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
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/freezer.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
55 #include <linux/reciprocal_div.h>
58 #include <asm/unistd.h>
61 * Scheduler clock - returns current time in nanosec units.
62 * This is default implementation.
63 * Architectures and sub-architectures can override this.
65 unsigned long long __attribute__((weak
)) sched_clock(void)
67 return (unsigned long long)jiffies
* (1000000000 / HZ
);
71 * Convert user-nice values [ -20 ... 0 ... 19 ]
72 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
75 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
76 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
77 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
80 * 'User priority' is the nice value converted to something we
81 * can work with better when scaling various scheduler parameters,
82 * it's a [ 0 ... 39 ] range.
84 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
85 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
86 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
89 * Some helpers for converting nanosecond timing to jiffy resolution
91 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
92 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
94 #define NICE_0_LOAD SCHED_LOAD_SCALE
95 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
98 * These are the 'tuning knobs' of the scheduler:
100 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
101 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
102 * Timeslices get refilled after they expire.
104 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
105 #define DEF_TIMESLICE (100 * HZ / 1000)
106 #define ON_RUNQUEUE_WEIGHT 30
107 #define CHILD_PENALTY 95
108 #define PARENT_PENALTY 100
109 #define EXIT_WEIGHT 3
110 #define PRIO_BONUS_RATIO 25
111 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
112 #define INTERACTIVE_DELTA 2
113 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
114 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
115 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
118 * If a task is 'interactive' then we reinsert it in the active
119 * array after it has expired its current timeslice. (it will not
120 * continue to run immediately, it will still roundrobin with
121 * other interactive tasks.)
123 * This part scales the interactivity limit depending on niceness.
125 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
126 * Here are a few examples of different nice levels:
128 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
129 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
130 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
131 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
132 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
134 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
135 * priority range a task can explore, a value of '1' means the
136 * task is rated interactive.)
138 * Ie. nice +19 tasks can never get 'interactive' enough to be
139 * reinserted into the active array. And only heavily CPU-hog nice -20
140 * tasks will be expired. Default nice 0 tasks are somewhere between,
141 * it takes some effort for them to get interactive, but it's not
145 #define CURRENT_BONUS(p) \
146 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
149 #define GRANULARITY (10 * HZ / 1000 ? : 1)
152 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
153 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
156 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
157 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
160 #define SCALE(v1,v1_max,v2_max) \
161 (v1) * (v2_max) / (v1_max)
164 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
167 #define TASK_INTERACTIVE(p) \
168 ((p)->prio <= (p)->static_prio - DELTA(p))
170 #define INTERACTIVE_SLEEP(p) \
171 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
172 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
174 #define TASK_PREEMPTS_CURR(p, rq) \
175 ((p)->prio < (rq)->curr->prio)
177 #define SCALE_PRIO(x, prio) \
178 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
180 static unsigned int static_prio_timeslice(int static_prio
)
182 if (static_prio
< NICE_TO_PRIO(0))
183 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
185 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
190 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
191 * Since cpu_power is a 'constant', we can use a reciprocal divide.
193 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
195 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
199 * Each time a sched group cpu_power is changed,
200 * we must compute its reciprocal value
202 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
204 sg
->__cpu_power
+= val
;
205 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
210 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
211 * to time slice values: [800ms ... 100ms ... 5ms]
213 * The higher a thread's priority, the bigger timeslices
214 * it gets during one round of execution. But even the lowest
215 * priority thread gets MIN_TIMESLICE worth of execution time.
218 static inline unsigned int task_timeslice(struct task_struct
*p
)
220 return static_prio_timeslice(p
->static_prio
);
223 static inline int rt_policy(int policy
)
225 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
230 static inline int task_has_rt_policy(struct task_struct
*p
)
232 return rt_policy(p
->policy
);
236 * This is the priority-queue data structure of the RT scheduling class:
238 struct rt_prio_array
{
239 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
240 struct list_head queue
[MAX_RT_PRIO
];
244 struct load_weight load
;
245 u64 load_update_start
, load_update_last
;
246 unsigned long delta_fair
, delta_exec
, delta_stat
;
249 /* CFS-related fields in a runqueue */
251 struct load_weight load
;
252 unsigned long nr_running
;
258 unsigned long wait_runtime_overruns
, wait_runtime_underruns
;
260 struct rb_root tasks_timeline
;
261 struct rb_node
*rb_leftmost
;
262 struct rb_node
*rb_load_balance_curr
;
263 #ifdef CONFIG_FAIR_GROUP_SCHED
264 /* 'curr' points to currently running entity on this cfs_rq.
265 * It is set to NULL otherwise (i.e when none are currently running).
267 struct sched_entity
*curr
;
268 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
270 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
271 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
272 * (like users, containers etc.)
274 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
275 * list is used during load balance.
277 struct list_head leaf_cfs_rq_list
; /* Better name : task_cfs_rq_list? */
281 /* Real-Time classes' related field in a runqueue: */
283 struct rt_prio_array active
;
284 int rt_load_balance_idx
;
285 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
289 * The prio-array type of the old scheduler:
292 unsigned int nr_active
;
293 DECLARE_BITMAP(bitmap
, MAX_PRIO
+1); /* include 1 bit for delimiter */
294 struct list_head queue
[MAX_PRIO
];
298 * This is the main, per-CPU runqueue data structure.
300 * Locking rule: those places that want to lock multiple runqueues
301 * (such as the load balancing or the thread migration code), lock
302 * acquire operations must be ordered by ascending &runqueue.
305 spinlock_t lock
; /* runqueue lock */
308 * nr_running and cpu_load should be in the same cacheline because
309 * remote CPUs use both these fields when doing load calculation.
311 unsigned long nr_running
;
312 unsigned long raw_weighted_load
;
313 #define CPU_LOAD_IDX_MAX 5
314 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
315 unsigned char idle_at_tick
;
317 unsigned char in_nohz_recently
;
319 struct load_stat ls
; /* capture load from *all* tasks on this cpu */
320 unsigned long nr_load_updates
;
324 #ifdef CONFIG_FAIR_GROUP_SCHED
325 struct list_head leaf_cfs_rq_list
; /* list of leaf cfs_rq on this cpu */
330 * This is part of a global counter where only the total sum
331 * over all CPUs matters. A task can increase this counter on
332 * one CPU and if it got migrated afterwards it may decrease
333 * it on another CPU. Always updated under the runqueue lock:
335 unsigned long nr_uninterruptible
;
337 unsigned long expired_timestamp
;
338 unsigned long long most_recent_timestamp
;
340 struct task_struct
*curr
, *idle
;
341 unsigned long next_balance
;
342 struct mm_struct
*prev_mm
;
344 struct prio_array
*active
, *expired
, arrays
[2];
345 int best_expired_prio
;
347 u64 clock
, prev_clock_raw
;
350 unsigned int clock_warps
, clock_overflows
;
351 unsigned int clock_unstable_events
;
353 struct sched_class
*load_balance_class
;
358 struct sched_domain
*sd
;
360 /* For active balancing */
363 int cpu
; /* cpu of this runqueue */
365 struct task_struct
*migration_thread
;
366 struct list_head migration_queue
;
369 #ifdef CONFIG_SCHEDSTATS
371 struct sched_info rq_sched_info
;
373 /* sys_sched_yield() stats */
374 unsigned long yld_exp_empty
;
375 unsigned long yld_act_empty
;
376 unsigned long yld_both_empty
;
377 unsigned long yld_cnt
;
379 /* schedule() stats */
380 unsigned long sched_switch
;
381 unsigned long sched_cnt
;
382 unsigned long sched_goidle
;
384 /* try_to_wake_up() stats */
385 unsigned long ttwu_cnt
;
386 unsigned long ttwu_local
;
388 struct lock_class_key rq_lock_key
;
391 static DEFINE_PER_CPU(struct rq
, runqueues
) ____cacheline_aligned_in_smp
;
392 static DEFINE_MUTEX(sched_hotcpu_mutex
);
394 static inline int cpu_of(struct rq
*rq
)
404 * Per-runqueue clock, as finegrained as the platform can give us:
406 static unsigned long long __rq_clock(struct rq
*rq
)
408 u64 prev_raw
= rq
->prev_clock_raw
;
409 u64 now
= sched_clock();
410 s64 delta
= now
- prev_raw
;
411 u64 clock
= rq
->clock
;
414 * Protect against sched_clock() occasionally going backwards:
416 if (unlikely(delta
< 0)) {
421 * Catch too large forward jumps too:
423 if (unlikely(delta
> 2*TICK_NSEC
)) {
425 rq
->clock_overflows
++;
427 if (unlikely(delta
> rq
->clock_max_delta
))
428 rq
->clock_max_delta
= delta
;
433 rq
->prev_clock_raw
= now
;
439 static inline unsigned long long rq_clock(struct rq
*rq
)
441 int this_cpu
= smp_processor_id();
443 if (this_cpu
== cpu_of(rq
))
444 return __rq_clock(rq
);
450 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
451 * See detach_destroy_domains: synchronize_sched for details.
453 * The domain tree of any CPU may only be accessed from within
454 * preempt-disabled sections.
456 #define for_each_domain(cpu, __sd) \
457 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
459 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
460 #define this_rq() (&__get_cpu_var(runqueues))
461 #define task_rq(p) cpu_rq(task_cpu(p))
462 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
464 #ifdef CONFIG_FAIR_GROUP_SCHED
465 /* Change a task's ->cfs_rq if it moves across CPUs */
466 static inline void set_task_cfs_rq(struct task_struct
*p
)
468 p
->se
.cfs_rq
= &task_rq(p
)->cfs
;
471 static inline void set_task_cfs_rq(struct task_struct
*p
)
476 #ifndef prepare_arch_switch
477 # define prepare_arch_switch(next) do { } while (0)
479 #ifndef finish_arch_switch
480 # define finish_arch_switch(prev) do { } while (0)
483 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
484 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
486 return rq
->curr
== p
;
489 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
493 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
495 #ifdef CONFIG_DEBUG_SPINLOCK
496 /* this is a valid case when another task releases the spinlock */
497 rq
->lock
.owner
= current
;
500 * If we are tracking spinlock dependencies then we have to
501 * fix up the runqueue lock - which gets 'carried over' from
504 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
506 spin_unlock_irq(&rq
->lock
);
509 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
510 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
515 return rq
->curr
== p
;
519 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
523 * We can optimise this out completely for !SMP, because the
524 * SMP rebalancing from interrupt is the only thing that cares
529 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
530 spin_unlock_irq(&rq
->lock
);
532 spin_unlock(&rq
->lock
);
536 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
540 * After ->oncpu is cleared, the task can be moved to a different CPU.
541 * We must ensure this doesn't happen until the switch is completely
547 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
551 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
554 * __task_rq_lock - lock the runqueue a given task resides on.
555 * Must be called interrupts disabled.
557 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
564 spin_lock(&rq
->lock
);
565 if (unlikely(rq
!= task_rq(p
))) {
566 spin_unlock(&rq
->lock
);
567 goto repeat_lock_task
;
573 * task_rq_lock - lock the runqueue a given task resides on and disable
574 * interrupts. Note the ordering: we can safely lookup the task_rq without
575 * explicitly disabling preemption.
577 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
583 local_irq_save(*flags
);
585 spin_lock(&rq
->lock
);
586 if (unlikely(rq
!= task_rq(p
))) {
587 spin_unlock_irqrestore(&rq
->lock
, *flags
);
588 goto repeat_lock_task
;
593 static inline void __task_rq_unlock(struct rq
*rq
)
596 spin_unlock(&rq
->lock
);
599 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
602 spin_unlock_irqrestore(&rq
->lock
, *flags
);
606 * this_rq_lock - lock this runqueue and disable interrupts.
608 static inline struct rq
*this_rq_lock(void)
615 spin_lock(&rq
->lock
);
621 * resched_task - mark a task 'to be rescheduled now'.
623 * On UP this means the setting of the need_resched flag, on SMP it
624 * might also involve a cross-CPU call to trigger the scheduler on
629 #ifndef tsk_is_polling
630 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
633 static void resched_task(struct task_struct
*p
)
637 assert_spin_locked(&task_rq(p
)->lock
);
639 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
642 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
645 if (cpu
== smp_processor_id())
648 /* NEED_RESCHED must be visible before we test polling */
650 if (!tsk_is_polling(p
))
651 smp_send_reschedule(cpu
);
654 static void resched_cpu(int cpu
)
656 struct rq
*rq
= cpu_rq(cpu
);
659 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
661 resched_task(cpu_curr(cpu
));
662 spin_unlock_irqrestore(&rq
->lock
, flags
);
665 static inline void resched_task(struct task_struct
*p
)
667 assert_spin_locked(&task_rq(p
)->lock
);
668 set_tsk_need_resched(p
);
672 #include "sched_stats.h"
674 static u64
div64_likely32(u64 divident
, unsigned long divisor
)
676 #if BITS_PER_LONG == 32
677 if (likely(divident
<= 0xffffffffULL
))
678 return (u32
)divident
/ divisor
;
679 do_div(divident
, divisor
);
683 return divident
/ divisor
;
687 #if BITS_PER_LONG == 32
688 # define WMULT_CONST (~0UL)
690 # define WMULT_CONST (1UL << 32)
693 #define WMULT_SHIFT 32
695 static inline unsigned long
696 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
697 struct load_weight
*lw
)
701 if (unlikely(!lw
->inv_weight
))
702 lw
->inv_weight
= WMULT_CONST
/ lw
->weight
;
704 tmp
= (u64
)delta_exec
* weight
;
706 * Check whether we'd overflow the 64-bit multiplication:
708 if (unlikely(tmp
> WMULT_CONST
)) {
709 tmp
= ((tmp
>> WMULT_SHIFT
/2) * lw
->inv_weight
)
712 tmp
= (tmp
* lw
->inv_weight
) >> WMULT_SHIFT
;
715 return (unsigned long)min(tmp
, (u64
)sysctl_sched_runtime_limit
);
718 static inline unsigned long
719 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
721 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
724 static void update_load_add(struct load_weight
*lw
, unsigned long inc
)
730 static void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
736 static void __update_curr_load(struct rq
*rq
, struct load_stat
*ls
)
738 if (rq
->curr
!= rq
->idle
&& ls
->load
.weight
) {
739 ls
->delta_exec
+= ls
->delta_stat
;
740 ls
->delta_fair
+= calc_delta_fair(ls
->delta_stat
, &ls
->load
);
746 * Update delta_exec, delta_fair fields for rq.
748 * delta_fair clock advances at a rate inversely proportional to
749 * total load (rq->ls.load.weight) on the runqueue, while
750 * delta_exec advances at the same rate as wall-clock (provided
753 * delta_exec / delta_fair is a measure of the (smoothened) load on this
754 * runqueue over any given interval. This (smoothened) load is used
755 * during load balance.
757 * This function is called /before/ updating rq->ls.load
758 * and when switching tasks.
760 static void update_curr_load(struct rq
*rq
, u64 now
)
762 struct load_stat
*ls
= &rq
->ls
;
765 start
= ls
->load_update_start
;
766 ls
->load_update_start
= now
;
767 ls
->delta_stat
+= now
- start
;
769 * Stagger updates to ls->delta_fair. Very frequent updates
772 if (ls
->delta_stat
>= sysctl_sched_stat_granularity
)
773 __update_curr_load(rq
, ls
);
777 * To aid in avoiding the subversion of "niceness" due to uneven distribution
778 * of tasks with abnormal "nice" values across CPUs the contribution that
779 * each task makes to its run queue's load is weighted according to its
780 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
781 * scaled version of the new time slice allocation that they receive on time
786 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
787 * If static_prio_timeslice() is ever changed to break this assumption then
788 * this code will need modification
790 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
791 #define LOAD_WEIGHT(lp) \
792 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
793 #define PRIO_TO_LOAD_WEIGHT(prio) \
794 LOAD_WEIGHT(static_prio_timeslice(prio))
795 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
796 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
799 inc_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
801 rq
->raw_weighted_load
+= p
->load_weight
;
805 dec_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
807 rq
->raw_weighted_load
-= p
->load_weight
;
810 static inline void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
813 inc_raw_weighted_load(rq
, p
);
816 static inline void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
819 dec_raw_weighted_load(rq
, p
);
822 static void set_load_weight(struct task_struct
*p
)
824 if (task_has_rt_policy(p
)) {
826 if (p
== task_rq(p
)->migration_thread
)
828 * The migration thread does the actual balancing.
829 * Giving its load any weight will skew balancing
835 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
837 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
841 * Adding/removing a task to/from a priority array:
843 static void dequeue_task(struct task_struct
*p
, struct prio_array
*array
)
846 list_del(&p
->run_list
);
847 if (list_empty(array
->queue
+ p
->prio
))
848 __clear_bit(p
->prio
, array
->bitmap
);
851 static void enqueue_task(struct task_struct
*p
, struct prio_array
*array
)
853 sched_info_queued(p
);
854 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
855 __set_bit(p
->prio
, array
->bitmap
);
861 * Put task to the end of the run list without the overhead of dequeue
862 * followed by enqueue.
864 static void requeue_task(struct task_struct
*p
, struct prio_array
*array
)
866 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
870 enqueue_task_head(struct task_struct
*p
, struct prio_array
*array
)
872 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
873 __set_bit(p
->prio
, array
->bitmap
);
879 * __normal_prio - return the priority that is based on the static
880 * priority but is modified by bonuses/penalties.
882 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
883 * into the -5 ... 0 ... +5 bonus/penalty range.
885 * We use 25% of the full 0...39 priority range so that:
887 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
888 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
890 * Both properties are important to certain workloads.
893 static inline int __normal_prio(struct task_struct
*p
)
897 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
899 prio
= p
->static_prio
- bonus
;
900 if (prio
< MAX_RT_PRIO
)
902 if (prio
> MAX_PRIO
-1)
908 * Calculate the expected normal priority: i.e. priority
909 * without taking RT-inheritance into account. Might be
910 * boosted by interactivity modifiers. Changes upon fork,
911 * setprio syscalls, and whenever the interactivity
912 * estimator recalculates.
914 static inline int normal_prio(struct task_struct
*p
)
918 if (task_has_rt_policy(p
))
919 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
921 prio
= __normal_prio(p
);
926 * Calculate the current priority, i.e. the priority
927 * taken into account by the scheduler. This value might
928 * be boosted by RT tasks, or might be boosted by
929 * interactivity modifiers. Will be RT if the task got
930 * RT-boosted. If not then it returns p->normal_prio.
932 static int effective_prio(struct task_struct
*p
)
934 p
->normal_prio
= normal_prio(p
);
936 * If we are RT tasks or we were boosted to RT priority,
937 * keep the priority unchanged. Otherwise, update priority
938 * to the normal priority:
940 if (!rt_prio(p
->prio
))
941 return p
->normal_prio
;
946 * __activate_task - move a task to the runqueue.
948 static void __activate_task(struct task_struct
*p
, struct rq
*rq
)
950 struct prio_array
*target
= rq
->active
;
953 target
= rq
->expired
;
954 enqueue_task(p
, target
);
955 inc_nr_running(p
, rq
);
959 * __activate_idle_task - move idle task to the _front_ of runqueue.
961 static inline void __activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
963 enqueue_task_head(p
, rq
->active
);
964 inc_nr_running(p
, rq
);
968 * Recalculate p->normal_prio and p->prio after having slept,
969 * updating the sleep-average too:
971 static int recalc_task_prio(struct task_struct
*p
, unsigned long long now
)
973 /* Caller must always ensure 'now >= p->timestamp' */
974 unsigned long sleep_time
= now
- p
->timestamp
;
979 if (likely(sleep_time
> 0)) {
981 * This ceiling is set to the lowest priority that would allow
982 * a task to be reinserted into the active array on timeslice
985 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
987 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
989 * Prevents user tasks from achieving best priority
990 * with one single large enough sleep.
992 p
->sleep_avg
= ceiling
;
995 * This code gives a bonus to interactive tasks.
997 * The boost works by updating the 'average sleep time'
998 * value here, based on ->timestamp. The more time a
999 * task spends sleeping, the higher the average gets -
1000 * and the higher the priority boost gets as well.
1002 p
->sleep_avg
+= sleep_time
;
1005 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
1006 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
1009 return effective_prio(p
);
1013 * activate_task - move a task to the runqueue and do priority recalculation
1015 * Update all the scheduling statistics stuff. (sleep average
1016 * calculation, priority modifiers, etc.)
1018 static void activate_task(struct task_struct
*p
, struct rq
*rq
, int local
)
1020 unsigned long long now
;
1025 now
= sched_clock();
1028 /* Compensate for drifting sched_clock */
1029 struct rq
*this_rq
= this_rq();
1030 now
= (now
- this_rq
->most_recent_timestamp
)
1031 + rq
->most_recent_timestamp
;
1036 * Sleep time is in units of nanosecs, so shift by 20 to get a
1037 * milliseconds-range estimation of the amount of time that the task
1040 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
1041 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1042 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(p
),
1043 (now
- p
->timestamp
) >> 20);
1046 p
->prio
= recalc_task_prio(p
, now
);
1049 __activate_task(p
, rq
);
1053 * deactivate_task - remove a task from the runqueue.
1055 static void deactivate_task(struct task_struct
*p
, struct rq
*rq
)
1057 dec_nr_running(p
, rq
);
1058 dequeue_task(p
, p
->array
);
1063 * task_curr - is this task currently executing on a CPU?
1064 * @p: the task in question.
1066 inline int task_curr(const struct task_struct
*p
)
1068 return cpu_curr(task_cpu(p
)) == p
;
1071 /* Used instead of source_load when we know the type == 0 */
1072 unsigned long weighted_cpuload(const int cpu
)
1074 return cpu_rq(cpu
)->raw_weighted_load
;
1079 void set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1081 task_thread_info(p
)->cpu
= cpu
;
1084 struct migration_req
{
1085 struct list_head list
;
1087 struct task_struct
*task
;
1090 struct completion done
;
1094 * The task's runqueue lock must be held.
1095 * Returns true if you have to wait for migration thread.
1098 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1100 struct rq
*rq
= task_rq(p
);
1103 * If the task is not on a runqueue (and not running), then
1104 * it is sufficient to simply update the task's cpu field.
1106 if (!p
->array
&& !task_running(rq
, p
)) {
1107 set_task_cpu(p
, dest_cpu
);
1111 init_completion(&req
->done
);
1113 req
->dest_cpu
= dest_cpu
;
1114 list_add(&req
->list
, &rq
->migration_queue
);
1120 * wait_task_inactive - wait for a thread to unschedule.
1122 * The caller must ensure that the task *will* unschedule sometime soon,
1123 * else this function might spin for a *long* time. This function can't
1124 * be called with interrupts off, or it may introduce deadlock with
1125 * smp_call_function() if an IPI is sent by the same process we are
1126 * waiting to become inactive.
1128 void wait_task_inactive(struct task_struct
*p
)
1130 unsigned long flags
;
1132 struct prio_array
*array
;
1137 * We do the initial early heuristics without holding
1138 * any task-queue locks at all. We'll only try to get
1139 * the runqueue lock when things look like they will
1145 * If the task is actively running on another CPU
1146 * still, just relax and busy-wait without holding
1149 * NOTE! Since we don't hold any locks, it's not
1150 * even sure that "rq" stays as the right runqueue!
1151 * But we don't care, since "task_running()" will
1152 * return false if the runqueue has changed and p
1153 * is actually now running somewhere else!
1155 while (task_running(rq
, p
))
1159 * Ok, time to look more closely! We need the rq
1160 * lock now, to be *sure*. If we're wrong, we'll
1161 * just go back and repeat.
1163 rq
= task_rq_lock(p
, &flags
);
1164 running
= task_running(rq
, p
);
1166 task_rq_unlock(rq
, &flags
);
1169 * Was it really running after all now that we
1170 * checked with the proper locks actually held?
1172 * Oops. Go back and try again..
1174 if (unlikely(running
)) {
1180 * It's not enough that it's not actively running,
1181 * it must be off the runqueue _entirely_, and not
1184 * So if it wa still runnable (but just not actively
1185 * running right now), it's preempted, and we should
1186 * yield - it could be a while.
1188 if (unlikely(array
)) {
1194 * Ahh, all good. It wasn't running, and it wasn't
1195 * runnable, which means that it will never become
1196 * running in the future either. We're all done!
1201 * kick_process - kick a running thread to enter/exit the kernel
1202 * @p: the to-be-kicked thread
1204 * Cause a process which is running on another CPU to enter
1205 * kernel-mode, without any delay. (to get signals handled.)
1207 * NOTE: this function doesnt have to take the runqueue lock,
1208 * because all it wants to ensure is that the remote task enters
1209 * the kernel. If the IPI races and the task has been migrated
1210 * to another CPU then no harm is done and the purpose has been
1213 void kick_process(struct task_struct
*p
)
1219 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1220 smp_send_reschedule(cpu
);
1225 * Return a low guess at the load of a migration-source cpu weighted
1226 * according to the scheduling class and "nice" value.
1228 * We want to under-estimate the load of migration sources, to
1229 * balance conservatively.
1231 static inline unsigned long source_load(int cpu
, int type
)
1233 struct rq
*rq
= cpu_rq(cpu
);
1236 return rq
->raw_weighted_load
;
1238 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1242 * Return a high guess at the load of a migration-target cpu weighted
1243 * according to the scheduling class and "nice" value.
1245 static inline unsigned long target_load(int cpu
, int type
)
1247 struct rq
*rq
= cpu_rq(cpu
);
1250 return rq
->raw_weighted_load
;
1252 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1256 * Return the average load per task on the cpu's run queue
1258 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1260 struct rq
*rq
= cpu_rq(cpu
);
1261 unsigned long n
= rq
->nr_running
;
1263 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1267 * find_idlest_group finds and returns the least busy CPU group within the
1270 static struct sched_group
*
1271 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1273 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1274 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1275 int load_idx
= sd
->forkexec_idx
;
1276 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1279 unsigned long load
, avg_load
;
1283 /* Skip over this group if it has no CPUs allowed */
1284 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1287 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1289 /* Tally up the load of all CPUs in the group */
1292 for_each_cpu_mask(i
, group
->cpumask
) {
1293 /* Bias balancing toward cpus of our domain */
1295 load
= source_load(i
, load_idx
);
1297 load
= target_load(i
, load_idx
);
1302 /* Adjust by relative CPU power of the group */
1303 avg_load
= sg_div_cpu_power(group
,
1304 avg_load
* SCHED_LOAD_SCALE
);
1307 this_load
= avg_load
;
1309 } else if (avg_load
< min_load
) {
1310 min_load
= avg_load
;
1314 group
= group
->next
;
1315 } while (group
!= sd
->groups
);
1317 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1323 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1326 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1329 unsigned long load
, min_load
= ULONG_MAX
;
1333 /* Traverse only the allowed CPUs */
1334 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1336 for_each_cpu_mask(i
, tmp
) {
1337 load
= weighted_cpuload(i
);
1339 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1349 * sched_balance_self: balance the current task (running on cpu) in domains
1350 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1353 * Balance, ie. select the least loaded group.
1355 * Returns the target CPU number, or the same CPU if no balancing is needed.
1357 * preempt must be disabled.
1359 static int sched_balance_self(int cpu
, int flag
)
1361 struct task_struct
*t
= current
;
1362 struct sched_domain
*tmp
, *sd
= NULL
;
1364 for_each_domain(cpu
, tmp
) {
1366 * If power savings logic is enabled for a domain, stop there.
1368 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1370 if (tmp
->flags
& flag
)
1376 struct sched_group
*group
;
1377 int new_cpu
, weight
;
1379 if (!(sd
->flags
& flag
)) {
1385 group
= find_idlest_group(sd
, t
, cpu
);
1391 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1392 if (new_cpu
== -1 || new_cpu
== cpu
) {
1393 /* Now try balancing at a lower domain level of cpu */
1398 /* Now try balancing at a lower domain level of new_cpu */
1401 weight
= cpus_weight(span
);
1402 for_each_domain(cpu
, tmp
) {
1403 if (weight
<= cpus_weight(tmp
->span
))
1405 if (tmp
->flags
& flag
)
1408 /* while loop will break here if sd == NULL */
1414 #endif /* CONFIG_SMP */
1417 * wake_idle() will wake a task on an idle cpu if task->cpu is
1418 * not idle and an idle cpu is available. The span of cpus to
1419 * search starts with cpus closest then further out as needed,
1420 * so we always favor a closer, idle cpu.
1422 * Returns the CPU we should wake onto.
1424 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1425 static int wake_idle(int cpu
, struct task_struct
*p
)
1428 struct sched_domain
*sd
;
1432 * If it is idle, then it is the best cpu to run this task.
1434 * This cpu is also the best, if it has more than one task already.
1435 * Siblings must be also busy(in most cases) as they didn't already
1436 * pickup the extra load from this cpu and hence we need not check
1437 * sibling runqueue info. This will avoid the checks and cache miss
1438 * penalities associated with that.
1440 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1443 for_each_domain(cpu
, sd
) {
1444 if (sd
->flags
& SD_WAKE_IDLE
) {
1445 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1446 for_each_cpu_mask(i
, tmp
) {
1457 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1464 * try_to_wake_up - wake up a thread
1465 * @p: the to-be-woken-up thread
1466 * @state: the mask of task states that can be woken
1467 * @sync: do a synchronous wakeup?
1469 * Put it on the run-queue if it's not already there. The "current"
1470 * thread is always on the run-queue (except when the actual
1471 * re-schedule is in progress), and as such you're allowed to do
1472 * the simpler "current->state = TASK_RUNNING" to mark yourself
1473 * runnable without the overhead of this.
1475 * returns failure only if the task is already active.
1477 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1479 int cpu
, this_cpu
, success
= 0;
1480 unsigned long flags
;
1484 struct sched_domain
*sd
, *this_sd
= NULL
;
1485 unsigned long load
, this_load
;
1489 rq
= task_rq_lock(p
, &flags
);
1490 old_state
= p
->state
;
1491 if (!(old_state
& state
))
1498 this_cpu
= smp_processor_id();
1501 if (unlikely(task_running(rq
, p
)))
1506 schedstat_inc(rq
, ttwu_cnt
);
1507 if (cpu
== this_cpu
) {
1508 schedstat_inc(rq
, ttwu_local
);
1512 for_each_domain(this_cpu
, sd
) {
1513 if (cpu_isset(cpu
, sd
->span
)) {
1514 schedstat_inc(sd
, ttwu_wake_remote
);
1520 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1524 * Check for affine wakeup and passive balancing possibilities.
1527 int idx
= this_sd
->wake_idx
;
1528 unsigned int imbalance
;
1530 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1532 load
= source_load(cpu
, idx
);
1533 this_load
= target_load(this_cpu
, idx
);
1535 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1537 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1538 unsigned long tl
= this_load
;
1539 unsigned long tl_per_task
;
1541 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1544 * If sync wakeup then subtract the (maximum possible)
1545 * effect of the currently running task from the load
1546 * of the current CPU:
1549 tl
-= current
->load_weight
;
1552 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1553 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1555 * This domain has SD_WAKE_AFFINE and
1556 * p is cache cold in this domain, and
1557 * there is no bad imbalance.
1559 schedstat_inc(this_sd
, ttwu_move_affine
);
1565 * Start passive balancing when half the imbalance_pct
1568 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1569 if (imbalance
*this_load
<= 100*load
) {
1570 schedstat_inc(this_sd
, ttwu_move_balance
);
1576 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1578 new_cpu
= wake_idle(new_cpu
, p
);
1579 if (new_cpu
!= cpu
) {
1580 set_task_cpu(p
, new_cpu
);
1581 task_rq_unlock(rq
, &flags
);
1582 /* might preempt at this point */
1583 rq
= task_rq_lock(p
, &flags
);
1584 old_state
= p
->state
;
1585 if (!(old_state
& state
))
1590 this_cpu
= smp_processor_id();
1595 #endif /* CONFIG_SMP */
1596 if (old_state
== TASK_UNINTERRUPTIBLE
)
1597 rq
->nr_uninterruptible
--;
1599 activate_task(p
, rq
, cpu
== this_cpu
);
1601 * Sync wakeups (i.e. those types of wakeups where the waker
1602 * has indicated that it will leave the CPU in short order)
1603 * don't trigger a preemption, if the woken up task will run on
1604 * this cpu. (in this case the 'I will reschedule' promise of
1605 * the waker guarantees that the freshly woken up task is going
1606 * to be considered on this CPU.)
1608 if (!sync
|| cpu
!= this_cpu
) {
1609 if (TASK_PREEMPTS_CURR(p
, rq
))
1610 resched_task(rq
->curr
);
1615 p
->state
= TASK_RUNNING
;
1617 task_rq_unlock(rq
, &flags
);
1622 int fastcall
wake_up_process(struct task_struct
*p
)
1624 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1625 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1627 EXPORT_SYMBOL(wake_up_process
);
1629 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1631 return try_to_wake_up(p
, state
, 0);
1634 static void task_running_tick(struct rq
*rq
, struct task_struct
*p
);
1636 * Perform scheduler related setup for a newly forked process p.
1637 * p is forked by current.
1639 void fastcall
sched_fork(struct task_struct
*p
, int clone_flags
)
1641 int cpu
= get_cpu();
1644 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1646 set_task_cpu(p
, cpu
);
1649 * We mark the process as running here, but have not actually
1650 * inserted it onto the runqueue yet. This guarantees that
1651 * nobody will actually run it, and a signal or other external
1652 * event cannot wake it up and insert it on the runqueue either.
1654 p
->state
= TASK_RUNNING
;
1657 * Make sure we do not leak PI boosting priority to the child:
1659 p
->prio
= current
->normal_prio
;
1661 INIT_LIST_HEAD(&p
->run_list
);
1663 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1664 if (unlikely(sched_info_on()))
1665 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1667 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1670 #ifdef CONFIG_PREEMPT
1671 /* Want to start with kernel preemption disabled. */
1672 task_thread_info(p
)->preempt_count
= 1;
1675 * Share the timeslice between parent and child, thus the
1676 * total amount of pending timeslices in the system doesn't change,
1677 * resulting in more scheduling fairness.
1679 local_irq_disable();
1680 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1682 * The remainder of the first timeslice might be recovered by
1683 * the parent if the child exits early enough.
1685 p
->first_time_slice
= 1;
1686 current
->time_slice
>>= 1;
1687 p
->timestamp
= sched_clock();
1688 if (unlikely(!current
->time_slice
)) {
1690 * This case is rare, it happens when the parent has only
1691 * a single jiffy left from its timeslice. Taking the
1692 * runqueue lock is not a problem.
1694 current
->time_slice
= 1;
1695 task_running_tick(cpu_rq(cpu
), current
);
1702 * wake_up_new_task - wake up a newly created task for the first time.
1704 * This function will do some initial scheduler statistics housekeeping
1705 * that must be done for every newly created context, then puts the task
1706 * on the runqueue and wakes it.
1708 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1710 struct rq
*rq
, *this_rq
;
1711 unsigned long flags
;
1714 rq
= task_rq_lock(p
, &flags
);
1715 BUG_ON(p
->state
!= TASK_RUNNING
);
1716 this_cpu
= smp_processor_id();
1720 * We decrease the sleep average of forking parents
1721 * and children as well, to keep max-interactive tasks
1722 * from forking tasks that are max-interactive. The parent
1723 * (current) is done further down, under its lock.
1725 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1726 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1728 p
->prio
= effective_prio(p
);
1730 if (likely(cpu
== this_cpu
)) {
1731 if (!(clone_flags
& CLONE_VM
)) {
1733 * The VM isn't cloned, so we're in a good position to
1734 * do child-runs-first in anticipation of an exec. This
1735 * usually avoids a lot of COW overhead.
1737 if (unlikely(!current
->array
))
1738 __activate_task(p
, rq
);
1740 p
->prio
= current
->prio
;
1741 p
->normal_prio
= current
->normal_prio
;
1742 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1743 p
->array
= current
->array
;
1744 p
->array
->nr_active
++;
1745 inc_nr_running(p
, rq
);
1749 /* Run child last */
1750 __activate_task(p
, rq
);
1752 * We skip the following code due to cpu == this_cpu
1754 * task_rq_unlock(rq, &flags);
1755 * this_rq = task_rq_lock(current, &flags);
1759 this_rq
= cpu_rq(this_cpu
);
1762 * Not the local CPU - must adjust timestamp. This should
1763 * get optimised away in the !CONFIG_SMP case.
1765 p
->timestamp
= (p
->timestamp
- this_rq
->most_recent_timestamp
)
1766 + rq
->most_recent_timestamp
;
1767 __activate_task(p
, rq
);
1768 if (TASK_PREEMPTS_CURR(p
, rq
))
1769 resched_task(rq
->curr
);
1772 * Parent and child are on different CPUs, now get the
1773 * parent runqueue to update the parent's ->sleep_avg:
1775 task_rq_unlock(rq
, &flags
);
1776 this_rq
= task_rq_lock(current
, &flags
);
1778 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1779 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1780 task_rq_unlock(this_rq
, &flags
);
1784 * prepare_task_switch - prepare to switch tasks
1785 * @rq: the runqueue preparing to switch
1786 * @next: the task we are going to switch to.
1788 * This is called with the rq lock held and interrupts off. It must
1789 * be paired with a subsequent finish_task_switch after the context
1792 * prepare_task_switch sets up locking and calls architecture specific
1795 static inline void prepare_task_switch(struct rq
*rq
, struct task_struct
*next
)
1797 prepare_lock_switch(rq
, next
);
1798 prepare_arch_switch(next
);
1802 * finish_task_switch - clean up after a task-switch
1803 * @rq: runqueue associated with task-switch
1804 * @prev: the thread we just switched away from.
1806 * finish_task_switch must be called after the context switch, paired
1807 * with a prepare_task_switch call before the context switch.
1808 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1809 * and do any other architecture-specific cleanup actions.
1811 * Note that we may have delayed dropping an mm in context_switch(). If
1812 * so, we finish that here outside of the runqueue lock. (Doing it
1813 * with the lock held can cause deadlocks; see schedule() for
1816 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1817 __releases(rq
->lock
)
1819 struct mm_struct
*mm
= rq
->prev_mm
;
1825 * A task struct has one reference for the use as "current".
1826 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1827 * schedule one last time. The schedule call will never return, and
1828 * the scheduled task must drop that reference.
1829 * The test for TASK_DEAD must occur while the runqueue locks are
1830 * still held, otherwise prev could be scheduled on another cpu, die
1831 * there before we look at prev->state, and then the reference would
1833 * Manfred Spraul <manfred@colorfullife.com>
1835 prev_state
= prev
->state
;
1836 finish_arch_switch(prev
);
1837 finish_lock_switch(rq
, prev
);
1840 if (unlikely(prev_state
== TASK_DEAD
)) {
1842 * Remove function-return probe instances associated with this
1843 * task and put them back on the free list.
1845 kprobe_flush_task(prev
);
1846 put_task_struct(prev
);
1851 * schedule_tail - first thing a freshly forked thread must call.
1852 * @prev: the thread we just switched away from.
1854 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1855 __releases(rq
->lock
)
1857 struct rq
*rq
= this_rq();
1859 finish_task_switch(rq
, prev
);
1860 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1861 /* In this case, finish_task_switch does not reenable preemption */
1864 if (current
->set_child_tid
)
1865 put_user(current
->pid
, current
->set_child_tid
);
1869 * context_switch - switch to the new MM and the new
1870 * thread's register state.
1872 static inline struct task_struct
*
1873 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1874 struct task_struct
*next
)
1876 struct mm_struct
*mm
= next
->mm
;
1877 struct mm_struct
*oldmm
= prev
->active_mm
;
1880 * For paravirt, this is coupled with an exit in switch_to to
1881 * combine the page table reload and the switch backend into
1884 arch_enter_lazy_cpu_mode();
1887 next
->active_mm
= oldmm
;
1888 atomic_inc(&oldmm
->mm_count
);
1889 enter_lazy_tlb(oldmm
, next
);
1891 switch_mm(oldmm
, mm
, next
);
1894 prev
->active_mm
= NULL
;
1895 WARN_ON(rq
->prev_mm
);
1896 rq
->prev_mm
= oldmm
;
1899 * Since the runqueue lock will be released by the next
1900 * task (which is an invalid locking op but in the case
1901 * of the scheduler it's an obvious special-case), so we
1902 * do an early lockdep release here:
1904 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1905 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1908 /* Here we just switch the register state and the stack. */
1909 switch_to(prev
, next
, prev
);
1915 * nr_running, nr_uninterruptible and nr_context_switches:
1917 * externally visible scheduler statistics: current number of runnable
1918 * threads, current number of uninterruptible-sleeping threads, total
1919 * number of context switches performed since bootup.
1921 unsigned long nr_running(void)
1923 unsigned long i
, sum
= 0;
1925 for_each_online_cpu(i
)
1926 sum
+= cpu_rq(i
)->nr_running
;
1931 unsigned long nr_uninterruptible(void)
1933 unsigned long i
, sum
= 0;
1935 for_each_possible_cpu(i
)
1936 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1939 * Since we read the counters lockless, it might be slightly
1940 * inaccurate. Do not allow it to go below zero though:
1942 if (unlikely((long)sum
< 0))
1948 unsigned long long nr_context_switches(void)
1951 unsigned long long sum
= 0;
1953 for_each_possible_cpu(i
)
1954 sum
+= cpu_rq(i
)->nr_switches
;
1959 unsigned long nr_iowait(void)
1961 unsigned long i
, sum
= 0;
1963 for_each_possible_cpu(i
)
1964 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1969 unsigned long nr_active(void)
1971 unsigned long i
, running
= 0, uninterruptible
= 0;
1973 for_each_online_cpu(i
) {
1974 running
+= cpu_rq(i
)->nr_running
;
1975 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1978 if (unlikely((long)uninterruptible
< 0))
1979 uninterruptible
= 0;
1981 return running
+ uninterruptible
;
1987 * Is this task likely cache-hot:
1990 task_hot(struct task_struct
*p
, unsigned long long now
, struct sched_domain
*sd
)
1992 return (long long)(now
- p
->last_ran
) < (long long)sd
->cache_hot_time
;
1996 * double_rq_lock - safely lock two runqueues
1998 * Note this does not disable interrupts like task_rq_lock,
1999 * you need to do so manually before calling.
2001 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2002 __acquires(rq1
->lock
)
2003 __acquires(rq2
->lock
)
2005 BUG_ON(!irqs_disabled());
2007 spin_lock(&rq1
->lock
);
2008 __acquire(rq2
->lock
); /* Fake it out ;) */
2011 spin_lock(&rq1
->lock
);
2012 spin_lock(&rq2
->lock
);
2014 spin_lock(&rq2
->lock
);
2015 spin_lock(&rq1
->lock
);
2021 * double_rq_unlock - safely unlock two runqueues
2023 * Note this does not restore interrupts like task_rq_unlock,
2024 * you need to do so manually after calling.
2026 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2027 __releases(rq1
->lock
)
2028 __releases(rq2
->lock
)
2030 spin_unlock(&rq1
->lock
);
2032 spin_unlock(&rq2
->lock
);
2034 __release(rq2
->lock
);
2038 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2040 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2041 __releases(this_rq
->lock
)
2042 __acquires(busiest
->lock
)
2043 __acquires(this_rq
->lock
)
2045 if (unlikely(!irqs_disabled())) {
2046 /* printk() doesn't work good under rq->lock */
2047 spin_unlock(&this_rq
->lock
);
2050 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2051 if (busiest
< this_rq
) {
2052 spin_unlock(&this_rq
->lock
);
2053 spin_lock(&busiest
->lock
);
2054 spin_lock(&this_rq
->lock
);
2056 spin_lock(&busiest
->lock
);
2061 * If dest_cpu is allowed for this process, migrate the task to it.
2062 * This is accomplished by forcing the cpu_allowed mask to only
2063 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2064 * the cpu_allowed mask is restored.
2066 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2068 struct migration_req req
;
2069 unsigned long flags
;
2072 rq
= task_rq_lock(p
, &flags
);
2073 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2074 || unlikely(cpu_is_offline(dest_cpu
)))
2077 /* force the process onto the specified CPU */
2078 if (migrate_task(p
, dest_cpu
, &req
)) {
2079 /* Need to wait for migration thread (might exit: take ref). */
2080 struct task_struct
*mt
= rq
->migration_thread
;
2082 get_task_struct(mt
);
2083 task_rq_unlock(rq
, &flags
);
2084 wake_up_process(mt
);
2085 put_task_struct(mt
);
2086 wait_for_completion(&req
.done
);
2091 task_rq_unlock(rq
, &flags
);
2095 * sched_exec - execve() is a valuable balancing opportunity, because at
2096 * this point the task has the smallest effective memory and cache footprint.
2098 void sched_exec(void)
2100 int new_cpu
, this_cpu
= get_cpu();
2101 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2103 if (new_cpu
!= this_cpu
)
2104 sched_migrate_task(current
, new_cpu
);
2108 * pull_task - move a task from a remote runqueue to the local runqueue.
2109 * Both runqueues must be locked.
2111 static void pull_task(struct rq
*src_rq
, struct prio_array
*src_array
,
2112 struct task_struct
*p
, struct rq
*this_rq
,
2113 struct prio_array
*this_array
, int this_cpu
)
2115 dequeue_task(p
, src_array
);
2116 dec_nr_running(p
, src_rq
);
2117 set_task_cpu(p
, this_cpu
);
2118 inc_nr_running(p
, this_rq
);
2119 enqueue_task(p
, this_array
);
2120 p
->timestamp
= (p
->timestamp
- src_rq
->most_recent_timestamp
)
2121 + this_rq
->most_recent_timestamp
;
2123 * Note that idle threads have a prio of MAX_PRIO, for this test
2124 * to be always true for them.
2126 if (TASK_PREEMPTS_CURR(p
, this_rq
))
2127 resched_task(this_rq
->curr
);
2131 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2134 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2135 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2139 * We do not migrate tasks that are:
2140 * 1) running (obviously), or
2141 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2142 * 3) are cache-hot on their current CPU.
2144 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2148 if (task_running(rq
, p
))
2152 * Aggressive migration if:
2153 * 1) task is cache cold, or
2154 * 2) too many balance attempts have failed.
2157 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2158 #ifdef CONFIG_SCHEDSTATS
2159 if (task_hot(p
, rq
->most_recent_timestamp
, sd
))
2160 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2165 if (task_hot(p
, rq
->most_recent_timestamp
, sd
))
2170 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2173 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2174 * load from busiest to this_rq, as part of a balancing operation within
2175 * "domain". Returns the number of tasks moved.
2177 * Called with both runqueues locked.
2179 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2180 unsigned long max_nr_move
, unsigned long max_load_move
,
2181 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2184 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, best_prio
,
2185 best_prio_seen
, skip_for_load
;
2186 struct prio_array
*array
, *dst_array
;
2187 struct list_head
*head
, *curr
;
2188 struct task_struct
*tmp
;
2191 if (max_nr_move
== 0 || max_load_move
== 0)
2194 rem_load_move
= max_load_move
;
2196 this_best_prio
= rq_best_prio(this_rq
);
2197 best_prio
= rq_best_prio(busiest
);
2199 * Enable handling of the case where there is more than one task
2200 * with the best priority. If the current running task is one
2201 * of those with prio==best_prio we know it won't be moved
2202 * and therefore it's safe to override the skip (based on load) of
2203 * any task we find with that prio.
2205 best_prio_seen
= best_prio
== busiest
->curr
->prio
;
2208 * We first consider expired tasks. Those will likely not be
2209 * executed in the near future, and they are most likely to
2210 * be cache-cold, thus switching CPUs has the least effect
2213 if (busiest
->expired
->nr_active
) {
2214 array
= busiest
->expired
;
2215 dst_array
= this_rq
->expired
;
2217 array
= busiest
->active
;
2218 dst_array
= this_rq
->active
;
2222 /* Start searching at priority 0: */
2226 idx
= sched_find_first_bit(array
->bitmap
);
2228 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2229 if (idx
>= MAX_PRIO
) {
2230 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2231 array
= busiest
->active
;
2232 dst_array
= this_rq
->active
;
2238 head
= array
->queue
+ idx
;
2241 tmp
= list_entry(curr
, struct task_struct
, run_list
);
2246 * To help distribute high priority tasks accross CPUs we don't
2247 * skip a task if it will be the highest priority task (i.e. smallest
2248 * prio value) on its new queue regardless of its load weight
2250 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2251 if (skip_for_load
&& idx
< this_best_prio
)
2252 skip_for_load
= !best_prio_seen
&& idx
== best_prio
;
2253 if (skip_for_load
||
2254 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2256 best_prio_seen
|= idx
== best_prio
;
2263 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2265 rem_load_move
-= tmp
->load_weight
;
2268 * We only want to steal up to the prescribed number of tasks
2269 * and the prescribed amount of weighted load.
2271 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2272 if (idx
< this_best_prio
)
2273 this_best_prio
= idx
;
2281 * Right now, this is the only place pull_task() is called,
2282 * so we can safely collect pull_task() stats here rather than
2283 * inside pull_task().
2285 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2288 *all_pinned
= pinned
;
2293 * find_busiest_group finds and returns the busiest CPU group within the
2294 * domain. It calculates and returns the amount of weighted load which
2295 * should be moved to restore balance via the imbalance parameter.
2297 static struct sched_group
*
2298 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2299 unsigned long *imbalance
, enum cpu_idle_type idle
, int *sd_idle
,
2300 cpumask_t
*cpus
, int *balance
)
2302 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2303 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2304 unsigned long max_pull
;
2305 unsigned long busiest_load_per_task
, busiest_nr_running
;
2306 unsigned long this_load_per_task
, this_nr_running
;
2308 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2309 int power_savings_balance
= 1;
2310 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2311 unsigned long min_nr_running
= ULONG_MAX
;
2312 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2315 max_load
= this_load
= total_load
= total_pwr
= 0;
2316 busiest_load_per_task
= busiest_nr_running
= 0;
2317 this_load_per_task
= this_nr_running
= 0;
2318 if (idle
== CPU_NOT_IDLE
)
2319 load_idx
= sd
->busy_idx
;
2320 else if (idle
== CPU_NEWLY_IDLE
)
2321 load_idx
= sd
->newidle_idx
;
2323 load_idx
= sd
->idle_idx
;
2326 unsigned long load
, group_capacity
;
2329 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2330 unsigned long sum_nr_running
, sum_weighted_load
;
2332 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2335 balance_cpu
= first_cpu(group
->cpumask
);
2337 /* Tally up the load of all CPUs in the group */
2338 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2340 for_each_cpu_mask(i
, group
->cpumask
) {
2343 if (!cpu_isset(i
, *cpus
))
2348 if (*sd_idle
&& !idle_cpu(i
))
2351 /* Bias balancing toward cpus of our domain */
2353 if (idle_cpu(i
) && !first_idle_cpu
) {
2358 load
= target_load(i
, load_idx
);
2360 load
= source_load(i
, load_idx
);
2363 sum_nr_running
+= rq
->nr_running
;
2364 sum_weighted_load
+= rq
->raw_weighted_load
;
2368 * First idle cpu or the first cpu(busiest) in this sched group
2369 * is eligible for doing load balancing at this and above
2372 if (local_group
&& balance_cpu
!= this_cpu
&& balance
) {
2377 total_load
+= avg_load
;
2378 total_pwr
+= group
->__cpu_power
;
2380 /* Adjust by relative CPU power of the group */
2381 avg_load
= sg_div_cpu_power(group
,
2382 avg_load
* SCHED_LOAD_SCALE
);
2384 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2387 this_load
= avg_load
;
2389 this_nr_running
= sum_nr_running
;
2390 this_load_per_task
= sum_weighted_load
;
2391 } else if (avg_load
> max_load
&&
2392 sum_nr_running
> group_capacity
) {
2393 max_load
= avg_load
;
2395 busiest_nr_running
= sum_nr_running
;
2396 busiest_load_per_task
= sum_weighted_load
;
2399 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2401 * Busy processors will not participate in power savings
2404 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2408 * If the local group is idle or completely loaded
2409 * no need to do power savings balance at this domain
2411 if (local_group
&& (this_nr_running
>= group_capacity
||
2413 power_savings_balance
= 0;
2416 * If a group is already running at full capacity or idle,
2417 * don't include that group in power savings calculations
2419 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2424 * Calculate the group which has the least non-idle load.
2425 * This is the group from where we need to pick up the load
2428 if ((sum_nr_running
< min_nr_running
) ||
2429 (sum_nr_running
== min_nr_running
&&
2430 first_cpu(group
->cpumask
) <
2431 first_cpu(group_min
->cpumask
))) {
2433 min_nr_running
= sum_nr_running
;
2434 min_load_per_task
= sum_weighted_load
/
2439 * Calculate the group which is almost near its
2440 * capacity but still has some space to pick up some load
2441 * from other group and save more power
2443 if (sum_nr_running
<= group_capacity
- 1) {
2444 if (sum_nr_running
> leader_nr_running
||
2445 (sum_nr_running
== leader_nr_running
&&
2446 first_cpu(group
->cpumask
) >
2447 first_cpu(group_leader
->cpumask
))) {
2448 group_leader
= group
;
2449 leader_nr_running
= sum_nr_running
;
2454 group
= group
->next
;
2455 } while (group
!= sd
->groups
);
2457 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2460 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2462 if (this_load
>= avg_load
||
2463 100*max_load
<= sd
->imbalance_pct
*this_load
)
2466 busiest_load_per_task
/= busiest_nr_running
;
2468 * We're trying to get all the cpus to the average_load, so we don't
2469 * want to push ourselves above the average load, nor do we wish to
2470 * reduce the max loaded cpu below the average load, as either of these
2471 * actions would just result in more rebalancing later, and ping-pong
2472 * tasks around. Thus we look for the minimum possible imbalance.
2473 * Negative imbalances (*we* are more loaded than anyone else) will
2474 * be counted as no imbalance for these purposes -- we can't fix that
2475 * by pulling tasks to us. Be careful of negative numbers as they'll
2476 * appear as very large values with unsigned longs.
2478 if (max_load
<= busiest_load_per_task
)
2482 * In the presence of smp nice balancing, certain scenarios can have
2483 * max load less than avg load(as we skip the groups at or below
2484 * its cpu_power, while calculating max_load..)
2486 if (max_load
< avg_load
) {
2488 goto small_imbalance
;
2491 /* Don't want to pull so many tasks that a group would go idle */
2492 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2494 /* How much load to actually move to equalise the imbalance */
2495 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2496 (avg_load
- this_load
) * this->__cpu_power
)
2500 * if *imbalance is less than the average load per runnable task
2501 * there is no gaurantee that any tasks will be moved so we'll have
2502 * a think about bumping its value to force at least one task to be
2505 if (*imbalance
< busiest_load_per_task
) {
2506 unsigned long tmp
, pwr_now
, pwr_move
;
2510 pwr_move
= pwr_now
= 0;
2512 if (this_nr_running
) {
2513 this_load_per_task
/= this_nr_running
;
2514 if (busiest_load_per_task
> this_load_per_task
)
2517 this_load_per_task
= SCHED_LOAD_SCALE
;
2519 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2520 *imbalance
= busiest_load_per_task
;
2525 * OK, we don't have enough imbalance to justify moving tasks,
2526 * however we may be able to increase total CPU power used by
2530 pwr_now
+= busiest
->__cpu_power
*
2531 min(busiest_load_per_task
, max_load
);
2532 pwr_now
+= this->__cpu_power
*
2533 min(this_load_per_task
, this_load
);
2534 pwr_now
/= SCHED_LOAD_SCALE
;
2536 /* Amount of load we'd subtract */
2537 tmp
= sg_div_cpu_power(busiest
,
2538 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2540 pwr_move
+= busiest
->__cpu_power
*
2541 min(busiest_load_per_task
, max_load
- tmp
);
2543 /* Amount of load we'd add */
2544 if (max_load
* busiest
->__cpu_power
<
2545 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2546 tmp
= sg_div_cpu_power(this,
2547 max_load
* busiest
->__cpu_power
);
2549 tmp
= sg_div_cpu_power(this,
2550 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2551 pwr_move
+= this->__cpu_power
*
2552 min(this_load_per_task
, this_load
+ tmp
);
2553 pwr_move
/= SCHED_LOAD_SCALE
;
2555 /* Move if we gain throughput */
2556 if (pwr_move
<= pwr_now
)
2559 *imbalance
= busiest_load_per_task
;
2565 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2566 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2569 if (this == group_leader
&& group_leader
!= group_min
) {
2570 *imbalance
= min_load_per_task
;
2580 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2583 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2584 unsigned long imbalance
, cpumask_t
*cpus
)
2586 struct rq
*busiest
= NULL
, *rq
;
2587 unsigned long max_load
= 0;
2590 for_each_cpu_mask(i
, group
->cpumask
) {
2592 if (!cpu_isset(i
, *cpus
))
2597 if (rq
->nr_running
== 1 && rq
->raw_weighted_load
> imbalance
)
2600 if (rq
->raw_weighted_load
> max_load
) {
2601 max_load
= rq
->raw_weighted_load
;
2610 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2611 * so long as it is large enough.
2613 #define MAX_PINNED_INTERVAL 512
2615 static inline unsigned long minus_1_or_zero(unsigned long n
)
2617 return n
> 0 ? n
- 1 : 0;
2621 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2622 * tasks if there is an imbalance.
2624 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2625 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2628 int nr_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2629 struct sched_group
*group
;
2630 unsigned long imbalance
;
2632 cpumask_t cpus
= CPU_MASK_ALL
;
2633 unsigned long flags
;
2636 * When power savings policy is enabled for the parent domain, idle
2637 * sibling can pick up load irrespective of busy siblings. In this case,
2638 * let the state of idle sibling percolate up as IDLE, instead of
2639 * portraying it as CPU_NOT_IDLE.
2641 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2642 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2645 schedstat_inc(sd
, lb_cnt
[idle
]);
2648 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2655 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2659 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2661 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2665 BUG_ON(busiest
== this_rq
);
2667 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2670 if (busiest
->nr_running
> 1) {
2672 * Attempt to move tasks. If find_busiest_group has found
2673 * an imbalance but busiest->nr_running <= 1, the group is
2674 * still unbalanced. nr_moved simply stays zero, so it is
2675 * correctly treated as an imbalance.
2677 local_irq_save(flags
);
2678 double_rq_lock(this_rq
, busiest
);
2679 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2680 minus_1_or_zero(busiest
->nr_running
),
2681 imbalance
, sd
, idle
, &all_pinned
);
2682 double_rq_unlock(this_rq
, busiest
);
2683 local_irq_restore(flags
);
2686 * some other cpu did the load balance for us.
2688 if (nr_moved
&& this_cpu
!= smp_processor_id())
2689 resched_cpu(this_cpu
);
2691 /* All tasks on this runqueue were pinned by CPU affinity */
2692 if (unlikely(all_pinned
)) {
2693 cpu_clear(cpu_of(busiest
), cpus
);
2694 if (!cpus_empty(cpus
))
2701 schedstat_inc(sd
, lb_failed
[idle
]);
2702 sd
->nr_balance_failed
++;
2704 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2706 spin_lock_irqsave(&busiest
->lock
, flags
);
2708 /* don't kick the migration_thread, if the curr
2709 * task on busiest cpu can't be moved to this_cpu
2711 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2712 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2714 goto out_one_pinned
;
2717 if (!busiest
->active_balance
) {
2718 busiest
->active_balance
= 1;
2719 busiest
->push_cpu
= this_cpu
;
2722 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2724 wake_up_process(busiest
->migration_thread
);
2727 * We've kicked active balancing, reset the failure
2730 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2733 sd
->nr_balance_failed
= 0;
2735 if (likely(!active_balance
)) {
2736 /* We were unbalanced, so reset the balancing interval */
2737 sd
->balance_interval
= sd
->min_interval
;
2740 * If we've begun active balancing, start to back off. This
2741 * case may not be covered by the all_pinned logic if there
2742 * is only 1 task on the busy runqueue (because we don't call
2745 if (sd
->balance_interval
< sd
->max_interval
)
2746 sd
->balance_interval
*= 2;
2749 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2750 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2755 schedstat_inc(sd
, lb_balanced
[idle
]);
2757 sd
->nr_balance_failed
= 0;
2760 /* tune up the balancing interval */
2761 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2762 (sd
->balance_interval
< sd
->max_interval
))
2763 sd
->balance_interval
*= 2;
2765 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2766 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2772 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2773 * tasks if there is an imbalance.
2775 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2776 * this_rq is locked.
2779 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2781 struct sched_group
*group
;
2782 struct rq
*busiest
= NULL
;
2783 unsigned long imbalance
;
2786 cpumask_t cpus
= CPU_MASK_ALL
;
2789 * When power savings policy is enabled for the parent domain, idle
2790 * sibling can pick up load irrespective of busy siblings. In this case,
2791 * let the state of idle sibling percolate up as IDLE, instead of
2792 * portraying it as CPU_NOT_IDLE.
2794 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2795 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2798 schedstat_inc(sd
, lb_cnt
[CPU_NEWLY_IDLE
]);
2800 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2801 &sd_idle
, &cpus
, NULL
);
2803 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2807 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2810 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2814 BUG_ON(busiest
== this_rq
);
2816 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2819 if (busiest
->nr_running
> 1) {
2820 /* Attempt to move tasks */
2821 double_lock_balance(this_rq
, busiest
);
2822 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2823 minus_1_or_zero(busiest
->nr_running
),
2824 imbalance
, sd
, CPU_NEWLY_IDLE
, NULL
);
2825 spin_unlock(&busiest
->lock
);
2828 cpu_clear(cpu_of(busiest
), cpus
);
2829 if (!cpus_empty(cpus
))
2835 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2836 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2837 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2840 sd
->nr_balance_failed
= 0;
2845 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2846 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2847 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2849 sd
->nr_balance_failed
= 0;
2855 * idle_balance is called by schedule() if this_cpu is about to become
2856 * idle. Attempts to pull tasks from other CPUs.
2858 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2860 struct sched_domain
*sd
;
2861 int pulled_task
= 0;
2862 unsigned long next_balance
= jiffies
+ 60 * HZ
;
2864 for_each_domain(this_cpu
, sd
) {
2865 unsigned long interval
;
2867 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2870 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2871 /* If we've pulled tasks over stop searching: */
2872 pulled_task
= load_balance_newidle(this_cpu
,
2875 interval
= msecs_to_jiffies(sd
->balance_interval
);
2876 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2877 next_balance
= sd
->last_balance
+ interval
;
2883 * We are going idle. next_balance may be set based on
2884 * a busy processor. So reset next_balance.
2886 this_rq
->next_balance
= next_balance
;
2890 * active_load_balance is run by migration threads. It pushes running tasks
2891 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2892 * running on each physical CPU where possible, and avoids physical /
2893 * logical imbalances.
2895 * Called with busiest_rq locked.
2897 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2899 int target_cpu
= busiest_rq
->push_cpu
;
2900 struct sched_domain
*sd
;
2901 struct rq
*target_rq
;
2903 /* Is there any task to move? */
2904 if (busiest_rq
->nr_running
<= 1)
2907 target_rq
= cpu_rq(target_cpu
);
2910 * This condition is "impossible", if it occurs
2911 * we need to fix it. Originally reported by
2912 * Bjorn Helgaas on a 128-cpu setup.
2914 BUG_ON(busiest_rq
== target_rq
);
2916 /* move a task from busiest_rq to target_rq */
2917 double_lock_balance(busiest_rq
, target_rq
);
2919 /* Search for an sd spanning us and the target CPU. */
2920 for_each_domain(target_cpu
, sd
) {
2921 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2922 cpu_isset(busiest_cpu
, sd
->span
))
2927 schedstat_inc(sd
, alb_cnt
);
2929 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2930 RTPRIO_TO_LOAD_WEIGHT(100), sd
, CPU_IDLE
,
2932 schedstat_inc(sd
, alb_pushed
);
2934 schedstat_inc(sd
, alb_failed
);
2936 spin_unlock(&target_rq
->lock
);
2939 static void update_load(struct rq
*this_rq
)
2941 unsigned long this_load
;
2942 unsigned int i
, scale
;
2944 this_load
= this_rq
->raw_weighted_load
;
2946 /* Update our load: */
2947 for (i
= 0, scale
= 1; i
< 3; i
++, scale
+= scale
) {
2948 unsigned long old_load
, new_load
;
2950 /* scale is effectively 1 << i now, and >> i divides by scale */
2952 old_load
= this_rq
->cpu_load
[i
];
2953 new_load
= this_load
;
2955 * Round up the averaging division if load is increasing. This
2956 * prevents us from getting stuck on 9 if the load is 10, for
2959 if (new_load
> old_load
)
2960 new_load
+= scale
-1;
2961 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2967 atomic_t load_balancer
;
2969 } nohz ____cacheline_aligned
= {
2970 .load_balancer
= ATOMIC_INIT(-1),
2971 .cpu_mask
= CPU_MASK_NONE
,
2975 * This routine will try to nominate the ilb (idle load balancing)
2976 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2977 * load balancing on behalf of all those cpus. If all the cpus in the system
2978 * go into this tickless mode, then there will be no ilb owner (as there is
2979 * no need for one) and all the cpus will sleep till the next wakeup event
2982 * For the ilb owner, tick is not stopped. And this tick will be used
2983 * for idle load balancing. ilb owner will still be part of
2986 * While stopping the tick, this cpu will become the ilb owner if there
2987 * is no other owner. And will be the owner till that cpu becomes busy
2988 * or if all cpus in the system stop their ticks at which point
2989 * there is no need for ilb owner.
2991 * When the ilb owner becomes busy, it nominates another owner, during the
2992 * next busy scheduler_tick()
2994 int select_nohz_load_balancer(int stop_tick
)
2996 int cpu
= smp_processor_id();
2999 cpu_set(cpu
, nohz
.cpu_mask
);
3000 cpu_rq(cpu
)->in_nohz_recently
= 1;
3003 * If we are going offline and still the leader, give up!
3005 if (cpu_is_offline(cpu
) &&
3006 atomic_read(&nohz
.load_balancer
) == cpu
) {
3007 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3012 /* time for ilb owner also to sleep */
3013 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3014 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3015 atomic_set(&nohz
.load_balancer
, -1);
3019 if (atomic_read(&nohz
.load_balancer
) == -1) {
3020 /* make me the ilb owner */
3021 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3023 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3026 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3029 cpu_clear(cpu
, nohz
.cpu_mask
);
3031 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3032 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3039 static DEFINE_SPINLOCK(balancing
);
3042 * It checks each scheduling domain to see if it is due to be balanced,
3043 * and initiates a balancing operation if so.
3045 * Balancing parameters are set up in arch_init_sched_domains.
3047 static inline void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3050 struct rq
*rq
= cpu_rq(cpu
);
3051 unsigned long interval
;
3052 struct sched_domain
*sd
;
3053 /* Earliest time when we have to do rebalance again */
3054 unsigned long next_balance
= jiffies
+ 60*HZ
;
3056 for_each_domain(cpu
, sd
) {
3057 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3060 interval
= sd
->balance_interval
;
3061 if (idle
!= CPU_IDLE
)
3062 interval
*= sd
->busy_factor
;
3064 /* scale ms to jiffies */
3065 interval
= msecs_to_jiffies(interval
);
3066 if (unlikely(!interval
))
3069 if (sd
->flags
& SD_SERIALIZE
) {
3070 if (!spin_trylock(&balancing
))
3074 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3075 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3077 * We've pulled tasks over so either we're no
3078 * longer idle, or one of our SMT siblings is
3081 idle
= CPU_NOT_IDLE
;
3083 sd
->last_balance
= jiffies
;
3085 if (sd
->flags
& SD_SERIALIZE
)
3086 spin_unlock(&balancing
);
3088 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3089 next_balance
= sd
->last_balance
+ interval
;
3092 * Stop the load balance at this level. There is another
3093 * CPU in our sched group which is doing load balancing more
3099 rq
->next_balance
= next_balance
;
3103 * run_rebalance_domains is triggered when needed from the scheduler tick.
3104 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3105 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3107 static void run_rebalance_domains(struct softirq_action
*h
)
3109 int local_cpu
= smp_processor_id();
3110 struct rq
*local_rq
= cpu_rq(local_cpu
);
3111 enum cpu_idle_type idle
= local_rq
->idle_at_tick
? CPU_IDLE
: CPU_NOT_IDLE
;
3113 rebalance_domains(local_cpu
, idle
);
3117 * If this cpu is the owner for idle load balancing, then do the
3118 * balancing on behalf of the other idle cpus whose ticks are
3121 if (local_rq
->idle_at_tick
&&
3122 atomic_read(&nohz
.load_balancer
) == local_cpu
) {
3123 cpumask_t cpus
= nohz
.cpu_mask
;
3127 cpu_clear(local_cpu
, cpus
);
3128 for_each_cpu_mask(balance_cpu
, cpus
) {
3130 * If this cpu gets work to do, stop the load balancing
3131 * work being done for other cpus. Next load
3132 * balancing owner will pick it up.
3137 rebalance_domains(balance_cpu
, CPU_IDLE
);
3139 rq
= cpu_rq(balance_cpu
);
3140 if (time_after(local_rq
->next_balance
, rq
->next_balance
))
3141 local_rq
->next_balance
= rq
->next_balance
;
3148 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3150 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3151 * idle load balancing owner or decide to stop the periodic load balancing,
3152 * if the whole system is idle.
3154 static inline void trigger_load_balance(int cpu
)
3156 struct rq
*rq
= cpu_rq(cpu
);
3159 * If we were in the nohz mode recently and busy at the current
3160 * scheduler tick, then check if we need to nominate new idle
3163 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3164 rq
->in_nohz_recently
= 0;
3166 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3167 cpu_clear(cpu
, nohz
.cpu_mask
);
3168 atomic_set(&nohz
.load_balancer
, -1);
3171 if (atomic_read(&nohz
.load_balancer
) == -1) {
3173 * simple selection for now: Nominate the
3174 * first cpu in the nohz list to be the next
3177 * TBD: Traverse the sched domains and nominate
3178 * the nearest cpu in the nohz.cpu_mask.
3180 int ilb
= first_cpu(nohz
.cpu_mask
);
3188 * If this cpu is idle and doing idle load balancing for all the
3189 * cpus with ticks stopped, is it time for that to stop?
3191 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3192 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3198 * If this cpu is idle and the idle load balancing is done by
3199 * someone else, then no need raise the SCHED_SOFTIRQ
3201 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3202 cpu_isset(cpu
, nohz
.cpu_mask
))
3205 if (time_after_eq(jiffies
, rq
->next_balance
))
3206 raise_softirq(SCHED_SOFTIRQ
);
3210 * on UP we do not need to balance between CPUs:
3212 static inline void idle_balance(int cpu
, struct rq
*rq
)
3217 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3219 EXPORT_PER_CPU_SYMBOL(kstat
);
3222 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3223 * that have not yet been banked in case the task is currently running.
3225 unsigned long long task_sched_runtime(struct task_struct
*p
)
3227 unsigned long flags
;
3231 rq
= task_rq_lock(p
, &flags
);
3232 ns
= p
->se
.sum_exec_runtime
;
3233 if (rq
->curr
== p
) {
3234 delta_exec
= rq_clock(rq
) - p
->se
.exec_start
;
3235 if ((s64
)delta_exec
> 0)
3238 task_rq_unlock(rq
, &flags
);
3244 * We place interactive tasks back into the active array, if possible.
3246 * To guarantee that this does not starve expired tasks we ignore the
3247 * interactivity of a task if the first expired task had to wait more
3248 * than a 'reasonable' amount of time. This deadline timeout is
3249 * load-dependent, as the frequency of array switched decreases with
3250 * increasing number of running tasks. We also ignore the interactivity
3251 * if a better static_prio task has expired:
3253 static inline int expired_starving(struct rq
*rq
)
3255 if (rq
->curr
->static_prio
> rq
->best_expired_prio
)
3257 if (!STARVATION_LIMIT
|| !rq
->expired_timestamp
)
3259 if (jiffies
- rq
->expired_timestamp
> STARVATION_LIMIT
* rq
->nr_running
)
3265 * Account user cpu time to a process.
3266 * @p: the process that the cpu time gets accounted to
3267 * @hardirq_offset: the offset to subtract from hardirq_count()
3268 * @cputime: the cpu time spent in user space since the last update
3270 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3272 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3275 p
->utime
= cputime_add(p
->utime
, cputime
);
3277 /* Add user time to cpustat. */
3278 tmp
= cputime_to_cputime64(cputime
);
3279 if (TASK_NICE(p
) > 0)
3280 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3282 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3286 * Account system cpu time to a process.
3287 * @p: the process that the cpu time gets accounted to
3288 * @hardirq_offset: the offset to subtract from hardirq_count()
3289 * @cputime: the cpu time spent in kernel space since the last update
3291 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3294 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3295 struct rq
*rq
= this_rq();
3298 p
->stime
= cputime_add(p
->stime
, cputime
);
3300 /* Add system time to cpustat. */
3301 tmp
= cputime_to_cputime64(cputime
);
3302 if (hardirq_count() - hardirq_offset
)
3303 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3304 else if (softirq_count())
3305 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3306 else if (p
!= rq
->idle
)
3307 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3308 else if (atomic_read(&rq
->nr_iowait
) > 0)
3309 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3311 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3312 /* Account for system time used */
3313 acct_update_integrals(p
);
3317 * Account for involuntary wait time.
3318 * @p: the process from which the cpu time has been stolen
3319 * @steal: the cpu time spent in involuntary wait
3321 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3323 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3324 cputime64_t tmp
= cputime_to_cputime64(steal
);
3325 struct rq
*rq
= this_rq();
3327 if (p
== rq
->idle
) {
3328 p
->stime
= cputime_add(p
->stime
, steal
);
3329 if (atomic_read(&rq
->nr_iowait
) > 0)
3330 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3332 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3334 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3337 static void task_running_tick(struct rq
*rq
, struct task_struct
*p
)
3339 if (p
->array
!= rq
->active
) {
3340 /* Task has expired but was not scheduled yet */
3341 set_tsk_need_resched(p
);
3344 spin_lock(&rq
->lock
);
3346 * The task was running during this tick - update the
3347 * time slice counter. Note: we do not update a thread's
3348 * priority until it either goes to sleep or uses up its
3349 * timeslice. This makes it possible for interactive tasks
3350 * to use up their timeslices at their highest priority levels.
3354 * RR tasks need a special form of timeslice management.
3355 * FIFO tasks have no timeslices.
3357 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
3358 p
->time_slice
= task_timeslice(p
);
3359 p
->first_time_slice
= 0;
3360 set_tsk_need_resched(p
);
3362 /* put it at the end of the queue: */
3363 requeue_task(p
, rq
->active
);
3367 if (!--p
->time_slice
) {
3368 dequeue_task(p
, rq
->active
);
3369 set_tsk_need_resched(p
);
3370 p
->prio
= effective_prio(p
);
3371 p
->time_slice
= task_timeslice(p
);
3372 p
->first_time_slice
= 0;
3374 if (!rq
->expired_timestamp
)
3375 rq
->expired_timestamp
= jiffies
;
3376 if (!TASK_INTERACTIVE(p
) || expired_starving(rq
)) {
3377 enqueue_task(p
, rq
->expired
);
3378 if (p
->static_prio
< rq
->best_expired_prio
)
3379 rq
->best_expired_prio
= p
->static_prio
;
3381 enqueue_task(p
, rq
->active
);
3384 * Prevent a too long timeslice allowing a task to monopolize
3385 * the CPU. We do this by splitting up the timeslice into
3388 * Note: this does not mean the task's timeslices expire or
3389 * get lost in any way, they just might be preempted by
3390 * another task of equal priority. (one with higher
3391 * priority would have preempted this task already.) We
3392 * requeue this task to the end of the list on this priority
3393 * level, which is in essence a round-robin of tasks with
3396 * This only applies to tasks in the interactive
3397 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3399 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
3400 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
3401 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
3402 (p
->array
== rq
->active
)) {
3404 requeue_task(p
, rq
->active
);
3405 set_tsk_need_resched(p
);
3409 spin_unlock(&rq
->lock
);
3413 * This function gets called by the timer code, with HZ frequency.
3414 * We call it with interrupts disabled.
3416 * It also gets called by the fork code, when changing the parent's
3419 void scheduler_tick(void)
3421 struct task_struct
*p
= current
;
3422 int cpu
= smp_processor_id();
3423 int idle_at_tick
= idle_cpu(cpu
);
3424 struct rq
*rq
= cpu_rq(cpu
);
3427 task_running_tick(rq
, p
);
3430 rq
->idle_at_tick
= idle_at_tick
;
3431 trigger_load_balance(cpu
);
3435 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3437 void fastcall
add_preempt_count(int val
)
3442 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3444 preempt_count() += val
;
3446 * Spinlock count overflowing soon?
3448 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3451 EXPORT_SYMBOL(add_preempt_count
);
3453 void fastcall
sub_preempt_count(int val
)
3458 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3461 * Is the spinlock portion underflowing?
3463 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3464 !(preempt_count() & PREEMPT_MASK
)))
3467 preempt_count() -= val
;
3469 EXPORT_SYMBOL(sub_preempt_count
);
3474 * schedule() is the main scheduler function.
3476 asmlinkage
void __sched
schedule(void)
3478 struct task_struct
*prev
, *next
;
3479 struct prio_array
*array
;
3480 struct list_head
*queue
;
3481 unsigned long long now
;
3482 unsigned long run_time
;
3488 * Test if we are atomic. Since do_exit() needs to call into
3489 * schedule() atomically, we ignore that path for now.
3490 * Otherwise, whine if we are scheduling when we should not be.
3492 if (unlikely(in_atomic() && !current
->exit_state
)) {
3493 printk(KERN_ERR
"BUG: scheduling while atomic: "
3495 current
->comm
, preempt_count(), current
->pid
);
3496 debug_show_held_locks(current
);
3497 if (irqs_disabled())
3498 print_irqtrace_events(current
);
3501 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3506 release_kernel_lock(prev
);
3507 need_resched_nonpreemptible
:
3511 * The idle thread is not allowed to schedule!
3512 * Remove this check after it has been exercised a bit.
3514 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3515 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3519 schedstat_inc(rq
, sched_cnt
);
3520 now
= sched_clock();
3521 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3522 run_time
= now
- prev
->timestamp
;
3523 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3526 run_time
= NS_MAX_SLEEP_AVG
;
3529 * Tasks charged proportionately less run_time at high sleep_avg to
3530 * delay them losing their interactive status
3532 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3534 spin_lock_irq(&rq
->lock
);
3536 switch_count
= &prev
->nivcsw
;
3537 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3538 switch_count
= &prev
->nvcsw
;
3539 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3540 unlikely(signal_pending(prev
))))
3541 prev
->state
= TASK_RUNNING
;
3543 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3544 rq
->nr_uninterruptible
++;
3545 deactivate_task(prev
, rq
);
3549 cpu
= smp_processor_id();
3550 if (unlikely(!rq
->nr_running
)) {
3551 idle_balance(cpu
, rq
);
3552 if (!rq
->nr_running
) {
3554 rq
->expired_timestamp
= 0;
3560 if (unlikely(!array
->nr_active
)) {
3562 * Switch the active and expired arrays.
3564 schedstat_inc(rq
, sched_switch
);
3565 rq
->active
= rq
->expired
;
3566 rq
->expired
= array
;
3568 rq
->expired_timestamp
= 0;
3569 rq
->best_expired_prio
= MAX_PRIO
;
3572 idx
= sched_find_first_bit(array
->bitmap
);
3573 queue
= array
->queue
+ idx
;
3574 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
3577 if (next
== rq
->idle
)
3578 schedstat_inc(rq
, sched_goidle
);
3580 prefetch_stack(next
);
3581 clear_tsk_need_resched(prev
);
3582 rcu_qsctr_inc(task_cpu(prev
));
3584 prev
->sleep_avg
-= run_time
;
3585 if ((long)prev
->sleep_avg
<= 0)
3586 prev
->sleep_avg
= 0;
3587 prev
->timestamp
= prev
->last_ran
= now
;
3589 sched_info_switch(prev
, next
);
3590 if (likely(prev
!= next
)) {
3591 next
->timestamp
= next
->last_ran
= now
;
3596 prepare_task_switch(rq
, next
);
3597 prev
= context_switch(rq
, prev
, next
);
3600 * this_rq must be evaluated again because prev may have moved
3601 * CPUs since it called schedule(), thus the 'rq' on its stack
3602 * frame will be invalid.
3604 finish_task_switch(this_rq(), prev
);
3606 spin_unlock_irq(&rq
->lock
);
3609 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3610 goto need_resched_nonpreemptible
;
3611 preempt_enable_no_resched();
3612 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3615 EXPORT_SYMBOL(schedule
);
3617 #ifdef CONFIG_PREEMPT
3619 * this is the entry point to schedule() from in-kernel preemption
3620 * off of preempt_enable. Kernel preemptions off return from interrupt
3621 * occur there and call schedule directly.
3623 asmlinkage
void __sched
preempt_schedule(void)
3625 struct thread_info
*ti
= current_thread_info();
3626 #ifdef CONFIG_PREEMPT_BKL
3627 struct task_struct
*task
= current
;
3628 int saved_lock_depth
;
3631 * If there is a non-zero preempt_count or interrupts are disabled,
3632 * we do not want to preempt the current task. Just return..
3634 if (likely(ti
->preempt_count
|| irqs_disabled()))
3638 add_preempt_count(PREEMPT_ACTIVE
);
3640 * We keep the big kernel semaphore locked, but we
3641 * clear ->lock_depth so that schedule() doesnt
3642 * auto-release the semaphore:
3644 #ifdef CONFIG_PREEMPT_BKL
3645 saved_lock_depth
= task
->lock_depth
;
3646 task
->lock_depth
= -1;
3649 #ifdef CONFIG_PREEMPT_BKL
3650 task
->lock_depth
= saved_lock_depth
;
3652 sub_preempt_count(PREEMPT_ACTIVE
);
3654 /* we could miss a preemption opportunity between schedule and now */
3656 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3659 EXPORT_SYMBOL(preempt_schedule
);
3662 * this is the entry point to schedule() from kernel preemption
3663 * off of irq context.
3664 * Note, that this is called and return with irqs disabled. This will
3665 * protect us against recursive calling from irq.
3667 asmlinkage
void __sched
preempt_schedule_irq(void)
3669 struct thread_info
*ti
= current_thread_info();
3670 #ifdef CONFIG_PREEMPT_BKL
3671 struct task_struct
*task
= current
;
3672 int saved_lock_depth
;
3674 /* Catch callers which need to be fixed */
3675 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3678 add_preempt_count(PREEMPT_ACTIVE
);
3680 * We keep the big kernel semaphore locked, but we
3681 * clear ->lock_depth so that schedule() doesnt
3682 * auto-release the semaphore:
3684 #ifdef CONFIG_PREEMPT_BKL
3685 saved_lock_depth
= task
->lock_depth
;
3686 task
->lock_depth
= -1;
3690 local_irq_disable();
3691 #ifdef CONFIG_PREEMPT_BKL
3692 task
->lock_depth
= saved_lock_depth
;
3694 sub_preempt_count(PREEMPT_ACTIVE
);
3696 /* we could miss a preemption opportunity between schedule and now */
3698 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3702 #endif /* CONFIG_PREEMPT */
3704 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3707 return try_to_wake_up(curr
->private, mode
, sync
);
3709 EXPORT_SYMBOL(default_wake_function
);
3712 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3713 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3714 * number) then we wake all the non-exclusive tasks and one exclusive task.
3716 * There are circumstances in which we can try to wake a task which has already
3717 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3718 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3720 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3721 int nr_exclusive
, int sync
, void *key
)
3723 struct list_head
*tmp
, *next
;
3725 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3726 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3727 unsigned flags
= curr
->flags
;
3729 if (curr
->func(curr
, mode
, sync
, key
) &&
3730 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3736 * __wake_up - wake up threads blocked on a waitqueue.
3738 * @mode: which threads
3739 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3740 * @key: is directly passed to the wakeup function
3742 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3743 int nr_exclusive
, void *key
)
3745 unsigned long flags
;
3747 spin_lock_irqsave(&q
->lock
, flags
);
3748 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3749 spin_unlock_irqrestore(&q
->lock
, flags
);
3751 EXPORT_SYMBOL(__wake_up
);
3754 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3756 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3758 __wake_up_common(q
, mode
, 1, 0, NULL
);
3762 * __wake_up_sync - wake up threads blocked on a waitqueue.
3764 * @mode: which threads
3765 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3767 * The sync wakeup differs that the waker knows that it will schedule
3768 * away soon, so while the target thread will be woken up, it will not
3769 * be migrated to another CPU - ie. the two threads are 'synchronized'
3770 * with each other. This can prevent needless bouncing between CPUs.
3772 * On UP it can prevent extra preemption.
3775 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3777 unsigned long flags
;
3783 if (unlikely(!nr_exclusive
))
3786 spin_lock_irqsave(&q
->lock
, flags
);
3787 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3788 spin_unlock_irqrestore(&q
->lock
, flags
);
3790 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3792 void fastcall
complete(struct completion
*x
)
3794 unsigned long flags
;
3796 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3798 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3800 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3802 EXPORT_SYMBOL(complete
);
3804 void fastcall
complete_all(struct completion
*x
)
3806 unsigned long flags
;
3808 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3809 x
->done
+= UINT_MAX
/2;
3810 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3812 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3814 EXPORT_SYMBOL(complete_all
);
3816 void fastcall __sched
wait_for_completion(struct completion
*x
)
3820 spin_lock_irq(&x
->wait
.lock
);
3822 DECLARE_WAITQUEUE(wait
, current
);
3824 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3825 __add_wait_queue_tail(&x
->wait
, &wait
);
3827 __set_current_state(TASK_UNINTERRUPTIBLE
);
3828 spin_unlock_irq(&x
->wait
.lock
);
3830 spin_lock_irq(&x
->wait
.lock
);
3832 __remove_wait_queue(&x
->wait
, &wait
);
3835 spin_unlock_irq(&x
->wait
.lock
);
3837 EXPORT_SYMBOL(wait_for_completion
);
3839 unsigned long fastcall __sched
3840 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3844 spin_lock_irq(&x
->wait
.lock
);
3846 DECLARE_WAITQUEUE(wait
, current
);
3848 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3849 __add_wait_queue_tail(&x
->wait
, &wait
);
3851 __set_current_state(TASK_UNINTERRUPTIBLE
);
3852 spin_unlock_irq(&x
->wait
.lock
);
3853 timeout
= schedule_timeout(timeout
);
3854 spin_lock_irq(&x
->wait
.lock
);
3856 __remove_wait_queue(&x
->wait
, &wait
);
3860 __remove_wait_queue(&x
->wait
, &wait
);
3864 spin_unlock_irq(&x
->wait
.lock
);
3867 EXPORT_SYMBOL(wait_for_completion_timeout
);
3869 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3875 spin_lock_irq(&x
->wait
.lock
);
3877 DECLARE_WAITQUEUE(wait
, current
);
3879 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3880 __add_wait_queue_tail(&x
->wait
, &wait
);
3882 if (signal_pending(current
)) {
3884 __remove_wait_queue(&x
->wait
, &wait
);
3887 __set_current_state(TASK_INTERRUPTIBLE
);
3888 spin_unlock_irq(&x
->wait
.lock
);
3890 spin_lock_irq(&x
->wait
.lock
);
3892 __remove_wait_queue(&x
->wait
, &wait
);
3896 spin_unlock_irq(&x
->wait
.lock
);
3900 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3902 unsigned long fastcall __sched
3903 wait_for_completion_interruptible_timeout(struct completion
*x
,
3904 unsigned long timeout
)
3908 spin_lock_irq(&x
->wait
.lock
);
3910 DECLARE_WAITQUEUE(wait
, current
);
3912 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3913 __add_wait_queue_tail(&x
->wait
, &wait
);
3915 if (signal_pending(current
)) {
3916 timeout
= -ERESTARTSYS
;
3917 __remove_wait_queue(&x
->wait
, &wait
);
3920 __set_current_state(TASK_INTERRUPTIBLE
);
3921 spin_unlock_irq(&x
->wait
.lock
);
3922 timeout
= schedule_timeout(timeout
);
3923 spin_lock_irq(&x
->wait
.lock
);
3925 __remove_wait_queue(&x
->wait
, &wait
);
3929 __remove_wait_queue(&x
->wait
, &wait
);
3933 spin_unlock_irq(&x
->wait
.lock
);
3936 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3939 #define SLEEP_ON_VAR \
3940 unsigned long flags; \
3941 wait_queue_t wait; \
3942 init_waitqueue_entry(&wait, current);
3944 #define SLEEP_ON_HEAD \
3945 spin_lock_irqsave(&q->lock,flags); \
3946 __add_wait_queue(q, &wait); \
3947 spin_unlock(&q->lock);
3949 #define SLEEP_ON_TAIL \
3950 spin_lock_irq(&q->lock); \
3951 __remove_wait_queue(q, &wait); \
3952 spin_unlock_irqrestore(&q->lock, flags);
3954 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3958 current
->state
= TASK_INTERRUPTIBLE
;
3964 EXPORT_SYMBOL(interruptible_sleep_on
);
3966 long fastcall __sched
3967 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3971 current
->state
= TASK_INTERRUPTIBLE
;
3974 timeout
= schedule_timeout(timeout
);
3979 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3981 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3985 current
->state
= TASK_UNINTERRUPTIBLE
;
3991 EXPORT_SYMBOL(sleep_on
);
3993 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3997 current
->state
= TASK_UNINTERRUPTIBLE
;
4000 timeout
= schedule_timeout(timeout
);
4006 EXPORT_SYMBOL(sleep_on_timeout
);
4008 #ifdef CONFIG_RT_MUTEXES
4011 * rt_mutex_setprio - set the current priority of a task
4013 * @prio: prio value (kernel-internal form)
4015 * This function changes the 'effective' priority of a task. It does
4016 * not touch ->normal_prio like __setscheduler().
4018 * Used by the rt_mutex code to implement priority inheritance logic.
4020 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4022 struct prio_array
*array
;
4023 unsigned long flags
;
4027 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4029 rq
= task_rq_lock(p
, &flags
);
4034 dequeue_task(p
, array
);
4039 * If changing to an RT priority then queue it
4040 * in the active array!
4044 enqueue_task(p
, array
);
4046 * Reschedule if we are currently running on this runqueue and
4047 * our priority decreased, or if we are not currently running on
4048 * this runqueue and our priority is higher than the current's
4050 if (task_running(rq
, p
)) {
4051 if (p
->prio
> oldprio
)
4052 resched_task(rq
->curr
);
4053 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4054 resched_task(rq
->curr
);
4056 task_rq_unlock(rq
, &flags
);
4061 void set_user_nice(struct task_struct
*p
, long nice
)
4063 struct prio_array
*array
;
4064 int old_prio
, delta
;
4065 unsigned long flags
;
4068 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4071 * We have to be careful, if called from sys_setpriority(),
4072 * the task might be in the middle of scheduling on another CPU.
4074 rq
= task_rq_lock(p
, &flags
);
4076 * The RT priorities are set via sched_setscheduler(), but we still
4077 * allow the 'normal' nice value to be set - but as expected
4078 * it wont have any effect on scheduling until the task is
4079 * not SCHED_NORMAL/SCHED_BATCH:
4081 if (task_has_rt_policy(p
)) {
4082 p
->static_prio
= NICE_TO_PRIO(nice
);
4087 dequeue_task(p
, array
);
4088 dec_raw_weighted_load(rq
, p
);
4091 p
->static_prio
= NICE_TO_PRIO(nice
);
4094 p
->prio
= effective_prio(p
);
4095 delta
= p
->prio
- old_prio
;
4098 enqueue_task(p
, array
);
4099 inc_raw_weighted_load(rq
, p
);
4101 * If the task increased its priority or is running and
4102 * lowered its priority, then reschedule its CPU:
4104 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4105 resched_task(rq
->curr
);
4108 task_rq_unlock(rq
, &flags
);
4110 EXPORT_SYMBOL(set_user_nice
);
4113 * can_nice - check if a task can reduce its nice value
4117 int can_nice(const struct task_struct
*p
, const int nice
)
4119 /* convert nice value [19,-20] to rlimit style value [1,40] */
4120 int nice_rlim
= 20 - nice
;
4122 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4123 capable(CAP_SYS_NICE
));
4126 #ifdef __ARCH_WANT_SYS_NICE
4129 * sys_nice - change the priority of the current process.
4130 * @increment: priority increment
4132 * sys_setpriority is a more generic, but much slower function that
4133 * does similar things.
4135 asmlinkage
long sys_nice(int increment
)
4140 * Setpriority might change our priority at the same moment.
4141 * We don't have to worry. Conceptually one call occurs first
4142 * and we have a single winner.
4144 if (increment
< -40)
4149 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4155 if (increment
< 0 && !can_nice(current
, nice
))
4158 retval
= security_task_setnice(current
, nice
);
4162 set_user_nice(current
, nice
);
4169 * task_prio - return the priority value of a given task.
4170 * @p: the task in question.
4172 * This is the priority value as seen by users in /proc.
4173 * RT tasks are offset by -200. Normal tasks are centered
4174 * around 0, value goes from -16 to +15.
4176 int task_prio(const struct task_struct
*p
)
4178 return p
->prio
- MAX_RT_PRIO
;
4182 * task_nice - return the nice value of a given task.
4183 * @p: the task in question.
4185 int task_nice(const struct task_struct
*p
)
4187 return TASK_NICE(p
);
4189 EXPORT_SYMBOL_GPL(task_nice
);
4192 * idle_cpu - is a given cpu idle currently?
4193 * @cpu: the processor in question.
4195 int idle_cpu(int cpu
)
4197 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4201 * idle_task - return the idle task for a given cpu.
4202 * @cpu: the processor in question.
4204 struct task_struct
*idle_task(int cpu
)
4206 return cpu_rq(cpu
)->idle
;
4210 * find_process_by_pid - find a process with a matching PID value.
4211 * @pid: the pid in question.
4213 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4215 return pid
? find_task_by_pid(pid
) : current
;
4218 /* Actually do priority change: must hold rq lock. */
4219 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
4224 p
->rt_priority
= prio
;
4225 p
->normal_prio
= normal_prio(p
);
4226 /* we are holding p->pi_lock already */
4227 p
->prio
= rt_mutex_getprio(p
);
4229 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4231 if (policy
== SCHED_BATCH
)
4237 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4238 * @p: the task in question.
4239 * @policy: new policy.
4240 * @param: structure containing the new RT priority.
4242 * NOTE that the task may be already dead.
4244 int sched_setscheduler(struct task_struct
*p
, int policy
,
4245 struct sched_param
*param
)
4247 int retval
, oldprio
, oldpolicy
= -1;
4248 struct prio_array
*array
;
4249 unsigned long flags
;
4252 /* may grab non-irq protected spin_locks */
4253 BUG_ON(in_interrupt());
4255 /* double check policy once rq lock held */
4257 policy
= oldpolicy
= p
->policy
;
4258 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4259 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
4262 * Valid priorities for SCHED_FIFO and SCHED_RR are
4263 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4266 if (param
->sched_priority
< 0 ||
4267 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4268 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4270 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4274 * Allow unprivileged RT tasks to decrease priority:
4276 if (!capable(CAP_SYS_NICE
)) {
4277 if (rt_policy(policy
)) {
4278 unsigned long rlim_rtprio
;
4279 unsigned long flags
;
4281 if (!lock_task_sighand(p
, &flags
))
4283 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4284 unlock_task_sighand(p
, &flags
);
4286 /* can't set/change the rt policy */
4287 if (policy
!= p
->policy
&& !rlim_rtprio
)
4290 /* can't increase priority */
4291 if (param
->sched_priority
> p
->rt_priority
&&
4292 param
->sched_priority
> rlim_rtprio
)
4296 /* can't change other user's priorities */
4297 if ((current
->euid
!= p
->euid
) &&
4298 (current
->euid
!= p
->uid
))
4302 retval
= security_task_setscheduler(p
, policy
, param
);
4306 * make sure no PI-waiters arrive (or leave) while we are
4307 * changing the priority of the task:
4309 spin_lock_irqsave(&p
->pi_lock
, flags
);
4311 * To be able to change p->policy safely, the apropriate
4312 * runqueue lock must be held.
4314 rq
= __task_rq_lock(p
);
4315 /* recheck policy now with rq lock held */
4316 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4317 policy
= oldpolicy
= -1;
4318 __task_rq_unlock(rq
);
4319 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4324 deactivate_task(p
, rq
);
4326 __setscheduler(p
, policy
, param
->sched_priority
);
4328 __activate_task(p
, rq
);
4330 * Reschedule if we are currently running on this runqueue and
4331 * our priority decreased, or if we are not currently running on
4332 * this runqueue and our priority is higher than the current's
4334 if (task_running(rq
, p
)) {
4335 if (p
->prio
> oldprio
)
4336 resched_task(rq
->curr
);
4337 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4338 resched_task(rq
->curr
);
4340 __task_rq_unlock(rq
);
4341 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4343 rt_mutex_adjust_pi(p
);
4347 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4350 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4352 struct sched_param lparam
;
4353 struct task_struct
*p
;
4356 if (!param
|| pid
< 0)
4358 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4363 p
= find_process_by_pid(pid
);
4365 retval
= sched_setscheduler(p
, policy
, &lparam
);
4372 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4373 * @pid: the pid in question.
4374 * @policy: new policy.
4375 * @param: structure containing the new RT priority.
4377 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4378 struct sched_param __user
*param
)
4380 /* negative values for policy are not valid */
4384 return do_sched_setscheduler(pid
, policy
, param
);
4388 * sys_sched_setparam - set/change the RT priority of a thread
4389 * @pid: the pid in question.
4390 * @param: structure containing the new RT priority.
4392 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4394 return do_sched_setscheduler(pid
, -1, param
);
4398 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4399 * @pid: the pid in question.
4401 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4403 struct task_struct
*p
;
4404 int retval
= -EINVAL
;
4410 read_lock(&tasklist_lock
);
4411 p
= find_process_by_pid(pid
);
4413 retval
= security_task_getscheduler(p
);
4417 read_unlock(&tasklist_lock
);
4424 * sys_sched_getscheduler - get the RT priority of a thread
4425 * @pid: the pid in question.
4426 * @param: structure containing the RT priority.
4428 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4430 struct sched_param lp
;
4431 struct task_struct
*p
;
4432 int retval
= -EINVAL
;
4434 if (!param
|| pid
< 0)
4437 read_lock(&tasklist_lock
);
4438 p
= find_process_by_pid(pid
);
4443 retval
= security_task_getscheduler(p
);
4447 lp
.sched_priority
= p
->rt_priority
;
4448 read_unlock(&tasklist_lock
);
4451 * This one might sleep, we cannot do it with a spinlock held ...
4453 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4459 read_unlock(&tasklist_lock
);
4463 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4465 cpumask_t cpus_allowed
;
4466 struct task_struct
*p
;
4469 mutex_lock(&sched_hotcpu_mutex
);
4470 read_lock(&tasklist_lock
);
4472 p
= find_process_by_pid(pid
);
4474 read_unlock(&tasklist_lock
);
4475 mutex_unlock(&sched_hotcpu_mutex
);
4480 * It is not safe to call set_cpus_allowed with the
4481 * tasklist_lock held. We will bump the task_struct's
4482 * usage count and then drop tasklist_lock.
4485 read_unlock(&tasklist_lock
);
4488 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4489 !capable(CAP_SYS_NICE
))
4492 retval
= security_task_setscheduler(p
, 0, NULL
);
4496 cpus_allowed
= cpuset_cpus_allowed(p
);
4497 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4498 retval
= set_cpus_allowed(p
, new_mask
);
4502 mutex_unlock(&sched_hotcpu_mutex
);
4506 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4507 cpumask_t
*new_mask
)
4509 if (len
< sizeof(cpumask_t
)) {
4510 memset(new_mask
, 0, sizeof(cpumask_t
));
4511 } else if (len
> sizeof(cpumask_t
)) {
4512 len
= sizeof(cpumask_t
);
4514 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4518 * sys_sched_setaffinity - set the cpu affinity of a process
4519 * @pid: pid of the process
4520 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4521 * @user_mask_ptr: user-space pointer to the new cpu mask
4523 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4524 unsigned long __user
*user_mask_ptr
)
4529 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4533 return sched_setaffinity(pid
, new_mask
);
4537 * Represents all cpu's present in the system
4538 * In systems capable of hotplug, this map could dynamically grow
4539 * as new cpu's are detected in the system via any platform specific
4540 * method, such as ACPI for e.g.
4543 cpumask_t cpu_present_map __read_mostly
;
4544 EXPORT_SYMBOL(cpu_present_map
);
4547 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4548 EXPORT_SYMBOL(cpu_online_map
);
4550 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4551 EXPORT_SYMBOL(cpu_possible_map
);
4554 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4556 struct task_struct
*p
;
4559 mutex_lock(&sched_hotcpu_mutex
);
4560 read_lock(&tasklist_lock
);
4563 p
= find_process_by_pid(pid
);
4567 retval
= security_task_getscheduler(p
);
4571 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4574 read_unlock(&tasklist_lock
);
4575 mutex_unlock(&sched_hotcpu_mutex
);
4583 * sys_sched_getaffinity - get the cpu affinity of a process
4584 * @pid: pid of the process
4585 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4586 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4588 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4589 unsigned long __user
*user_mask_ptr
)
4594 if (len
< sizeof(cpumask_t
))
4597 ret
= sched_getaffinity(pid
, &mask
);
4601 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4604 return sizeof(cpumask_t
);
4608 * sys_sched_yield - yield the current processor to other threads.
4610 * This function yields the current CPU by moving the calling thread
4611 * to the expired array. If there are no other threads running on this
4612 * CPU then this function will return.
4614 asmlinkage
long sys_sched_yield(void)
4616 struct rq
*rq
= this_rq_lock();
4617 struct prio_array
*array
= current
->array
, *target
= rq
->expired
;
4619 schedstat_inc(rq
, yld_cnt
);
4621 * We implement yielding by moving the task into the expired
4624 * (special rule: RT tasks will just roundrobin in the active
4627 if (rt_task(current
))
4628 target
= rq
->active
;
4630 if (array
->nr_active
== 1) {
4631 schedstat_inc(rq
, yld_act_empty
);
4632 if (!rq
->expired
->nr_active
)
4633 schedstat_inc(rq
, yld_both_empty
);
4634 } else if (!rq
->expired
->nr_active
)
4635 schedstat_inc(rq
, yld_exp_empty
);
4637 if (array
!= target
) {
4638 dequeue_task(current
, array
);
4639 enqueue_task(current
, target
);
4642 * requeue_task is cheaper so perform that if possible.
4644 requeue_task(current
, array
);
4647 * Since we are going to call schedule() anyway, there's
4648 * no need to preempt or enable interrupts:
4650 __release(rq
->lock
);
4651 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4652 _raw_spin_unlock(&rq
->lock
);
4653 preempt_enable_no_resched();
4660 static void __cond_resched(void)
4662 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4663 __might_sleep(__FILE__
, __LINE__
);
4666 * The BKS might be reacquired before we have dropped
4667 * PREEMPT_ACTIVE, which could trigger a second
4668 * cond_resched() call.
4671 add_preempt_count(PREEMPT_ACTIVE
);
4673 sub_preempt_count(PREEMPT_ACTIVE
);
4674 } while (need_resched());
4677 int __sched
cond_resched(void)
4679 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4680 system_state
== SYSTEM_RUNNING
) {
4686 EXPORT_SYMBOL(cond_resched
);
4689 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4690 * call schedule, and on return reacquire the lock.
4692 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4693 * operations here to prevent schedule() from being called twice (once via
4694 * spin_unlock(), once by hand).
4696 int cond_resched_lock(spinlock_t
*lock
)
4700 if (need_lockbreak(lock
)) {
4706 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4707 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4708 _raw_spin_unlock(lock
);
4709 preempt_enable_no_resched();
4716 EXPORT_SYMBOL(cond_resched_lock
);
4718 int __sched
cond_resched_softirq(void)
4720 BUG_ON(!in_softirq());
4722 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4730 EXPORT_SYMBOL(cond_resched_softirq
);
4733 * yield - yield the current processor to other threads.
4735 * This is a shortcut for kernel-space yielding - it marks the
4736 * thread runnable and calls sys_sched_yield().
4738 void __sched
yield(void)
4740 set_current_state(TASK_RUNNING
);
4743 EXPORT_SYMBOL(yield
);
4746 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4747 * that process accounting knows that this is a task in IO wait state.
4749 * But don't do that if it is a deliberate, throttling IO wait (this task
4750 * has set its backing_dev_info: the queue against which it should throttle)
4752 void __sched
io_schedule(void)
4754 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4756 delayacct_blkio_start();
4757 atomic_inc(&rq
->nr_iowait
);
4759 atomic_dec(&rq
->nr_iowait
);
4760 delayacct_blkio_end();
4762 EXPORT_SYMBOL(io_schedule
);
4764 long __sched
io_schedule_timeout(long timeout
)
4766 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4769 delayacct_blkio_start();
4770 atomic_inc(&rq
->nr_iowait
);
4771 ret
= schedule_timeout(timeout
);
4772 atomic_dec(&rq
->nr_iowait
);
4773 delayacct_blkio_end();
4778 * sys_sched_get_priority_max - return maximum RT priority.
4779 * @policy: scheduling class.
4781 * this syscall returns the maximum rt_priority that can be used
4782 * by a given scheduling class.
4784 asmlinkage
long sys_sched_get_priority_max(int policy
)
4791 ret
= MAX_USER_RT_PRIO
-1;
4802 * sys_sched_get_priority_min - return minimum RT priority.
4803 * @policy: scheduling class.
4805 * this syscall returns the minimum rt_priority that can be used
4806 * by a given scheduling class.
4808 asmlinkage
long sys_sched_get_priority_min(int policy
)
4825 * sys_sched_rr_get_interval - return the default timeslice of a process.
4826 * @pid: pid of the process.
4827 * @interval: userspace pointer to the timeslice value.
4829 * this syscall writes the default timeslice value of a given process
4830 * into the user-space timespec buffer. A value of '0' means infinity.
4833 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4835 struct task_struct
*p
;
4836 int retval
= -EINVAL
;
4843 read_lock(&tasklist_lock
);
4844 p
= find_process_by_pid(pid
);
4848 retval
= security_task_getscheduler(p
);
4852 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4853 0 : task_timeslice(p
), &t
);
4854 read_unlock(&tasklist_lock
);
4855 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4859 read_unlock(&tasklist_lock
);
4863 static const char stat_nam
[] = "RSDTtZX";
4865 static void show_task(struct task_struct
*p
)
4867 unsigned long free
= 0;
4870 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4871 printk("%-13.13s %c", p
->comm
,
4872 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4873 #if (BITS_PER_LONG == 32)
4874 if (state
== TASK_RUNNING
)
4875 printk(" running ");
4877 printk(" %08lX ", thread_saved_pc(p
));
4879 if (state
== TASK_RUNNING
)
4880 printk(" running task ");
4882 printk(" %016lx ", thread_saved_pc(p
));
4884 #ifdef CONFIG_DEBUG_STACK_USAGE
4886 unsigned long *n
= end_of_stack(p
);
4889 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4892 printk("%5lu %5d %6d", free
, p
->pid
, p
->parent
->pid
);
4894 printk(" (L-TLB)\n");
4896 printk(" (NOTLB)\n");
4898 if (state
!= TASK_RUNNING
)
4899 show_stack(p
, NULL
);
4902 void show_state_filter(unsigned long state_filter
)
4904 struct task_struct
*g
, *p
;
4906 #if (BITS_PER_LONG == 32)
4909 printk(" task PC stack pid father child younger older\n");
4913 printk(" task PC stack pid father child younger older\n");
4915 read_lock(&tasklist_lock
);
4916 do_each_thread(g
, p
) {
4918 * reset the NMI-timeout, listing all files on a slow
4919 * console might take alot of time:
4921 touch_nmi_watchdog();
4922 if (!state_filter
|| (p
->state
& state_filter
))
4924 } while_each_thread(g
, p
);
4926 touch_all_softlockup_watchdogs();
4928 read_unlock(&tasklist_lock
);
4930 * Only show locks if all tasks are dumped:
4932 if (state_filter
== -1)
4933 debug_show_all_locks();
4936 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4942 * init_idle - set up an idle thread for a given CPU
4943 * @idle: task in question
4944 * @cpu: cpu the idle task belongs to
4946 * NOTE: this function does not set the idle thread's NEED_RESCHED
4947 * flag, to make booting more robust.
4949 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4951 struct rq
*rq
= cpu_rq(cpu
);
4952 unsigned long flags
;
4954 idle
->timestamp
= sched_clock();
4955 idle
->sleep_avg
= 0;
4957 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4958 idle
->state
= TASK_RUNNING
;
4959 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4960 set_task_cpu(idle
, cpu
);
4962 spin_lock_irqsave(&rq
->lock
, flags
);
4963 rq
->curr
= rq
->idle
= idle
;
4964 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4967 spin_unlock_irqrestore(&rq
->lock
, flags
);
4969 /* Set the preempt count _outside_ the spinlocks! */
4970 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4971 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4973 task_thread_info(idle
)->preempt_count
= 0;
4978 * In a system that switches off the HZ timer nohz_cpu_mask
4979 * indicates which cpus entered this state. This is used
4980 * in the rcu update to wait only for active cpus. For system
4981 * which do not switch off the HZ timer nohz_cpu_mask should
4982 * always be CPU_MASK_NONE.
4984 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4988 * This is how migration works:
4990 * 1) we queue a struct migration_req structure in the source CPU's
4991 * runqueue and wake up that CPU's migration thread.
4992 * 2) we down() the locked semaphore => thread blocks.
4993 * 3) migration thread wakes up (implicitly it forces the migrated
4994 * thread off the CPU)
4995 * 4) it gets the migration request and checks whether the migrated
4996 * task is still in the wrong runqueue.
4997 * 5) if it's in the wrong runqueue then the migration thread removes
4998 * it and puts it into the right queue.
4999 * 6) migration thread up()s the semaphore.
5000 * 7) we wake up and the migration is done.
5004 * Change a given task's CPU affinity. Migrate the thread to a
5005 * proper CPU and schedule it away if the CPU it's executing on
5006 * is removed from the allowed bitmask.
5008 * NOTE: the caller must have a valid reference to the task, the
5009 * task must not exit() & deallocate itself prematurely. The
5010 * call is not atomic; no spinlocks may be held.
5012 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5014 struct migration_req req
;
5015 unsigned long flags
;
5019 rq
= task_rq_lock(p
, &flags
);
5020 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5025 p
->cpus_allowed
= new_mask
;
5026 /* Can the task run on the task's current CPU? If so, we're done */
5027 if (cpu_isset(task_cpu(p
), new_mask
))
5030 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5031 /* Need help from migration thread: drop lock and wait. */
5032 task_rq_unlock(rq
, &flags
);
5033 wake_up_process(rq
->migration_thread
);
5034 wait_for_completion(&req
.done
);
5035 tlb_migrate_finish(p
->mm
);
5039 task_rq_unlock(rq
, &flags
);
5043 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5046 * Move (not current) task off this cpu, onto dest cpu. We're doing
5047 * this because either it can't run here any more (set_cpus_allowed()
5048 * away from this CPU, or CPU going down), or because we're
5049 * attempting to rebalance this task on exec (sched_exec).
5051 * So we race with normal scheduler movements, but that's OK, as long
5052 * as the task is no longer on this CPU.
5054 * Returns non-zero if task was successfully migrated.
5056 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5058 struct rq
*rq_dest
, *rq_src
;
5061 if (unlikely(cpu_is_offline(dest_cpu
)))
5064 rq_src
= cpu_rq(src_cpu
);
5065 rq_dest
= cpu_rq(dest_cpu
);
5067 double_rq_lock(rq_src
, rq_dest
);
5068 /* Already moved. */
5069 if (task_cpu(p
) != src_cpu
)
5071 /* Affinity changed (again). */
5072 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5075 set_task_cpu(p
, dest_cpu
);
5078 * Sync timestamp with rq_dest's before activating.
5079 * The same thing could be achieved by doing this step
5080 * afterwards, and pretending it was a local activate.
5081 * This way is cleaner and logically correct.
5083 p
->timestamp
= p
->timestamp
- rq_src
->most_recent_timestamp
5084 + rq_dest
->most_recent_timestamp
;
5085 deactivate_task(p
, rq_src
);
5086 __activate_task(p
, rq_dest
);
5087 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
5088 resched_task(rq_dest
->curr
);
5092 double_rq_unlock(rq_src
, rq_dest
);
5097 * migration_thread - this is a highprio system thread that performs
5098 * thread migration by bumping thread off CPU then 'pushing' onto
5101 static int migration_thread(void *data
)
5103 int cpu
= (long)data
;
5107 BUG_ON(rq
->migration_thread
!= current
);
5109 set_current_state(TASK_INTERRUPTIBLE
);
5110 while (!kthread_should_stop()) {
5111 struct migration_req
*req
;
5112 struct list_head
*head
;
5116 spin_lock_irq(&rq
->lock
);
5118 if (cpu_is_offline(cpu
)) {
5119 spin_unlock_irq(&rq
->lock
);
5123 if (rq
->active_balance
) {
5124 active_load_balance(rq
, cpu
);
5125 rq
->active_balance
= 0;
5128 head
= &rq
->migration_queue
;
5130 if (list_empty(head
)) {
5131 spin_unlock_irq(&rq
->lock
);
5133 set_current_state(TASK_INTERRUPTIBLE
);
5136 req
= list_entry(head
->next
, struct migration_req
, list
);
5137 list_del_init(head
->next
);
5139 spin_unlock(&rq
->lock
);
5140 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5143 complete(&req
->done
);
5145 __set_current_state(TASK_RUNNING
);
5149 /* Wait for kthread_stop */
5150 set_current_state(TASK_INTERRUPTIBLE
);
5151 while (!kthread_should_stop()) {
5153 set_current_state(TASK_INTERRUPTIBLE
);
5155 __set_current_state(TASK_RUNNING
);
5159 #ifdef CONFIG_HOTPLUG_CPU
5161 * Figure out where task on dead CPU should go, use force if neccessary.
5162 * NOTE: interrupts should be disabled by the caller
5164 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5166 unsigned long flags
;
5173 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5174 cpus_and(mask
, mask
, p
->cpus_allowed
);
5175 dest_cpu
= any_online_cpu(mask
);
5177 /* On any allowed CPU? */
5178 if (dest_cpu
== NR_CPUS
)
5179 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5181 /* No more Mr. Nice Guy. */
5182 if (dest_cpu
== NR_CPUS
) {
5183 rq
= task_rq_lock(p
, &flags
);
5184 cpus_setall(p
->cpus_allowed
);
5185 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5186 task_rq_unlock(rq
, &flags
);
5189 * Don't tell them about moving exiting tasks or
5190 * kernel threads (both mm NULL), since they never
5193 if (p
->mm
&& printk_ratelimit())
5194 printk(KERN_INFO
"process %d (%s) no "
5195 "longer affine to cpu%d\n",
5196 p
->pid
, p
->comm
, dead_cpu
);
5198 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5203 * While a dead CPU has no uninterruptible tasks queued at this point,
5204 * it might still have a nonzero ->nr_uninterruptible counter, because
5205 * for performance reasons the counter is not stricly tracking tasks to
5206 * their home CPUs. So we just add the counter to another CPU's counter,
5207 * to keep the global sum constant after CPU-down:
5209 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5211 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5212 unsigned long flags
;
5214 local_irq_save(flags
);
5215 double_rq_lock(rq_src
, rq_dest
);
5216 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5217 rq_src
->nr_uninterruptible
= 0;
5218 double_rq_unlock(rq_src
, rq_dest
);
5219 local_irq_restore(flags
);
5222 /* Run through task list and migrate tasks from the dead cpu. */
5223 static void migrate_live_tasks(int src_cpu
)
5225 struct task_struct
*p
, *t
;
5227 write_lock_irq(&tasklist_lock
);
5229 do_each_thread(t
, p
) {
5233 if (task_cpu(p
) == src_cpu
)
5234 move_task_off_dead_cpu(src_cpu
, p
);
5235 } while_each_thread(t
, p
);
5237 write_unlock_irq(&tasklist_lock
);
5240 /* Schedules idle task to be the next runnable task on current CPU.
5241 * It does so by boosting its priority to highest possible and adding it to
5242 * the _front_ of the runqueue. Used by CPU offline code.
5244 void sched_idle_next(void)
5246 int this_cpu
= smp_processor_id();
5247 struct rq
*rq
= cpu_rq(this_cpu
);
5248 struct task_struct
*p
= rq
->idle
;
5249 unsigned long flags
;
5251 /* cpu has to be offline */
5252 BUG_ON(cpu_online(this_cpu
));
5255 * Strictly not necessary since rest of the CPUs are stopped by now
5256 * and interrupts disabled on the current cpu.
5258 spin_lock_irqsave(&rq
->lock
, flags
);
5260 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5262 /* Add idle task to the _front_ of its priority queue: */
5263 __activate_idle_task(p
, rq
);
5265 spin_unlock_irqrestore(&rq
->lock
, flags
);
5269 * Ensures that the idle task is using init_mm right before its cpu goes
5272 void idle_task_exit(void)
5274 struct mm_struct
*mm
= current
->active_mm
;
5276 BUG_ON(cpu_online(smp_processor_id()));
5279 switch_mm(mm
, &init_mm
, current
);
5283 /* called under rq->lock with disabled interrupts */
5284 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5286 struct rq
*rq
= cpu_rq(dead_cpu
);
5288 /* Must be exiting, otherwise would be on tasklist. */
5289 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5291 /* Cannot have done final schedule yet: would have vanished. */
5292 BUG_ON(p
->state
== TASK_DEAD
);
5297 * Drop lock around migration; if someone else moves it,
5298 * that's OK. No task can be added to this CPU, so iteration is
5300 * NOTE: interrupts should be left disabled --dev@
5302 spin_unlock(&rq
->lock
);
5303 move_task_off_dead_cpu(dead_cpu
, p
);
5304 spin_lock(&rq
->lock
);
5309 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5310 static void migrate_dead_tasks(unsigned int dead_cpu
)
5312 struct rq
*rq
= cpu_rq(dead_cpu
);
5313 unsigned int arr
, i
;
5315 for (arr
= 0; arr
< 2; arr
++) {
5316 for (i
= 0; i
< MAX_PRIO
; i
++) {
5317 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
5319 while (!list_empty(list
))
5320 migrate_dead(dead_cpu
, list_entry(list
->next
,
5321 struct task_struct
, run_list
));
5325 #endif /* CONFIG_HOTPLUG_CPU */
5328 * migration_call - callback that gets triggered when a CPU is added.
5329 * Here we can start up the necessary migration thread for the new CPU.
5331 static int __cpuinit
5332 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5334 struct task_struct
*p
;
5335 int cpu
= (long)hcpu
;
5336 unsigned long flags
;
5340 case CPU_LOCK_ACQUIRE
:
5341 mutex_lock(&sched_hotcpu_mutex
);
5344 case CPU_UP_PREPARE
:
5345 case CPU_UP_PREPARE_FROZEN
:
5346 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
5349 p
->flags
|= PF_NOFREEZE
;
5350 kthread_bind(p
, cpu
);
5351 /* Must be high prio: stop_machine expects to yield to it. */
5352 rq
= task_rq_lock(p
, &flags
);
5353 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5354 task_rq_unlock(rq
, &flags
);
5355 cpu_rq(cpu
)->migration_thread
= p
;
5359 case CPU_ONLINE_FROZEN
:
5360 /* Strictly unneccessary, as first user will wake it. */
5361 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5364 #ifdef CONFIG_HOTPLUG_CPU
5365 case CPU_UP_CANCELED
:
5366 case CPU_UP_CANCELED_FROZEN
:
5367 if (!cpu_rq(cpu
)->migration_thread
)
5369 /* Unbind it from offline cpu so it can run. Fall thru. */
5370 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5371 any_online_cpu(cpu_online_map
));
5372 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5373 cpu_rq(cpu
)->migration_thread
= NULL
;
5377 case CPU_DEAD_FROZEN
:
5378 migrate_live_tasks(cpu
);
5380 kthread_stop(rq
->migration_thread
);
5381 rq
->migration_thread
= NULL
;
5382 /* Idle task back to normal (off runqueue, low prio) */
5383 rq
= task_rq_lock(rq
->idle
, &flags
);
5384 deactivate_task(rq
->idle
, rq
);
5385 rq
->idle
->static_prio
= MAX_PRIO
;
5386 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
5387 migrate_dead_tasks(cpu
);
5388 task_rq_unlock(rq
, &flags
);
5389 migrate_nr_uninterruptible(rq
);
5390 BUG_ON(rq
->nr_running
!= 0);
5392 /* No need to migrate the tasks: it was best-effort if
5393 * they didn't take sched_hotcpu_mutex. Just wake up
5394 * the requestors. */
5395 spin_lock_irq(&rq
->lock
);
5396 while (!list_empty(&rq
->migration_queue
)) {
5397 struct migration_req
*req
;
5399 req
= list_entry(rq
->migration_queue
.next
,
5400 struct migration_req
, list
);
5401 list_del_init(&req
->list
);
5402 complete(&req
->done
);
5404 spin_unlock_irq(&rq
->lock
);
5407 case CPU_LOCK_RELEASE
:
5408 mutex_unlock(&sched_hotcpu_mutex
);
5414 /* Register at highest priority so that task migration (migrate_all_tasks)
5415 * happens before everything else.
5417 static struct notifier_block __cpuinitdata migration_notifier
= {
5418 .notifier_call
= migration_call
,
5422 int __init
migration_init(void)
5424 void *cpu
= (void *)(long)smp_processor_id();
5427 /* Start one for the boot CPU: */
5428 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5429 BUG_ON(err
== NOTIFY_BAD
);
5430 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5431 register_cpu_notifier(&migration_notifier
);
5439 /* Number of possible processor ids */
5440 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5441 EXPORT_SYMBOL(nr_cpu_ids
);
5443 #undef SCHED_DOMAIN_DEBUG
5444 #ifdef SCHED_DOMAIN_DEBUG
5445 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5450 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5454 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5459 struct sched_group
*group
= sd
->groups
;
5460 cpumask_t groupmask
;
5462 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5463 cpus_clear(groupmask
);
5466 for (i
= 0; i
< level
+ 1; i
++)
5468 printk("domain %d: ", level
);
5470 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5471 printk("does not load-balance\n");
5473 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5478 printk("span %s\n", str
);
5480 if (!cpu_isset(cpu
, sd
->span
))
5481 printk(KERN_ERR
"ERROR: domain->span does not contain "
5483 if (!cpu_isset(cpu
, group
->cpumask
))
5484 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5488 for (i
= 0; i
< level
+ 2; i
++)
5494 printk(KERN_ERR
"ERROR: group is NULL\n");
5498 if (!group
->__cpu_power
) {
5500 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5504 if (!cpus_weight(group
->cpumask
)) {
5506 printk(KERN_ERR
"ERROR: empty group\n");
5509 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5511 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5514 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5516 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5519 group
= group
->next
;
5520 } while (group
!= sd
->groups
);
5523 if (!cpus_equal(sd
->span
, groupmask
))
5524 printk(KERN_ERR
"ERROR: groups don't span "
5532 if (!cpus_subset(groupmask
, sd
->span
))
5533 printk(KERN_ERR
"ERROR: parent span is not a superset "
5534 "of domain->span\n");
5539 # define sched_domain_debug(sd, cpu) do { } while (0)
5542 static int sd_degenerate(struct sched_domain
*sd
)
5544 if (cpus_weight(sd
->span
) == 1)
5547 /* Following flags need at least 2 groups */
5548 if (sd
->flags
& (SD_LOAD_BALANCE
|
5549 SD_BALANCE_NEWIDLE
|
5553 SD_SHARE_PKG_RESOURCES
)) {
5554 if (sd
->groups
!= sd
->groups
->next
)
5558 /* Following flags don't use groups */
5559 if (sd
->flags
& (SD_WAKE_IDLE
|
5568 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5570 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5572 if (sd_degenerate(parent
))
5575 if (!cpus_equal(sd
->span
, parent
->span
))
5578 /* Does parent contain flags not in child? */
5579 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5580 if (cflags
& SD_WAKE_AFFINE
)
5581 pflags
&= ~SD_WAKE_BALANCE
;
5582 /* Flags needing groups don't count if only 1 group in parent */
5583 if (parent
->groups
== parent
->groups
->next
) {
5584 pflags
&= ~(SD_LOAD_BALANCE
|
5585 SD_BALANCE_NEWIDLE
|
5589 SD_SHARE_PKG_RESOURCES
);
5591 if (~cflags
& pflags
)
5598 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5599 * hold the hotplug lock.
5601 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5603 struct rq
*rq
= cpu_rq(cpu
);
5604 struct sched_domain
*tmp
;
5606 /* Remove the sched domains which do not contribute to scheduling. */
5607 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5608 struct sched_domain
*parent
= tmp
->parent
;
5611 if (sd_parent_degenerate(tmp
, parent
)) {
5612 tmp
->parent
= parent
->parent
;
5614 parent
->parent
->child
= tmp
;
5618 if (sd
&& sd_degenerate(sd
)) {
5624 sched_domain_debug(sd
, cpu
);
5626 rcu_assign_pointer(rq
->sd
, sd
);
5629 /* cpus with isolated domains */
5630 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5632 /* Setup the mask of cpus configured for isolated domains */
5633 static int __init
isolated_cpu_setup(char *str
)
5635 int ints
[NR_CPUS
], i
;
5637 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5638 cpus_clear(cpu_isolated_map
);
5639 for (i
= 1; i
<= ints
[0]; i
++)
5640 if (ints
[i
] < NR_CPUS
)
5641 cpu_set(ints
[i
], cpu_isolated_map
);
5645 __setup ("isolcpus=", isolated_cpu_setup
);
5648 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5649 * to a function which identifies what group(along with sched group) a CPU
5650 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5651 * (due to the fact that we keep track of groups covered with a cpumask_t).
5653 * init_sched_build_groups will build a circular linked list of the groups
5654 * covered by the given span, and will set each group's ->cpumask correctly,
5655 * and ->cpu_power to 0.
5658 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5659 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5660 struct sched_group
**sg
))
5662 struct sched_group
*first
= NULL
, *last
= NULL
;
5663 cpumask_t covered
= CPU_MASK_NONE
;
5666 for_each_cpu_mask(i
, span
) {
5667 struct sched_group
*sg
;
5668 int group
= group_fn(i
, cpu_map
, &sg
);
5671 if (cpu_isset(i
, covered
))
5674 sg
->cpumask
= CPU_MASK_NONE
;
5675 sg
->__cpu_power
= 0;
5677 for_each_cpu_mask(j
, span
) {
5678 if (group_fn(j
, cpu_map
, NULL
) != group
)
5681 cpu_set(j
, covered
);
5682 cpu_set(j
, sg
->cpumask
);
5693 #define SD_NODES_PER_DOMAIN 16
5698 * find_next_best_node - find the next node to include in a sched_domain
5699 * @node: node whose sched_domain we're building
5700 * @used_nodes: nodes already in the sched_domain
5702 * Find the next node to include in a given scheduling domain. Simply
5703 * finds the closest node not already in the @used_nodes map.
5705 * Should use nodemask_t.
5707 static int find_next_best_node(int node
, unsigned long *used_nodes
)
5709 int i
, n
, val
, min_val
, best_node
= 0;
5713 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5714 /* Start at @node */
5715 n
= (node
+ i
) % MAX_NUMNODES
;
5717 if (!nr_cpus_node(n
))
5720 /* Skip already used nodes */
5721 if (test_bit(n
, used_nodes
))
5724 /* Simple min distance search */
5725 val
= node_distance(node
, n
);
5727 if (val
< min_val
) {
5733 set_bit(best_node
, used_nodes
);
5738 * sched_domain_node_span - get a cpumask for a node's sched_domain
5739 * @node: node whose cpumask we're constructing
5740 * @size: number of nodes to include in this span
5742 * Given a node, construct a good cpumask for its sched_domain to span. It
5743 * should be one that prevents unnecessary balancing, but also spreads tasks
5746 static cpumask_t
sched_domain_node_span(int node
)
5748 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
5749 cpumask_t span
, nodemask
;
5753 bitmap_zero(used_nodes
, MAX_NUMNODES
);
5755 nodemask
= node_to_cpumask(node
);
5756 cpus_or(span
, span
, nodemask
);
5757 set_bit(node
, used_nodes
);
5759 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5760 int next_node
= find_next_best_node(node
, used_nodes
);
5762 nodemask
= node_to_cpumask(next_node
);
5763 cpus_or(span
, span
, nodemask
);
5770 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5773 * SMT sched-domains:
5775 #ifdef CONFIG_SCHED_SMT
5776 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
5777 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
5779 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
5780 struct sched_group
**sg
)
5783 *sg
= &per_cpu(sched_group_cpus
, cpu
);
5789 * multi-core sched-domains:
5791 #ifdef CONFIG_SCHED_MC
5792 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
5793 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
5796 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5797 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5798 struct sched_group
**sg
)
5801 cpumask_t mask
= cpu_sibling_map
[cpu
];
5802 cpus_and(mask
, mask
, *cpu_map
);
5803 group
= first_cpu(mask
);
5805 *sg
= &per_cpu(sched_group_core
, group
);
5808 #elif defined(CONFIG_SCHED_MC)
5809 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
5810 struct sched_group
**sg
)
5813 *sg
= &per_cpu(sched_group_core
, cpu
);
5818 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
5819 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
5821 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
5822 struct sched_group
**sg
)
5825 #ifdef CONFIG_SCHED_MC
5826 cpumask_t mask
= cpu_coregroup_map(cpu
);
5827 cpus_and(mask
, mask
, *cpu_map
);
5828 group
= first_cpu(mask
);
5829 #elif defined(CONFIG_SCHED_SMT)
5830 cpumask_t mask
= cpu_sibling_map
[cpu
];
5831 cpus_and(mask
, mask
, *cpu_map
);
5832 group
= first_cpu(mask
);
5837 *sg
= &per_cpu(sched_group_phys
, group
);
5843 * The init_sched_build_groups can't handle what we want to do with node
5844 * groups, so roll our own. Now each node has its own list of groups which
5845 * gets dynamically allocated.
5847 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
5848 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
5850 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
5851 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
5853 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
5854 struct sched_group
**sg
)
5856 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
5859 cpus_and(nodemask
, nodemask
, *cpu_map
);
5860 group
= first_cpu(nodemask
);
5863 *sg
= &per_cpu(sched_group_allnodes
, group
);
5867 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
5869 struct sched_group
*sg
= group_head
;
5875 for_each_cpu_mask(j
, sg
->cpumask
) {
5876 struct sched_domain
*sd
;
5878 sd
= &per_cpu(phys_domains
, j
);
5879 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
5881 * Only add "power" once for each
5887 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
5890 if (sg
!= group_head
)
5896 /* Free memory allocated for various sched_group structures */
5897 static void free_sched_groups(const cpumask_t
*cpu_map
)
5901 for_each_cpu_mask(cpu
, *cpu_map
) {
5902 struct sched_group
**sched_group_nodes
5903 = sched_group_nodes_bycpu
[cpu
];
5905 if (!sched_group_nodes
)
5908 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5909 cpumask_t nodemask
= node_to_cpumask(i
);
5910 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
5912 cpus_and(nodemask
, nodemask
, *cpu_map
);
5913 if (cpus_empty(nodemask
))
5923 if (oldsg
!= sched_group_nodes
[i
])
5926 kfree(sched_group_nodes
);
5927 sched_group_nodes_bycpu
[cpu
] = NULL
;
5931 static void free_sched_groups(const cpumask_t
*cpu_map
)
5937 * Initialize sched groups cpu_power.
5939 * cpu_power indicates the capacity of sched group, which is used while
5940 * distributing the load between different sched groups in a sched domain.
5941 * Typically cpu_power for all the groups in a sched domain will be same unless
5942 * there are asymmetries in the topology. If there are asymmetries, group
5943 * having more cpu_power will pickup more load compared to the group having
5946 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5947 * the maximum number of tasks a group can handle in the presence of other idle
5948 * or lightly loaded groups in the same sched domain.
5950 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5952 struct sched_domain
*child
;
5953 struct sched_group
*group
;
5955 WARN_ON(!sd
|| !sd
->groups
);
5957 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
5962 sd
->groups
->__cpu_power
= 0;
5965 * For perf policy, if the groups in child domain share resources
5966 * (for example cores sharing some portions of the cache hierarchy
5967 * or SMT), then set this domain groups cpu_power such that each group
5968 * can handle only one task, when there are other idle groups in the
5969 * same sched domain.
5971 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
5973 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
5974 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
5979 * add cpu_power of each child group to this groups cpu_power
5981 group
= child
->groups
;
5983 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
5984 group
= group
->next
;
5985 } while (group
!= child
->groups
);
5989 * Build sched domains for a given set of cpus and attach the sched domains
5990 * to the individual cpus
5992 static int build_sched_domains(const cpumask_t
*cpu_map
)
5995 struct sched_domain
*sd
;
5997 struct sched_group
**sched_group_nodes
= NULL
;
5998 int sd_allnodes
= 0;
6001 * Allocate the per-node list of sched groups
6003 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6005 if (!sched_group_nodes
) {
6006 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6009 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6013 * Set up domains for cpus specified by the cpu_map.
6015 for_each_cpu_mask(i
, *cpu_map
) {
6016 struct sched_domain
*sd
= NULL
, *p
;
6017 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6019 cpus_and(nodemask
, nodemask
, *cpu_map
);
6022 if (cpus_weight(*cpu_map
)
6023 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6024 sd
= &per_cpu(allnodes_domains
, i
);
6025 *sd
= SD_ALLNODES_INIT
;
6026 sd
->span
= *cpu_map
;
6027 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6033 sd
= &per_cpu(node_domains
, i
);
6035 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6039 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6043 sd
= &per_cpu(phys_domains
, i
);
6045 sd
->span
= nodemask
;
6049 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6051 #ifdef CONFIG_SCHED_MC
6053 sd
= &per_cpu(core_domains
, i
);
6055 sd
->span
= cpu_coregroup_map(i
);
6056 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6059 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6062 #ifdef CONFIG_SCHED_SMT
6064 sd
= &per_cpu(cpu_domains
, i
);
6065 *sd
= SD_SIBLING_INIT
;
6066 sd
->span
= cpu_sibling_map
[i
];
6067 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6070 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6074 #ifdef CONFIG_SCHED_SMT
6075 /* Set up CPU (sibling) groups */
6076 for_each_cpu_mask(i
, *cpu_map
) {
6077 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6078 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6079 if (i
!= first_cpu(this_sibling_map
))
6082 init_sched_build_groups(this_sibling_map
, cpu_map
, &cpu_to_cpu_group
);
6086 #ifdef CONFIG_SCHED_MC
6087 /* Set up multi-core groups */
6088 for_each_cpu_mask(i
, *cpu_map
) {
6089 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6090 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6091 if (i
!= first_cpu(this_core_map
))
6093 init_sched_build_groups(this_core_map
, cpu_map
, &cpu_to_core_group
);
6098 /* Set up physical groups */
6099 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6100 cpumask_t nodemask
= node_to_cpumask(i
);
6102 cpus_and(nodemask
, nodemask
, *cpu_map
);
6103 if (cpus_empty(nodemask
))
6106 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6110 /* Set up node groups */
6112 init_sched_build_groups(*cpu_map
, cpu_map
, &cpu_to_allnodes_group
);
6114 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6115 /* Set up node groups */
6116 struct sched_group
*sg
, *prev
;
6117 cpumask_t nodemask
= node_to_cpumask(i
);
6118 cpumask_t domainspan
;
6119 cpumask_t covered
= CPU_MASK_NONE
;
6122 cpus_and(nodemask
, nodemask
, *cpu_map
);
6123 if (cpus_empty(nodemask
)) {
6124 sched_group_nodes
[i
] = NULL
;
6128 domainspan
= sched_domain_node_span(i
);
6129 cpus_and(domainspan
, domainspan
, *cpu_map
);
6131 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6133 printk(KERN_WARNING
"Can not alloc domain group for "
6137 sched_group_nodes
[i
] = sg
;
6138 for_each_cpu_mask(j
, nodemask
) {
6139 struct sched_domain
*sd
;
6140 sd
= &per_cpu(node_domains
, j
);
6143 sg
->__cpu_power
= 0;
6144 sg
->cpumask
= nodemask
;
6146 cpus_or(covered
, covered
, nodemask
);
6149 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6150 cpumask_t tmp
, notcovered
;
6151 int n
= (i
+ j
) % MAX_NUMNODES
;
6153 cpus_complement(notcovered
, covered
);
6154 cpus_and(tmp
, notcovered
, *cpu_map
);
6155 cpus_and(tmp
, tmp
, domainspan
);
6156 if (cpus_empty(tmp
))
6159 nodemask
= node_to_cpumask(n
);
6160 cpus_and(tmp
, tmp
, nodemask
);
6161 if (cpus_empty(tmp
))
6164 sg
= kmalloc_node(sizeof(struct sched_group
),
6168 "Can not alloc domain group for node %d\n", j
);
6171 sg
->__cpu_power
= 0;
6173 sg
->next
= prev
->next
;
6174 cpus_or(covered
, covered
, tmp
);
6181 /* Calculate CPU power for physical packages and nodes */
6182 #ifdef CONFIG_SCHED_SMT
6183 for_each_cpu_mask(i
, *cpu_map
) {
6184 sd
= &per_cpu(cpu_domains
, i
);
6185 init_sched_groups_power(i
, sd
);
6188 #ifdef CONFIG_SCHED_MC
6189 for_each_cpu_mask(i
, *cpu_map
) {
6190 sd
= &per_cpu(core_domains
, i
);
6191 init_sched_groups_power(i
, sd
);
6195 for_each_cpu_mask(i
, *cpu_map
) {
6196 sd
= &per_cpu(phys_domains
, i
);
6197 init_sched_groups_power(i
, sd
);
6201 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6202 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6205 struct sched_group
*sg
;
6207 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6208 init_numa_sched_groups_power(sg
);
6212 /* Attach the domains */
6213 for_each_cpu_mask(i
, *cpu_map
) {
6214 struct sched_domain
*sd
;
6215 #ifdef CONFIG_SCHED_SMT
6216 sd
= &per_cpu(cpu_domains
, i
);
6217 #elif defined(CONFIG_SCHED_MC)
6218 sd
= &per_cpu(core_domains
, i
);
6220 sd
= &per_cpu(phys_domains
, i
);
6222 cpu_attach_domain(sd
, i
);
6229 free_sched_groups(cpu_map
);
6234 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6236 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6238 cpumask_t cpu_default_map
;
6242 * Setup mask for cpus without special case scheduling requirements.
6243 * For now this just excludes isolated cpus, but could be used to
6244 * exclude other special cases in the future.
6246 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6248 err
= build_sched_domains(&cpu_default_map
);
6253 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6255 free_sched_groups(cpu_map
);
6259 * Detach sched domains from a group of cpus specified in cpu_map
6260 * These cpus will now be attached to the NULL domain
6262 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6266 for_each_cpu_mask(i
, *cpu_map
)
6267 cpu_attach_domain(NULL
, i
);
6268 synchronize_sched();
6269 arch_destroy_sched_domains(cpu_map
);
6273 * Partition sched domains as specified by the cpumasks below.
6274 * This attaches all cpus from the cpumasks to the NULL domain,
6275 * waits for a RCU quiescent period, recalculates sched
6276 * domain information and then attaches them back to the
6277 * correct sched domains
6278 * Call with hotplug lock held
6280 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6282 cpumask_t change_map
;
6285 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6286 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6287 cpus_or(change_map
, *partition1
, *partition2
);
6289 /* Detach sched domains from all of the affected cpus */
6290 detach_destroy_domains(&change_map
);
6291 if (!cpus_empty(*partition1
))
6292 err
= build_sched_domains(partition1
);
6293 if (!err
&& !cpus_empty(*partition2
))
6294 err
= build_sched_domains(partition2
);
6299 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6300 int arch_reinit_sched_domains(void)
6304 mutex_lock(&sched_hotcpu_mutex
);
6305 detach_destroy_domains(&cpu_online_map
);
6306 err
= arch_init_sched_domains(&cpu_online_map
);
6307 mutex_unlock(&sched_hotcpu_mutex
);
6312 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6316 if (buf
[0] != '0' && buf
[0] != '1')
6320 sched_smt_power_savings
= (buf
[0] == '1');
6322 sched_mc_power_savings
= (buf
[0] == '1');
6324 ret
= arch_reinit_sched_domains();
6326 return ret
? ret
: count
;
6329 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6333 #ifdef CONFIG_SCHED_SMT
6335 err
= sysfs_create_file(&cls
->kset
.kobj
,
6336 &attr_sched_smt_power_savings
.attr
);
6338 #ifdef CONFIG_SCHED_MC
6339 if (!err
&& mc_capable())
6340 err
= sysfs_create_file(&cls
->kset
.kobj
,
6341 &attr_sched_mc_power_savings
.attr
);
6347 #ifdef CONFIG_SCHED_MC
6348 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6350 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6352 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6353 const char *buf
, size_t count
)
6355 return sched_power_savings_store(buf
, count
, 0);
6357 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6358 sched_mc_power_savings_store
);
6361 #ifdef CONFIG_SCHED_SMT
6362 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6364 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6366 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6367 const char *buf
, size_t count
)
6369 return sched_power_savings_store(buf
, count
, 1);
6371 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6372 sched_smt_power_savings_store
);
6376 * Force a reinitialization of the sched domains hierarchy. The domains
6377 * and groups cannot be updated in place without racing with the balancing
6378 * code, so we temporarily attach all running cpus to the NULL domain
6379 * which will prevent rebalancing while the sched domains are recalculated.
6381 static int update_sched_domains(struct notifier_block
*nfb
,
6382 unsigned long action
, void *hcpu
)
6385 case CPU_UP_PREPARE
:
6386 case CPU_UP_PREPARE_FROZEN
:
6387 case CPU_DOWN_PREPARE
:
6388 case CPU_DOWN_PREPARE_FROZEN
:
6389 detach_destroy_domains(&cpu_online_map
);
6392 case CPU_UP_CANCELED
:
6393 case CPU_UP_CANCELED_FROZEN
:
6394 case CPU_DOWN_FAILED
:
6395 case CPU_DOWN_FAILED_FROZEN
:
6397 case CPU_ONLINE_FROZEN
:
6399 case CPU_DEAD_FROZEN
:
6401 * Fall through and re-initialise the domains.
6408 /* The hotplug lock is already held by cpu_up/cpu_down */
6409 arch_init_sched_domains(&cpu_online_map
);
6414 void __init
sched_init_smp(void)
6416 cpumask_t non_isolated_cpus
;
6418 mutex_lock(&sched_hotcpu_mutex
);
6419 arch_init_sched_domains(&cpu_online_map
);
6420 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6421 if (cpus_empty(non_isolated_cpus
))
6422 cpu_set(smp_processor_id(), non_isolated_cpus
);
6423 mutex_unlock(&sched_hotcpu_mutex
);
6424 /* XXX: Theoretical race here - CPU may be hotplugged now */
6425 hotcpu_notifier(update_sched_domains
, 0);
6427 /* Move init over to a non-isolated CPU */
6428 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6432 void __init
sched_init_smp(void)
6435 #endif /* CONFIG_SMP */
6437 int in_sched_functions(unsigned long addr
)
6439 /* Linker adds these: start and end of __sched functions */
6440 extern char __sched_text_start
[], __sched_text_end
[];
6442 return in_lock_functions(addr
) ||
6443 (addr
>= (unsigned long)__sched_text_start
6444 && addr
< (unsigned long)__sched_text_end
);
6447 void __init
sched_init(void)
6450 int highest_cpu
= 0;
6452 for_each_possible_cpu(i
) {
6453 struct prio_array
*array
;
6457 spin_lock_init(&rq
->lock
);
6458 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6460 rq
->active
= rq
->arrays
;
6461 rq
->expired
= rq
->arrays
+ 1;
6462 rq
->best_expired_prio
= MAX_PRIO
;
6466 for (j
= 1; j
< 3; j
++)
6467 rq
->cpu_load
[j
] = 0;
6468 rq
->active_balance
= 0;
6471 rq
->migration_thread
= NULL
;
6472 INIT_LIST_HEAD(&rq
->migration_queue
);
6474 atomic_set(&rq
->nr_iowait
, 0);
6476 for (j
= 0; j
< 2; j
++) {
6477 array
= rq
->arrays
+ j
;
6478 for (k
= 0; k
< MAX_PRIO
; k
++) {
6479 INIT_LIST_HEAD(array
->queue
+ k
);
6480 __clear_bit(k
, array
->bitmap
);
6482 // delimiter for bitsearch
6483 __set_bit(MAX_PRIO
, array
->bitmap
);
6488 set_load_weight(&init_task
);
6491 nr_cpu_ids
= highest_cpu
+ 1;
6492 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6495 #ifdef CONFIG_RT_MUTEXES
6496 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6500 * The boot idle thread does lazy MMU switching as well:
6502 atomic_inc(&init_mm
.mm_count
);
6503 enter_lazy_tlb(&init_mm
, current
);
6506 * Make us the idle thread. Technically, schedule() should not be
6507 * called from this thread, however somewhere below it might be,
6508 * but because we are the idle thread, we just pick up running again
6509 * when this runqueue becomes "idle".
6511 init_idle(current
, smp_processor_id());
6514 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6515 void __might_sleep(char *file
, int line
)
6518 static unsigned long prev_jiffy
; /* ratelimiting */
6520 if ((in_atomic() || irqs_disabled()) &&
6521 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6522 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6524 prev_jiffy
= jiffies
;
6525 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6526 " context at %s:%d\n", file
, line
);
6527 printk("in_atomic():%d, irqs_disabled():%d\n",
6528 in_atomic(), irqs_disabled());
6529 debug_show_held_locks(current
);
6530 if (irqs_disabled())
6531 print_irqtrace_events(current
);
6536 EXPORT_SYMBOL(__might_sleep
);
6539 #ifdef CONFIG_MAGIC_SYSRQ
6540 void normalize_rt_tasks(void)
6542 struct prio_array
*array
;
6543 struct task_struct
*g
, *p
;
6544 unsigned long flags
;
6547 read_lock_irq(&tasklist_lock
);
6549 do_each_thread(g
, p
) {
6553 spin_lock_irqsave(&p
->pi_lock
, flags
);
6554 rq
= __task_rq_lock(p
);
6558 deactivate_task(p
, task_rq(p
));
6559 __setscheduler(p
, SCHED_NORMAL
, 0);
6561 __activate_task(p
, task_rq(p
));
6562 resched_task(rq
->curr
);
6565 __task_rq_unlock(rq
);
6566 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6567 } while_each_thread(g
, p
);
6569 read_unlock_irq(&tasklist_lock
);
6572 #endif /* CONFIG_MAGIC_SYSRQ */
6576 * These functions are only useful for the IA64 MCA handling.
6578 * They can only be called when the whole system has been
6579 * stopped - every CPU needs to be quiescent, and no scheduling
6580 * activity can take place. Using them for anything else would
6581 * be a serious bug, and as a result, they aren't even visible
6582 * under any other configuration.
6586 * curr_task - return the current task for a given cpu.
6587 * @cpu: the processor in question.
6589 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6591 struct task_struct
*curr_task(int cpu
)
6593 return cpu_curr(cpu
);
6597 * set_curr_task - set the current task for a given cpu.
6598 * @cpu: the processor in question.
6599 * @p: the task pointer to set.
6601 * Description: This function must only be used when non-maskable interrupts
6602 * are serviced on a separate stack. It allows the architecture to switch the
6603 * notion of the current task on a cpu in a non-blocking manner. This function
6604 * must be called with all CPU's synchronized, and interrupts disabled, the
6605 * and caller must save the original value of the current task (see
6606 * curr_task() above) and restore that value before reenabling interrupts and
6607 * re-starting the system.
6609 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6611 void set_curr_task(int cpu
, struct task_struct
*p
)