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
];
175 struct load_weight load
;
178 /* CFS-related fields in a runqueue */
180 struct load_weight load
;
181 unsigned long nr_running
;
188 unsigned long wait_runtime_overruns
, wait_runtime_underruns
;
190 struct rb_root tasks_timeline
;
191 struct rb_node
*rb_leftmost
;
192 struct rb_node
*rb_load_balance_curr
;
193 /* 'curr' points to currently running entity on this cfs_rq.
194 * It is set to NULL otherwise (i.e when none are currently running).
196 struct sched_entity
*curr
;
197 #ifdef CONFIG_FAIR_GROUP_SCHED
198 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
200 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
201 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
202 * (like users, containers etc.)
204 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
205 * list is used during load balance.
207 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
211 /* Real-Time classes' related field in a runqueue: */
213 struct rt_prio_array active
;
214 int rt_load_balance_idx
;
215 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
219 * This is the main, per-CPU runqueue data structure.
221 * Locking rule: those places that want to lock multiple runqueues
222 * (such as the load balancing or the thread migration code), lock
223 * acquire operations must be ordered by ascending &runqueue.
226 spinlock_t lock
; /* runqueue lock */
229 * nr_running and cpu_load should be in the same cacheline because
230 * remote CPUs use both these fields when doing load calculation.
232 unsigned long nr_running
;
233 #define CPU_LOAD_IDX_MAX 5
234 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
235 unsigned char idle_at_tick
;
237 unsigned char in_nohz_recently
;
239 struct load_stat ls
; /* capture load from *all* tasks on this cpu */
240 unsigned long nr_load_updates
;
244 #ifdef CONFIG_FAIR_GROUP_SCHED
245 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
250 * This is part of a global counter where only the total sum
251 * over all CPUs matters. A task can increase this counter on
252 * one CPU and if it got migrated afterwards it may decrease
253 * it on another CPU. Always updated under the runqueue lock:
255 unsigned long nr_uninterruptible
;
257 struct task_struct
*curr
, *idle
;
258 unsigned long next_balance
;
259 struct mm_struct
*prev_mm
;
261 u64 clock
, prev_clock_raw
;
264 unsigned int clock_warps
, clock_overflows
;
266 unsigned int clock_deep_idle_events
;
272 struct sched_domain
*sd
;
274 /* For active balancing */
277 int cpu
; /* cpu of this runqueue */
279 struct task_struct
*migration_thread
;
280 struct list_head migration_queue
;
283 #ifdef CONFIG_SCHEDSTATS
285 struct sched_info rq_sched_info
;
287 /* sys_sched_yield() stats */
288 unsigned long yld_exp_empty
;
289 unsigned long yld_act_empty
;
290 unsigned long yld_both_empty
;
291 unsigned long yld_cnt
;
293 /* schedule() stats */
294 unsigned long sched_switch
;
295 unsigned long sched_cnt
;
296 unsigned long sched_goidle
;
298 /* try_to_wake_up() stats */
299 unsigned long ttwu_cnt
;
300 unsigned long ttwu_local
;
302 struct lock_class_key rq_lock_key
;
305 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
306 static DEFINE_MUTEX(sched_hotcpu_mutex
);
308 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
310 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
313 static inline int cpu_of(struct rq
*rq
)
323 * Update the per-runqueue clock, as finegrained as the platform can give
324 * us, but without assuming monotonicity, etc.:
326 static void __update_rq_clock(struct rq
*rq
)
328 u64 prev_raw
= rq
->prev_clock_raw
;
329 u64 now
= sched_clock();
330 s64 delta
= now
- prev_raw
;
331 u64 clock
= rq
->clock
;
333 #ifdef CONFIG_SCHED_DEBUG
334 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
337 * Protect against sched_clock() occasionally going backwards:
339 if (unlikely(delta
< 0)) {
344 * Catch too large forward jumps too:
346 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
347 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
348 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
351 rq
->clock_overflows
++;
353 if (unlikely(delta
> rq
->clock_max_delta
))
354 rq
->clock_max_delta
= delta
;
359 rq
->prev_clock_raw
= now
;
363 static void update_rq_clock(struct rq
*rq
)
365 if (likely(smp_processor_id() == cpu_of(rq
)))
366 __update_rq_clock(rq
);
370 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
371 * See detach_destroy_domains: synchronize_sched for details.
373 * The domain tree of any CPU may only be accessed from within
374 * preempt-disabled sections.
376 #define for_each_domain(cpu, __sd) \
377 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
379 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
380 #define this_rq() (&__get_cpu_var(runqueues))
381 #define task_rq(p) cpu_rq(task_cpu(p))
382 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
385 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
386 * clock constructed from sched_clock():
388 unsigned long long cpu_clock(int cpu
)
390 unsigned long long now
;
394 local_irq_save(flags
);
398 local_irq_restore(flags
);
403 #ifdef CONFIG_FAIR_GROUP_SCHED
404 /* Change a task's ->cfs_rq if it moves across CPUs */
405 static inline void set_task_cfs_rq(struct task_struct
*p
)
407 p
->se
.cfs_rq
= &task_rq(p
)->cfs
;
410 static inline void set_task_cfs_rq(struct task_struct
*p
)
415 #ifndef prepare_arch_switch
416 # define prepare_arch_switch(next) do { } while (0)
418 #ifndef finish_arch_switch
419 # define finish_arch_switch(prev) do { } while (0)
422 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
423 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
425 return rq
->curr
== p
;
428 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
432 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
434 #ifdef CONFIG_DEBUG_SPINLOCK
435 /* this is a valid case when another task releases the spinlock */
436 rq
->lock
.owner
= current
;
439 * If we are tracking spinlock dependencies then we have to
440 * fix up the runqueue lock - which gets 'carried over' from
443 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
445 spin_unlock_irq(&rq
->lock
);
448 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
449 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
454 return rq
->curr
== p
;
458 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
462 * We can optimise this out completely for !SMP, because the
463 * SMP rebalancing from interrupt is the only thing that cares
468 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
469 spin_unlock_irq(&rq
->lock
);
471 spin_unlock(&rq
->lock
);
475 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
479 * After ->oncpu is cleared, the task can be moved to a different CPU.
480 * We must ensure this doesn't happen until the switch is completely
486 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
490 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
493 * __task_rq_lock - lock the runqueue a given task resides on.
494 * Must be called interrupts disabled.
496 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
503 spin_lock(&rq
->lock
);
504 if (unlikely(rq
!= task_rq(p
))) {
505 spin_unlock(&rq
->lock
);
506 goto repeat_lock_task
;
512 * task_rq_lock - lock the runqueue a given task resides on and disable
513 * interrupts. Note the ordering: we can safely lookup the task_rq without
514 * explicitly disabling preemption.
516 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
522 local_irq_save(*flags
);
524 spin_lock(&rq
->lock
);
525 if (unlikely(rq
!= task_rq(p
))) {
526 spin_unlock_irqrestore(&rq
->lock
, *flags
);
527 goto repeat_lock_task
;
532 static inline void __task_rq_unlock(struct rq
*rq
)
535 spin_unlock(&rq
->lock
);
538 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
541 spin_unlock_irqrestore(&rq
->lock
, *flags
);
545 * this_rq_lock - lock this runqueue and disable interrupts.
547 static inline struct rq
*this_rq_lock(void)
554 spin_lock(&rq
->lock
);
560 * We are going deep-idle (irqs are disabled):
562 void sched_clock_idle_sleep_event(void)
564 struct rq
*rq
= cpu_rq(smp_processor_id());
566 spin_lock(&rq
->lock
);
567 __update_rq_clock(rq
);
568 spin_unlock(&rq
->lock
);
569 rq
->clock_deep_idle_events
++;
571 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
574 * We just idled delta nanoseconds (called with irqs disabled):
576 void sched_clock_idle_wakeup_event(u64 delta_ns
)
578 struct rq
*rq
= cpu_rq(smp_processor_id());
579 u64 now
= sched_clock();
581 rq
->idle_clock
+= delta_ns
;
583 * Override the previous timestamp and ignore all
584 * sched_clock() deltas that occured while we idled,
585 * and use the PM-provided delta_ns to advance the
588 spin_lock(&rq
->lock
);
589 rq
->prev_clock_raw
= now
;
590 rq
->clock
+= delta_ns
;
591 spin_unlock(&rq
->lock
);
593 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
596 * resched_task - mark a task 'to be rescheduled now'.
598 * On UP this means the setting of the need_resched flag, on SMP it
599 * might also involve a cross-CPU call to trigger the scheduler on
604 #ifndef tsk_is_polling
605 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
608 static void resched_task(struct task_struct
*p
)
612 assert_spin_locked(&task_rq(p
)->lock
);
614 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
617 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
620 if (cpu
== smp_processor_id())
623 /* NEED_RESCHED must be visible before we test polling */
625 if (!tsk_is_polling(p
))
626 smp_send_reschedule(cpu
);
629 static void resched_cpu(int cpu
)
631 struct rq
*rq
= cpu_rq(cpu
);
634 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
636 resched_task(cpu_curr(cpu
));
637 spin_unlock_irqrestore(&rq
->lock
, flags
);
640 static inline void resched_task(struct task_struct
*p
)
642 assert_spin_locked(&task_rq(p
)->lock
);
643 set_tsk_need_resched(p
);
647 static u64
div64_likely32(u64 divident
, unsigned long divisor
)
649 #if BITS_PER_LONG == 32
650 if (likely(divident
<= 0xffffffffULL
))
651 return (u32
)divident
/ divisor
;
652 do_div(divident
, divisor
);
656 return divident
/ divisor
;
660 #if BITS_PER_LONG == 32
661 # define WMULT_CONST (~0UL)
663 # define WMULT_CONST (1UL << 32)
666 #define WMULT_SHIFT 32
669 * Shift right and round:
671 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
674 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
675 struct load_weight
*lw
)
679 if (unlikely(!lw
->inv_weight
))
680 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
682 tmp
= (u64
)delta_exec
* weight
;
684 * Check whether we'd overflow the 64-bit multiplication:
686 if (unlikely(tmp
> WMULT_CONST
))
687 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
690 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
692 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
695 static inline unsigned long
696 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
698 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
701 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
704 lw
->inv_weight
= WMULT_CONST
/ lw
->weight
;
707 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
710 if (likely(lw
->weight
))
711 lw
->inv_weight
= WMULT_CONST
/ lw
->weight
;
715 * To aid in avoiding the subversion of "niceness" due to uneven distribution
716 * of tasks with abnormal "nice" values across CPUs the contribution that
717 * each task makes to its run queue's load is weighted according to its
718 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
719 * scaled version of the new time slice allocation that they receive on time
723 #define WEIGHT_IDLEPRIO 2
724 #define WMULT_IDLEPRIO (1 << 31)
727 * Nice levels are multiplicative, with a gentle 10% change for every
728 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
729 * nice 1, it will get ~10% less CPU time than another CPU-bound task
730 * that remained on nice 0.
732 * The "10% effect" is relative and cumulative: from _any_ nice level,
733 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
734 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
735 * If a task goes up by ~10% and another task goes down by ~10% then
736 * the relative distance between them is ~25%.)
738 static const int prio_to_weight
[40] = {
739 /* -20 */ 88761, 71755, 56483, 46273, 36291,
740 /* -15 */ 29154, 23254, 18705, 14949, 11916,
741 /* -10 */ 9548, 7620, 6100, 4904, 3906,
742 /* -5 */ 3121, 2501, 1991, 1586, 1277,
743 /* 0 */ 1024, 820, 655, 526, 423,
744 /* 5 */ 335, 272, 215, 172, 137,
745 /* 10 */ 110, 87, 70, 56, 45,
746 /* 15 */ 36, 29, 23, 18, 15,
750 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
752 * In cases where the weight does not change often, we can use the
753 * precalculated inverse to speed up arithmetics by turning divisions
754 * into multiplications:
756 static const u32 prio_to_wmult
[40] = {
757 /* -20 */ 48388, 59856, 76040, 92818, 118348,
758 /* -15 */ 147320, 184698, 229616, 287308, 360437,
759 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
760 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
761 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
762 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
763 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
764 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
767 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
770 * runqueue iterator, to support SMP load-balancing between different
771 * scheduling classes, without having to expose their internal data
772 * structures to the load-balancing proper:
776 struct task_struct
*(*start
)(void *);
777 struct task_struct
*(*next
)(void *);
780 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
781 unsigned long max_nr_move
, unsigned long max_load_move
,
782 struct sched_domain
*sd
, enum cpu_idle_type idle
,
783 int *all_pinned
, unsigned long *load_moved
,
784 int *this_best_prio
, struct rq_iterator
*iterator
);
786 #include "sched_stats.h"
787 #include "sched_rt.c"
788 #include "sched_fair.c"
789 #include "sched_idletask.c"
790 #ifdef CONFIG_SCHED_DEBUG
791 # include "sched_debug.c"
794 #define sched_class_highest (&rt_sched_class)
797 * Update delta_exec, delta_fair fields for rq.
799 * delta_fair clock advances at a rate inversely proportional to
800 * total load (rq->ls.load.weight) on the runqueue, while
801 * delta_exec advances at the same rate as wall-clock (provided
804 * delta_exec / delta_fair is a measure of the (smoothened) load on this
805 * runqueue over any given interval. This (smoothened) load is used
806 * during load balance.
808 * This function is called /before/ updating rq->ls.load
809 * and when switching tasks.
811 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
813 update_load_add(&rq
->ls
.load
, p
->se
.load
.weight
);
816 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
818 update_load_sub(&rq
->ls
.load
, p
->se
.load
.weight
);
821 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
827 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
833 static void set_load_weight(struct task_struct
*p
)
835 p
->se
.wait_runtime
= 0;
837 if (task_has_rt_policy(p
)) {
838 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
839 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
844 * SCHED_IDLE tasks get minimal weight:
846 if (p
->policy
== SCHED_IDLE
) {
847 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
848 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
852 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
853 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
856 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
858 sched_info_queued(p
);
859 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
863 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
865 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
870 * __normal_prio - return the priority that is based on the static prio
872 static inline int __normal_prio(struct task_struct
*p
)
874 return p
->static_prio
;
878 * Calculate the expected normal priority: i.e. priority
879 * without taking RT-inheritance into account. Might be
880 * boosted by interactivity modifiers. Changes upon fork,
881 * setprio syscalls, and whenever the interactivity
882 * estimator recalculates.
884 static inline int normal_prio(struct task_struct
*p
)
888 if (task_has_rt_policy(p
))
889 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
891 prio
= __normal_prio(p
);
896 * Calculate the current priority, i.e. the priority
897 * taken into account by the scheduler. This value might
898 * be boosted by RT tasks, or might be boosted by
899 * interactivity modifiers. Will be RT if the task got
900 * RT-boosted. If not then it returns p->normal_prio.
902 static int effective_prio(struct task_struct
*p
)
904 p
->normal_prio
= normal_prio(p
);
906 * If we are RT tasks or we were boosted to RT priority,
907 * keep the priority unchanged. Otherwise, update priority
908 * to the normal priority:
910 if (!rt_prio(p
->prio
))
911 return p
->normal_prio
;
916 * activate_task - move a task to the runqueue.
918 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
920 if (p
->state
== TASK_UNINTERRUPTIBLE
)
921 rq
->nr_uninterruptible
--;
923 enqueue_task(rq
, p
, wakeup
);
924 inc_nr_running(p
, rq
);
928 * activate_idle_task - move idle task to the _front_ of runqueue.
930 static inline void activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
934 if (p
->state
== TASK_UNINTERRUPTIBLE
)
935 rq
->nr_uninterruptible
--;
937 enqueue_task(rq
, p
, 0);
938 inc_nr_running(p
, rq
);
942 * deactivate_task - remove a task from the runqueue.
944 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
946 if (p
->state
== TASK_UNINTERRUPTIBLE
)
947 rq
->nr_uninterruptible
++;
949 dequeue_task(rq
, p
, sleep
);
950 dec_nr_running(p
, rq
);
954 * task_curr - is this task currently executing on a CPU?
955 * @p: the task in question.
957 inline int task_curr(const struct task_struct
*p
)
959 return cpu_curr(task_cpu(p
)) == p
;
962 /* Used instead of source_load when we know the type == 0 */
963 unsigned long weighted_cpuload(const int cpu
)
965 return cpu_rq(cpu
)->ls
.load
.weight
;
968 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
971 task_thread_info(p
)->cpu
= cpu
;
978 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
980 int old_cpu
= task_cpu(p
);
981 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
982 u64 clock_offset
, fair_clock_offset
;
984 clock_offset
= old_rq
->clock
- new_rq
->clock
;
985 fair_clock_offset
= old_rq
->cfs
.fair_clock
- new_rq
->cfs
.fair_clock
;
987 if (p
->se
.wait_start_fair
)
988 p
->se
.wait_start_fair
-= fair_clock_offset
;
989 if (p
->se
.sleep_start_fair
)
990 p
->se
.sleep_start_fair
-= fair_clock_offset
;
992 #ifdef CONFIG_SCHEDSTATS
993 if (p
->se
.wait_start
)
994 p
->se
.wait_start
-= clock_offset
;
995 if (p
->se
.sleep_start
)
996 p
->se
.sleep_start
-= clock_offset
;
997 if (p
->se
.block_start
)
998 p
->se
.block_start
-= clock_offset
;
1001 __set_task_cpu(p
, new_cpu
);
1004 struct migration_req
{
1005 struct list_head list
;
1007 struct task_struct
*task
;
1010 struct completion done
;
1014 * The task's runqueue lock must be held.
1015 * Returns true if you have to wait for migration thread.
1018 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1020 struct rq
*rq
= task_rq(p
);
1023 * If the task is not on a runqueue (and not running), then
1024 * it is sufficient to simply update the task's cpu field.
1026 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1027 set_task_cpu(p
, dest_cpu
);
1031 init_completion(&req
->done
);
1033 req
->dest_cpu
= dest_cpu
;
1034 list_add(&req
->list
, &rq
->migration_queue
);
1040 * wait_task_inactive - wait for a thread to unschedule.
1042 * The caller must ensure that the task *will* unschedule sometime soon,
1043 * else this function might spin for a *long* time. This function can't
1044 * be called with interrupts off, or it may introduce deadlock with
1045 * smp_call_function() if an IPI is sent by the same process we are
1046 * waiting to become inactive.
1048 void wait_task_inactive(struct task_struct
*p
)
1050 unsigned long flags
;
1056 * We do the initial early heuristics without holding
1057 * any task-queue locks at all. We'll only try to get
1058 * the runqueue lock when things look like they will
1064 * If the task is actively running on another CPU
1065 * still, just relax and busy-wait without holding
1068 * NOTE! Since we don't hold any locks, it's not
1069 * even sure that "rq" stays as the right runqueue!
1070 * But we don't care, since "task_running()" will
1071 * return false if the runqueue has changed and p
1072 * is actually now running somewhere else!
1074 while (task_running(rq
, p
))
1078 * Ok, time to look more closely! We need the rq
1079 * lock now, to be *sure*. If we're wrong, we'll
1080 * just go back and repeat.
1082 rq
= task_rq_lock(p
, &flags
);
1083 running
= task_running(rq
, p
);
1084 on_rq
= p
->se
.on_rq
;
1085 task_rq_unlock(rq
, &flags
);
1088 * Was it really running after all now that we
1089 * checked with the proper locks actually held?
1091 * Oops. Go back and try again..
1093 if (unlikely(running
)) {
1099 * It's not enough that it's not actively running,
1100 * it must be off the runqueue _entirely_, and not
1103 * So if it wa still runnable (but just not actively
1104 * running right now), it's preempted, and we should
1105 * yield - it could be a while.
1107 if (unlikely(on_rq
)) {
1113 * Ahh, all good. It wasn't running, and it wasn't
1114 * runnable, which means that it will never become
1115 * running in the future either. We're all done!
1120 * kick_process - kick a running thread to enter/exit the kernel
1121 * @p: the to-be-kicked thread
1123 * Cause a process which is running on another CPU to enter
1124 * kernel-mode, without any delay. (to get signals handled.)
1126 * NOTE: this function doesnt have to take the runqueue lock,
1127 * because all it wants to ensure is that the remote task enters
1128 * the kernel. If the IPI races and the task has been migrated
1129 * to another CPU then no harm is done and the purpose has been
1132 void kick_process(struct task_struct
*p
)
1138 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1139 smp_send_reschedule(cpu
);
1144 * Return a low guess at the load of a migration-source cpu weighted
1145 * according to the scheduling class and "nice" value.
1147 * We want to under-estimate the load of migration sources, to
1148 * balance conservatively.
1150 static inline unsigned long source_load(int cpu
, int type
)
1152 struct rq
*rq
= cpu_rq(cpu
);
1153 unsigned long total
= weighted_cpuload(cpu
);
1158 return min(rq
->cpu_load
[type
-1], total
);
1162 * Return a high guess at the load of a migration-target cpu weighted
1163 * according to the scheduling class and "nice" value.
1165 static inline unsigned long target_load(int cpu
, int type
)
1167 struct rq
*rq
= cpu_rq(cpu
);
1168 unsigned long total
= weighted_cpuload(cpu
);
1173 return max(rq
->cpu_load
[type
-1], total
);
1177 * Return the average load per task on the cpu's run queue
1179 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1181 struct rq
*rq
= cpu_rq(cpu
);
1182 unsigned long total
= weighted_cpuload(cpu
);
1183 unsigned long n
= rq
->nr_running
;
1185 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1189 * find_idlest_group finds and returns the least busy CPU group within the
1192 static struct sched_group
*
1193 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1195 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1196 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1197 int load_idx
= sd
->forkexec_idx
;
1198 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1201 unsigned long load
, avg_load
;
1205 /* Skip over this group if it has no CPUs allowed */
1206 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1209 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1211 /* Tally up the load of all CPUs in the group */
1214 for_each_cpu_mask(i
, group
->cpumask
) {
1215 /* Bias balancing toward cpus of our domain */
1217 load
= source_load(i
, load_idx
);
1219 load
= target_load(i
, load_idx
);
1224 /* Adjust by relative CPU power of the group */
1225 avg_load
= sg_div_cpu_power(group
,
1226 avg_load
* SCHED_LOAD_SCALE
);
1229 this_load
= avg_load
;
1231 } else if (avg_load
< min_load
) {
1232 min_load
= avg_load
;
1236 group
= group
->next
;
1237 } while (group
!= sd
->groups
);
1239 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1245 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1248 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1251 unsigned long load
, min_load
= ULONG_MAX
;
1255 /* Traverse only the allowed CPUs */
1256 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1258 for_each_cpu_mask(i
, tmp
) {
1259 load
= weighted_cpuload(i
);
1261 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1271 * sched_balance_self: balance the current task (running on cpu) in domains
1272 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1275 * Balance, ie. select the least loaded group.
1277 * Returns the target CPU number, or the same CPU if no balancing is needed.
1279 * preempt must be disabled.
1281 static int sched_balance_self(int cpu
, int flag
)
1283 struct task_struct
*t
= current
;
1284 struct sched_domain
*tmp
, *sd
= NULL
;
1286 for_each_domain(cpu
, tmp
) {
1288 * If power savings logic is enabled for a domain, stop there.
1290 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1292 if (tmp
->flags
& flag
)
1298 struct sched_group
*group
;
1299 int new_cpu
, weight
;
1301 if (!(sd
->flags
& flag
)) {
1307 group
= find_idlest_group(sd
, t
, cpu
);
1313 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1314 if (new_cpu
== -1 || new_cpu
== cpu
) {
1315 /* Now try balancing at a lower domain level of cpu */
1320 /* Now try balancing at a lower domain level of new_cpu */
1323 weight
= cpus_weight(span
);
1324 for_each_domain(cpu
, tmp
) {
1325 if (weight
<= cpus_weight(tmp
->span
))
1327 if (tmp
->flags
& flag
)
1330 /* while loop will break here if sd == NULL */
1336 #endif /* CONFIG_SMP */
1339 * wake_idle() will wake a task on an idle cpu if task->cpu is
1340 * not idle and an idle cpu is available. The span of cpus to
1341 * search starts with cpus closest then further out as needed,
1342 * so we always favor a closer, idle cpu.
1344 * Returns the CPU we should wake onto.
1346 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1347 static int wake_idle(int cpu
, struct task_struct
*p
)
1350 struct sched_domain
*sd
;
1354 * If it is idle, then it is the best cpu to run this task.
1356 * This cpu is also the best, if it has more than one task already.
1357 * Siblings must be also busy(in most cases) as they didn't already
1358 * pickup the extra load from this cpu and hence we need not check
1359 * sibling runqueue info. This will avoid the checks and cache miss
1360 * penalities associated with that.
1362 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1365 for_each_domain(cpu
, sd
) {
1366 if (sd
->flags
& SD_WAKE_IDLE
) {
1367 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1368 for_each_cpu_mask(i
, tmp
) {
1379 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1386 * try_to_wake_up - wake up a thread
1387 * @p: the to-be-woken-up thread
1388 * @state: the mask of task states that can be woken
1389 * @sync: do a synchronous wakeup?
1391 * Put it on the run-queue if it's not already there. The "current"
1392 * thread is always on the run-queue (except when the actual
1393 * re-schedule is in progress), and as such you're allowed to do
1394 * the simpler "current->state = TASK_RUNNING" to mark yourself
1395 * runnable without the overhead of this.
1397 * returns failure only if the task is already active.
1399 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1401 int cpu
, this_cpu
, success
= 0;
1402 unsigned long flags
;
1406 struct sched_domain
*sd
, *this_sd
= NULL
;
1407 unsigned long load
, this_load
;
1411 rq
= task_rq_lock(p
, &flags
);
1412 old_state
= p
->state
;
1413 if (!(old_state
& state
))
1420 this_cpu
= smp_processor_id();
1423 if (unlikely(task_running(rq
, p
)))
1428 schedstat_inc(rq
, ttwu_cnt
);
1429 if (cpu
== this_cpu
) {
1430 schedstat_inc(rq
, ttwu_local
);
1434 for_each_domain(this_cpu
, sd
) {
1435 if (cpu_isset(cpu
, sd
->span
)) {
1436 schedstat_inc(sd
, ttwu_wake_remote
);
1442 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1446 * Check for affine wakeup and passive balancing possibilities.
1449 int idx
= this_sd
->wake_idx
;
1450 unsigned int imbalance
;
1452 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1454 load
= source_load(cpu
, idx
);
1455 this_load
= target_load(this_cpu
, idx
);
1457 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1459 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1460 unsigned long tl
= this_load
;
1461 unsigned long tl_per_task
;
1463 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1466 * If sync wakeup then subtract the (maximum possible)
1467 * effect of the currently running task from the load
1468 * of the current CPU:
1471 tl
-= current
->se
.load
.weight
;
1474 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1475 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1477 * This domain has SD_WAKE_AFFINE and
1478 * p is cache cold in this domain, and
1479 * there is no bad imbalance.
1481 schedstat_inc(this_sd
, ttwu_move_affine
);
1487 * Start passive balancing when half the imbalance_pct
1490 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1491 if (imbalance
*this_load
<= 100*load
) {
1492 schedstat_inc(this_sd
, ttwu_move_balance
);
1498 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1500 new_cpu
= wake_idle(new_cpu
, p
);
1501 if (new_cpu
!= cpu
) {
1502 set_task_cpu(p
, new_cpu
);
1503 task_rq_unlock(rq
, &flags
);
1504 /* might preempt at this point */
1505 rq
= task_rq_lock(p
, &flags
);
1506 old_state
= p
->state
;
1507 if (!(old_state
& state
))
1512 this_cpu
= smp_processor_id();
1517 #endif /* CONFIG_SMP */
1518 update_rq_clock(rq
);
1519 activate_task(rq
, p
, 1);
1521 * Sync wakeups (i.e. those types of wakeups where the waker
1522 * has indicated that it will leave the CPU in short order)
1523 * don't trigger a preemption, if the woken up task will run on
1524 * this cpu. (in this case the 'I will reschedule' promise of
1525 * the waker guarantees that the freshly woken up task is going
1526 * to be considered on this CPU.)
1528 if (!sync
|| cpu
!= this_cpu
)
1529 check_preempt_curr(rq
, p
);
1533 p
->state
= TASK_RUNNING
;
1535 task_rq_unlock(rq
, &flags
);
1540 int fastcall
wake_up_process(struct task_struct
*p
)
1542 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1543 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1545 EXPORT_SYMBOL(wake_up_process
);
1547 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1549 return try_to_wake_up(p
, state
, 0);
1553 * Perform scheduler related setup for a newly forked process p.
1554 * p is forked by current.
1556 * __sched_fork() is basic setup used by init_idle() too:
1558 static void __sched_fork(struct task_struct
*p
)
1560 p
->se
.wait_start_fair
= 0;
1561 p
->se
.exec_start
= 0;
1562 p
->se
.sum_exec_runtime
= 0;
1563 p
->se
.prev_sum_exec_runtime
= 0;
1564 p
->se
.wait_runtime
= 0;
1565 p
->se
.sleep_start_fair
= 0;
1567 #ifdef CONFIG_SCHEDSTATS
1568 p
->se
.wait_start
= 0;
1569 p
->se
.sum_wait_runtime
= 0;
1570 p
->se
.sum_sleep_runtime
= 0;
1571 p
->se
.sleep_start
= 0;
1572 p
->se
.block_start
= 0;
1573 p
->se
.sleep_max
= 0;
1574 p
->se
.block_max
= 0;
1576 p
->se
.slice_max
= 0;
1578 p
->se
.wait_runtime_overruns
= 0;
1579 p
->se
.wait_runtime_underruns
= 0;
1582 INIT_LIST_HEAD(&p
->run_list
);
1585 #ifdef CONFIG_PREEMPT_NOTIFIERS
1586 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1590 * We mark the process as running here, but have not actually
1591 * inserted it onto the runqueue yet. This guarantees that
1592 * nobody will actually run it, and a signal or other external
1593 * event cannot wake it up and insert it on the runqueue either.
1595 p
->state
= TASK_RUNNING
;
1599 * fork()/clone()-time setup:
1601 void sched_fork(struct task_struct
*p
, int clone_flags
)
1603 int cpu
= get_cpu();
1608 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1610 __set_task_cpu(p
, cpu
);
1613 * Make sure we do not leak PI boosting priority to the child:
1615 p
->prio
= current
->normal_prio
;
1617 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1618 if (likely(sched_info_on()))
1619 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1621 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1624 #ifdef CONFIG_PREEMPT
1625 /* Want to start with kernel preemption disabled. */
1626 task_thread_info(p
)->preempt_count
= 1;
1632 * wake_up_new_task - wake up a newly created task for the first time.
1634 * This function will do some initial scheduler statistics housekeeping
1635 * that must be done for every newly created context, then puts the task
1636 * on the runqueue and wakes it.
1638 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1640 unsigned long flags
;
1644 rq
= task_rq_lock(p
, &flags
);
1645 BUG_ON(p
->state
!= TASK_RUNNING
);
1646 this_cpu
= smp_processor_id(); /* parent's CPU */
1647 update_rq_clock(rq
);
1649 p
->prio
= effective_prio(p
);
1651 if (rt_prio(p
->prio
))
1652 p
->sched_class
= &rt_sched_class
;
1654 p
->sched_class
= &fair_sched_class
;
1656 if (task_cpu(p
) != this_cpu
|| !p
->sched_class
->task_new
||
1657 !current
->se
.on_rq
) {
1658 activate_task(rq
, p
, 0);
1661 * Let the scheduling class do new task startup
1662 * management (if any):
1664 p
->sched_class
->task_new(rq
, p
);
1665 inc_nr_running(p
, rq
);
1667 check_preempt_curr(rq
, p
);
1668 task_rq_unlock(rq
, &flags
);
1671 #ifdef CONFIG_PREEMPT_NOTIFIERS
1674 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1675 * @notifier: notifier struct to register
1677 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1679 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1681 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1684 * preempt_notifier_unregister - no longer interested in preemption notifications
1685 * @notifier: notifier struct to unregister
1687 * This is safe to call from within a preemption notifier.
1689 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1691 hlist_del(¬ifier
->link
);
1693 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1695 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1697 struct preempt_notifier
*notifier
;
1698 struct hlist_node
*node
;
1700 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1701 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1705 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1706 struct task_struct
*next
)
1708 struct preempt_notifier
*notifier
;
1709 struct hlist_node
*node
;
1711 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1712 notifier
->ops
->sched_out(notifier
, next
);
1717 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1722 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1723 struct task_struct
*next
)
1730 * prepare_task_switch - prepare to switch tasks
1731 * @rq: the runqueue preparing to switch
1732 * @prev: the current task that is being switched out
1733 * @next: the task we are going to switch to.
1735 * This is called with the rq lock held and interrupts off. It must
1736 * be paired with a subsequent finish_task_switch after the context
1739 * prepare_task_switch sets up locking and calls architecture specific
1743 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1744 struct task_struct
*next
)
1746 fire_sched_out_preempt_notifiers(prev
, next
);
1747 prepare_lock_switch(rq
, next
);
1748 prepare_arch_switch(next
);
1752 * finish_task_switch - clean up after a task-switch
1753 * @rq: runqueue associated with task-switch
1754 * @prev: the thread we just switched away from.
1756 * finish_task_switch must be called after the context switch, paired
1757 * with a prepare_task_switch call before the context switch.
1758 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1759 * and do any other architecture-specific cleanup actions.
1761 * Note that we may have delayed dropping an mm in context_switch(). If
1762 * so, we finish that here outside of the runqueue lock. (Doing it
1763 * with the lock held can cause deadlocks; see schedule() for
1766 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1767 __releases(rq
->lock
)
1769 struct mm_struct
*mm
= rq
->prev_mm
;
1775 * A task struct has one reference for the use as "current".
1776 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1777 * schedule one last time. The schedule call will never return, and
1778 * the scheduled task must drop that reference.
1779 * The test for TASK_DEAD must occur while the runqueue locks are
1780 * still held, otherwise prev could be scheduled on another cpu, die
1781 * there before we look at prev->state, and then the reference would
1783 * Manfred Spraul <manfred@colorfullife.com>
1785 prev_state
= prev
->state
;
1786 finish_arch_switch(prev
);
1787 finish_lock_switch(rq
, prev
);
1788 fire_sched_in_preempt_notifiers(current
);
1791 if (unlikely(prev_state
== TASK_DEAD
)) {
1793 * Remove function-return probe instances associated with this
1794 * task and put them back on the free list.
1796 kprobe_flush_task(prev
);
1797 put_task_struct(prev
);
1802 * schedule_tail - first thing a freshly forked thread must call.
1803 * @prev: the thread we just switched away from.
1805 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1806 __releases(rq
->lock
)
1808 struct rq
*rq
= this_rq();
1810 finish_task_switch(rq
, prev
);
1811 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1812 /* In this case, finish_task_switch does not reenable preemption */
1815 if (current
->set_child_tid
)
1816 put_user(current
->pid
, current
->set_child_tid
);
1820 * context_switch - switch to the new MM and the new
1821 * thread's register state.
1824 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1825 struct task_struct
*next
)
1827 struct mm_struct
*mm
, *oldmm
;
1829 prepare_task_switch(rq
, prev
, next
);
1831 oldmm
= prev
->active_mm
;
1833 * For paravirt, this is coupled with an exit in switch_to to
1834 * combine the page table reload and the switch backend into
1837 arch_enter_lazy_cpu_mode();
1839 if (unlikely(!mm
)) {
1840 next
->active_mm
= oldmm
;
1841 atomic_inc(&oldmm
->mm_count
);
1842 enter_lazy_tlb(oldmm
, next
);
1844 switch_mm(oldmm
, mm
, next
);
1846 if (unlikely(!prev
->mm
)) {
1847 prev
->active_mm
= NULL
;
1848 rq
->prev_mm
= oldmm
;
1851 * Since the runqueue lock will be released by the next
1852 * task (which is an invalid locking op but in the case
1853 * of the scheduler it's an obvious special-case), so we
1854 * do an early lockdep release here:
1856 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1857 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1860 /* Here we just switch the register state and the stack. */
1861 switch_to(prev
, next
, prev
);
1865 * this_rq must be evaluated again because prev may have moved
1866 * CPUs since it called schedule(), thus the 'rq' on its stack
1867 * frame will be invalid.
1869 finish_task_switch(this_rq(), prev
);
1873 * nr_running, nr_uninterruptible and nr_context_switches:
1875 * externally visible scheduler statistics: current number of runnable
1876 * threads, current number of uninterruptible-sleeping threads, total
1877 * number of context switches performed since bootup.
1879 unsigned long nr_running(void)
1881 unsigned long i
, sum
= 0;
1883 for_each_online_cpu(i
)
1884 sum
+= cpu_rq(i
)->nr_running
;
1889 unsigned long nr_uninterruptible(void)
1891 unsigned long i
, sum
= 0;
1893 for_each_possible_cpu(i
)
1894 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1897 * Since we read the counters lockless, it might be slightly
1898 * inaccurate. Do not allow it to go below zero though:
1900 if (unlikely((long)sum
< 0))
1906 unsigned long long nr_context_switches(void)
1909 unsigned long long sum
= 0;
1911 for_each_possible_cpu(i
)
1912 sum
+= cpu_rq(i
)->nr_switches
;
1917 unsigned long nr_iowait(void)
1919 unsigned long i
, sum
= 0;
1921 for_each_possible_cpu(i
)
1922 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1927 unsigned long nr_active(void)
1929 unsigned long i
, running
= 0, uninterruptible
= 0;
1931 for_each_online_cpu(i
) {
1932 running
+= cpu_rq(i
)->nr_running
;
1933 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1936 if (unlikely((long)uninterruptible
< 0))
1937 uninterruptible
= 0;
1939 return running
+ uninterruptible
;
1943 * Update rq->cpu_load[] statistics. This function is usually called every
1944 * scheduler tick (TICK_NSEC).
1946 static void update_cpu_load(struct rq
*this_rq
)
1948 unsigned long this_load
= this_rq
->ls
.load
.weight
;
1951 this_rq
->nr_load_updates
++;
1953 /* Update our load: */
1954 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
1955 unsigned long old_load
, new_load
;
1957 /* scale is effectively 1 << i now, and >> i divides by scale */
1959 old_load
= this_rq
->cpu_load
[i
];
1960 new_load
= this_load
;
1962 * Round up the averaging division if load is increasing. This
1963 * prevents us from getting stuck on 9 if the load is 10, for
1966 if (new_load
> old_load
)
1967 new_load
+= scale
-1;
1968 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
1975 * double_rq_lock - safely lock two runqueues
1977 * Note this does not disable interrupts like task_rq_lock,
1978 * you need to do so manually before calling.
1980 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1981 __acquires(rq1
->lock
)
1982 __acquires(rq2
->lock
)
1984 BUG_ON(!irqs_disabled());
1986 spin_lock(&rq1
->lock
);
1987 __acquire(rq2
->lock
); /* Fake it out ;) */
1990 spin_lock(&rq1
->lock
);
1991 spin_lock(&rq2
->lock
);
1993 spin_lock(&rq2
->lock
);
1994 spin_lock(&rq1
->lock
);
1997 update_rq_clock(rq1
);
1998 update_rq_clock(rq2
);
2002 * double_rq_unlock - safely unlock two runqueues
2004 * Note this does not restore interrupts like task_rq_unlock,
2005 * you need to do so manually after calling.
2007 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2008 __releases(rq1
->lock
)
2009 __releases(rq2
->lock
)
2011 spin_unlock(&rq1
->lock
);
2013 spin_unlock(&rq2
->lock
);
2015 __release(rq2
->lock
);
2019 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2021 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2022 __releases(this_rq
->lock
)
2023 __acquires(busiest
->lock
)
2024 __acquires(this_rq
->lock
)
2026 if (unlikely(!irqs_disabled())) {
2027 /* printk() doesn't work good under rq->lock */
2028 spin_unlock(&this_rq
->lock
);
2031 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2032 if (busiest
< this_rq
) {
2033 spin_unlock(&this_rq
->lock
);
2034 spin_lock(&busiest
->lock
);
2035 spin_lock(&this_rq
->lock
);
2037 spin_lock(&busiest
->lock
);
2042 * If dest_cpu is allowed for this process, migrate the task to it.
2043 * This is accomplished by forcing the cpu_allowed mask to only
2044 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2045 * the cpu_allowed mask is restored.
2047 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2049 struct migration_req req
;
2050 unsigned long flags
;
2053 rq
= task_rq_lock(p
, &flags
);
2054 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2055 || unlikely(cpu_is_offline(dest_cpu
)))
2058 /* force the process onto the specified CPU */
2059 if (migrate_task(p
, dest_cpu
, &req
)) {
2060 /* Need to wait for migration thread (might exit: take ref). */
2061 struct task_struct
*mt
= rq
->migration_thread
;
2063 get_task_struct(mt
);
2064 task_rq_unlock(rq
, &flags
);
2065 wake_up_process(mt
);
2066 put_task_struct(mt
);
2067 wait_for_completion(&req
.done
);
2072 task_rq_unlock(rq
, &flags
);
2076 * sched_exec - execve() is a valuable balancing opportunity, because at
2077 * this point the task has the smallest effective memory and cache footprint.
2079 void sched_exec(void)
2081 int new_cpu
, this_cpu
= get_cpu();
2082 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2084 if (new_cpu
!= this_cpu
)
2085 sched_migrate_task(current
, new_cpu
);
2089 * pull_task - move a task from a remote runqueue to the local runqueue.
2090 * Both runqueues must be locked.
2092 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2093 struct rq
*this_rq
, int this_cpu
)
2095 deactivate_task(src_rq
, p
, 0);
2096 set_task_cpu(p
, this_cpu
);
2097 activate_task(this_rq
, p
, 0);
2099 * Note that idle threads have a prio of MAX_PRIO, for this test
2100 * to be always true for them.
2102 check_preempt_curr(this_rq
, p
);
2106 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2109 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2110 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2114 * We do not migrate tasks that are:
2115 * 1) running (obviously), or
2116 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2117 * 3) are cache-hot on their current CPU.
2119 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2123 if (task_running(rq
, p
))
2129 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2130 unsigned long max_nr_move
, unsigned long max_load_move
,
2131 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2132 int *all_pinned
, unsigned long *load_moved
,
2133 int *this_best_prio
, struct rq_iterator
*iterator
)
2135 int pulled
= 0, pinned
= 0, skip_for_load
;
2136 struct task_struct
*p
;
2137 long rem_load_move
= max_load_move
;
2139 if (max_nr_move
== 0 || max_load_move
== 0)
2145 * Start the load-balancing iterator:
2147 p
= iterator
->start(iterator
->arg
);
2152 * To help distribute high priority tasks accross CPUs we don't
2153 * skip a task if it will be the highest priority task (i.e. smallest
2154 * prio value) on its new queue regardless of its load weight
2156 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2157 SCHED_LOAD_SCALE_FUZZ
;
2158 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2159 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2160 p
= iterator
->next(iterator
->arg
);
2164 pull_task(busiest
, p
, this_rq
, this_cpu
);
2166 rem_load_move
-= p
->se
.load
.weight
;
2169 * We only want to steal up to the prescribed number of tasks
2170 * and the prescribed amount of weighted load.
2172 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2173 if (p
->prio
< *this_best_prio
)
2174 *this_best_prio
= p
->prio
;
2175 p
= iterator
->next(iterator
->arg
);
2180 * Right now, this is the only place pull_task() is called,
2181 * so we can safely collect pull_task() stats here rather than
2182 * inside pull_task().
2184 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2187 *all_pinned
= pinned
;
2188 *load_moved
= max_load_move
- rem_load_move
;
2193 * move_tasks tries to move up to max_load_move weighted load from busiest to
2194 * this_rq, as part of a balancing operation within domain "sd".
2195 * Returns 1 if successful and 0 otherwise.
2197 * Called with both runqueues locked.
2199 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2200 unsigned long max_load_move
,
2201 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2204 struct sched_class
*class = sched_class_highest
;
2205 unsigned long total_load_moved
= 0;
2206 int this_best_prio
= this_rq
->curr
->prio
;
2210 class->load_balance(this_rq
, this_cpu
, busiest
,
2211 ULONG_MAX
, max_load_move
- total_load_moved
,
2212 sd
, idle
, all_pinned
, &this_best_prio
);
2213 class = class->next
;
2214 } while (class && max_load_move
> total_load_moved
);
2216 return total_load_moved
> 0;
2220 * move_one_task tries to move exactly one task from busiest to this_rq, as
2221 * part of active balancing operations within "domain".
2222 * Returns 1 if successful and 0 otherwise.
2224 * Called with both runqueues locked.
2226 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2227 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2229 struct sched_class
*class;
2230 int this_best_prio
= MAX_PRIO
;
2232 for (class = sched_class_highest
; class; class = class->next
)
2233 if (class->load_balance(this_rq
, this_cpu
, busiest
,
2234 1, ULONG_MAX
, sd
, idle
, NULL
,
2242 * find_busiest_group finds and returns the busiest CPU group within the
2243 * domain. It calculates and returns the amount of weighted load which
2244 * should be moved to restore balance via the imbalance parameter.
2246 static struct sched_group
*
2247 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2248 unsigned long *imbalance
, enum cpu_idle_type idle
,
2249 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2251 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2252 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2253 unsigned long max_pull
;
2254 unsigned long busiest_load_per_task
, busiest_nr_running
;
2255 unsigned long this_load_per_task
, this_nr_running
;
2257 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2258 int power_savings_balance
= 1;
2259 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2260 unsigned long min_nr_running
= ULONG_MAX
;
2261 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2264 max_load
= this_load
= total_load
= total_pwr
= 0;
2265 busiest_load_per_task
= busiest_nr_running
= 0;
2266 this_load_per_task
= this_nr_running
= 0;
2267 if (idle
== CPU_NOT_IDLE
)
2268 load_idx
= sd
->busy_idx
;
2269 else if (idle
== CPU_NEWLY_IDLE
)
2270 load_idx
= sd
->newidle_idx
;
2272 load_idx
= sd
->idle_idx
;
2275 unsigned long load
, group_capacity
;
2278 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2279 unsigned long sum_nr_running
, sum_weighted_load
;
2281 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2284 balance_cpu
= first_cpu(group
->cpumask
);
2286 /* Tally up the load of all CPUs in the group */
2287 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2289 for_each_cpu_mask(i
, group
->cpumask
) {
2292 if (!cpu_isset(i
, *cpus
))
2297 if (*sd_idle
&& rq
->nr_running
)
2300 /* Bias balancing toward cpus of our domain */
2302 if (idle_cpu(i
) && !first_idle_cpu
) {
2307 load
= target_load(i
, load_idx
);
2309 load
= source_load(i
, load_idx
);
2312 sum_nr_running
+= rq
->nr_running
;
2313 sum_weighted_load
+= weighted_cpuload(i
);
2317 * First idle cpu or the first cpu(busiest) in this sched group
2318 * is eligible for doing load balancing at this and above
2319 * domains. In the newly idle case, we will allow all the cpu's
2320 * to do the newly idle load balance.
2322 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2323 balance_cpu
!= this_cpu
&& balance
) {
2328 total_load
+= avg_load
;
2329 total_pwr
+= group
->__cpu_power
;
2331 /* Adjust by relative CPU power of the group */
2332 avg_load
= sg_div_cpu_power(group
,
2333 avg_load
* SCHED_LOAD_SCALE
);
2335 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2338 this_load
= avg_load
;
2340 this_nr_running
= sum_nr_running
;
2341 this_load_per_task
= sum_weighted_load
;
2342 } else if (avg_load
> max_load
&&
2343 sum_nr_running
> group_capacity
) {
2344 max_load
= avg_load
;
2346 busiest_nr_running
= sum_nr_running
;
2347 busiest_load_per_task
= sum_weighted_load
;
2350 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2352 * Busy processors will not participate in power savings
2355 if (idle
== CPU_NOT_IDLE
||
2356 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2360 * If the local group is idle or completely loaded
2361 * no need to do power savings balance at this domain
2363 if (local_group
&& (this_nr_running
>= group_capacity
||
2365 power_savings_balance
= 0;
2368 * If a group is already running at full capacity or idle,
2369 * don't include that group in power savings calculations
2371 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2376 * Calculate the group which has the least non-idle load.
2377 * This is the group from where we need to pick up the load
2380 if ((sum_nr_running
< min_nr_running
) ||
2381 (sum_nr_running
== min_nr_running
&&
2382 first_cpu(group
->cpumask
) <
2383 first_cpu(group_min
->cpumask
))) {
2385 min_nr_running
= sum_nr_running
;
2386 min_load_per_task
= sum_weighted_load
/
2391 * Calculate the group which is almost near its
2392 * capacity but still has some space to pick up some load
2393 * from other group and save more power
2395 if (sum_nr_running
<= group_capacity
- 1) {
2396 if (sum_nr_running
> leader_nr_running
||
2397 (sum_nr_running
== leader_nr_running
&&
2398 first_cpu(group
->cpumask
) >
2399 first_cpu(group_leader
->cpumask
))) {
2400 group_leader
= group
;
2401 leader_nr_running
= sum_nr_running
;
2406 group
= group
->next
;
2407 } while (group
!= sd
->groups
);
2409 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2412 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2414 if (this_load
>= avg_load
||
2415 100*max_load
<= sd
->imbalance_pct
*this_load
)
2418 busiest_load_per_task
/= busiest_nr_running
;
2420 * We're trying to get all the cpus to the average_load, so we don't
2421 * want to push ourselves above the average load, nor do we wish to
2422 * reduce the max loaded cpu below the average load, as either of these
2423 * actions would just result in more rebalancing later, and ping-pong
2424 * tasks around. Thus we look for the minimum possible imbalance.
2425 * Negative imbalances (*we* are more loaded than anyone else) will
2426 * be counted as no imbalance for these purposes -- we can't fix that
2427 * by pulling tasks to us. Be careful of negative numbers as they'll
2428 * appear as very large values with unsigned longs.
2430 if (max_load
<= busiest_load_per_task
)
2434 * In the presence of smp nice balancing, certain scenarios can have
2435 * max load less than avg load(as we skip the groups at or below
2436 * its cpu_power, while calculating max_load..)
2438 if (max_load
< avg_load
) {
2440 goto small_imbalance
;
2443 /* Don't want to pull so many tasks that a group would go idle */
2444 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2446 /* How much load to actually move to equalise the imbalance */
2447 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2448 (avg_load
- this_load
) * this->__cpu_power
)
2452 * if *imbalance is less than the average load per runnable task
2453 * there is no gaurantee that any tasks will be moved so we'll have
2454 * a think about bumping its value to force at least one task to be
2457 if (*imbalance
< busiest_load_per_task
) {
2458 unsigned long tmp
, pwr_now
, pwr_move
;
2462 pwr_move
= pwr_now
= 0;
2464 if (this_nr_running
) {
2465 this_load_per_task
/= this_nr_running
;
2466 if (busiest_load_per_task
> this_load_per_task
)
2469 this_load_per_task
= SCHED_LOAD_SCALE
;
2471 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2472 busiest_load_per_task
* imbn
) {
2473 *imbalance
= busiest_load_per_task
;
2478 * OK, we don't have enough imbalance to justify moving tasks,
2479 * however we may be able to increase total CPU power used by
2483 pwr_now
+= busiest
->__cpu_power
*
2484 min(busiest_load_per_task
, max_load
);
2485 pwr_now
+= this->__cpu_power
*
2486 min(this_load_per_task
, this_load
);
2487 pwr_now
/= SCHED_LOAD_SCALE
;
2489 /* Amount of load we'd subtract */
2490 tmp
= sg_div_cpu_power(busiest
,
2491 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2493 pwr_move
+= busiest
->__cpu_power
*
2494 min(busiest_load_per_task
, max_load
- tmp
);
2496 /* Amount of load we'd add */
2497 if (max_load
* busiest
->__cpu_power
<
2498 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2499 tmp
= sg_div_cpu_power(this,
2500 max_load
* busiest
->__cpu_power
);
2502 tmp
= sg_div_cpu_power(this,
2503 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2504 pwr_move
+= this->__cpu_power
*
2505 min(this_load_per_task
, this_load
+ tmp
);
2506 pwr_move
/= SCHED_LOAD_SCALE
;
2508 /* Move if we gain throughput */
2509 if (pwr_move
> pwr_now
)
2510 *imbalance
= busiest_load_per_task
;
2516 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2517 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2520 if (this == group_leader
&& group_leader
!= group_min
) {
2521 *imbalance
= min_load_per_task
;
2531 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2534 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2535 unsigned long imbalance
, cpumask_t
*cpus
)
2537 struct rq
*busiest
= NULL
, *rq
;
2538 unsigned long max_load
= 0;
2541 for_each_cpu_mask(i
, group
->cpumask
) {
2544 if (!cpu_isset(i
, *cpus
))
2548 wl
= weighted_cpuload(i
);
2550 if (rq
->nr_running
== 1 && wl
> imbalance
)
2553 if (wl
> max_load
) {
2563 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2564 * so long as it is large enough.
2566 #define MAX_PINNED_INTERVAL 512
2569 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2570 * tasks if there is an imbalance.
2572 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2573 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2576 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2577 struct sched_group
*group
;
2578 unsigned long imbalance
;
2580 cpumask_t cpus
= CPU_MASK_ALL
;
2581 unsigned long flags
;
2584 * When power savings policy is enabled for the parent domain, idle
2585 * sibling can pick up load irrespective of busy siblings. In this case,
2586 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2587 * portraying it as CPU_NOT_IDLE.
2589 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2590 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2593 schedstat_inc(sd
, lb_cnt
[idle
]);
2596 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2603 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2607 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2609 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2613 BUG_ON(busiest
== this_rq
);
2615 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2618 if (busiest
->nr_running
> 1) {
2620 * Attempt to move tasks. If find_busiest_group has found
2621 * an imbalance but busiest->nr_running <= 1, the group is
2622 * still unbalanced. ld_moved simply stays zero, so it is
2623 * correctly treated as an imbalance.
2625 local_irq_save(flags
);
2626 double_rq_lock(this_rq
, busiest
);
2627 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2628 imbalance
, sd
, idle
, &all_pinned
);
2629 double_rq_unlock(this_rq
, busiest
);
2630 local_irq_restore(flags
);
2633 * some other cpu did the load balance for us.
2635 if (ld_moved
&& this_cpu
!= smp_processor_id())
2636 resched_cpu(this_cpu
);
2638 /* All tasks on this runqueue were pinned by CPU affinity */
2639 if (unlikely(all_pinned
)) {
2640 cpu_clear(cpu_of(busiest
), cpus
);
2641 if (!cpus_empty(cpus
))
2648 schedstat_inc(sd
, lb_failed
[idle
]);
2649 sd
->nr_balance_failed
++;
2651 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2653 spin_lock_irqsave(&busiest
->lock
, flags
);
2655 /* don't kick the migration_thread, if the curr
2656 * task on busiest cpu can't be moved to this_cpu
2658 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2659 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2661 goto out_one_pinned
;
2664 if (!busiest
->active_balance
) {
2665 busiest
->active_balance
= 1;
2666 busiest
->push_cpu
= this_cpu
;
2669 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2671 wake_up_process(busiest
->migration_thread
);
2674 * We've kicked active balancing, reset the failure
2677 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2680 sd
->nr_balance_failed
= 0;
2682 if (likely(!active_balance
)) {
2683 /* We were unbalanced, so reset the balancing interval */
2684 sd
->balance_interval
= sd
->min_interval
;
2687 * If we've begun active balancing, start to back off. This
2688 * case may not be covered by the all_pinned logic if there
2689 * is only 1 task on the busy runqueue (because we don't call
2692 if (sd
->balance_interval
< sd
->max_interval
)
2693 sd
->balance_interval
*= 2;
2696 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2697 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2702 schedstat_inc(sd
, lb_balanced
[idle
]);
2704 sd
->nr_balance_failed
= 0;
2707 /* tune up the balancing interval */
2708 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2709 (sd
->balance_interval
< sd
->max_interval
))
2710 sd
->balance_interval
*= 2;
2712 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2713 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2719 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2720 * tasks if there is an imbalance.
2722 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2723 * this_rq is locked.
2726 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2728 struct sched_group
*group
;
2729 struct rq
*busiest
= NULL
;
2730 unsigned long imbalance
;
2734 cpumask_t cpus
= CPU_MASK_ALL
;
2737 * When power savings policy is enabled for the parent domain, idle
2738 * sibling can pick up load irrespective of busy siblings. In this case,
2739 * let the state of idle sibling percolate up as IDLE, instead of
2740 * portraying it as CPU_NOT_IDLE.
2742 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2743 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2746 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2748 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2749 &sd_idle
, &cpus
, NULL
);
2751 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2755 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2758 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2762 BUG_ON(busiest
== this_rq
);
2764 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2767 if (busiest
->nr_running
> 1) {
2768 /* Attempt to move tasks */
2769 double_lock_balance(this_rq
, busiest
);
2770 /* this_rq->clock is already updated */
2771 update_rq_clock(busiest
);
2772 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2773 imbalance
, sd
, CPU_NEWLY_IDLE
,
2775 spin_unlock(&busiest
->lock
);
2777 if (unlikely(all_pinned
)) {
2778 cpu_clear(cpu_of(busiest
), cpus
);
2779 if (!cpus_empty(cpus
))
2785 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2786 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2787 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2790 sd
->nr_balance_failed
= 0;
2795 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2796 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2797 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2799 sd
->nr_balance_failed
= 0;
2805 * idle_balance is called by schedule() if this_cpu is about to become
2806 * idle. Attempts to pull tasks from other CPUs.
2808 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2810 struct sched_domain
*sd
;
2811 int pulled_task
= -1;
2812 unsigned long next_balance
= jiffies
+ HZ
;
2814 for_each_domain(this_cpu
, sd
) {
2815 unsigned long interval
;
2817 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2820 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2821 /* If we've pulled tasks over stop searching: */
2822 pulled_task
= load_balance_newidle(this_cpu
,
2825 interval
= msecs_to_jiffies(sd
->balance_interval
);
2826 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2827 next_balance
= sd
->last_balance
+ interval
;
2831 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
2833 * We are going idle. next_balance may be set based on
2834 * a busy processor. So reset next_balance.
2836 this_rq
->next_balance
= next_balance
;
2841 * active_load_balance is run by migration threads. It pushes running tasks
2842 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2843 * running on each physical CPU where possible, and avoids physical /
2844 * logical imbalances.
2846 * Called with busiest_rq locked.
2848 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2850 int target_cpu
= busiest_rq
->push_cpu
;
2851 struct sched_domain
*sd
;
2852 struct rq
*target_rq
;
2854 /* Is there any task to move? */
2855 if (busiest_rq
->nr_running
<= 1)
2858 target_rq
= cpu_rq(target_cpu
);
2861 * This condition is "impossible", if it occurs
2862 * we need to fix it. Originally reported by
2863 * Bjorn Helgaas on a 128-cpu setup.
2865 BUG_ON(busiest_rq
== target_rq
);
2867 /* move a task from busiest_rq to target_rq */
2868 double_lock_balance(busiest_rq
, target_rq
);
2869 update_rq_clock(busiest_rq
);
2870 update_rq_clock(target_rq
);
2872 /* Search for an sd spanning us and the target CPU. */
2873 for_each_domain(target_cpu
, sd
) {
2874 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2875 cpu_isset(busiest_cpu
, sd
->span
))
2880 schedstat_inc(sd
, alb_cnt
);
2882 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
2884 schedstat_inc(sd
, alb_pushed
);
2886 schedstat_inc(sd
, alb_failed
);
2888 spin_unlock(&target_rq
->lock
);
2893 atomic_t load_balancer
;
2895 } nohz ____cacheline_aligned
= {
2896 .load_balancer
= ATOMIC_INIT(-1),
2897 .cpu_mask
= CPU_MASK_NONE
,
2901 * This routine will try to nominate the ilb (idle load balancing)
2902 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2903 * load balancing on behalf of all those cpus. If all the cpus in the system
2904 * go into this tickless mode, then there will be no ilb owner (as there is
2905 * no need for one) and all the cpus will sleep till the next wakeup event
2908 * For the ilb owner, tick is not stopped. And this tick will be used
2909 * for idle load balancing. ilb owner will still be part of
2912 * While stopping the tick, this cpu will become the ilb owner if there
2913 * is no other owner. And will be the owner till that cpu becomes busy
2914 * or if all cpus in the system stop their ticks at which point
2915 * there is no need for ilb owner.
2917 * When the ilb owner becomes busy, it nominates another owner, during the
2918 * next busy scheduler_tick()
2920 int select_nohz_load_balancer(int stop_tick
)
2922 int cpu
= smp_processor_id();
2925 cpu_set(cpu
, nohz
.cpu_mask
);
2926 cpu_rq(cpu
)->in_nohz_recently
= 1;
2929 * If we are going offline and still the leader, give up!
2931 if (cpu_is_offline(cpu
) &&
2932 atomic_read(&nohz
.load_balancer
) == cpu
) {
2933 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2938 /* time for ilb owner also to sleep */
2939 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
2940 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2941 atomic_set(&nohz
.load_balancer
, -1);
2945 if (atomic_read(&nohz
.load_balancer
) == -1) {
2946 /* make me the ilb owner */
2947 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
2949 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
2952 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
2955 cpu_clear(cpu
, nohz
.cpu_mask
);
2957 if (atomic_read(&nohz
.load_balancer
) == cpu
)
2958 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
2965 static DEFINE_SPINLOCK(balancing
);
2968 * It checks each scheduling domain to see if it is due to be balanced,
2969 * and initiates a balancing operation if so.
2971 * Balancing parameters are set up in arch_init_sched_domains.
2973 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
2976 struct rq
*rq
= cpu_rq(cpu
);
2977 unsigned long interval
;
2978 struct sched_domain
*sd
;
2979 /* Earliest time when we have to do rebalance again */
2980 unsigned long next_balance
= jiffies
+ 60*HZ
;
2981 int update_next_balance
= 0;
2983 for_each_domain(cpu
, sd
) {
2984 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2987 interval
= sd
->balance_interval
;
2988 if (idle
!= CPU_IDLE
)
2989 interval
*= sd
->busy_factor
;
2991 /* scale ms to jiffies */
2992 interval
= msecs_to_jiffies(interval
);
2993 if (unlikely(!interval
))
2995 if (interval
> HZ
*NR_CPUS
/10)
2996 interval
= HZ
*NR_CPUS
/10;
2999 if (sd
->flags
& SD_SERIALIZE
) {
3000 if (!spin_trylock(&balancing
))
3004 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3005 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3007 * We've pulled tasks over so either we're no
3008 * longer idle, or one of our SMT siblings is
3011 idle
= CPU_NOT_IDLE
;
3013 sd
->last_balance
= jiffies
;
3015 if (sd
->flags
& SD_SERIALIZE
)
3016 spin_unlock(&balancing
);
3018 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3019 next_balance
= sd
->last_balance
+ interval
;
3020 update_next_balance
= 1;
3024 * Stop the load balance at this level. There is another
3025 * CPU in our sched group which is doing load balancing more
3033 * next_balance will be updated only when there is a need.
3034 * When the cpu is attached to null domain for ex, it will not be
3037 if (likely(update_next_balance
))
3038 rq
->next_balance
= next_balance
;
3042 * run_rebalance_domains is triggered when needed from the scheduler tick.
3043 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3044 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3046 static void run_rebalance_domains(struct softirq_action
*h
)
3048 int this_cpu
= smp_processor_id();
3049 struct rq
*this_rq
= cpu_rq(this_cpu
);
3050 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3051 CPU_IDLE
: CPU_NOT_IDLE
;
3053 rebalance_domains(this_cpu
, idle
);
3057 * If this cpu is the owner for idle load balancing, then do the
3058 * balancing on behalf of the other idle cpus whose ticks are
3061 if (this_rq
->idle_at_tick
&&
3062 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3063 cpumask_t cpus
= nohz
.cpu_mask
;
3067 cpu_clear(this_cpu
, cpus
);
3068 for_each_cpu_mask(balance_cpu
, cpus
) {
3070 * If this cpu gets work to do, stop the load balancing
3071 * work being done for other cpus. Next load
3072 * balancing owner will pick it up.
3077 rebalance_domains(balance_cpu
, CPU_IDLE
);
3079 rq
= cpu_rq(balance_cpu
);
3080 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3081 this_rq
->next_balance
= rq
->next_balance
;
3088 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3090 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3091 * idle load balancing owner or decide to stop the periodic load balancing,
3092 * if the whole system is idle.
3094 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3098 * If we were in the nohz mode recently and busy at the current
3099 * scheduler tick, then check if we need to nominate new idle
3102 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3103 rq
->in_nohz_recently
= 0;
3105 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3106 cpu_clear(cpu
, nohz
.cpu_mask
);
3107 atomic_set(&nohz
.load_balancer
, -1);
3110 if (atomic_read(&nohz
.load_balancer
) == -1) {
3112 * simple selection for now: Nominate the
3113 * first cpu in the nohz list to be the next
3116 * TBD: Traverse the sched domains and nominate
3117 * the nearest cpu in the nohz.cpu_mask.
3119 int ilb
= first_cpu(nohz
.cpu_mask
);
3127 * If this cpu is idle and doing idle load balancing for all the
3128 * cpus with ticks stopped, is it time for that to stop?
3130 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3131 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3137 * If this cpu is idle and the idle load balancing is done by
3138 * someone else, then no need raise the SCHED_SOFTIRQ
3140 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3141 cpu_isset(cpu
, nohz
.cpu_mask
))
3144 if (time_after_eq(jiffies
, rq
->next_balance
))
3145 raise_softirq(SCHED_SOFTIRQ
);
3148 #else /* CONFIG_SMP */
3151 * on UP we do not need to balance between CPUs:
3153 static inline void idle_balance(int cpu
, struct rq
*rq
)
3157 /* Avoid "used but not defined" warning on UP */
3158 static int balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3159 unsigned long max_nr_move
, unsigned long max_load_move
,
3160 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3161 int *all_pinned
, unsigned long *load_moved
,
3162 int *this_best_prio
, struct rq_iterator
*iterator
)
3171 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3173 EXPORT_PER_CPU_SYMBOL(kstat
);
3176 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3177 * that have not yet been banked in case the task is currently running.
3179 unsigned long long task_sched_runtime(struct task_struct
*p
)
3181 unsigned long flags
;
3185 rq
= task_rq_lock(p
, &flags
);
3186 ns
= p
->se
.sum_exec_runtime
;
3187 if (rq
->curr
== p
) {
3188 update_rq_clock(rq
);
3189 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3190 if ((s64
)delta_exec
> 0)
3193 task_rq_unlock(rq
, &flags
);
3199 * Account user cpu time to a process.
3200 * @p: the process that the cpu time gets accounted to
3201 * @hardirq_offset: the offset to subtract from hardirq_count()
3202 * @cputime: the cpu time spent in user space since the last update
3204 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3206 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3209 p
->utime
= cputime_add(p
->utime
, cputime
);
3211 /* Add user time to cpustat. */
3212 tmp
= cputime_to_cputime64(cputime
);
3213 if (TASK_NICE(p
) > 0)
3214 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3216 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3220 * Account system cpu time to a process.
3221 * @p: the process that the cpu time gets accounted to
3222 * @hardirq_offset: the offset to subtract from hardirq_count()
3223 * @cputime: the cpu time spent in kernel space since the last update
3225 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3228 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3229 struct rq
*rq
= this_rq();
3232 p
->stime
= cputime_add(p
->stime
, cputime
);
3234 /* Add system time to cpustat. */
3235 tmp
= cputime_to_cputime64(cputime
);
3236 if (hardirq_count() - hardirq_offset
)
3237 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3238 else if (softirq_count())
3239 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3240 else if (p
!= rq
->idle
)
3241 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3242 else if (atomic_read(&rq
->nr_iowait
) > 0)
3243 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3245 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3246 /* Account for system time used */
3247 acct_update_integrals(p
);
3251 * Account for involuntary wait time.
3252 * @p: the process from which the cpu time has been stolen
3253 * @steal: the cpu time spent in involuntary wait
3255 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3257 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3258 cputime64_t tmp
= cputime_to_cputime64(steal
);
3259 struct rq
*rq
= this_rq();
3261 if (p
== rq
->idle
) {
3262 p
->stime
= cputime_add(p
->stime
, steal
);
3263 if (atomic_read(&rq
->nr_iowait
) > 0)
3264 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3266 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3268 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3272 * This function gets called by the timer code, with HZ frequency.
3273 * We call it with interrupts disabled.
3275 * It also gets called by the fork code, when changing the parent's
3278 void scheduler_tick(void)
3280 int cpu
= smp_processor_id();
3281 struct rq
*rq
= cpu_rq(cpu
);
3282 struct task_struct
*curr
= rq
->curr
;
3283 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3285 spin_lock(&rq
->lock
);
3286 __update_rq_clock(rq
);
3288 * Let rq->clock advance by at least TICK_NSEC:
3290 if (unlikely(rq
->clock
< next_tick
))
3291 rq
->clock
= next_tick
;
3292 rq
->tick_timestamp
= rq
->clock
;
3293 update_cpu_load(rq
);
3294 if (curr
!= rq
->idle
) /* FIXME: needed? */
3295 curr
->sched_class
->task_tick(rq
, curr
);
3296 spin_unlock(&rq
->lock
);
3299 rq
->idle_at_tick
= idle_cpu(cpu
);
3300 trigger_load_balance(rq
, cpu
);
3304 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3306 void fastcall
add_preempt_count(int val
)
3311 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3313 preempt_count() += val
;
3315 * Spinlock count overflowing soon?
3317 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3320 EXPORT_SYMBOL(add_preempt_count
);
3322 void fastcall
sub_preempt_count(int val
)
3327 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3330 * Is the spinlock portion underflowing?
3332 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3333 !(preempt_count() & PREEMPT_MASK
)))
3336 preempt_count() -= val
;
3338 EXPORT_SYMBOL(sub_preempt_count
);
3343 * Print scheduling while atomic bug:
3345 static noinline
void __schedule_bug(struct task_struct
*prev
)
3347 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d\n",
3348 prev
->comm
, preempt_count(), prev
->pid
);
3349 debug_show_held_locks(prev
);
3350 if (irqs_disabled())
3351 print_irqtrace_events(prev
);
3356 * Various schedule()-time debugging checks and statistics:
3358 static inline void schedule_debug(struct task_struct
*prev
)
3361 * Test if we are atomic. Since do_exit() needs to call into
3362 * schedule() atomically, we ignore that path for now.
3363 * Otherwise, whine if we are scheduling when we should not be.
3365 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3366 __schedule_bug(prev
);
3368 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3370 schedstat_inc(this_rq(), sched_cnt
);
3374 * Pick up the highest-prio task:
3376 static inline struct task_struct
*
3377 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3379 struct sched_class
*class;
3380 struct task_struct
*p
;
3383 * Optimization: we know that if all tasks are in
3384 * the fair class we can call that function directly:
3386 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3387 p
= fair_sched_class
.pick_next_task(rq
);
3392 class = sched_class_highest
;
3394 p
= class->pick_next_task(rq
);
3398 * Will never be NULL as the idle class always
3399 * returns a non-NULL p:
3401 class = class->next
;
3406 * schedule() is the main scheduler function.
3408 asmlinkage
void __sched
schedule(void)
3410 struct task_struct
*prev
, *next
;
3417 cpu
= smp_processor_id();
3421 switch_count
= &prev
->nivcsw
;
3423 release_kernel_lock(prev
);
3424 need_resched_nonpreemptible
:
3426 schedule_debug(prev
);
3428 spin_lock_irq(&rq
->lock
);
3429 clear_tsk_need_resched(prev
);
3430 __update_rq_clock(rq
);
3432 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3433 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3434 unlikely(signal_pending(prev
)))) {
3435 prev
->state
= TASK_RUNNING
;
3437 deactivate_task(rq
, prev
, 1);
3439 switch_count
= &prev
->nvcsw
;
3442 if (unlikely(!rq
->nr_running
))
3443 idle_balance(cpu
, rq
);
3445 prev
->sched_class
->put_prev_task(rq
, prev
);
3446 next
= pick_next_task(rq
, prev
);
3448 sched_info_switch(prev
, next
);
3450 if (likely(prev
!= next
)) {
3455 context_switch(rq
, prev
, next
); /* unlocks the rq */
3457 spin_unlock_irq(&rq
->lock
);
3459 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3460 cpu
= smp_processor_id();
3462 goto need_resched_nonpreemptible
;
3464 preempt_enable_no_resched();
3465 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3468 EXPORT_SYMBOL(schedule
);
3470 #ifdef CONFIG_PREEMPT
3472 * this is the entry point to schedule() from in-kernel preemption
3473 * off of preempt_enable. Kernel preemptions off return from interrupt
3474 * occur there and call schedule directly.
3476 asmlinkage
void __sched
preempt_schedule(void)
3478 struct thread_info
*ti
= current_thread_info();
3479 #ifdef CONFIG_PREEMPT_BKL
3480 struct task_struct
*task
= current
;
3481 int saved_lock_depth
;
3484 * If there is a non-zero preempt_count or interrupts are disabled,
3485 * we do not want to preempt the current task. Just return..
3487 if (likely(ti
->preempt_count
|| irqs_disabled()))
3491 add_preempt_count(PREEMPT_ACTIVE
);
3493 * We keep the big kernel semaphore locked, but we
3494 * clear ->lock_depth so that schedule() doesnt
3495 * auto-release the semaphore:
3497 #ifdef CONFIG_PREEMPT_BKL
3498 saved_lock_depth
= task
->lock_depth
;
3499 task
->lock_depth
= -1;
3502 #ifdef CONFIG_PREEMPT_BKL
3503 task
->lock_depth
= saved_lock_depth
;
3505 sub_preempt_count(PREEMPT_ACTIVE
);
3507 /* we could miss a preemption opportunity between schedule and now */
3509 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3512 EXPORT_SYMBOL(preempt_schedule
);
3515 * this is the entry point to schedule() from kernel preemption
3516 * off of irq context.
3517 * Note, that this is called and return with irqs disabled. This will
3518 * protect us against recursive calling from irq.
3520 asmlinkage
void __sched
preempt_schedule_irq(void)
3522 struct thread_info
*ti
= current_thread_info();
3523 #ifdef CONFIG_PREEMPT_BKL
3524 struct task_struct
*task
= current
;
3525 int saved_lock_depth
;
3527 /* Catch callers which need to be fixed */
3528 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3531 add_preempt_count(PREEMPT_ACTIVE
);
3533 * We keep the big kernel semaphore locked, but we
3534 * clear ->lock_depth so that schedule() doesnt
3535 * auto-release the semaphore:
3537 #ifdef CONFIG_PREEMPT_BKL
3538 saved_lock_depth
= task
->lock_depth
;
3539 task
->lock_depth
= -1;
3543 local_irq_disable();
3544 #ifdef CONFIG_PREEMPT_BKL
3545 task
->lock_depth
= saved_lock_depth
;
3547 sub_preempt_count(PREEMPT_ACTIVE
);
3549 /* we could miss a preemption opportunity between schedule and now */
3551 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3555 #endif /* CONFIG_PREEMPT */
3557 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3560 return try_to_wake_up(curr
->private, mode
, sync
);
3562 EXPORT_SYMBOL(default_wake_function
);
3565 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3566 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3567 * number) then we wake all the non-exclusive tasks and one exclusive task.
3569 * There are circumstances in which we can try to wake a task which has already
3570 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3571 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3573 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3574 int nr_exclusive
, int sync
, void *key
)
3576 wait_queue_t
*curr
, *next
;
3578 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3579 unsigned flags
= curr
->flags
;
3581 if (curr
->func(curr
, mode
, sync
, key
) &&
3582 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3588 * __wake_up - wake up threads blocked on a waitqueue.
3590 * @mode: which threads
3591 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3592 * @key: is directly passed to the wakeup function
3594 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3595 int nr_exclusive
, void *key
)
3597 unsigned long flags
;
3599 spin_lock_irqsave(&q
->lock
, flags
);
3600 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3601 spin_unlock_irqrestore(&q
->lock
, flags
);
3603 EXPORT_SYMBOL(__wake_up
);
3606 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3608 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3610 __wake_up_common(q
, mode
, 1, 0, NULL
);
3614 * __wake_up_sync - wake up threads blocked on a waitqueue.
3616 * @mode: which threads
3617 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3619 * The sync wakeup differs that the waker knows that it will schedule
3620 * away soon, so while the target thread will be woken up, it will not
3621 * be migrated to another CPU - ie. the two threads are 'synchronized'
3622 * with each other. This can prevent needless bouncing between CPUs.
3624 * On UP it can prevent extra preemption.
3627 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3629 unsigned long flags
;
3635 if (unlikely(!nr_exclusive
))
3638 spin_lock_irqsave(&q
->lock
, flags
);
3639 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3640 spin_unlock_irqrestore(&q
->lock
, flags
);
3642 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3644 void fastcall
complete(struct completion
*x
)
3646 unsigned long flags
;
3648 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3650 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3652 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3654 EXPORT_SYMBOL(complete
);
3656 void fastcall
complete_all(struct completion
*x
)
3658 unsigned long flags
;
3660 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3661 x
->done
+= UINT_MAX
/2;
3662 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3664 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3666 EXPORT_SYMBOL(complete_all
);
3668 void fastcall __sched
wait_for_completion(struct completion
*x
)
3672 spin_lock_irq(&x
->wait
.lock
);
3674 DECLARE_WAITQUEUE(wait
, current
);
3676 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3677 __add_wait_queue_tail(&x
->wait
, &wait
);
3679 __set_current_state(TASK_UNINTERRUPTIBLE
);
3680 spin_unlock_irq(&x
->wait
.lock
);
3682 spin_lock_irq(&x
->wait
.lock
);
3684 __remove_wait_queue(&x
->wait
, &wait
);
3687 spin_unlock_irq(&x
->wait
.lock
);
3689 EXPORT_SYMBOL(wait_for_completion
);
3691 unsigned long fastcall __sched
3692 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3696 spin_lock_irq(&x
->wait
.lock
);
3698 DECLARE_WAITQUEUE(wait
, current
);
3700 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3701 __add_wait_queue_tail(&x
->wait
, &wait
);
3703 __set_current_state(TASK_UNINTERRUPTIBLE
);
3704 spin_unlock_irq(&x
->wait
.lock
);
3705 timeout
= schedule_timeout(timeout
);
3706 spin_lock_irq(&x
->wait
.lock
);
3708 __remove_wait_queue(&x
->wait
, &wait
);
3712 __remove_wait_queue(&x
->wait
, &wait
);
3716 spin_unlock_irq(&x
->wait
.lock
);
3719 EXPORT_SYMBOL(wait_for_completion_timeout
);
3721 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3727 spin_lock_irq(&x
->wait
.lock
);
3729 DECLARE_WAITQUEUE(wait
, current
);
3731 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3732 __add_wait_queue_tail(&x
->wait
, &wait
);
3734 if (signal_pending(current
)) {
3736 __remove_wait_queue(&x
->wait
, &wait
);
3739 __set_current_state(TASK_INTERRUPTIBLE
);
3740 spin_unlock_irq(&x
->wait
.lock
);
3742 spin_lock_irq(&x
->wait
.lock
);
3744 __remove_wait_queue(&x
->wait
, &wait
);
3748 spin_unlock_irq(&x
->wait
.lock
);
3752 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3754 unsigned long fastcall __sched
3755 wait_for_completion_interruptible_timeout(struct completion
*x
,
3756 unsigned long timeout
)
3760 spin_lock_irq(&x
->wait
.lock
);
3762 DECLARE_WAITQUEUE(wait
, current
);
3764 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3765 __add_wait_queue_tail(&x
->wait
, &wait
);
3767 if (signal_pending(current
)) {
3768 timeout
= -ERESTARTSYS
;
3769 __remove_wait_queue(&x
->wait
, &wait
);
3772 __set_current_state(TASK_INTERRUPTIBLE
);
3773 spin_unlock_irq(&x
->wait
.lock
);
3774 timeout
= schedule_timeout(timeout
);
3775 spin_lock_irq(&x
->wait
.lock
);
3777 __remove_wait_queue(&x
->wait
, &wait
);
3781 __remove_wait_queue(&x
->wait
, &wait
);
3785 spin_unlock_irq(&x
->wait
.lock
);
3788 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3791 sleep_on_head(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3793 spin_lock_irqsave(&q
->lock
, *flags
);
3794 __add_wait_queue(q
, wait
);
3795 spin_unlock(&q
->lock
);
3799 sleep_on_tail(wait_queue_head_t
*q
, wait_queue_t
*wait
, unsigned long *flags
)
3801 spin_lock_irq(&q
->lock
);
3802 __remove_wait_queue(q
, wait
);
3803 spin_unlock_irqrestore(&q
->lock
, *flags
);
3806 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3808 unsigned long flags
;
3811 init_waitqueue_entry(&wait
, current
);
3813 current
->state
= TASK_INTERRUPTIBLE
;
3815 sleep_on_head(q
, &wait
, &flags
);
3817 sleep_on_tail(q
, &wait
, &flags
);
3819 EXPORT_SYMBOL(interruptible_sleep_on
);
3822 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3824 unsigned long flags
;
3827 init_waitqueue_entry(&wait
, current
);
3829 current
->state
= TASK_INTERRUPTIBLE
;
3831 sleep_on_head(q
, &wait
, &flags
);
3832 timeout
= schedule_timeout(timeout
);
3833 sleep_on_tail(q
, &wait
, &flags
);
3837 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3839 void __sched
sleep_on(wait_queue_head_t
*q
)
3841 unsigned long flags
;
3844 init_waitqueue_entry(&wait
, current
);
3846 current
->state
= TASK_UNINTERRUPTIBLE
;
3848 sleep_on_head(q
, &wait
, &flags
);
3850 sleep_on_tail(q
, &wait
, &flags
);
3852 EXPORT_SYMBOL(sleep_on
);
3854 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3856 unsigned long flags
;
3859 init_waitqueue_entry(&wait
, current
);
3861 current
->state
= TASK_UNINTERRUPTIBLE
;
3863 sleep_on_head(q
, &wait
, &flags
);
3864 timeout
= schedule_timeout(timeout
);
3865 sleep_on_tail(q
, &wait
, &flags
);
3869 EXPORT_SYMBOL(sleep_on_timeout
);
3871 #ifdef CONFIG_RT_MUTEXES
3874 * rt_mutex_setprio - set the current priority of a task
3876 * @prio: prio value (kernel-internal form)
3878 * This function changes the 'effective' priority of a task. It does
3879 * not touch ->normal_prio like __setscheduler().
3881 * Used by the rt_mutex code to implement priority inheritance logic.
3883 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3885 unsigned long flags
;
3889 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3891 rq
= task_rq_lock(p
, &flags
);
3892 update_rq_clock(rq
);
3895 on_rq
= p
->se
.on_rq
;
3897 dequeue_task(rq
, p
, 0);
3900 p
->sched_class
= &rt_sched_class
;
3902 p
->sched_class
= &fair_sched_class
;
3907 enqueue_task(rq
, p
, 0);
3909 * Reschedule if we are currently running on this runqueue and
3910 * our priority decreased, or if we are not currently running on
3911 * this runqueue and our priority is higher than the current's
3913 if (task_running(rq
, p
)) {
3914 if (p
->prio
> oldprio
)
3915 resched_task(rq
->curr
);
3917 check_preempt_curr(rq
, p
);
3920 task_rq_unlock(rq
, &flags
);
3925 void set_user_nice(struct task_struct
*p
, long nice
)
3927 int old_prio
, delta
, on_rq
;
3928 unsigned long flags
;
3931 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3934 * We have to be careful, if called from sys_setpriority(),
3935 * the task might be in the middle of scheduling on another CPU.
3937 rq
= task_rq_lock(p
, &flags
);
3938 update_rq_clock(rq
);
3940 * The RT priorities are set via sched_setscheduler(), but we still
3941 * allow the 'normal' nice value to be set - but as expected
3942 * it wont have any effect on scheduling until the task is
3943 * SCHED_FIFO/SCHED_RR:
3945 if (task_has_rt_policy(p
)) {
3946 p
->static_prio
= NICE_TO_PRIO(nice
);
3949 on_rq
= p
->se
.on_rq
;
3951 dequeue_task(rq
, p
, 0);
3955 p
->static_prio
= NICE_TO_PRIO(nice
);
3958 p
->prio
= effective_prio(p
);
3959 delta
= p
->prio
- old_prio
;
3962 enqueue_task(rq
, p
, 0);
3965 * If the task increased its priority or is running and
3966 * lowered its priority, then reschedule its CPU:
3968 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3969 resched_task(rq
->curr
);
3972 task_rq_unlock(rq
, &flags
);
3974 EXPORT_SYMBOL(set_user_nice
);
3977 * can_nice - check if a task can reduce its nice value
3981 int can_nice(const struct task_struct
*p
, const int nice
)
3983 /* convert nice value [19,-20] to rlimit style value [1,40] */
3984 int nice_rlim
= 20 - nice
;
3986 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
3987 capable(CAP_SYS_NICE
));
3990 #ifdef __ARCH_WANT_SYS_NICE
3993 * sys_nice - change the priority of the current process.
3994 * @increment: priority increment
3996 * sys_setpriority is a more generic, but much slower function that
3997 * does similar things.
3999 asmlinkage
long sys_nice(int increment
)
4004 * Setpriority might change our priority at the same moment.
4005 * We don't have to worry. Conceptually one call occurs first
4006 * and we have a single winner.
4008 if (increment
< -40)
4013 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4019 if (increment
< 0 && !can_nice(current
, nice
))
4022 retval
= security_task_setnice(current
, nice
);
4026 set_user_nice(current
, nice
);
4033 * task_prio - return the priority value of a given task.
4034 * @p: the task in question.
4036 * This is the priority value as seen by users in /proc.
4037 * RT tasks are offset by -200. Normal tasks are centered
4038 * around 0, value goes from -16 to +15.
4040 int task_prio(const struct task_struct
*p
)
4042 return p
->prio
- MAX_RT_PRIO
;
4046 * task_nice - return the nice value of a given task.
4047 * @p: the task in question.
4049 int task_nice(const struct task_struct
*p
)
4051 return TASK_NICE(p
);
4053 EXPORT_SYMBOL_GPL(task_nice
);
4056 * idle_cpu - is a given cpu idle currently?
4057 * @cpu: the processor in question.
4059 int idle_cpu(int cpu
)
4061 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4065 * idle_task - return the idle task for a given cpu.
4066 * @cpu: the processor in question.
4068 struct task_struct
*idle_task(int cpu
)
4070 return cpu_rq(cpu
)->idle
;
4074 * find_process_by_pid - find a process with a matching PID value.
4075 * @pid: the pid in question.
4077 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4079 return pid
? find_task_by_pid(pid
) : current
;
4082 /* Actually do priority change: must hold rq lock. */
4084 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4086 BUG_ON(p
->se
.on_rq
);
4089 switch (p
->policy
) {
4093 p
->sched_class
= &fair_sched_class
;
4097 p
->sched_class
= &rt_sched_class
;
4101 p
->rt_priority
= prio
;
4102 p
->normal_prio
= normal_prio(p
);
4103 /* we are holding p->pi_lock already */
4104 p
->prio
= rt_mutex_getprio(p
);
4109 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4110 * @p: the task in question.
4111 * @policy: new policy.
4112 * @param: structure containing the new RT priority.
4114 * NOTE that the task may be already dead.
4116 int sched_setscheduler(struct task_struct
*p
, int policy
,
4117 struct sched_param
*param
)
4119 int retval
, oldprio
, oldpolicy
= -1, on_rq
;
4120 unsigned long flags
;
4123 /* may grab non-irq protected spin_locks */
4124 BUG_ON(in_interrupt());
4126 /* double check policy once rq lock held */
4128 policy
= oldpolicy
= p
->policy
;
4129 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4130 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4131 policy
!= SCHED_IDLE
)
4134 * Valid priorities for SCHED_FIFO and SCHED_RR are
4135 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4136 * SCHED_BATCH and SCHED_IDLE is 0.
4138 if (param
->sched_priority
< 0 ||
4139 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4140 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4142 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4146 * Allow unprivileged RT tasks to decrease priority:
4148 if (!capable(CAP_SYS_NICE
)) {
4149 if (rt_policy(policy
)) {
4150 unsigned long rlim_rtprio
;
4152 if (!lock_task_sighand(p
, &flags
))
4154 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4155 unlock_task_sighand(p
, &flags
);
4157 /* can't set/change the rt policy */
4158 if (policy
!= p
->policy
&& !rlim_rtprio
)
4161 /* can't increase priority */
4162 if (param
->sched_priority
> p
->rt_priority
&&
4163 param
->sched_priority
> rlim_rtprio
)
4167 * Like positive nice levels, dont allow tasks to
4168 * move out of SCHED_IDLE either:
4170 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4173 /* can't change other user's priorities */
4174 if ((current
->euid
!= p
->euid
) &&
4175 (current
->euid
!= p
->uid
))
4179 retval
= security_task_setscheduler(p
, policy
, param
);
4183 * make sure no PI-waiters arrive (or leave) while we are
4184 * changing the priority of the task:
4186 spin_lock_irqsave(&p
->pi_lock
, flags
);
4188 * To be able to change p->policy safely, the apropriate
4189 * runqueue lock must be held.
4191 rq
= __task_rq_lock(p
);
4192 /* recheck policy now with rq lock held */
4193 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4194 policy
= oldpolicy
= -1;
4195 __task_rq_unlock(rq
);
4196 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4199 update_rq_clock(rq
);
4200 on_rq
= p
->se
.on_rq
;
4202 deactivate_task(rq
, p
, 0);
4204 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4206 activate_task(rq
, p
, 0);
4208 * Reschedule if we are currently running on this runqueue and
4209 * our priority decreased, or if we are not currently running on
4210 * this runqueue and our priority is higher than the current's
4212 if (task_running(rq
, p
)) {
4213 if (p
->prio
> oldprio
)
4214 resched_task(rq
->curr
);
4216 check_preempt_curr(rq
, p
);
4219 __task_rq_unlock(rq
);
4220 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4222 rt_mutex_adjust_pi(p
);
4226 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4229 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4231 struct sched_param lparam
;
4232 struct task_struct
*p
;
4235 if (!param
|| pid
< 0)
4237 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4242 p
= find_process_by_pid(pid
);
4244 retval
= sched_setscheduler(p
, policy
, &lparam
);
4251 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4252 * @pid: the pid in question.
4253 * @policy: new policy.
4254 * @param: structure containing the new RT priority.
4256 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4257 struct sched_param __user
*param
)
4259 /* negative values for policy are not valid */
4263 return do_sched_setscheduler(pid
, policy
, param
);
4267 * sys_sched_setparam - set/change the RT priority of a thread
4268 * @pid: the pid in question.
4269 * @param: structure containing the new RT priority.
4271 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4273 return do_sched_setscheduler(pid
, -1, param
);
4277 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4278 * @pid: the pid in question.
4280 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4282 struct task_struct
*p
;
4283 int retval
= -EINVAL
;
4289 read_lock(&tasklist_lock
);
4290 p
= find_process_by_pid(pid
);
4292 retval
= security_task_getscheduler(p
);
4296 read_unlock(&tasklist_lock
);
4303 * sys_sched_getscheduler - get the RT priority of a thread
4304 * @pid: the pid in question.
4305 * @param: structure containing the RT priority.
4307 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4309 struct sched_param lp
;
4310 struct task_struct
*p
;
4311 int retval
= -EINVAL
;
4313 if (!param
|| pid
< 0)
4316 read_lock(&tasklist_lock
);
4317 p
= find_process_by_pid(pid
);
4322 retval
= security_task_getscheduler(p
);
4326 lp
.sched_priority
= p
->rt_priority
;
4327 read_unlock(&tasklist_lock
);
4330 * This one might sleep, we cannot do it with a spinlock held ...
4332 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4338 read_unlock(&tasklist_lock
);
4342 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4344 cpumask_t cpus_allowed
;
4345 struct task_struct
*p
;
4348 mutex_lock(&sched_hotcpu_mutex
);
4349 read_lock(&tasklist_lock
);
4351 p
= find_process_by_pid(pid
);
4353 read_unlock(&tasklist_lock
);
4354 mutex_unlock(&sched_hotcpu_mutex
);
4359 * It is not safe to call set_cpus_allowed with the
4360 * tasklist_lock held. We will bump the task_struct's
4361 * usage count and then drop tasklist_lock.
4364 read_unlock(&tasklist_lock
);
4367 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4368 !capable(CAP_SYS_NICE
))
4371 retval
= security_task_setscheduler(p
, 0, NULL
);
4375 cpus_allowed
= cpuset_cpus_allowed(p
);
4376 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4377 retval
= set_cpus_allowed(p
, new_mask
);
4381 mutex_unlock(&sched_hotcpu_mutex
);
4385 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4386 cpumask_t
*new_mask
)
4388 if (len
< sizeof(cpumask_t
)) {
4389 memset(new_mask
, 0, sizeof(cpumask_t
));
4390 } else if (len
> sizeof(cpumask_t
)) {
4391 len
= sizeof(cpumask_t
);
4393 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4397 * sys_sched_setaffinity - set the cpu affinity of a process
4398 * @pid: pid of the process
4399 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4400 * @user_mask_ptr: user-space pointer to the new cpu mask
4402 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4403 unsigned long __user
*user_mask_ptr
)
4408 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4412 return sched_setaffinity(pid
, new_mask
);
4416 * Represents all cpu's present in the system
4417 * In systems capable of hotplug, this map could dynamically grow
4418 * as new cpu's are detected in the system via any platform specific
4419 * method, such as ACPI for e.g.
4422 cpumask_t cpu_present_map __read_mostly
;
4423 EXPORT_SYMBOL(cpu_present_map
);
4426 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4427 EXPORT_SYMBOL(cpu_online_map
);
4429 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4430 EXPORT_SYMBOL(cpu_possible_map
);
4433 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4435 struct task_struct
*p
;
4438 mutex_lock(&sched_hotcpu_mutex
);
4439 read_lock(&tasklist_lock
);
4442 p
= find_process_by_pid(pid
);
4446 retval
= security_task_getscheduler(p
);
4450 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4453 read_unlock(&tasklist_lock
);
4454 mutex_unlock(&sched_hotcpu_mutex
);
4460 * sys_sched_getaffinity - get the cpu affinity of a process
4461 * @pid: pid of the process
4462 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4463 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4465 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4466 unsigned long __user
*user_mask_ptr
)
4471 if (len
< sizeof(cpumask_t
))
4474 ret
= sched_getaffinity(pid
, &mask
);
4478 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4481 return sizeof(cpumask_t
);
4485 * sys_sched_yield - yield the current processor to other threads.
4487 * This function yields the current CPU to other tasks. If there are no
4488 * other threads running on this CPU then this function will return.
4490 asmlinkage
long sys_sched_yield(void)
4492 struct rq
*rq
= this_rq_lock();
4494 schedstat_inc(rq
, yld_cnt
);
4495 current
->sched_class
->yield_task(rq
, current
);
4498 * Since we are going to call schedule() anyway, there's
4499 * no need to preempt or enable interrupts:
4501 __release(rq
->lock
);
4502 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4503 _raw_spin_unlock(&rq
->lock
);
4504 preempt_enable_no_resched();
4511 static void __cond_resched(void)
4513 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4514 __might_sleep(__FILE__
, __LINE__
);
4517 * The BKS might be reacquired before we have dropped
4518 * PREEMPT_ACTIVE, which could trigger a second
4519 * cond_resched() call.
4522 add_preempt_count(PREEMPT_ACTIVE
);
4524 sub_preempt_count(PREEMPT_ACTIVE
);
4525 } while (need_resched());
4528 int __sched
cond_resched(void)
4530 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4531 system_state
== SYSTEM_RUNNING
) {
4537 EXPORT_SYMBOL(cond_resched
);
4540 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4541 * call schedule, and on return reacquire the lock.
4543 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4544 * operations here to prevent schedule() from being called twice (once via
4545 * spin_unlock(), once by hand).
4547 int cond_resched_lock(spinlock_t
*lock
)
4551 if (need_lockbreak(lock
)) {
4557 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4558 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4559 _raw_spin_unlock(lock
);
4560 preempt_enable_no_resched();
4567 EXPORT_SYMBOL(cond_resched_lock
);
4569 int __sched
cond_resched_softirq(void)
4571 BUG_ON(!in_softirq());
4573 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4581 EXPORT_SYMBOL(cond_resched_softirq
);
4584 * yield - yield the current processor to other threads.
4586 * This is a shortcut for kernel-space yielding - it marks the
4587 * thread runnable and calls sys_sched_yield().
4589 void __sched
yield(void)
4591 set_current_state(TASK_RUNNING
);
4594 EXPORT_SYMBOL(yield
);
4597 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4598 * that process accounting knows that this is a task in IO wait state.
4600 * But don't do that if it is a deliberate, throttling IO wait (this task
4601 * has set its backing_dev_info: the queue against which it should throttle)
4603 void __sched
io_schedule(void)
4605 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4607 delayacct_blkio_start();
4608 atomic_inc(&rq
->nr_iowait
);
4610 atomic_dec(&rq
->nr_iowait
);
4611 delayacct_blkio_end();
4613 EXPORT_SYMBOL(io_schedule
);
4615 long __sched
io_schedule_timeout(long timeout
)
4617 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4620 delayacct_blkio_start();
4621 atomic_inc(&rq
->nr_iowait
);
4622 ret
= schedule_timeout(timeout
);
4623 atomic_dec(&rq
->nr_iowait
);
4624 delayacct_blkio_end();
4629 * sys_sched_get_priority_max - return maximum RT priority.
4630 * @policy: scheduling class.
4632 * this syscall returns the maximum rt_priority that can be used
4633 * by a given scheduling class.
4635 asmlinkage
long sys_sched_get_priority_max(int policy
)
4642 ret
= MAX_USER_RT_PRIO
-1;
4654 * sys_sched_get_priority_min - return minimum RT priority.
4655 * @policy: scheduling class.
4657 * this syscall returns the minimum rt_priority that can be used
4658 * by a given scheduling class.
4660 asmlinkage
long sys_sched_get_priority_min(int policy
)
4678 * sys_sched_rr_get_interval - return the default timeslice of a process.
4679 * @pid: pid of the process.
4680 * @interval: userspace pointer to the timeslice value.
4682 * this syscall writes the default timeslice value of a given process
4683 * into the user-space timespec buffer. A value of '0' means infinity.
4686 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4688 struct task_struct
*p
;
4689 int retval
= -EINVAL
;
4696 read_lock(&tasklist_lock
);
4697 p
= find_process_by_pid(pid
);
4701 retval
= security_task_getscheduler(p
);
4705 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4706 0 : static_prio_timeslice(p
->static_prio
), &t
);
4707 read_unlock(&tasklist_lock
);
4708 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4712 read_unlock(&tasklist_lock
);
4716 static const char stat_nam
[] = "RSDTtZX";
4718 static void show_task(struct task_struct
*p
)
4720 unsigned long free
= 0;
4723 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4724 printk("%-13.13s %c", p
->comm
,
4725 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4726 #if BITS_PER_LONG == 32
4727 if (state
== TASK_RUNNING
)
4728 printk(" running ");
4730 printk(" %08lx ", thread_saved_pc(p
));
4732 if (state
== TASK_RUNNING
)
4733 printk(" running task ");
4735 printk(" %016lx ", thread_saved_pc(p
));
4737 #ifdef CONFIG_DEBUG_STACK_USAGE
4739 unsigned long *n
= end_of_stack(p
);
4742 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4745 printk("%5lu %5d %6d\n", free
, p
->pid
, p
->parent
->pid
);
4747 if (state
!= TASK_RUNNING
)
4748 show_stack(p
, NULL
);
4751 void show_state_filter(unsigned long state_filter
)
4753 struct task_struct
*g
, *p
;
4755 #if BITS_PER_LONG == 32
4757 " task PC stack pid father\n");
4760 " task PC stack pid father\n");
4762 read_lock(&tasklist_lock
);
4763 do_each_thread(g
, p
) {
4765 * reset the NMI-timeout, listing all files on a slow
4766 * console might take alot of time:
4768 touch_nmi_watchdog();
4769 if (!state_filter
|| (p
->state
& state_filter
))
4771 } while_each_thread(g
, p
);
4773 touch_all_softlockup_watchdogs();
4775 #ifdef CONFIG_SCHED_DEBUG
4776 sysrq_sched_debug_show();
4778 read_unlock(&tasklist_lock
);
4780 * Only show locks if all tasks are dumped:
4782 if (state_filter
== -1)
4783 debug_show_all_locks();
4786 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4788 idle
->sched_class
= &idle_sched_class
;
4792 * init_idle - set up an idle thread for a given CPU
4793 * @idle: task in question
4794 * @cpu: cpu the idle task belongs to
4796 * NOTE: this function does not set the idle thread's NEED_RESCHED
4797 * flag, to make booting more robust.
4799 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4801 struct rq
*rq
= cpu_rq(cpu
);
4802 unsigned long flags
;
4805 idle
->se
.exec_start
= sched_clock();
4807 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4808 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4809 __set_task_cpu(idle
, cpu
);
4811 spin_lock_irqsave(&rq
->lock
, flags
);
4812 rq
->curr
= rq
->idle
= idle
;
4813 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4816 spin_unlock_irqrestore(&rq
->lock
, flags
);
4818 /* Set the preempt count _outside_ the spinlocks! */
4819 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4820 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4822 task_thread_info(idle
)->preempt_count
= 0;
4825 * The idle tasks have their own, simple scheduling class:
4827 idle
->sched_class
= &idle_sched_class
;
4831 * In a system that switches off the HZ timer nohz_cpu_mask
4832 * indicates which cpus entered this state. This is used
4833 * in the rcu update to wait only for active cpus. For system
4834 * which do not switch off the HZ timer nohz_cpu_mask should
4835 * always be CPU_MASK_NONE.
4837 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4841 * This is how migration works:
4843 * 1) we queue a struct migration_req structure in the source CPU's
4844 * runqueue and wake up that CPU's migration thread.
4845 * 2) we down() the locked semaphore => thread blocks.
4846 * 3) migration thread wakes up (implicitly it forces the migrated
4847 * thread off the CPU)
4848 * 4) it gets the migration request and checks whether the migrated
4849 * task is still in the wrong runqueue.
4850 * 5) if it's in the wrong runqueue then the migration thread removes
4851 * it and puts it into the right queue.
4852 * 6) migration thread up()s the semaphore.
4853 * 7) we wake up and the migration is done.
4857 * Change a given task's CPU affinity. Migrate the thread to a
4858 * proper CPU and schedule it away if the CPU it's executing on
4859 * is removed from the allowed bitmask.
4861 * NOTE: the caller must have a valid reference to the task, the
4862 * task must not exit() & deallocate itself prematurely. The
4863 * call is not atomic; no spinlocks may be held.
4865 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
4867 struct migration_req req
;
4868 unsigned long flags
;
4872 rq
= task_rq_lock(p
, &flags
);
4873 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
4878 p
->cpus_allowed
= new_mask
;
4879 /* Can the task run on the task's current CPU? If so, we're done */
4880 if (cpu_isset(task_cpu(p
), new_mask
))
4883 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
4884 /* Need help from migration thread: drop lock and wait. */
4885 task_rq_unlock(rq
, &flags
);
4886 wake_up_process(rq
->migration_thread
);
4887 wait_for_completion(&req
.done
);
4888 tlb_migrate_finish(p
->mm
);
4892 task_rq_unlock(rq
, &flags
);
4896 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
4899 * Move (not current) task off this cpu, onto dest cpu. We're doing
4900 * this because either it can't run here any more (set_cpus_allowed()
4901 * away from this CPU, or CPU going down), or because we're
4902 * attempting to rebalance this task on exec (sched_exec).
4904 * So we race with normal scheduler movements, but that's OK, as long
4905 * as the task is no longer on this CPU.
4907 * Returns non-zero if task was successfully migrated.
4909 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4911 struct rq
*rq_dest
, *rq_src
;
4914 if (unlikely(cpu_is_offline(dest_cpu
)))
4917 rq_src
= cpu_rq(src_cpu
);
4918 rq_dest
= cpu_rq(dest_cpu
);
4920 double_rq_lock(rq_src
, rq_dest
);
4921 /* Already moved. */
4922 if (task_cpu(p
) != src_cpu
)
4924 /* Affinity changed (again). */
4925 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
4928 on_rq
= p
->se
.on_rq
;
4930 deactivate_task(rq_src
, p
, 0);
4932 set_task_cpu(p
, dest_cpu
);
4934 activate_task(rq_dest
, p
, 0);
4935 check_preempt_curr(rq_dest
, p
);
4939 double_rq_unlock(rq_src
, rq_dest
);
4944 * migration_thread - this is a highprio system thread that performs
4945 * thread migration by bumping thread off CPU then 'pushing' onto
4948 static int migration_thread(void *data
)
4950 int cpu
= (long)data
;
4954 BUG_ON(rq
->migration_thread
!= current
);
4956 set_current_state(TASK_INTERRUPTIBLE
);
4957 while (!kthread_should_stop()) {
4958 struct migration_req
*req
;
4959 struct list_head
*head
;
4961 spin_lock_irq(&rq
->lock
);
4963 if (cpu_is_offline(cpu
)) {
4964 spin_unlock_irq(&rq
->lock
);
4968 if (rq
->active_balance
) {
4969 active_load_balance(rq
, cpu
);
4970 rq
->active_balance
= 0;
4973 head
= &rq
->migration_queue
;
4975 if (list_empty(head
)) {
4976 spin_unlock_irq(&rq
->lock
);
4978 set_current_state(TASK_INTERRUPTIBLE
);
4981 req
= list_entry(head
->next
, struct migration_req
, list
);
4982 list_del_init(head
->next
);
4984 spin_unlock(&rq
->lock
);
4985 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
4988 complete(&req
->done
);
4990 __set_current_state(TASK_RUNNING
);
4994 /* Wait for kthread_stop */
4995 set_current_state(TASK_INTERRUPTIBLE
);
4996 while (!kthread_should_stop()) {
4998 set_current_state(TASK_INTERRUPTIBLE
);
5000 __set_current_state(TASK_RUNNING
);
5004 #ifdef CONFIG_HOTPLUG_CPU
5006 * Figure out where task on dead CPU should go, use force if neccessary.
5007 * NOTE: interrupts should be disabled by the caller
5009 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5011 unsigned long flags
;
5018 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5019 cpus_and(mask
, mask
, p
->cpus_allowed
);
5020 dest_cpu
= any_online_cpu(mask
);
5022 /* On any allowed CPU? */
5023 if (dest_cpu
== NR_CPUS
)
5024 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5026 /* No more Mr. Nice Guy. */
5027 if (dest_cpu
== NR_CPUS
) {
5028 rq
= task_rq_lock(p
, &flags
);
5029 cpus_setall(p
->cpus_allowed
);
5030 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5031 task_rq_unlock(rq
, &flags
);
5034 * Don't tell them about moving exiting tasks or
5035 * kernel threads (both mm NULL), since they never
5038 if (p
->mm
&& printk_ratelimit())
5039 printk(KERN_INFO
"process %d (%s) no "
5040 "longer affine to cpu%d\n",
5041 p
->pid
, p
->comm
, dead_cpu
);
5043 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5048 * While a dead CPU has no uninterruptible tasks queued at this point,
5049 * it might still have a nonzero ->nr_uninterruptible counter, because
5050 * for performance reasons the counter is not stricly tracking tasks to
5051 * their home CPUs. So we just add the counter to another CPU's counter,
5052 * to keep the global sum constant after CPU-down:
5054 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5056 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5057 unsigned long flags
;
5059 local_irq_save(flags
);
5060 double_rq_lock(rq_src
, rq_dest
);
5061 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5062 rq_src
->nr_uninterruptible
= 0;
5063 double_rq_unlock(rq_src
, rq_dest
);
5064 local_irq_restore(flags
);
5067 /* Run through task list and migrate tasks from the dead cpu. */
5068 static void migrate_live_tasks(int src_cpu
)
5070 struct task_struct
*p
, *t
;
5072 write_lock_irq(&tasklist_lock
);
5074 do_each_thread(t
, p
) {
5078 if (task_cpu(p
) == src_cpu
)
5079 move_task_off_dead_cpu(src_cpu
, p
);
5080 } while_each_thread(t
, p
);
5082 write_unlock_irq(&tasklist_lock
);
5086 * Schedules idle task to be the next runnable task on current CPU.
5087 * It does so by boosting its priority to highest possible and adding it to
5088 * the _front_ of the runqueue. Used by CPU offline code.
5090 void sched_idle_next(void)
5092 int this_cpu
= smp_processor_id();
5093 struct rq
*rq
= cpu_rq(this_cpu
);
5094 struct task_struct
*p
= rq
->idle
;
5095 unsigned long flags
;
5097 /* cpu has to be offline */
5098 BUG_ON(cpu_online(this_cpu
));
5101 * Strictly not necessary since rest of the CPUs are stopped by now
5102 * and interrupts disabled on the current cpu.
5104 spin_lock_irqsave(&rq
->lock
, flags
);
5106 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5108 /* Add idle task to the _front_ of its priority queue: */
5109 activate_idle_task(p
, rq
);
5111 spin_unlock_irqrestore(&rq
->lock
, flags
);
5115 * Ensures that the idle task is using init_mm right before its cpu goes
5118 void idle_task_exit(void)
5120 struct mm_struct
*mm
= current
->active_mm
;
5122 BUG_ON(cpu_online(smp_processor_id()));
5125 switch_mm(mm
, &init_mm
, current
);
5129 /* called under rq->lock with disabled interrupts */
5130 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5132 struct rq
*rq
= cpu_rq(dead_cpu
);
5134 /* Must be exiting, otherwise would be on tasklist. */
5135 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5137 /* Cannot have done final schedule yet: would have vanished. */
5138 BUG_ON(p
->state
== TASK_DEAD
);
5143 * Drop lock around migration; if someone else moves it,
5144 * that's OK. No task can be added to this CPU, so iteration is
5146 * NOTE: interrupts should be left disabled --dev@
5148 spin_unlock(&rq
->lock
);
5149 move_task_off_dead_cpu(dead_cpu
, p
);
5150 spin_lock(&rq
->lock
);
5155 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5156 static void migrate_dead_tasks(unsigned int dead_cpu
)
5158 struct rq
*rq
= cpu_rq(dead_cpu
);
5159 struct task_struct
*next
;
5162 if (!rq
->nr_running
)
5164 update_rq_clock(rq
);
5165 next
= pick_next_task(rq
, rq
->curr
);
5168 migrate_dead(dead_cpu
, next
);
5172 #endif /* CONFIG_HOTPLUG_CPU */
5174 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5176 static struct ctl_table sd_ctl_dir
[] = {
5178 .procname
= "sched_domain",
5184 static struct ctl_table sd_ctl_root
[] = {
5186 .ctl_name
= CTL_KERN
,
5187 .procname
= "kernel",
5189 .child
= sd_ctl_dir
,
5194 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5196 struct ctl_table
*entry
=
5197 kmalloc(n
* sizeof(struct ctl_table
), GFP_KERNEL
);
5200 memset(entry
, 0, n
* sizeof(struct ctl_table
));
5206 set_table_entry(struct ctl_table
*entry
,
5207 const char *procname
, void *data
, int maxlen
,
5208 mode_t mode
, proc_handler
*proc_handler
)
5210 entry
->procname
= procname
;
5212 entry
->maxlen
= maxlen
;
5214 entry
->proc_handler
= proc_handler
;
5217 static struct ctl_table
*
5218 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5220 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5222 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5223 sizeof(long), 0644, proc_doulongvec_minmax
);
5224 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5225 sizeof(long), 0644, proc_doulongvec_minmax
);
5226 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5227 sizeof(int), 0644, proc_dointvec_minmax
);
5228 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5229 sizeof(int), 0644, proc_dointvec_minmax
);
5230 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5231 sizeof(int), 0644, proc_dointvec_minmax
);
5232 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5233 sizeof(int), 0644, proc_dointvec_minmax
);
5234 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5235 sizeof(int), 0644, proc_dointvec_minmax
);
5236 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5237 sizeof(int), 0644, proc_dointvec_minmax
);
5238 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5239 sizeof(int), 0644, proc_dointvec_minmax
);
5240 set_table_entry(&table
[10], "cache_nice_tries",
5241 &sd
->cache_nice_tries
,
5242 sizeof(int), 0644, proc_dointvec_minmax
);
5243 set_table_entry(&table
[12], "flags", &sd
->flags
,
5244 sizeof(int), 0644, proc_dointvec_minmax
);
5249 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5251 struct ctl_table
*entry
, *table
;
5252 struct sched_domain
*sd
;
5253 int domain_num
= 0, i
;
5256 for_each_domain(cpu
, sd
)
5258 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5261 for_each_domain(cpu
, sd
) {
5262 snprintf(buf
, 32, "domain%d", i
);
5263 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5265 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5272 static struct ctl_table_header
*sd_sysctl_header
;
5273 static void init_sched_domain_sysctl(void)
5275 int i
, cpu_num
= num_online_cpus();
5276 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5279 sd_ctl_dir
[0].child
= entry
;
5281 for (i
= 0; i
< cpu_num
; i
++, entry
++) {
5282 snprintf(buf
, 32, "cpu%d", i
);
5283 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5285 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5287 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5290 static void init_sched_domain_sysctl(void)
5296 * migration_call - callback that gets triggered when a CPU is added.
5297 * Here we can start up the necessary migration thread for the new CPU.
5299 static int __cpuinit
5300 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5302 struct task_struct
*p
;
5303 int cpu
= (long)hcpu
;
5304 unsigned long flags
;
5308 case CPU_LOCK_ACQUIRE
:
5309 mutex_lock(&sched_hotcpu_mutex
);
5312 case CPU_UP_PREPARE
:
5313 case CPU_UP_PREPARE_FROZEN
:
5314 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5317 kthread_bind(p
, cpu
);
5318 /* Must be high prio: stop_machine expects to yield to it. */
5319 rq
= task_rq_lock(p
, &flags
);
5320 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5321 task_rq_unlock(rq
, &flags
);
5322 cpu_rq(cpu
)->migration_thread
= p
;
5326 case CPU_ONLINE_FROZEN
:
5327 /* Strictly unneccessary, as first user will wake it. */
5328 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5331 #ifdef CONFIG_HOTPLUG_CPU
5332 case CPU_UP_CANCELED
:
5333 case CPU_UP_CANCELED_FROZEN
:
5334 if (!cpu_rq(cpu
)->migration_thread
)
5336 /* Unbind it from offline cpu so it can run. Fall thru. */
5337 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5338 any_online_cpu(cpu_online_map
));
5339 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5340 cpu_rq(cpu
)->migration_thread
= NULL
;
5344 case CPU_DEAD_FROZEN
:
5345 migrate_live_tasks(cpu
);
5347 kthread_stop(rq
->migration_thread
);
5348 rq
->migration_thread
= NULL
;
5349 /* Idle task back to normal (off runqueue, low prio) */
5350 rq
= task_rq_lock(rq
->idle
, &flags
);
5351 update_rq_clock(rq
);
5352 deactivate_task(rq
, rq
->idle
, 0);
5353 rq
->idle
->static_prio
= MAX_PRIO
;
5354 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5355 rq
->idle
->sched_class
= &idle_sched_class
;
5356 migrate_dead_tasks(cpu
);
5357 task_rq_unlock(rq
, &flags
);
5358 migrate_nr_uninterruptible(rq
);
5359 BUG_ON(rq
->nr_running
!= 0);
5361 /* No need to migrate the tasks: it was best-effort if
5362 * they didn't take sched_hotcpu_mutex. Just wake up
5363 * the requestors. */
5364 spin_lock_irq(&rq
->lock
);
5365 while (!list_empty(&rq
->migration_queue
)) {
5366 struct migration_req
*req
;
5368 req
= list_entry(rq
->migration_queue
.next
,
5369 struct migration_req
, list
);
5370 list_del_init(&req
->list
);
5371 complete(&req
->done
);
5373 spin_unlock_irq(&rq
->lock
);
5376 case CPU_LOCK_RELEASE
:
5377 mutex_unlock(&sched_hotcpu_mutex
);
5383 /* Register at highest priority so that task migration (migrate_all_tasks)
5384 * happens before everything else.
5386 static struct notifier_block __cpuinitdata migration_notifier
= {
5387 .notifier_call
= migration_call
,
5391 int __init
migration_init(void)
5393 void *cpu
= (void *)(long)smp_processor_id();
5396 /* Start one for the boot CPU: */
5397 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5398 BUG_ON(err
== NOTIFY_BAD
);
5399 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5400 register_cpu_notifier(&migration_notifier
);
5408 /* Number of possible processor ids */
5409 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5410 EXPORT_SYMBOL(nr_cpu_ids
);
5412 #undef SCHED_DOMAIN_DEBUG
5413 #ifdef SCHED_DOMAIN_DEBUG
5414 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5419 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5423 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5428 struct sched_group
*group
= sd
->groups
;
5429 cpumask_t groupmask
;
5431 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5432 cpus_clear(groupmask
);
5435 for (i
= 0; i
< level
+ 1; i
++)
5437 printk("domain %d: ", level
);
5439 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5440 printk("does not load-balance\n");
5442 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5447 printk("span %s\n", str
);
5449 if (!cpu_isset(cpu
, sd
->span
))
5450 printk(KERN_ERR
"ERROR: domain->span does not contain "
5452 if (!cpu_isset(cpu
, group
->cpumask
))
5453 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5457 for (i
= 0; i
< level
+ 2; i
++)
5463 printk(KERN_ERR
"ERROR: group is NULL\n");
5467 if (!group
->__cpu_power
) {
5469 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5473 if (!cpus_weight(group
->cpumask
)) {
5475 printk(KERN_ERR
"ERROR: empty group\n");
5478 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5480 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5483 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5485 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5488 group
= group
->next
;
5489 } while (group
!= sd
->groups
);
5492 if (!cpus_equal(sd
->span
, groupmask
))
5493 printk(KERN_ERR
"ERROR: groups don't span "
5501 if (!cpus_subset(groupmask
, sd
->span
))
5502 printk(KERN_ERR
"ERROR: parent span is not a superset "
5503 "of domain->span\n");
5508 # define sched_domain_debug(sd, cpu) do { } while (0)
5511 static int sd_degenerate(struct sched_domain
*sd
)
5513 if (cpus_weight(sd
->span
) == 1)
5516 /* Following flags need at least 2 groups */
5517 if (sd
->flags
& (SD_LOAD_BALANCE
|
5518 SD_BALANCE_NEWIDLE
|
5522 SD_SHARE_PKG_RESOURCES
)) {
5523 if (sd
->groups
!= sd
->groups
->next
)
5527 /* Following flags don't use groups */
5528 if (sd
->flags
& (SD_WAKE_IDLE
|
5537 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5539 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5541 if (sd_degenerate(parent
))
5544 if (!cpus_equal(sd
->span
, parent
->span
))
5547 /* Does parent contain flags not in child? */
5548 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5549 if (cflags
& SD_WAKE_AFFINE
)
5550 pflags
&= ~SD_WAKE_BALANCE
;
5551 /* Flags needing groups don't count if only 1 group in parent */
5552 if (parent
->groups
== parent
->groups
->next
) {
5553 pflags
&= ~(SD_LOAD_BALANCE
|
5554 SD_BALANCE_NEWIDLE
|
5558 SD_SHARE_PKG_RESOURCES
);
5560 if (~cflags
& pflags
)
5567 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5568 * hold the hotplug lock.
5570 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5572 struct rq
*rq
= cpu_rq(cpu
);
5573 struct sched_domain
*tmp
;
5575 /* Remove the sched domains which do not contribute to scheduling. */
5576 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5577 struct sched_domain
*parent
= tmp
->parent
;
5580 if (sd_parent_degenerate(tmp
, parent
)) {
5581 tmp
->parent
= parent
->parent
;
5583 parent
->parent
->child
= tmp
;
5587 if (sd
&& sd_degenerate(sd
)) {
5593 sched_domain_debug(sd
, cpu
);
5595 rcu_assign_pointer(rq
->sd
, sd
);
5598 /* cpus with isolated domains */
5599 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5601 /* Setup the mask of cpus configured for isolated domains */
5602 static int __init
isolated_cpu_setup(char *str
)
5604 int ints
[NR_CPUS
], i
;
5606 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5607 cpus_clear(cpu_isolated_map
);
5608 for (i
= 1; i
<= ints
[0]; i
++)
5609 if (ints
[i
] < NR_CPUS
)
5610 cpu_set(ints
[i
], cpu_isolated_map
);
5614 __setup ("isolcpus=", isolated_cpu_setup
);
5617 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5618 * to a function which identifies what group(along with sched group) a CPU
5619 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5620 * (due to the fact that we keep track of groups covered with a cpumask_t).
5622 * init_sched_build_groups will build a circular linked list of the groups
5623 * covered by the given span, and will set each group's ->cpumask correctly,
5624 * and ->cpu_power to 0.
5627 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5628 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5629 struct sched_group
**sg
))
5631 struct sched_group
*first
= NULL
, *last
= NULL
;
5632 cpumask_t covered
= CPU_MASK_NONE
;
5635 for_each_cpu_mask(i
, span
) {
5636 struct sched_group
*sg
;
5637 int group
= group_fn(i
, cpu_map
, &sg
);
5640 if (cpu_isset(i
, covered
))
5643 sg
->cpumask
= CPU_MASK_NONE
;
5644 sg
->__cpu_power
= 0;
5646 for_each_cpu_mask(j
, span
) {
5647 if (group_fn(j
, cpu_map
, NULL
) != group
)
5650 cpu_set(j
, covered
);
5651 cpu_set(j
, sg
->cpumask
);
5662 #define SD_NODES_PER_DOMAIN 16
5667 * find_next_best_node - find the next node to include in a sched_domain
5668 * @node: node whose sched_domain we're building
5669 * @used_nodes: nodes already in the sched_domain
5671 * Find the next node to include in a given scheduling domain. Simply
5672 * finds the closest node not already in the @used_nodes map.
5674 * Should use nodemask_t.
5676 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5678 int i
, n
, val
, min_val
, best_node
= 0;
5682 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5683 /* Start at @node */
5684 n
= (node
+ i
) % MAX_NUMNODES
;
5686 if (!nr_cpus_node(n
))
5689 /* Skip already used nodes */
5690 if (test_bit(n
, used_nodes
))
5693 /* Simple min distance search */
5694 val
= node_distance(node
, n
);
5696 if (val
< min_val
) {
5702 set_bit(best_node
, used_nodes
);
5707 * sched_domain_node_span - get a cpumask for a node's sched_domain
5708 * @node: node whose cpumask we're constructing
5709 * @size: number of nodes to include in this span
5711 * Given a node, construct a good cpumask for its sched_domain to span. It
5712 * should be one that prevents unnecessary balancing, but also spreads tasks
5715 static cpumask_t
sched_domain_node_span(int node
)
5717 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5718 cpumask_t span
, nodemask
;
5722 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5724 nodemask
= node_to_cpumask(node
);
5725 cpus_or(span
, span
, nodemask
);
5726 set_bit(node
, used_nodes
);
5728 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5729 int next_node
= find_next_best_node(node
, used_nodes
);
5731 nodemask
= node_to_cpumask(next_node
);
5732 cpus_or(span
, span
, nodemask
);
5739 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5742 * SMT sched-domains:
5744 #ifdef CONFIG_SCHED_SMT
5745 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5746 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5748 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5749 struct sched_group
**sg
)
5752 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5758 * multi-core sched-domains:
5760 #ifdef CONFIG_SCHED_MC
5761 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5762 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5765 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5766 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5767 struct sched_group
**sg
)
5770 cpumask_t mask
= cpu_sibling_map
[cpu
];
5771 cpus_and(mask
, mask
, *cpu_map
);
5772 group
= first_cpu(mask
);
5774 *sg
= &per_cpu(sched_group_core
, group
);
5777 #elif defined(CONFIG_SCHED_MC)
5778 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5779 struct sched_group
**sg
)
5782 *sg
= &per_cpu(sched_group_core
, cpu
);
5787 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5788 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5790 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5791 struct sched_group
**sg
)
5794 #ifdef CONFIG_SCHED_MC
5795 cpumask_t mask
= cpu_coregroup_map(cpu
);
5796 cpus_and(mask
, mask
, *cpu_map
);
5797 group
= first_cpu(mask
);
5798 #elif defined(CONFIG_SCHED_SMT)
5799 cpumask_t mask
= cpu_sibling_map
[cpu
];
5800 cpus_and(mask
, mask
, *cpu_map
);
5801 group
= first_cpu(mask
);
5806 *sg
= &per_cpu(sched_group_phys
, group
);
5812 * The init_sched_build_groups can't handle what we want to do with node
5813 * groups, so roll our own. Now each node has its own list of groups which
5814 * gets dynamically allocated.
5816 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5817 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5819 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5820 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5822 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5823 struct sched_group
**sg
)
5825 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5828 cpus_and(nodemask
, nodemask
, *cpu_map
);
5829 group
= first_cpu(nodemask
);
5832 *sg
= &per_cpu(sched_group_allnodes
, group
);
5836 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5838 struct sched_group
*sg
= group_head
;
5844 for_each_cpu_mask(j
, sg
->cpumask
) {
5845 struct sched_domain
*sd
;
5847 sd
= &per_cpu(phys_domains
, j
);
5848 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5850 * Only add "power" once for each
5856 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5859 if (sg
!= group_head
)
5865 /* Free memory allocated for various sched_group structures */
5866 static void free_sched_groups(const cpumask_t
*cpu_map
)
5870 for_each_cpu_mask(cpu
, *cpu_map
) {
5871 struct sched_group
**sched_group_nodes
5872 = sched_group_nodes_bycpu
[cpu
];
5874 if (!sched_group_nodes
)
5877 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5878 cpumask_t nodemask
= node_to_cpumask(i
);
5879 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5881 cpus_and(nodemask
, nodemask
, *cpu_map
);
5882 if (cpus_empty(nodemask
))
5892 if (oldsg
!= sched_group_nodes
[i
])
5895 kfree(sched_group_nodes
);
5896 sched_group_nodes_bycpu
[cpu
] = NULL
;
5900 static void free_sched_groups(const cpumask_t
*cpu_map
)
5906 * Initialize sched groups cpu_power.
5908 * cpu_power indicates the capacity of sched group, which is used while
5909 * distributing the load between different sched groups in a sched domain.
5910 * Typically cpu_power for all the groups in a sched domain will be same unless
5911 * there are asymmetries in the topology. If there are asymmetries, group
5912 * having more cpu_power will pickup more load compared to the group having
5915 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5916 * the maximum number of tasks a group can handle in the presence of other idle
5917 * or lightly loaded groups in the same sched domain.
5919 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5921 struct sched_domain
*child
;
5922 struct sched_group
*group
;
5924 WARN_ON(!sd
|| !sd
->groups
);
5926 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5931 sd
->groups
->__cpu_power
= 0;
5934 * For perf policy, if the groups in child domain share resources
5935 * (for example cores sharing some portions of the cache hierarchy
5936 * or SMT), then set this domain groups cpu_power such that each group
5937 * can handle only one task, when there are other idle groups in the
5938 * same sched domain.
5940 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5942 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5943 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
5948 * add cpu_power of each child group to this groups cpu_power
5950 group
= child
->groups
;
5952 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
5953 group
= group
->next
;
5954 } while (group
!= child
->groups
);
5958 * Build sched domains for a given set of cpus and attach the sched domains
5959 * to the individual cpus
5961 static int build_sched_domains(const cpumask_t
*cpu_map
)
5965 struct sched_group
**sched_group_nodes
= NULL
;
5966 int sd_allnodes
= 0;
5969 * Allocate the per-node list of sched groups
5971 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
5973 if (!sched_group_nodes
) {
5974 printk(KERN_WARNING
"Can not alloc sched group node list\n");
5977 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
5981 * Set up domains for cpus specified by the cpu_map.
5983 for_each_cpu_mask(i
, *cpu_map
) {
5984 struct sched_domain
*sd
= NULL
, *p
;
5985 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
5987 cpus_and(nodemask
, nodemask
, *cpu_map
);
5990 if (cpus_weight(*cpu_map
) >
5991 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
5992 sd
= &per_cpu(allnodes_domains
, i
);
5993 *sd
= SD_ALLNODES_INIT
;
5994 sd
->span
= *cpu_map
;
5995 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6001 sd
= &per_cpu(node_domains
, i
);
6003 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6007 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6011 sd
= &per_cpu(phys_domains
, i
);
6013 sd
->span
= nodemask
;
6017 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6019 #ifdef CONFIG_SCHED_MC
6021 sd
= &per_cpu(core_domains
, i
);
6023 sd
->span
= cpu_coregroup_map(i
);
6024 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6027 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6030 #ifdef CONFIG_SCHED_SMT
6032 sd
= &per_cpu(cpu_domains
, i
);
6033 *sd
= SD_SIBLING_INIT
;
6034 sd
->span
= cpu_sibling_map
[i
];
6035 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6038 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6042 #ifdef CONFIG_SCHED_SMT
6043 /* Set up CPU (sibling) groups */
6044 for_each_cpu_mask(i
, *cpu_map
) {
6045 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6046 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6047 if (i
!= first_cpu(this_sibling_map
))
6050 init_sched_build_groups(this_sibling_map
, cpu_map
,
6055 #ifdef CONFIG_SCHED_MC
6056 /* Set up multi-core groups */
6057 for_each_cpu_mask(i
, *cpu_map
) {
6058 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6059 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6060 if (i
!= first_cpu(this_core_map
))
6062 init_sched_build_groups(this_core_map
, cpu_map
,
6063 &cpu_to_core_group
);
6067 /* Set up physical groups */
6068 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6069 cpumask_t nodemask
= node_to_cpumask(i
);
6071 cpus_and(nodemask
, nodemask
, *cpu_map
);
6072 if (cpus_empty(nodemask
))
6075 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6079 /* Set up node groups */
6081 init_sched_build_groups(*cpu_map
, cpu_map
,
6082 &cpu_to_allnodes_group
);
6084 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6085 /* Set up node groups */
6086 struct sched_group
*sg
, *prev
;
6087 cpumask_t nodemask
= node_to_cpumask(i
);
6088 cpumask_t domainspan
;
6089 cpumask_t covered
= CPU_MASK_NONE
;
6092 cpus_and(nodemask
, nodemask
, *cpu_map
);
6093 if (cpus_empty(nodemask
)) {
6094 sched_group_nodes
[i
] = NULL
;
6098 domainspan
= sched_domain_node_span(i
);
6099 cpus_and(domainspan
, domainspan
, *cpu_map
);
6101 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6103 printk(KERN_WARNING
"Can not alloc domain group for "
6107 sched_group_nodes
[i
] = sg
;
6108 for_each_cpu_mask(j
, nodemask
) {
6109 struct sched_domain
*sd
;
6111 sd
= &per_cpu(node_domains
, j
);
6114 sg
->__cpu_power
= 0;
6115 sg
->cpumask
= nodemask
;
6117 cpus_or(covered
, covered
, nodemask
);
6120 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6121 cpumask_t tmp
, notcovered
;
6122 int n
= (i
+ j
) % MAX_NUMNODES
;
6124 cpus_complement(notcovered
, covered
);
6125 cpus_and(tmp
, notcovered
, *cpu_map
);
6126 cpus_and(tmp
, tmp
, domainspan
);
6127 if (cpus_empty(tmp
))
6130 nodemask
= node_to_cpumask(n
);
6131 cpus_and(tmp
, tmp
, nodemask
);
6132 if (cpus_empty(tmp
))
6135 sg
= kmalloc_node(sizeof(struct sched_group
),
6139 "Can not alloc domain group for node %d\n", j
);
6142 sg
->__cpu_power
= 0;
6144 sg
->next
= prev
->next
;
6145 cpus_or(covered
, covered
, tmp
);
6152 /* Calculate CPU power for physical packages and nodes */
6153 #ifdef CONFIG_SCHED_SMT
6154 for_each_cpu_mask(i
, *cpu_map
) {
6155 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6157 init_sched_groups_power(i
, sd
);
6160 #ifdef CONFIG_SCHED_MC
6161 for_each_cpu_mask(i
, *cpu_map
) {
6162 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6164 init_sched_groups_power(i
, sd
);
6168 for_each_cpu_mask(i
, *cpu_map
) {
6169 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6171 init_sched_groups_power(i
, sd
);
6175 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6176 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6179 struct sched_group
*sg
;
6181 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6182 init_numa_sched_groups_power(sg
);
6186 /* Attach the domains */
6187 for_each_cpu_mask(i
, *cpu_map
) {
6188 struct sched_domain
*sd
;
6189 #ifdef CONFIG_SCHED_SMT
6190 sd
= &per_cpu(cpu_domains
, i
);
6191 #elif defined(CONFIG_SCHED_MC)
6192 sd
= &per_cpu(core_domains
, i
);
6194 sd
= &per_cpu(phys_domains
, i
);
6196 cpu_attach_domain(sd
, i
);
6203 free_sched_groups(cpu_map
);
6208 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6210 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6212 cpumask_t cpu_default_map
;
6216 * Setup mask for cpus without special case scheduling requirements.
6217 * For now this just excludes isolated cpus, but could be used to
6218 * exclude other special cases in the future.
6220 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6222 err
= build_sched_domains(&cpu_default_map
);
6227 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6229 free_sched_groups(cpu_map
);
6233 * Detach sched domains from a group of cpus specified in cpu_map
6234 * These cpus will now be attached to the NULL domain
6236 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6240 for_each_cpu_mask(i
, *cpu_map
)
6241 cpu_attach_domain(NULL
, i
);
6242 synchronize_sched();
6243 arch_destroy_sched_domains(cpu_map
);
6247 * Partition sched domains as specified by the cpumasks below.
6248 * This attaches all cpus from the cpumasks to the NULL domain,
6249 * waits for a RCU quiescent period, recalculates sched
6250 * domain information and then attaches them back to the
6251 * correct sched domains
6252 * Call with hotplug lock held
6254 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6256 cpumask_t change_map
;
6259 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6260 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6261 cpus_or(change_map
, *partition1
, *partition2
);
6263 /* Detach sched domains from all of the affected cpus */
6264 detach_destroy_domains(&change_map
);
6265 if (!cpus_empty(*partition1
))
6266 err
= build_sched_domains(partition1
);
6267 if (!err
&& !cpus_empty(*partition2
))
6268 err
= build_sched_domains(partition2
);
6273 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6274 static int arch_reinit_sched_domains(void)
6278 mutex_lock(&sched_hotcpu_mutex
);
6279 detach_destroy_domains(&cpu_online_map
);
6280 err
= arch_init_sched_domains(&cpu_online_map
);
6281 mutex_unlock(&sched_hotcpu_mutex
);
6286 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6290 if (buf
[0] != '0' && buf
[0] != '1')
6294 sched_smt_power_savings
= (buf
[0] == '1');
6296 sched_mc_power_savings
= (buf
[0] == '1');
6298 ret
= arch_reinit_sched_domains();
6300 return ret
? ret
: count
;
6303 #ifdef CONFIG_SCHED_MC
6304 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6306 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6308 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6309 const char *buf
, size_t count
)
6311 return sched_power_savings_store(buf
, count
, 0);
6313 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6314 sched_mc_power_savings_store
);
6317 #ifdef CONFIG_SCHED_SMT
6318 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6320 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6322 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6323 const char *buf
, size_t count
)
6325 return sched_power_savings_store(buf
, count
, 1);
6327 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6328 sched_smt_power_savings_store
);
6331 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6335 #ifdef CONFIG_SCHED_SMT
6337 err
= sysfs_create_file(&cls
->kset
.kobj
,
6338 &attr_sched_smt_power_savings
.attr
);
6340 #ifdef CONFIG_SCHED_MC
6341 if (!err
&& mc_capable())
6342 err
= sysfs_create_file(&cls
->kset
.kobj
,
6343 &attr_sched_mc_power_savings
.attr
);
6350 * Force a reinitialization of the sched domains hierarchy. The domains
6351 * and groups cannot be updated in place without racing with the balancing
6352 * code, so we temporarily attach all running cpus to the NULL domain
6353 * which will prevent rebalancing while the sched domains are recalculated.
6355 static int update_sched_domains(struct notifier_block
*nfb
,
6356 unsigned long action
, void *hcpu
)
6359 case CPU_UP_PREPARE
:
6360 case CPU_UP_PREPARE_FROZEN
:
6361 case CPU_DOWN_PREPARE
:
6362 case CPU_DOWN_PREPARE_FROZEN
:
6363 detach_destroy_domains(&cpu_online_map
);
6366 case CPU_UP_CANCELED
:
6367 case CPU_UP_CANCELED_FROZEN
:
6368 case CPU_DOWN_FAILED
:
6369 case CPU_DOWN_FAILED_FROZEN
:
6371 case CPU_ONLINE_FROZEN
:
6373 case CPU_DEAD_FROZEN
:
6375 * Fall through and re-initialise the domains.
6382 /* The hotplug lock is already held by cpu_up/cpu_down */
6383 arch_init_sched_domains(&cpu_online_map
);
6388 void __init
sched_init_smp(void)
6390 cpumask_t non_isolated_cpus
;
6392 mutex_lock(&sched_hotcpu_mutex
);
6393 arch_init_sched_domains(&cpu_online_map
);
6394 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6395 if (cpus_empty(non_isolated_cpus
))
6396 cpu_set(smp_processor_id(), non_isolated_cpus
);
6397 mutex_unlock(&sched_hotcpu_mutex
);
6398 /* XXX: Theoretical race here - CPU may be hotplugged now */
6399 hotcpu_notifier(update_sched_domains
, 0);
6401 init_sched_domain_sysctl();
6403 /* Move init over to a non-isolated CPU */
6404 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6408 void __init
sched_init_smp(void)
6411 #endif /* CONFIG_SMP */
6413 int in_sched_functions(unsigned long addr
)
6415 /* Linker adds these: start and end of __sched functions */
6416 extern char __sched_text_start
[], __sched_text_end
[];
6418 return in_lock_functions(addr
) ||
6419 (addr
>= (unsigned long)__sched_text_start
6420 && addr
< (unsigned long)__sched_text_end
);
6423 static inline void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6425 cfs_rq
->tasks_timeline
= RB_ROOT
;
6426 cfs_rq
->fair_clock
= 1;
6427 #ifdef CONFIG_FAIR_GROUP_SCHED
6432 void __init
sched_init(void)
6434 int highest_cpu
= 0;
6438 * Link up the scheduling class hierarchy:
6440 rt_sched_class
.next
= &fair_sched_class
;
6441 fair_sched_class
.next
= &idle_sched_class
;
6442 idle_sched_class
.next
= NULL
;
6444 for_each_possible_cpu(i
) {
6445 struct rt_prio_array
*array
;
6449 spin_lock_init(&rq
->lock
);
6450 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6453 init_cfs_rq(&rq
->cfs
, rq
);
6454 #ifdef CONFIG_FAIR_GROUP_SCHED
6455 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6456 list_add(&rq
->cfs
.leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
6459 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6460 rq
->cpu_load
[j
] = 0;
6463 rq
->active_balance
= 0;
6464 rq
->next_balance
= jiffies
;
6467 rq
->migration_thread
= NULL
;
6468 INIT_LIST_HEAD(&rq
->migration_queue
);
6470 atomic_set(&rq
->nr_iowait
, 0);
6472 array
= &rq
->rt
.active
;
6473 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6474 INIT_LIST_HEAD(array
->queue
+ j
);
6475 __clear_bit(j
, array
->bitmap
);
6478 /* delimiter for bitsearch: */
6479 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6482 set_load_weight(&init_task
);
6484 #ifdef CONFIG_PREEMPT_NOTIFIERS
6485 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6489 nr_cpu_ids
= highest_cpu
+ 1;
6490 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6493 #ifdef CONFIG_RT_MUTEXES
6494 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6498 * The boot idle thread does lazy MMU switching as well:
6500 atomic_inc(&init_mm
.mm_count
);
6501 enter_lazy_tlb(&init_mm
, current
);
6504 * Make us the idle thread. Technically, schedule() should not be
6505 * called from this thread, however somewhere below it might be,
6506 * but because we are the idle thread, we just pick up running again
6507 * when this runqueue becomes "idle".
6509 init_idle(current
, smp_processor_id());
6511 * During early bootup we pretend to be a normal task:
6513 current
->sched_class
= &fair_sched_class
;
6516 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6517 void __might_sleep(char *file
, int line
)
6520 static unsigned long prev_jiffy
; /* ratelimiting */
6522 if ((in_atomic() || irqs_disabled()) &&
6523 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6524 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6526 prev_jiffy
= jiffies
;
6527 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6528 " context at %s:%d\n", file
, line
);
6529 printk("in_atomic():%d, irqs_disabled():%d\n",
6530 in_atomic(), irqs_disabled());
6531 debug_show_held_locks(current
);
6532 if (irqs_disabled())
6533 print_irqtrace_events(current
);
6538 EXPORT_SYMBOL(__might_sleep
);
6541 #ifdef CONFIG_MAGIC_SYSRQ
6542 void normalize_rt_tasks(void)
6544 struct task_struct
*g
, *p
;
6545 unsigned long flags
;
6549 read_lock_irq(&tasklist_lock
);
6550 do_each_thread(g
, p
) {
6552 p
->se
.wait_runtime
= 0;
6553 p
->se
.exec_start
= 0;
6554 p
->se
.wait_start_fair
= 0;
6555 p
->se
.sleep_start_fair
= 0;
6556 #ifdef CONFIG_SCHEDSTATS
6557 p
->se
.wait_start
= 0;
6558 p
->se
.sleep_start
= 0;
6559 p
->se
.block_start
= 0;
6561 task_rq(p
)->cfs
.fair_clock
= 0;
6562 task_rq(p
)->clock
= 0;
6566 * Renice negative nice level userspace
6569 if (TASK_NICE(p
) < 0 && p
->mm
)
6570 set_user_nice(p
, 0);
6574 spin_lock_irqsave(&p
->pi_lock
, flags
);
6575 rq
= __task_rq_lock(p
);
6578 * Do not touch the migration thread:
6580 if (p
== rq
->migration_thread
)
6584 update_rq_clock(rq
);
6585 on_rq
= p
->se
.on_rq
;
6587 deactivate_task(rq
, p
, 0);
6588 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6590 activate_task(rq
, p
, 0);
6591 resched_task(rq
->curr
);
6596 __task_rq_unlock(rq
);
6597 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6598 } while_each_thread(g
, p
);
6600 read_unlock_irq(&tasklist_lock
);
6603 #endif /* CONFIG_MAGIC_SYSRQ */
6607 * These functions are only useful for the IA64 MCA handling.
6609 * They can only be called when the whole system has been
6610 * stopped - every CPU needs to be quiescent, and no scheduling
6611 * activity can take place. Using them for anything else would
6612 * be a serious bug, and as a result, they aren't even visible
6613 * under any other configuration.
6617 * curr_task - return the current task for a given cpu.
6618 * @cpu: the processor in question.
6620 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6622 struct task_struct
*curr_task(int cpu
)
6624 return cpu_curr(cpu
);
6628 * set_curr_task - set the current task for a given cpu.
6629 * @cpu: the processor in question.
6630 * @p: the task pointer to set.
6632 * Description: This function must only be used when non-maskable interrupts
6633 * are serviced on a separate stack. It allows the architecture to switch the
6634 * notion of the current task on a cpu in a non-blocking manner. This function
6635 * must be called with all CPU's synchronized, and interrupts disabled, the
6636 * and caller must save the original value of the current task (see
6637 * curr_task() above) and restore that value before reenabling interrupts and
6638 * re-starting the system.
6640 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6642 void set_curr_task(int cpu
, struct task_struct
*p
)