= security_task_setnice(current
, nice
);
4018 set_user_nice(current
, nice
);
4025 * task_prio - return the priority value of a given task.
4026 * @p: the task in question.
4028 * This is the priority value as seen by users in /proc.
4029 * RT tasks are offset by -200. Normal tasks are centered
4030 * around 0, value goes from -16 to +15.
4032 int task_prio(const struct task_struct
*p
)
4034 return p
->prio
- MAX_RT_PRIO
;
4038 * task_nice - return the nice value of a given task.
4039 * @p: the task in question.
4041 int task_nice(const struct task_struct
*p
)
4043 return TASK_NICE(p
);
4045 EXPORT_SYMBOL_GPL(task_nice
);
4048 * idle_cpu - is a given cpu idle currently?
4049 * @cpu: the processor in question.
4051 int idle_cpu(int cpu
)
4053 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4057 * idle_task - return the idle task for a given cpu.
4058 * @cpu: the processor in question.
4060 struct task_struct
*idle_task(int cpu
)
4062 return cpu_rq(cpu
)->idle
;
4066 * find_process_by_pid - find a process with a matching PID value.
4067 * @pid: the pid in question.
4069 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4071 return pid
? find_task_by_pid(pid
) : current
;
4074 /* Actually do priority change: must hold rq lock. */
4076 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4078 BUG_ON(p
->se
.on_rq
);
4081 switch (p
->policy
) {
4085 p
->sched_class
= &fair_sched_class
;
4089 p
->sched_class
= &rt_sched_class
;
4093 p
->rt_priority
= prio
;
4094 p
->normal_prio
= normal_prio(p
);
4095 /* we are holding p->pi_lock already */
4096 p
->prio
= rt_mutex_getprio(p
);
4101 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4102 * @p: the task in question.
4103 * @policy: new policy.
4104 * @param: structure containing the new RT priority.
4106 * NOTE that the task may be already dead.
4108 int sched_setscheduler(struct task_struct
*p
, int policy
,
4109 struct sched_param
*param
)
4111 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4112 unsigned long flags
;
4115 /* may grab non-irq protected spin_locks */
4116 BUG_ON(in_interrupt());
4118 /* double check policy once rq lock held */
4120 policy
= oldpolicy
= p
->policy
;
4121 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4122 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4123 policy
!= SCHED_IDLE
)
4126 * Valid priorities for SCHED_FIFO and SCHED_RR are
4127 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4128 * SCHED_BATCH and SCHED_IDLE is 0.
4130 if (param
->sched_priority
< 0 ||
4131 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4132 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4134 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4138 * Allow unprivileged RT tasks to decrease priority:
4140 if (!capable(CAP_SYS_NICE
)) {
4141 if (rt_policy(policy
)) {
4142 unsigned long rlim_rtprio
;
4144 if (!lock_task_sighand(p
, &flags
))
4146 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4147 unlock_task_sighand(p
, &flags
);
4149 /* can't set/change the rt policy */
4150 if (policy
!= p
->policy
&& !rlim_rtprio
)
4153 /* can't increase priority */
4154 if (param
->sched_priority
> p
->rt_priority
&&
4155 param
->sched_priority
> rlim_rtprio
)
4159 * Like positive nice levels, dont allow tasks to
4160 * move out of SCHED_IDLE either:
4162 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4165 /* can't change other user's priorities */
4166 if ((current
->euid
!= p
->euid
) &&
4167 (current
->euid
!= p
->uid
))
4171 retval
= security_task_setscheduler(p
, policy
, param
);
4175 * make sure no PI-waiters arrive (or leave) while we are
4176 * changing the priority of the task:
4178 spin_lock_irqsave(&p
->pi_lock
, flags
);
4180 * To be able to change p->policy safely, the apropriate
4181 * runqueue lock must be held.
4183 rq
= __task_rq_lock(p
);
4184 /* recheck policy now with rq lock held */
4185 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4186 policy
= oldpolicy
= -1;
4187 __task_rq_unlock(rq
);
4188 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4191 update_rq_clock(rq
);
4192 on_rq
= p
->se
.on_rq
;
4194 deactivate_task(rq
, p
, 0);
4196 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4198 activate_task(rq
, p
, 0);
4200 * Reschedule if we are currently running on this runqueue and
4201 * our priority decreased, or if we are not currently running on
4202 * this runqueue and our priority is higher than the current's
4204 if (task_running(rq
, p
)) {
4205 if (p
->prio
> oldprio
)
4206 resched_task(rq
->curr
);
4208 check_preempt_curr(rq
, p
);
4211 __task_rq_unlock(rq
);
4212 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4214 rt_mutex_adjust_pi(p
);
4218 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4221 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4223 struct sched_param lparam
;
4224 struct task_struct
*p
;
4227 if (!param
|| pid
< 0)
4229 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4234 p
= find_process_by_pid(pid
);
4236 retval
= sched_setscheduler(p
, policy
, &lparam
);
4243 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4244 * @pid: the pid in question.
4245 * @policy: new policy.
4246 * @param: structure containing the new RT priority.
4248 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4249 struct sched_param __user
*param
)
4251 /* negative values for policy are not valid */
4255 return do_sched_setscheduler(pid
, policy
, param
);
4259 * sys_sched_setparam - set/change the RT priority of a thread
4260 * @pid: the pid in question.
4261 * @param: structure containing the new RT priority.
4263 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4265 return do_sched_setscheduler(pid
, -1, param
);
4269 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4270 * @pid: the pid in question.
4272 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4274 struct task_struct
*p
;
4275 int retval
= -EINVAL
;
4281 read_lock(&tasklist_lock
);
4282 p
= find_process_by_pid(pid
);
4284 retval
= security_task_getscheduler(p
);
4288 read_unlock(&tasklist_lock
);
4295 * sys_sched_getscheduler - get the RT priority of a thread
4296 * @pid: the pid in question.
4297 * @param: structure containing the RT priority.
4299 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4301 struct sched_param lp
;
4302 struct task_struct
*p
;
4303 int retval
= -EINVAL
;
4305 if (!param
|| pid
< 0)
4308 read_lock(&tasklist_lock
);
4309 p
= find_process_by_pid(pid
);
4314 retval
= security_task_getscheduler(p
);
4318 lp
.sched_priority
= p
->rt_priority
;
4319 read_unlock(&tasklist_lock
);
4322 * This one might sleep, we cannot do it with a spinlock held ...
4324 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4330 read_unlock(&tasklist_lock
);
4334 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4336 cpumask_t cpus_allowed
;
4337 struct task_struct
*p
;
4340 mutex_lock(&sched_hotcpu_mutex
);
4341 read_lock(&tasklist_lock
);
4343 p
= find_process_by_pid(pid
);
4345 read_unlock(&tasklist_lock
);
4346 mutex_unlock(&sched_hotcpu_mutex
);
4351 * It is not safe to call set_cpus_allowed with the
4352 * tasklist_lock held. We will bump the task_struct's
4353 * usage count and then drop tasklist_lock.
4356 read_unlock(&tasklist_lock
);
4359 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4360 !capable(CAP_SYS_NICE
))
4363 retval
= security_task_setscheduler(p
, 0, NULL
);
4367 cpus_allowed
= cpuset_cpus_allowed(p
);
4368 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4369 retval
= set_cpus_allowed(p
, new_mask
);
4373 mutex_unlock(&sched_hotcpu_mutex
);
4377 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4378 cpumask_t
*new_mask
)
4380 if (len
< sizeof(cpumask_t
)) {
4381 memset(new_mask
, 0, sizeof(cpumask_t
));
4382 } else if (len
> sizeof(cpumask_t
)) {
4383 len
= sizeof(cpumask_t
);
4385 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4389 * sys_sched_setaffinity - set the cpu affinity of a process
4390 * @pid: pid of the process
4391 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4392 * @user_mask_ptr: user-space pointer to the new cpu mask
4394 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4395 unsigned long __user
*user_mask_ptr
)
4400 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4404 return sched_setaffinity(pid
, new_mask
);
4408 * Represents all cpu's present in the system
4409 * In systems capable of hotplug, this map could dynamically grow
4410 * as new cpu's are detected in the system via any platform specific
4411 * method, such as ACPI for e.g.
4414 cpumask_t cpu_present_map __read_mostly
;
4415 EXPORT_SYMBOL(cpu_present_map
);
4418 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4419 EXPORT_SYMBOL(cpu_online_map
);
4421 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4422 EXPORT_SYMBOL(cpu_possible_map
);
4425 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4427 struct task_struct
*p
;
4430 mutex_lock(&sched_hotcpu_mutex
);
4431 read_lock(&tasklist_lock
);
4434 p
= find_process_by_pid(pid
);
4438 retval
= security_task_getscheduler(p
);
4442 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4445 read_unlock(&tasklist_lock
);
4446 mutex_unlock(&sched_hotcpu_mutex
);
4452 * sys_sched_getaffinity - get the cpu affinity of a process
4453 * @pid: pid of the process
4454 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4455 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4457 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4458 unsigned long __user
*user_mask_ptr
)
4463 if (len
< sizeof(cpumask_t
))
4466 ret
= sched_getaffinity(pid
, &mask
);
4470 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4473 return sizeof(cpumask_t
);
4477 * sys_sched_yield - yield the current processor to other threads.
4479 * This function yields the current CPU to other tasks. If there are no
4480 * other threads running on this CPU then this function will return.
4482 asmlinkage
long sys_sched_yield(void)
4484 struct rq
*rq
= this_rq_lock();
4486 schedstat_inc(rq
, yld_cnt
);
4487 current
->sched_class
->yield_task(rq
, current
);
4490 * Since we are going to call schedule() anyway, there's
4491 * no need to preempt or enable interrupts:
4493 __release(rq
->lock
);
4494 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4495 _raw_spin_unlock(&rq
->lock
);
4496 preempt_enable_no_resched();
4503 static void __cond_resched(void)
4505 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4506 __might_sleep(__FILE__
, __LINE__
);
4509 * The BKS might be reacquired before we have dropped
4510 * PREEMPT_ACTIVE, which could trigger a second
4511 * cond_resched() call.
4514 add_preempt_count(PREEMPT_ACTIVE
);
4516 sub_preempt_count(PREEMPT_ACTIVE
);
4517 } while (need_resched());
4520 int __sched
cond_resched(void)
4522 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4523 system_state
== SYSTEM_RUNNING
) {
4529 EXPORT_SYMBOL(cond_resched
);
4532 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4533 * call schedule, and on return reacquire the lock.
4535 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4536 * operations here to prevent schedule() from being called twice (once via
4537 * spin_unlock(), once by hand).
4539 int cond_resched_lock(spinlock_t
*lock
)
4543 if (need_lockbreak(lock
)) {
4549 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4550 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4551 _raw_spin_unlock(lock
);
4552 preempt_enable_no_resched();
4559 EXPORT_SYMBOL(cond_resched_lock
);
4561 int __sched
cond_resched_softirq(void)
4563 BUG_ON(!in_softirq());
4565 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4573 EXPORT_SYMBOL(cond_resched_softirq
);
4576 * yield - yield the current processor to other threads.
4578 * This is a shortcut for kernel-space yielding - it marks the
4579 * thread runnable and calls sys_sched_yield().
4581 void __sched
yield(void)
4583 set_current_state(TASK_RUNNING
);
4586 EXPORT_SYMBOL(yield
);
4589 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4590 * that process accounting knows that this is a task in IO wait state.
4592 * But don't do that if it is a deliberate, throttling IO wait (this task
4593 * has set its backing_dev_info: the queue against which it should throttle)
4595 void __sched
io_schedule(void)
4597 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4599 delayacct_blkio_start();
4600 atomic_inc(&rq
->nr_iowait
);
4602 atomic_dec(&rq
->nr_iowait
);
4603 delayacct_blkio_end();
4605 EXPORT_SYMBOL(io_schedule
);
4607 long __sched
io_schedule_timeout(long timeout
)
4609 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4612 delayacct_blkio_start();
4613 atomic_inc(&rq
->nr_iowait
);
4614 ret
= schedule_timeout(timeout
);
4615 atomic_dec(&rq
->nr_iowait
);
4616 delayacct_blkio_end();
4621 * sys_sched_get_priority_max - return maximum RT priority.
4622 * @policy: scheduling class.
4624 * this syscall returns the maximum rt_priority that can be used
4625 * by a given scheduling class.
4627 asmlinkage
long sys_sched_get_priority_max(int policy
)
4634 ret
= MAX_USER_RT_PRIO
-1;
4646 * sys_sched_get_priority_min - return minimum RT priority.
4647 * @policy: scheduling class.
4649 * this syscall returns the minimum rt_priority that can be used
4650 * by a given scheduling class.
4652 asmlinkage
long sys_sched_get_priority_min(int policy
)
4670 * sys_sched_rr_get_interval - return the default timeslice of a process.
4671 * @pid: pid of the process.
4672 * @interval: userspace pointer to the timeslice value.
4674 * this syscall writes the default timeslice value of a given process
4675 * into the user-space timespec buffer. A value of '0' means infinity.
4678 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4680 struct task_struct
*p
;
4681 int retval
= -EINVAL
;
4688 read_lock(&tasklist_lock
);
4689 p
= find_process_by_pid(pid
);
4693 retval
= security_task_getscheduler(p
);
4697 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4698 0 : static_prio_timeslice(p
->static_prio
), &t
);
4699 read_unlock(&tasklist_lock
);
4700 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4704 read_unlock(&tasklist_lock
);
4708 static const char stat_nam
[] = "RSDTtZX";
4710 static void show_task(struct task_struct
*p
)
4712 unsigned long free
= 0;
4715 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4716 printk("%-13.13s %c", p
->comm
,
4717 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4718 #if BITS_PER_LONG == 32
4719 if (state
== TASK_RUNNING
)
4720 printk(" running ");
4722 printk(" %08lx ", thread_saved_pc(p
));
4724 if (state
== TASK_RUNNING
)
4725 printk(" running task ");
4727 printk(" %016lx ", thread_saved_pc(p
));
4729 #ifdef CONFIG_DEBUG_STACK_USAGE
4731 unsigned long *n
= end_of_stack(p
);
4734 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4737 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4739 if (state
!= TASK_RUNNING
)
4740 show_stack(p
, NULL
);
4743 void show_state_filter(unsigned long state_filter
)
4745 struct task_struct
*g
, *p
;
4747 #if BITS_PER_LONG == 32
4749 " task PC stack pid father\n");
4752 " task PC stack pid father\n");
4754 read_lock(&tasklist_lock
);
4755 do_each_thread(g
, p
) {
4757 * reset the NMI-timeout, listing all files on a slow
4758 * console might take alot of time:
4760 touch_nmi_watchdog();
4761 if (!state_filter
|| (p
->state
& state_filter
))
4763 } while_each_thread(g
, p
);
4765 touch_all_softlockup_watchdogs();
4767 #ifdef CONFIG_SCHED_DEBUG
4768 sysrq_sched_debug_show();
4770 read_unlock(&tasklist_lock
);
4772 * Only show locks if all tasks are dumped:
4774 if (state_filter
== -1)
4775 debug_show_all_locks();
4778 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4780 idle
->sched_class
= &idle_sched_class
;
4784 * init_idle - set up an idle thread for a given CPU
4785 * @idle: task in question
4786 * @cpu: cpu the idle task belongs to
4788 * NOTE: this function does not set the idle thread's NEED_RESCHED
4789 * flag, to make booting more robust.
4791 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4793 struct rq
*rq
= cpu_rq(cpu
);
4794 unsigned long flags
;
4797 idle
->se
.exec_start
= sched_clock();
4799 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4800 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4801 __set_task_cpu(idle
, cpu
);
4803 spin_lock_irqsave(&rq
->lock
, flags
);
4804 rq
->curr
= rq
->idle
= idle
;
4805 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4808 spin_unlock_irqrestore(&rq
->lock
, flags
);
4810 /* Set the preempt count _outside_ the spinlocks! */
4811 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4812 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4814 task_thread_info(idle
)->preempt_count
= 0;
4817 * The idle tasks have their own, simple scheduling class:
4819 idle
->sched_class
= &idle_sched_class
;
4823 * In a system that switches off the HZ timer nohz_cpu_mask
4824 * indicates which cpus entered this state. This is used
4825 * in the rcu update to wait only for active cpus. For system
4826 * which do not switch off the HZ timer nohz_cpu_mask should
4827 * always be CPU_MASK_NONE.
4829 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4833 * This is how migration works:
4835 * 1) we queue a struct migration_req structure in the source CPU's
4836 * runqueue and wake up that CPU's migration thread.
4837 * 2) we down() the locked semaphore => thread blocks.
4838 * 3) migration thread wakes up (implicitly it forces the migrated
4839 * thread off the CPU)
4840 * 4) it gets the migration request and checks whether the migrated
4841 * task is still in the wrong runqueue.
4842 * 5) if it's in the wrong runqueue then the migration thread removes
4843 * it and puts it into the right queue.
4844 * 6) migration thread up()s the semaphore.
4845 * 7) we wake up and the migration is done.
4849 * Change a given task's CPU affinity. Migrate the thread to a
4850 * proper CPU and schedule it away if the CPU it's executing on
4851 * is removed from the allowed bitmask.
4853 * NOTE: the caller must have a valid reference to the task, the
4854 * task must not exit() & deallocate itself prematurely. The
4855 * call is not atomic; no spinlocks may be held.
4857 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4859 struct migration_req req
;
4860 unsigned long flags
;
4864 rq
= task_rq_lock(p
, &flags
);
4865 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4870 p
->cpus_allowed
= new_mask
;
4871 /* Can the task run on the task's current CPU? If so, we're done */
4872 if (cpu_isset(task_cpu(p
), new_mask
))
4875 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4876 /* Need help from migration thread: drop lock and wait. */
4877 task_rq_unlock(rq
, &flags
);
4878 wake_up_process(rq
->migration_thread
);
4879 wait_for_completion(&req
.done
);
4880 tlb_migrate_finish(p
->mm
);
4884 task_rq_unlock(rq
, &flags
);
4888 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4891 * Move (not current) task off this cpu, onto dest cpu. We're doing
4892 * this because either it can't run here any more (set_cpus_allowed()
4893 * away from this CPU, or CPU going down), or because we're
4894 * attempting to rebalance this task on exec (sched_exec).
4896 * So we race with normal scheduler movements, but that's OK, as long
4897 * as the task is no longer on this CPU.
4899 * Returns non-zero if task was successfully migrated.
4901 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4903 struct rq
*rq_dest
, *rq_src
;
4906 if (unlikely(cpu_is_offline(dest_cpu
)))
4909 rq_src
= cpu_rq(src_cpu
);
4910 rq_dest
= cpu_rq(dest_cpu
);
4912 double_rq_lock(rq_src
, rq_dest
);
4913 /* Already moved. */
4914 if (task_cpu(p
) != src_cpu
)
4916 /* Affinity changed (again). */
4917 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4920 on_rq
= p
->se
.on_rq
;
4922 deactivate_task(rq_src
, p
, 0);
4924 set_task_cpu(p
, dest_cpu
);
4926 activate_task(rq_dest
, p
, 0);
4927 check_preempt_curr(rq_dest
, p
);
4931 double_rq_unlock(rq_src
, rq_dest
);
4936 * migration_thread - this is a highprio system thread that performs
4937 * thread migration by bumping thread off CPU then 'pushing' onto
4940 static int migration_thread(void *data
)
4942 int cpu
= (long)data
;
4946 BUG_ON(rq
->migration_thread
!= current
);
4948 set_current_state(TASK_INTERRUPTIBLE
);
4949 while (!kthread_should_stop()) {
4950 struct migration_req
*req
;
4951 struct list_head
*head
;
4953 spin_lock_irq(&rq
->lock
);
4955 if (cpu_is_offline(cpu
)) {
4956 spin_unlock_irq(&rq
->lock
);
4960 if (rq
->active_balance
) {
4961 active_load_balance(rq
, cpu
);
4962 rq
->active_balance
= 0;
4965 head
= &rq
->migration_queue
;
4967 if (list_empty(head
)) {
4968 spin_unlock_irq(&rq
->lock
);
4970 set_current_state(TASK_INTERRUPTIBLE
);
4973 req
= list_entry(head
->next
, struct migration_req
, list
);
4974 list_del_init(head
->next
);
4976 spin_unlock(&rq
->lock
);
4977 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4980 complete(&req
->done
);
4982 __set_current_state(TASK_RUNNING
);
4986 /* Wait for kthread_stop */
4987 set_current_state(TASK_INTERRUPTIBLE
);
4988 while (!kthread_should_stop()) {
4990 set_current_state(TASK_INTERRUPTIBLE
);
4992 __set_current_state(TASK_RUNNING
);
4996 #ifdef CONFIG_HOTPLUG_CPU
4998 * Figure out where task on dead CPU should go, use force if neccessary.
4999 * NOTE: interrupts should be disabled by the caller
5001 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5003 unsigned long flags
;
5010 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5011 cpus_and(mask
, mask
, p
->cpus_allowed
);
5012 dest_cpu
= any_online_cpu(mask
);
5014 /* On any allowed CPU? */
5015 if (dest_cpu
== NR_CPUS
)
5016 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5018 /* No more Mr. Nice Guy. */
5019 if (dest_cpu
== NR_CPUS
) {
5020 rq
= task_rq_lock(p
, &flags
);
5021 cpus_setall(p
->cpus_allowed
);
5022 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5023 task_rq_unlock(rq
, &flags
);
5026 * Don't tell them about moving exiting tasks or
5027 * kernel threads (both mm NULL), since they never
5030 if (p
->mm
&& printk_ratelimit())
5031 printk(KERN_INFO
"process %d (%s) no "
5032 "longer affine to cpu%d\n",
5033 p
->pid
, p
->comm
, dead_cpu
);
5035 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5040 * While a dead CPU has no uninterruptible tasks queued at this point,
5041 * it might still have a nonzero ->nr_uninterruptible counter, because
5042 * for performance reasons the counter is not stricly tracking tasks to
5043 * their home CPUs. So we just add the counter to another CPU's counter,
5044 * to keep the global sum constant after CPU-down:
5046 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5048 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5049 unsigned long flags
;
5051 local_irq_save(flags
);
5052 double_rq_lock(rq_src
, rq_dest
);
5053 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5054 rq_src
->nr_uninterruptible
= 0;
5055 double_rq_unlock(rq_src
, rq_dest
);
5056 local_irq_restore(flags
);
5059 /* Run through task list and migrate tasks from the dead cpu. */
5060 static void migrate_live_tasks(int src_cpu
)
5062 struct task_struct
*p
, *t
;
5064 write_lock_irq(&tasklist_lock
);
5066 do_each_thread(t
, p
) {
5070 if (task_cpu(p
) == src_cpu
)
5071 move_task_off_dead_cpu(src_cpu
, p
);
5072 } while_each_thread(t
, p
);
5074 write_unlock_irq(&tasklist_lock
);
5078 * Schedules idle task to be the next runnable task on current CPU.
5079 * It does so by boosting its priority to highest possible and adding it to
5080 * the _front_ of the runqueue. Used by CPU offline code.
5082 void sched_idle_next(void)
5084 int this_cpu
= smp_processor_id();
5085 struct rq
*rq
= cpu_rq(this_cpu
);
5086 struct task_struct
*p
= rq
->idle
;
5087 unsigned long flags
;
5089 /* cpu has to be offline */
5090 BUG_ON(cpu_online(this_cpu
));
5093 * Strictly not necessary since rest of the CPUs are stopped by now
5094 * and interrupts disabled on the current cpu.
5096 spin_lock_irqsave(&rq
->lock
, flags
);
5098 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5100 /* Add idle task to the _front_ of its priority queue: */
5101 activate_idle_task(p
, rq
);
5103 spin_unlock_irqrestore(&rq
->lock
, flags
);
5107 * Ensures that the idle task is using init_mm right before its cpu goes
5110 void idle_task_exit(void)
5112 struct mm_struct
*mm
= current
->active_mm
;
5114 BUG_ON(cpu_online(smp_processor_id()));
5117 switch_mm(mm
, &init_mm
, current
);
5121 /* called under rq->lock with disabled interrupts */
5122 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5124 struct rq
*rq
= cpu_rq(dead_cpu
);
5126 /* Must be exiting, otherwise would be on tasklist. */
5127 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5129 /* Cannot have done final schedule yet: would have vanished. */
5130 BUG_ON(p
->state
== TASK_DEAD
);
5135 * Drop lock around migration; if someone else moves it,
5136 * that's OK. No task can be added to this CPU, so iteration is
5138 * NOTE: interrupts should be left disabled --dev@
5140 spin_unlock(&rq
->lock
);
5141 move_task_off_dead_cpu(dead_cpu
, p
);
5142 spin_lock(&rq
->lock
);
5147 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5148 static void migrate_dead_tasks(unsigned int dead_cpu
)
5150 struct rq
*rq
= cpu_rq(dead_cpu
);
5151 struct task_struct
*next
;
5154 if (!rq
->nr_running
)
5156 update_rq_clock(rq
);
5157 next
= pick_next_task(rq
, rq
->curr
);
5160 migrate_dead(dead_cpu
, next
);
5164 #endif /* CONFIG_HOTPLUG_CPU */
5166 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5168 static struct ctl_table sd_ctl_dir
[] = {
5170 .procname
= "sched_domain",
5176 static struct ctl_table sd_ctl_root
[] = {
5178 .ctl_name
= CTL_KERN
,
5179 .procname
= "kernel",
5181 .child
= sd_ctl_dir
,
5186 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5188 struct ctl_table
*entry
=
5189 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5192 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5198 set_table_entry(struct ctl_table
*entry
,
5199 const char *procname
, void *data
, int maxlen
,
5200 mode_t mode
, proc_handler
*proc_handler
)
5202 entry
->procname
= procname
;
5204 entry
->maxlen
= maxlen
;
5206 entry
->proc_handler
= proc_handler
;
5209 static struct ctl_table
*
5210 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5212 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5214 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5215 sizeof(long), 0644, proc_doulongvec_minmax
);
5216 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5217 sizeof(long), 0644, proc_doulongvec_minmax
);
5218 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5219 sizeof(int), 0644, proc_dointvec_minmax
);
5220 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5221 sizeof(int), 0644, proc_dointvec_minmax
);
5222 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5223 sizeof(int), 0644, proc_dointvec_minmax
);
5224 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5225 sizeof(int), 0644, proc_dointvec_minmax
);
5226 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5227 sizeof(int), 0644, proc_dointvec_minmax
);
5228 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5229 sizeof(int), 0644, proc_dointvec_minmax
);
5230 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5231 sizeof(int), 0644, proc_dointvec_minmax
);
5232 set_table_entry(&table
[10], "cache_nice_tries",
5233 &sd
->cache_nice_tries
,
5234 sizeof(int), 0644, proc_dointvec_minmax
);
5235 set_table_entry(&table
[12], "flags", &sd
->flags
,
5236 sizeof(int), 0644, proc_dointvec_minmax
);
5241 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5243 struct ctl_table
*entry
, *table
;
5244 struct sched_domain
*sd
;
5245 int domain_num
= 0, i
;
5248 for_each_domain(cpu
, sd
)
5250 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5253 for_each_domain(cpu
, sd
) {
5254 snprintf(buf
, 32, "domain%d", i
);
5255 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5257 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5264 static struct ctl_table_header
*sd_sysctl_header
;
5265 static void init_sched_domain_sysctl(void)
5267 int i
, cpu_num
= num_online_cpus();
5268 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5271 sd_ctl_dir
[0].child
= entry
;
5273 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5274 snprintf(buf
, 32, "cpu%d", i
);
5275 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5277 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5279 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5282 static void init_sched_domain_sysctl(void)
5288 * migration_call - callback that gets triggered when a CPU is added.
5289 * Here we can start up the necessary migration thread for the new CPU.
5291 static int __cpuinit
5292 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5294 struct task_struct
*p
;
5295 int cpu
= (long)hcpu
;
5296 unsigned long flags
;
5300 case CPU_LOCK_ACQUIRE
:
5301 mutex_lock(&sched_hotcpu_mutex
);
5304 case CPU_UP_PREPARE
:
5305 case CPU_UP_PREPARE_FROZEN
:
5306 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5309 kthread_bind(p
, cpu
);
5310 /* Must be high prio: stop_machine expects to yield to it. */
5311 rq
= task_rq_lock(p
, &flags
);
5312 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5313 task_rq_unlock(rq
, &flags
);
5314 cpu_rq(cpu
)->migration_thread
= p
;
5318 case CPU_ONLINE_FROZEN
:
5319 /* Strictly unneccessary, as first user will wake it. */
5320 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5323 #ifdef CONFIG_HOTPLUG_CPU
5324 case CPU_UP_CANCELED
:
5325 case CPU_UP_CANCELED_FROZEN
:
5326 if (!cpu_rq(cpu
)->migration_thread
)
5328 /* Unbind it from offline cpu so it can run. Fall thru. */
5329 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5330 any_online_cpu(cpu_online_map
));
5331 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5332 cpu_rq(cpu
)->migration_thread
= NULL
;
5336 case CPU_DEAD_FROZEN
:
5337 migrate_live_tasks(cpu
);
5339 kthread_stop(rq
->migration_thread
);
5340 rq
->migration_thread
= NULL
;
5341 /* Idle task back to normal (off runqueue, low prio) */
5342 rq
= task_rq_lock(rq
->idle
, &flags
);
5343 update_rq_clock(rq
);
5344 deactivate_task(rq
, rq
->idle
, 0);
5345 rq
->idle
->static_prio
= MAX_PRIO
;
5346 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5347 rq
->idle
->sched_class
= &idle_sched_class
;
5348 migrate_dead_tasks(cpu
);
5349 task_rq_unlock(rq
, &flags
);
5350 migrate_nr_uninterruptible(rq
);
5351 BUG_ON(rq
->nr_running
!= 0);
5353 /* No need to migrate the tasks: it was best-effort if
5354 * they didn't take sched_hotcpu_mutex. Just wake up
5355 * the requestors. */
5356 spin_lock_irq(&rq
->lock
);
5357 while (!list_empty(&rq
->migration_queue
)) {
5358 struct migration_req
*req
;
5360 req
= list_entry(rq
->migration_queue
.next
,
5361 struct migration_req
, list
);
5362 list_del_init(&req
->list
);
5363 complete(&req
->done
);
5365 spin_unlock_irq(&rq
->lock
);
5368 case CPU_LOCK_RELEASE
:
5369 mutex_unlock(&sched_hotcpu_mutex
);
5375 /* Register at highest priority so that task migration (migrate_all_tasks)
5376 * happens before everything else.
5378 static struct notifier_block __cpuinitdata migration_notifier
= {
5379 .notifier_call
= migration_call
,
5383 int __init
migration_init(void)
5385 void *cpu
= (void *)(long)smp_processor_id();
5388 /* Start one for the boot CPU: */
5389 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5390 BUG_ON(err
== NOTIFY_BAD
);
5391 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5392 register_cpu_notifier(&migration_notifier
);
5400 /* Number of possible processor ids */
5401 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5402 EXPORT_SYMBOL(nr_cpu_ids
);
5404 #undef SCHED_DOMAIN_DEBUG
5405 #ifdef SCHED_DOMAIN_DEBUG
5406 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5411 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5415 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5420 struct sched_group
*group
= sd
->groups
;
5421 cpumask_t groupmask
;
5423 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5424 cpus_clear(groupmask
);
5427 for (i
= 0; i
< level
+ 1; i
++)
5429 printk("domain %d: ", level
);
5431 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5432 printk("does not load-balance\n");
5434 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5439 printk("span %s\n", str
);
5441 if (!cpu_isset(cpu
, sd
->span
))
5442 printk(KERN_ERR
"ERROR: domain->span does not contain "
5444 if (!cpu_isset(cpu
, group
->cpumask
))
5445 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5449 for (i
= 0; i
< level
+ 2; i
++)
5455 printk(KERN_ERR
"ERROR: group is NULL\n");
5459 if (!group
->__cpu_power
) {
5461 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5465 if (!cpus_weight(group
->cpumask
)) {
5467 printk(KERN_ERR
"ERROR: empty group\n");
5470 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5472 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5475 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5477 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5480 group
= group
->next
;
5481 } while (group
!= sd
->groups
);
5484 if (!cpus_equal(sd
->span
, groupmask
))
5485 printk(KERN_ERR
"ERROR: groups don't span "
5493 if (!cpus_subset(groupmask
, sd
->span
))
5494 printk(KERN_ERR
"ERROR: parent span is not a superset "
5495 "of domain->span\n");
5500 # define sched_domain_debug(sd, cpu) do { } while (0)
5503 static int sd_degenerate(struct sched_domain
*sd
)
5505 if (cpus_weight(sd
->span
) == 1)
5508 /* Following flags need at least 2 groups */
5509 if (sd
->flags
& (SD_LOAD_BALANCE
|
5510 SD_BALANCE_NEWIDLE
|
5514 SD_SHARE_PKG_RESOURCES
)) {
5515 if (sd
->groups
!= sd
->groups
->next
)
5519 /* Following flags don't use groups */
5520 if (sd
->flags
& (SD_WAKE_IDLE
|
5529 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5531 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5533 if (sd_degenerate(parent
))
5536 if (!cpus_equal(sd
->span
, parent
->span
))
5539 /* Does parent contain flags not in child? */
5540 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5541 if (cflags
& SD_WAKE_AFFINE
)
5542 pflags
&= ~SD_WAKE_BALANCE
;
5543 /* Flags needing groups don't count if only 1 group in parent */
5544 if (parent
->groups
== parent
->groups
->next
) {
5545 pflags
&= ~(SD_LOAD_BALANCE
|
5546 SD_BALANCE_NEWIDLE
|
5550 SD_SHARE_PKG_RESOURCES
);
5552 if (~cflags
& pflags
)
5559 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5560 * hold the hotplug lock.
5562 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5564 struct rq
*rq
= cpu_rq(cpu
);
5565 struct sched_domain
*tmp
;
5567 /* Remove the sched domains which do not contribute to scheduling. */
5568 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5569 struct sched_domain
*parent
= tmp
->parent
;
5572 if (sd_parent_degenerate(tmp
, parent
)) {
5573 tmp
->parent
= parent
->parent
;
5575 parent
->parent
->child
= tmp
;
5579 if (sd
&& sd_degenerate(sd
)) {
5585 sched_domain_debug(sd
, cpu
);
5587 rcu_assign_pointer(rq
->sd
, sd
);
5590 /* cpus with isolated domains */
5591 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5593 /* Setup the mask of cpus configured for isolated domains */
5594 static int __init
isolated_cpu_setup(char *str
)
5596 int ints
[NR_CPUS
], i
;
5598 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5599 cpus_clear(cpu_isolated_map
);
5600 for (i
= 1; i
<= ints
[0]; i
++)
5601 if (ints
[i
] < NR_CPUS
)
5602 cpu_set(ints
[i
], cpu_isolated_map
);
5606 __setup ("isolcpus=", isolated_cpu_setup
);
5609 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5610 * to a function which identifies what group(along with sched group) a CPU
5611 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5612 * (due to the fact that we keep track of groups covered with a cpumask_t).
5614 * init_sched_build_groups will build a circular linked list of the groups
5615 * covered by the given span, and will set each group's ->cpumask correctly,
5616 * and ->cpu_power to 0.
5619 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5620 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5621 struct sched_group
**sg
))
5623 struct sched_group
*first
= NULL
, *last
= NULL
;
5624 cpumask_t covered
= CPU_MASK_NONE
;
5627 for_each_cpu_mask(i
, span
) {
5628 struct sched_group
*sg
;
5629 int group
= group_fn(i
, cpu_map
, &sg
);
5632 if (cpu_isset(i
, covered
))
5635 sg
->cpumask
= CPU_MASK_NONE
;
5636 sg
->__cpu_power
= 0;
5638 for_each_cpu_mask(j
, span
) {
5639 if (group_fn(j
, cpu_map
, NULL
) != group
)
5642 cpu_set(j
, covered
);
5643 cpu_set(j
, sg
->cpumask
);
5654 #define SD_NODES_PER_DOMAIN 16
5659 * find_next_best_node - find the next node to include in a sched_domain
5660 * @node: node whose sched_domain we're building
5661 * @used_nodes: nodes already in the sched_domain
5663 * Find the next node to include in a given scheduling domain. Simply
5664 * finds the closest node not already in the @used_nodes map.
5666 * Should use nodemask_t.
5668 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5670 int i
, n
, val
, min_val
, best_node
= 0;
5674 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5675 /* Start at @node */
5676 n
= (node
+ i
) % MAX_NUMNODES
;
5678 if (!nr_cpus_node(n
))
5681 /* Skip already used nodes */
5682 if (test_bit(n
, used_nodes
))
5685 /* Simple min distance search */
5686 val
= node_distance(node
, n
);
5688 if (val
< min_val
) {
5694 set_bit(best_node
, used_nodes
);
5699 * sched_domain_node_span - get a cpumask for a node's sched_domain
5700 * @node: node whose cpumask we're constructing
5701 * @size: number of nodes to include in this span
5703 * Given a node, construct a good cpumask for its sched_domain to span. It
5704 * should be one that prevents unnecessary balancing, but also spreads tasks
5707 static cpumask_t
sched_domain_node_span(int node
)
5709 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5710 cpumask_t span
, nodemask
;
5714 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5716 nodemask
= node_to_cpumask(node
);
5717 cpus_or(span
, span
, nodemask
);
5718 set_bit(node
, used_nodes
);
5720 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5721 int next_node
= find_next_best_node(node
, used_nodes
);
5723 nodemask
= node_to_cpumask(next_node
);
5724 cpus_or(span
, span
, nodemask
);
5731 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5734 * SMT sched-domains:
5736 #ifdef CONFIG_SCHED_SMT
5737 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5738 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5740 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5741 struct sched_group
**sg
)
5744 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5750 * multi-core sched-domains:
5752 #ifdef CONFIG_SCHED_MC
5753 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5754 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5757 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5758 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5759 struct sched_group
**sg
)
5762 cpumask_t mask
= cpu_sibling_map
[cpu
];
5763 cpus_and(mask
, mask
, *cpu_map
);
5764 group
= first_cpu(mask
);
5766 *sg
= &per_cpu(sched_group_core
, group
);
5769 #elif defined(CONFIG_SCHED_MC)
5770 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5771 struct sched_group
**sg
)
5774 *sg
= &per_cpu(sched_group_core
, cpu
);
5779 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5780 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5782 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5783 struct sched_group
**sg
)
5786 #ifdef CONFIG_SCHED_MC
5787 cpumask_t mask
= cpu_coregroup_map(cpu
);
5788 cpus_and(mask
, mask
, *cpu_map
);
5789 group
= first_cpu(mask
);
5790 #elif defined(CONFIG_SCHED_SMT)
5791 cpumask_t mask
= cpu_sibling_map
[cpu
];
5792 cpus_and(mask
, mask
, *cpu_map
);
5793 group
= first_cpu(mask
);
5798 *sg
= &per_cpu(sched_group_phys
, group
);
5804 * The init_sched_build_groups can't handle what we want to do with node
5805 * groups, so roll our own. Now each node has its own list of groups which
5806 * gets dynamically allocated.
5808 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5809 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5811 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5812 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5814 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5815 struct sched_group
**sg
)
5817 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5820 cpus_and(nodemask
, nodemask
, *cpu_map
);
5821 group
= first_cpu(nodemask
);
5824 *sg
= &per_cpu(sched_group_allnodes
, group
);
5828 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5830 struct sched_group
*sg
= group_head
;
5836 for_each_cpu_mask(j
, sg
->cpumask
) {
5837 struct sched_domain
*sd
;
5839 sd
= &per_cpu(phys_domains
, j
);
5840 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5842 * Only add "power" once for each
5848 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5851 if (sg
!= group_head
)
5857 /* Free memory allocated for various sched_group structures */
5858 static void free_sched_groups(const cpumask_t
*cpu_map
)
5862 for_each_cpu_mask(cpu
, *cpu_map
) {
5863 struct sched_group
**sched_group_nodes
5864 = sched_group_nodes_bycpu
[cpu
];
5866 if (!sched_group_nodes
)
5869 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5870 cpumask_t nodemask
= node_to_cpumask(i
);
5871 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5873 cpus_and(nodemask
, nodemask
, *cpu_map
);
5874 if (cpus_empty(nodemask
))
5884 if (oldsg
!= sched_group_nodes
[i
])
5887 kfree(sched_group_nodes
);
5888 sched_group_nodes_bycpu
[cpu
] = NULL
;
5892 static void free_sched_groups(const cpumask_t
*cpu_map
)
5898 * Initialize sched groups cpu_power.
5900 * cpu_power indicates the capacity of sched group, which is used while
5901 * distributing the load between different sched groups in a sched domain.
5902 * Typically cpu_power for all the groups in a sched domain will be same unless
5903 * there are asymmetries in the topology. If there are asymmetries, group
5904 * having more cpu_power will pickup more load compared to the group having
5907 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5908 * the maximum number of tasks a group can handle in the presence of other idle
5909 * or lightly loaded groups in the same sched domain.
5911 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5913 struct sched_domain
*child
;
5914 struct sched_group
*group
;
5916 WARN_ON(!sd
|| !sd
->groups
);
5918 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5923 sd
->groups
->__cpu_power
= 0;
5926 * For perf policy, if the groups in child domain share resources
5927 * (for example cores sharing some portions of the cache hierarchy
5928 * or SMT), then set this domain groups cpu_power such that each group
5929 * can handle only one task, when there are other idle groups in the
5930 * same sched domain.
5932 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5934 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5935 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
5940 * add cpu_power of each child group to this groups cpu_power
5942 group
= child
->groups
;
5944 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
5945 group
= group
->next
;
5946 } while (group
!= child
->groups
);
5950 * Build sched domains for a given set of cpus and attach the sched domains
5951 * to the individual cpus
5953 static int build_sched_domains(const cpumask_t
*cpu_map
)
5957 struct sched_group
**sched_group_nodes
= NULL
;
5958 int sd_allnodes
= 0;
5961 * Allocate the per-node list of sched groups
5963 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5965 if (!sched_group_nodes
) {
5966 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5969 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5973 * Set up domains for cpus specified by the cpu_map.
5975 for_each_cpu_mask(i
, *cpu_map
) {
5976 struct sched_domain
*sd
= NULL
, *p
;
5977 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5979 cpus_and(nodemask
, nodemask
, *cpu_map
);
5982 if (cpus_weight(*cpu_map
) >
5983 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5984 sd
= &per_cpu(allnodes_domains
, i
);
5985 *sd
= SD_ALLNODES_INIT
;
5986 sd
->span
= *cpu_map
;
5987 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
5993 sd
= &per_cpu(node_domains
, i
);
5995 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
5999 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6003 sd
= &per_cpu(phys_domains
, i
);
6005 sd
->span
= nodemask
;
6009 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6011 #ifdef CONFIG_SCHED_MC
6013 sd
= &per_cpu(core_domains
, i
);
6015 sd
->span
= cpu_coregroup_map(i
);
6016 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6019 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6022 #ifdef CONFIG_SCHED_SMT
6024 sd
= &per_cpu(cpu_domains
, i
);
6025 *sd
= SD_SIBLING_INIT
;
6026 sd
->span
= cpu_sibling_map
[i
];
6027 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6030 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6034 #ifdef CONFIG_SCHED_SMT
6035 /* Set up CPU (sibling) groups */
6036 for_each_cpu_mask(i
, *cpu_map
) {
6037 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6038 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6039 if (i
!= first_cpu(this_sibling_map
))
6042 init_sched_build_groups(this_sibling_map
, cpu_map
,
6047 #ifdef CONFIG_SCHED_MC
6048 /* Set up multi-core groups */
6049 for_each_cpu_mask(i
, *cpu_map
) {
6050 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6051 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6052 if (i
!= first_cpu(this_core_map
))
6054 init_sched_build_groups(this_core_map
, cpu_map
,
6055 &cpu_to_core_group
);
6059 /* Set up physical groups */
6060 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6061 cpumask_t nodemask
= node_to_cpumask(i
);
6063 cpus_and(nodemask
, nodemask
, *cpu_map
);
6064 if (cpus_empty(nodemask
))
6067 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6071 /* Set up node groups */
6073 init_sched_build_groups(*cpu_map
, cpu_map
,
6074 &cpu_to_allnodes_group
);
6076 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6077 /* Set up node groups */
6078 struct sched_group
*sg
, *prev
;
6079 cpumask_t nodemask
= node_to_cpumask(i
);
6080 cpumask_t domainspan
;
6081 cpumask_t covered
= CPU_MASK_NONE
;
6084 cpus_and(nodemask
, nodemask
, *cpu_map
);
6085 if (cpus_empty(nodemask
)) {
6086 sched_group_nodes
[i
] = NULL
;
6090 domainspan
= sched_domain_node_span(i
);
6091 cpus_and(domainspan
, domainspan
, *cpu_map
);
6093 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6095 printk(KERN_WARNING
"Can not alloc domain group for "
6099 sched_group_nodes
[i
] = sg
;
6100 for_each_cpu_mask(j
, nodemask
) {
6101 struct sched_domain
*sd
;
6103 sd
= &per_cpu(node_domains
, j
);
6106 sg
->__cpu_power
= 0;
6107 sg
->cpumask
= nodemask
;
6109 cpus_or(covered
, covered
, nodemask
);
6112 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6113 cpumask_t tmp
, notcovered
;
6114 int n
= (i
+ j
) % MAX_NUMNODES
;
6116 cpus_complement(notcovered
, covered
);
6117 cpus_and(tmp
, notcovered
, *cpu_map
);
6118 cpus_and(tmp
, tmp
, domainspan
);
6119 if (cpus_empty(tmp
))
6122 nodemask
= node_to_cpumask(n
);
6123 cpus_and(tmp
, tmp
, nodemask
);
6124 if (cpus_empty(tmp
))
6127 sg
= kmalloc_node(sizeof(struct sched_group
),
6131 "Can not alloc domain group for node %d\n", j
);
6134 sg
->__cpu_power
= 0;
6136 sg
->next
= prev
->next
;
6137 cpus_or(covered
, covered
, tmp
);
6144 /* Calculate CPU power for physical packages and nodes */
6145 #ifdef CONFIG_SCHED_SMT
6146 for_each_cpu_mask(i
, *cpu_map
) {
6147 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6149 init_sched_groups_power(i
, sd
);
6152 #ifdef CONFIG_SCHED_MC
6153 for_each_cpu_mask(i
, *cpu_map
) {
6154 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6156 init_sched_groups_power(i
, sd
);
6160 for_each_cpu_mask(i
, *cpu_map
) {
6161 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6163 init_sched_groups_power(i
, sd
);
6167 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6168 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6171 struct sched_group
*sg
;
6173 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6174 init_numa_sched_groups_power(sg
);
6178 /* Attach the domains */
6179 for_each_cpu_mask(i
, *cpu_map
) {
6180 struct sched_domain
*sd
;
6181 #ifdef CONFIG_SCHED_SMT
6182 sd
= &per_cpu(cpu_domains
, i
);
6183 #elif defined(CONFIG_SCHED_MC)
6184 sd
= &per_cpu(core_domains
, i
);
6186 sd
= &per_cpu(phys_domains
, i
);
6188 cpu_attach_domain(sd
, i
);
6195 free_sched_groups(cpu_map
);
6200 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6202 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6204 cpumask_t cpu_default_map
;
6208 * Setup mask for cpus without special case scheduling requirements.
6209 * For now this just excludes isolated cpus, but could be used to
6210 * exclude other special cases in the future.
6212 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6214 err
= build_sched_domains(&cpu_default_map
);
6219 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6221 free_sched_groups(cpu_map
);
6225 * Detach sched domains from a group of cpus specified in cpu_map
6226 * These cpus will now be attached to the NULL domain
6228 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6232 for_each_cpu_mask(i
, *cpu_map
)
6233 cpu_attach_domain(NULL
, i
);
6234 synchronize_sched();
6235 arch_destroy_sched_domains(cpu_map
);
6239 * Partition sched domains as specified by the cpumasks below.
6240 * This attaches all cpus from the cpumasks to the NULL domain,
6241 * waits for a RCU quiescent period, recalculates sched
6242 * domain information and then attaches them back to the
6243 * correct sched domains
6244 * Call with hotplug lock held
6246 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6248 cpumask_t change_map
;
6251 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6252 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6253 cpus_or(change_map
, *partition1
, *partition2
);
6255 /* Detach sched domains from all of the affected cpus */
6256 detach_destroy_domains(&change_map
);
6257 if (!cpus_empty(*partition1
))
6258 err
= build_sched_domains(partition1
);
6259 if (!err
&& !cpus_empty(*partition2
))
6260 err
= build_sched_domains(partition2
);
6265 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6266 static int arch_reinit_sched_domains(void)
6270 mutex_lock(&sched_hotcpu_mutex
);
6271 detach_destroy_domains(&cpu_online_map
);
6272 err
= arch_init_sched_domains(&cpu_online_map
);
6273 mutex_unlock(&sched_hotcpu_mutex
);
6278 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6282 if (buf
[0] != '0' && buf
[0] != '1')
6286 sched_smt_power_savings
= (buf
[0] == '1');
6288 sched_mc_power_savings
= (buf
[0] == '1');
6290 ret
= arch_reinit_sched_domains();
6292 return ret
? ret
: count
;
6295 #ifdef CONFIG_SCHED_MC
6296 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6298 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6300 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6301 const char *buf
, size_t count
)
6303 return sched_power_savings_store(buf
, count
, 0);
6305 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6306 sched_mc_power_savings_store
);
6309 #ifdef CONFIG_SCHED_SMT
6310 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6312 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6314 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6315 const char *buf
, size_t count
)
6317 return sched_power_savings_store(buf
, count
, 1);
6319 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6320 sched_smt_power_savings_store
);
6323 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6327 #ifdef CONFIG_SCHED_SMT
6329 err
= sysfs_create_file(&cls
->kset
.kobj
,
6330 &attr_sched_smt_power_savings
.attr
);
6332 #ifdef CONFIG_SCHED_MC
6333 if (!err
&& mc_capable())
6334 err
= sysfs_create_file(&cls
->kset
.kobj
,
6335 &attr_sched_mc_power_savings
.attr
);
6342 * Force a reinitialization of the sched domains hierarchy. The domains
6343 * and groups cannot be updated in place without racing with the balancing
6344 * code, so we temporarily attach all running cpus to the NULL domain
6345 * which will prevent rebalancing while the sched domains are recalculated.
6347 static int update_sched_domains(struct notifier_block
*nfb
,
6348 unsigned long action
, void *hcpu
)
6351 case CPU_UP_PREPARE
:
6352 case CPU_UP_PREPARE_FROZEN
:
6353 case CPU_DOWN_PREPARE
:
6354 case CPU_DOWN_PREPARE_FROZEN
:
6355 detach_destroy_domains(&cpu_online_map
);
6358 case CPU_UP_CANCELED
:
6359 case CPU_UP_CANCELED_FROZEN
:
6360 case CPU_DOWN_FAILED
:
6361 case CPU_DOWN_FAILED_FROZEN
:
6363 case CPU_ONLINE_FROZEN
:
6365 case CPU_DEAD_FROZEN
:
6367 * Fall through and re-initialise the domains.
6374 /* The hotplug lock is already held by cpu_up/cpu_down */
6375 arch_init_sched_domains(&cpu_online_map
);
6380 void __init
sched_init_smp(void)
6382 cpumask_t non_isolated_cpus
;
6384 mutex_lock(&sched_hotcpu_mutex
);
6385 arch_init_sched_domains(&cpu_online_map
);
6386 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6387 if (cpus_empty(non_isolated_cpus
))
6388 cpu_set(smp_processor_id(), non_isolated_cpus
);
6389 mutex_unlock(&sched_hotcpu_mutex
);
6390 /* XXX: Theoretical race here - CPU may be hotplugged now */
6391 hotcpu_notifier(update_sched_domains
, 0);
6393 init_sched_domain_sysctl();
6395 /* Move init over to a non-isolated CPU */
6396 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6400 void __init
sched_init_smp(void)
6403 #endif /* CONFIG_SMP */
6405 int in_sched_functions(unsigned long addr
)
6407 /* Linker adds these: start and end of __sched functions */
6408 extern char __sched_text_start
[], __sched_text_end
[];
6410 return in_lock_functions(addr
) ||
6411 (addr
>= (unsigned long)__sched_text_start
6412 && addr
< (unsigned long)__sched_text_end
);
6415 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6417 cfs_rq
->tasks_timeline
= RB_ROOT
;
6418 #ifdef CONFIG_FAIR_GROUP_SCHED
6423 void __init
sched_init(void)
6425 int highest_cpu
= 0;
6429 * Link up the scheduling class hierarchy:
6431 rt_sched_class
.next
= &fair_sched_class
;
6432 fair_sched_class
.next
= &idle_sched_class
;
6433 idle_sched_class
.next
= NULL
;
6435 for_each_possible_cpu(i
) {
6436 struct rt_prio_array
*array
;
6440 spin_lock_init(&rq
->lock
);
6441 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6444 init_cfs_rq(&rq
->cfs
, rq
);
6445 #ifdef CONFIG_FAIR_GROUP_SCHED
6446 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6447 list_add(&rq
->cfs
.leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6450 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6451 rq
->cpu_load
[j
] = 0;
6454 rq
->active_balance
= 0;
6455 rq
->next_balance
= jiffies
;
6458 rq
->migration_thread
= NULL
;
6459 INIT_LIST_HEAD(&rq
->migration_queue
);
6461 atomic_set(&rq
->nr_iowait
, 0);
6463 array
= &rq
->rt
.active
;
6464 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6465 INIT_LIST_HEAD(array
->queue
+ j
);
6466 __clear_bit(j
, array
->bitmap
);
6469 /* delimiter for bitsearch: */
6470 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6473 set_load_weight(&init_task
);
6475 #ifdef CONFIG_PREEMPT_NOTIFIERS
6476 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6480 nr_cpu_ids
= highest_cpu
+ 1;
6481 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6484 #ifdef CONFIG_RT_MUTEXES
6485 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6489 * The boot idle thread does lazy MMU switching as well:
6491 atomic_inc(&init_mm
.mm_count
);
6492 enter_lazy_tlb(&init_mm
, current
);
6495 * Make us the idle thread. Technically, schedule() should not be
6496 * called from this thread, however somewhere below it might be,
6497 * but because we are the idle thread, we just pick up running again
6498 * when this runqueue becomes "idle".
6500 init_idle(current
, smp_processor_id());
6502 * During early bootup we pretend to be a normal task:
6504 current
->sched_class
= &fair_sched_class
;
6507 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6508 void __might_sleep(char *file
, int line
)
6511 static unsigned long prev_jiffy
; /* ratelimiting */
6513 if ((in_atomic() || irqs_disabled()) &&
6514 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6515 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6517 prev_jiffy
= jiffies
;
6518 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6519 " context at %s:%d\n", file
, line
);
6520 printk("in_atomic():%d, irqs_disabled():%d\n",
6521 in_atomic(), irqs_disabled());
6522 debug_show_held_locks(current
);
6523 if (irqs_disabled())
6524 print_irqtrace_events(current
);
6529 EXPORT_SYMBOL(__might_sleep
);
6532 #ifdef CONFIG_MAGIC_SYSRQ
6533 void normalize_rt_tasks(void)
6535 struct task_struct
*g
, *p
;
6536 unsigned long flags
;
6540 read_lock_irq(&tasklist_lock
);
6541 do_each_thread(g
, p
) {
6543 p
->se
.exec_start
= 0;
6544 #ifdef CONFIG_SCHEDSTATS
6545 p
->se
.wait_start
= 0;
6546 p
->se
.sleep_start
= 0;
6547 p
->se
.block_start
= 0;
6549 task_rq(p
)->clock
= 0;
6553 * Renice negative nice level userspace
6556 if (TASK_NICE(p
) < 0 && p
->mm
)
6557 set_user_nice(p
, 0);
6561 spin_lock_irqsave(&p
->pi_lock
, flags
);
6562 rq
= __task_rq_lock(p
);
6565 * Do not touch the migration thread:
6567 if (p
== rq
->migration_thread
)
6571 update_rq_clock(rq
);
6572 on_rq
= p
->se
.on_rq
;
6574 deactivate_task(rq
, p
, 0);
6575 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6577 activate_task(rq
, p
, 0);
6578 resched_task(rq
->curr
);
6583 __task_rq_unlock(rq
);
6584 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6585 } while_each_thread(g
, p
);
6587 read_unlock_irq(&tasklist_lock
);
6590 #endif /* CONFIG_MAGIC_SYSRQ */
6594 * These functions are only useful for the IA64 MCA handling.
6596 * They can only be called when the whole system has been
6597 * stopped - every CPU needs to be quiescent, and no scheduling
6598 * activity can take place. Using them for anything else would
6599 * be a serious bug, and as a result, they aren't even visible
6600 * under any other configuration.
6604 * curr_task - return the current task for a given cpu.
6605 * @cpu: the processor in question.
6607 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6609 struct task_struct
*curr_task(int cpu
)
6611 return cpu_curr(cpu
);
6615 * set_curr_task - set the current task for a given cpu.
6616 * @cpu: the processor in question.
6617 * @p: the task pointer to set.
6619 * Description: This function must only be used when non-maskable interrupts
6620 * are serviced on a separate stack. It allows the architecture to switch the
6621 * notion of the current task on a cpu in a non-blocking manner. This function
6622 * must be called with all CPU's synchronized, and interrupts disabled, the
6623 * and caller must save the original value of the current task (see
6624 * curr_task() above) and restore that value before reenabling interrupts and
6625 * re-starting the system.
6627 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6629 void set_curr_task(int cpu
, struct task_struct
*p
)