4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
64 #include <linux/pagemap.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __attribute__((weak)) sched_clock(void)
75 return (unsigned long long)jiffies * (1000000000 / HZ);
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Some helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
100 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
109 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
110 * Timeslices get refilled after they expire.
112 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
122 return reciprocal_divide(load, sg->reciprocal_cpu_power);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
131 sg->__cpu_power += val;
132 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
136 #define SCALE_PRIO(x, prio) \
137 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
140 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
141 * to time slice values: [800ms ... 100ms ... 5ms]
143 static unsigned int static_prio_timeslice(int static_prio)
145 if (static_prio == NICE_TO_PRIO(19))
148 if (static_prio < NICE_TO_PRIO(0))
149 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
151 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
154 static inline int rt_policy(int policy)
156 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
161 static inline int task_has_rt_policy(struct task_struct *p)
163 return rt_policy(p->policy);
167 * This is the priority-queue data structure of the RT scheduling class:
169 struct rt_prio_array {
170 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
171 struct list_head queue[MAX_RT_PRIO];
175 struct load_weight load;
176 u64 load_update_start, load_update_last;
177 unsigned long delta_fair, delta_exec, delta_stat;
180 /* CFS-related fields in a runqueue */
182 struct load_weight load;
183 unsigned long nr_running;
189 unsigned long wait_runtime_overruns, wait_runtime_underruns;
191 struct rb_root tasks_timeline;
192 struct rb_node *rb_leftmost;
193 struct rb_node *rb_load_balance_curr;
194 #ifdef CONFIG_FAIR_GROUP_SCHED
195 /* 'curr' points to currently running entity on this cfs_rq.
196 * It is set to NULL otherwise (i.e when none are currently running).
198 struct sched_entity *curr;
199 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
201 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
202 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
203 * (like users, containers etc.)
205 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
206 * list is used during load balance.
208 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
212 /* Real-Time classes' related field in a runqueue: */
214 struct rt_prio_array active;
215 int rt_load_balance_idx;
216 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
220 * This is the main, per-CPU runqueue data structure.
222 * Locking rule: those places that want to lock multiple runqueues
223 * (such as the load balancing or the thread migration code), lock
224 * acquire operations must be ordered by ascending &runqueue.
227 spinlock_t lock; /* runqueue lock */
230 * nr_running and cpu_load should be in the same cacheline because
231 * remote CPUs use both these fields when doing load calculation.
233 unsigned long nr_running;
234 #define CPU_LOAD_IDX_MAX 5
235 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
236 unsigned char idle_at_tick;
238 unsigned char in_nohz_recently;
240 struct load_stat ls; /* capture load from *all* tasks on this cpu */
241 unsigned long nr_load_updates;
245 #ifdef CONFIG_FAIR_GROUP_SCHED
246 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
251 * This is part of a global counter where only the total sum
252 * over all CPUs matters. A task can increase this counter on
253 * one CPU and if it got migrated afterwards it may decrease
254 * it on another CPU. Always updated under the runqueue lock:
256 unsigned long nr_uninterruptible;
258 struct task_struct *curr, *idle;
259 unsigned long next_balance;
260 struct mm_struct *prev_mm;
262 u64 clock, prev_clock_raw;
265 unsigned int clock_warps, clock_overflows;
267 unsigned int clock_deep_idle_events;
273 struct sched_domain *sd;
275 /* For active balancing */
278 int cpu; /* cpu of this runqueue */
280 struct task_struct *migration_thread;
281 struct list_head migration_queue;
284 #ifdef CONFIG_SCHEDSTATS
286 struct sched_info rq_sched_info;
288 /* sys_sched_yield() stats */
289 unsigned long yld_exp_empty;
290 unsigned long yld_act_empty;
291 unsigned long yld_both_empty;
292 unsigned long yld_cnt;
294 /* schedule() stats */
295 unsigned long sched_switch;
296 unsigned long sched_cnt;
297 unsigned long sched_goidle;
299 /* try_to_wake_up() stats */
300 unsigned long ttwu_cnt;
301 unsigned long ttwu_local;
303 struct lock_class_key rq_lock_key;
306 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
307 static DEFINE_MUTEX(sched_hotcpu_mutex);
309 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
311 rq->curr->sched_class->check_preempt_curr(rq, p);
314 static inline int cpu_of(struct rq *rq)
324 * Update the per-runqueue clock, as finegrained as the platform can give
325 * us, but without assuming monotonicity, etc.:
327 static void __update_rq_clock(struct rq *rq)
329 u64 prev_raw = rq->prev_clock_raw;
330 u64 now = sched_clock();
331 s64 delta = now - prev_raw;
332 u64 clock = rq->clock;
334 #ifdef CONFIG_SCHED_DEBUG
335 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
338 * Protect against sched_clock() occasionally going backwards:
340 if (unlikely(delta < 0)) {
345 * Catch too large forward jumps too:
347 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
348 if (clock < rq->tick_timestamp + TICK_NSEC)
349 clock = rq->tick_timestamp + TICK_NSEC;
352 rq->clock_overflows++;
354 if (unlikely(delta > rq->clock_max_delta))
355 rq->clock_max_delta = delta;
360 rq->prev_clock_raw = now;
364 static void update_rq_clock(struct rq *rq)
366 if (likely(smp_processor_id() == cpu_of(rq)))
367 __update_rq_clock(rq);
371 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
372 * See detach_destroy_domains: synchronize_sched for details.
374 * The domain tree of any CPU may only be accessed from within
375 * preempt-disabled sections.
377 #define for_each_domain(cpu, __sd) \
378 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
380 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
381 #define this_rq() (&__get_cpu_var(runqueues))
382 #define task_rq(p) cpu_rq(task_cpu(p))
383 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
386 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
387 * clock constructed from sched_clock():
389 unsigned long long cpu_clock(int cpu)
391 unsigned long long now;
395 local_irq_save(flags);
399 local_irq_restore(flags);
404 #ifdef CONFIG_FAIR_GROUP_SCHED
405 /* Change a task's ->cfs_rq if it moves across CPUs */
406 static inline void set_task_cfs_rq(struct task_struct *p)
408 p->se.cfs_rq = &task_rq(p)->cfs;
411 static inline void set_task_cfs_rq(struct task_struct *p)
416 #ifndef prepare_arch_switch
417 # define prepare_arch_switch(next) do { } while (0)
419 #ifndef finish_arch_switch
420 # define finish_arch_switch(prev) do { } while (0)
423 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
424 static inline int task_running(struct rq *rq, struct task_struct *p)
426 return rq->curr == p;
429 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
433 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
435 #ifdef CONFIG_DEBUG_SPINLOCK
436 /* this is a valid case when another task releases the spinlock */
437 rq->lock.owner = current;
440 * If we are tracking spinlock dependencies then we have to
441 * fix up the runqueue lock - which gets 'carried over' from
444 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
446 spin_unlock_irq(&rq->lock);
449 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
450 static inline int task_running(struct rq *rq, struct task_struct *p)
455 return rq->curr == p;
459 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
463 * We can optimise this out completely for !SMP, because the
464 * SMP rebalancing from interrupt is the only thing that cares
469 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
470 spin_unlock_irq(&rq->lock);
472 spin_unlock(&rq->lock);
476 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
480 * After ->oncpu is cleared, the task can be moved to a different CPU.
481 * We must ensure this doesn't happen until the switch is completely
487 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
491 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
494 * __task_rq_lock - lock the runqueue a given task resides on.
495 * Must be called interrupts disabled.
497 static inline struct rq *__task_rq_lock(struct task_struct *p)
504 spin_lock(&rq->lock);
505 if (unlikely(rq != task_rq(p))) {
506 spin_unlock(&rq->lock);
507 goto repeat_lock_task;
513 * task_rq_lock - lock the runqueue a given task resides on and disable
514 * interrupts. Note the ordering: we can safely lookup the task_rq without
515 * explicitly disabling preemption.
517 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
523 local_irq_save(*flags);
525 spin_lock(&rq->lock);
526 if (unlikely(rq != task_rq(p))) {
527 spin_unlock_irqrestore(&rq->lock, *flags);
528 goto repeat_lock_task;
533 static inline void __task_rq_unlock(struct rq *rq)
536 spin_unlock(&rq->lock);
539 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
542 spin_unlock_irqrestore(&rq->lock, *flags);
546 * this_rq_lock - lock this runqueue and disable interrupts.
548 static inline struct rq *this_rq_lock(void)
555 spin_lock(&rq->lock);
561 * We are going deep-idle (irqs are disabled):
563 void sched_clock_idle_sleep_event(void)
565 struct rq *rq = cpu_rq(smp_processor_id());
567 spin_lock(&rq->lock);
568 __update_rq_clock(rq);
569 spin_unlock(&rq->lock);
570 rq->clock_deep_idle_events++;
572 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
575 * We just idled delta nanoseconds (called with irqs disabled):
577 void sched_clock_idle_wakeup_event(u64 delta_ns)
579 struct rq *rq = cpu_rq(smp_processor_id());
580 u64 now = sched_clock();
582 rq->idle_clock += delta_ns;
584 * Override the previous timestamp and ignore all
585 * sched_clock() deltas that occured while we idled,
586 * and use the PM-provided delta_ns to advance the
589 spin_lock(&rq->lock);
590 rq->prev_clock_raw = now;
591 rq->clock += delta_ns;
592 spin_unlock(&rq->lock);
594 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
597 * resched_task - mark a task 'to be rescheduled now'.
599 * On UP this means the setting of the need_resched flag, on SMP it
600 * might also involve a cross-CPU call to trigger the scheduler on
605 #ifndef tsk_is_polling
606 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
609 static void resched_task(struct task_struct *p)
613 assert_spin_locked(&task_rq(p)->lock);
615 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
618 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
621 if (cpu == smp_processor_id())
624 /* NEED_RESCHED must be visible before we test polling */
626 if (!tsk_is_polling(p))
627 smp_send_reschedule(cpu);
630 static void resched_cpu(int cpu)
632 struct rq *rq = cpu_rq(cpu);
635 if (!spin_trylock_irqsave(&rq->lock, flags))
637 resched_task(cpu_curr(cpu));
638 spin_unlock_irqrestore(&rq->lock, flags);
641 static inline void resched_task(struct task_struct *p)
643 assert_spin_locked(&task_rq(p)->lock);
644 set_tsk_need_resched(p);
648 static u64 div64_likely32(u64 divident, unsigned long divisor)
650 #if BITS_PER_LONG == 32
651 if (likely(divident <= 0xffffffffULL))
652 return (u32)divident / divisor;
653 do_div(divident, divisor);
657 return divident / divisor;
661 #if BITS_PER_LONG == 32
662 # define WMULT_CONST (~0UL)
664 # define WMULT_CONST (1UL << 32)
667 #define WMULT_SHIFT 32
670 * Shift right and round:
672 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
675 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
676 struct load_weight *lw)
680 if (unlikely(!lw->inv_weight))
681 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
683 tmp = (u64)delta_exec * weight;
685 * Check whether we'd overflow the 64-bit multiplication:
687 if (unlikely(tmp > WMULT_CONST))
688 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
691 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
693 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
696 static inline unsigned long
697 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
699 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
702 static void update_load_add(struct load_weight *lw, unsigned long inc)
708 static void update_load_sub(struct load_weight *lw, unsigned long dec)
715 * To aid in avoiding the subversion of "niceness" due to uneven distribution
716 * of tasks with abnormal "nice" values across CPUs the contribution that
717 * each task makes to its run queue's load is weighted according to its
718 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
719 * scaled version of the new time slice allocation that they receive on time
723 #define WEIGHT_IDLEPRIO 2
724 #define WMULT_IDLEPRIO (1 << 31)
727 * Nice levels are multiplicative, with a gentle 10% change for every
728 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
729 * nice 1, it will get ~10% less CPU time than another CPU-bound task
730 * that remained on nice 0.
732 * The "10% effect" is relative and cumulative: from _any_ nice level,
733 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
734 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
735 * If a task goes up by ~10% and another task goes down by ~10% then
736 * the relative distance between them is ~25%.)
738 static const int prio_to_weight[40] = {
739 /* -20 */ 88761, 71755, 56483, 46273, 36291,
740 /* -15 */ 29154, 23254, 18705, 14949, 11916,
741 /* -10 */ 9548, 7620, 6100, 4904, 3906,
742 /* -5 */ 3121, 2501, 1991, 1586, 1277,
743 /* 0 */ 1024, 820, 655, 526, 423,
744 /* 5 */ 335, 272, 215, 172, 137,
745 /* 10 */ 110, 87, 70, 56, 45,
746 /* 15 */ 36, 29, 23, 18, 15,
750 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
752 * In cases where the weight does not change often, we can use the
753 * precalculated inverse to speed up arithmetics by turning divisions
754 * into multiplications:
756 static const u32 prio_to_wmult[40] = {
757 /* -20 */ 48388, 59856, 76040, 92818, 118348,
758 /* -15 */ 147320, 184698, 229616, 287308, 360437,
759 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
760 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
761 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
762 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
763 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
764 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
767 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
770 * runqueue iterator, to support SMP load-balancing between different
771 * scheduling classes, without having to expose their internal data
772 * structures to the load-balancing proper:
776 struct task_struct *(*start)(void *);
777 struct task_struct *(*next)(void *);
780 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
781 unsigned long max_nr_move, unsigned long max_load_move,
782 struct sched_domain *sd, enum cpu_idle_type idle,
783 int *all_pinned, unsigned long *load_moved,
784 int *this_best_prio, struct rq_iterator *iterator);
786 #include "sched_stats.h"
787 #include "sched_rt.c"
788 #include "sched_fair.c"
789 #include "sched_idletask.c"
790 #ifdef CONFIG_SCHED_DEBUG
791 # include "sched_debug.c"
794 #define sched_class_highest (&rt_sched_class)
796 static void __update_curr_load(struct rq *rq, struct load_stat *ls)
798 if (rq->curr != rq->idle && ls->load.weight) {
799 ls->delta_exec += ls->delta_stat;
800 ls->delta_fair += calc_delta_fair(ls->delta_stat, &ls->load);
806 * Update delta_exec, delta_fair fields for rq.
808 * delta_fair clock advances at a rate inversely proportional to
809 * total load (rq->ls.load.weight) on the runqueue, while
810 * delta_exec advances at the same rate as wall-clock (provided
813 * delta_exec / delta_fair is a measure of the (smoothened) load on this
814 * runqueue over any given interval. This (smoothened) load is used
815 * during load balance.
817 * This function is called /before/ updating rq->ls.load
818 * and when switching tasks.
820 static void update_curr_load(struct rq *rq)
822 struct load_stat *ls = &rq->ls;
825 start = ls->load_update_start;
826 ls->load_update_start = rq->clock;
827 ls->delta_stat += rq->clock - start;
829 * Stagger updates to ls->delta_fair. Very frequent updates
832 if (ls->delta_stat >= sysctl_sched_stat_granularity)
833 __update_curr_load(rq, ls);
836 static inline void inc_load(struct rq *rq, const struct task_struct *p)
838 update_curr_load(rq);
839 update_load_add(&rq->ls.load, p->se.load.weight);
842 static inline void dec_load(struct rq *rq, const struct task_struct *p)
844 update_curr_load(rq);
845 update_load_sub(&rq->ls.load, p->se.load.weight);
848 static void inc_nr_running(struct task_struct *p, struct rq *rq)
854 static void dec_nr_running(struct task_struct *p, struct rq *rq)
860 static void set_load_weight(struct task_struct *p)
862 p->se.wait_runtime = 0;
864 if (task_has_rt_policy(p)) {
865 p->se.load.weight = prio_to_weight[0] * 2;
866 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
871 * SCHED_IDLE tasks get minimal weight:
873 if (p->policy == SCHED_IDLE) {
874 p->se.load.weight = WEIGHT_IDLEPRIO;
875 p->se.load.inv_weight = WMULT_IDLEPRIO;
879 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
880 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
883 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
885 sched_info_queued(p);
886 p->sched_class->enqueue_task(rq, p, wakeup);
890 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
892 p->sched_class->dequeue_task(rq, p, sleep);
897 * __normal_prio - return the priority that is based on the static prio
899 static inline int __normal_prio(struct task_struct *p)
901 return p->static_prio;
905 * Calculate the expected normal priority: i.e. priority
906 * without taking RT-inheritance into account. Might be
907 * boosted by interactivity modifiers. Changes upon fork,
908 * setprio syscalls, and whenever the interactivity
909 * estimator recalculates.
911 static inline int normal_prio(struct task_struct *p)
915 if (task_has_rt_policy(p))
916 prio = MAX_RT_PRIO-1 - p->rt_priority;
918 prio = __normal_prio(p);
923 * Calculate the current priority, i.e. the priority
924 * taken into account by the scheduler. This value might
925 * be boosted by RT tasks, or might be boosted by
926 * interactivity modifiers. Will be RT if the task got
927 * RT-boosted. If not then it returns p->normal_prio.
929 static int effective_prio(struct task_struct *p)
931 p->normal_prio = normal_prio(p);
933 * If we are RT tasks or we were boosted to RT priority,
934 * keep the priority unchanged. Otherwise, update priority
935 * to the normal priority:
937 if (!rt_prio(p->prio))
938 return p->normal_prio;
943 * activate_task - move a task to the runqueue.
945 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
947 if (p->state == TASK_UNINTERRUPTIBLE)
948 rq->nr_uninterruptible--;
950 enqueue_task(rq, p, wakeup);
951 inc_nr_running(p, rq);
955 * activate_idle_task - move idle task to the _front_ of runqueue.
957 static inline void activate_idle_task(struct task_struct *p, struct rq *rq)
961 if (p->state == TASK_UNINTERRUPTIBLE)
962 rq->nr_uninterruptible--;
964 enqueue_task(rq, p, 0);
965 inc_nr_running(p, rq);
969 * deactivate_task - remove a task from the runqueue.
971 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
973 if (p->state == TASK_UNINTERRUPTIBLE)
974 rq->nr_uninterruptible++;
976 dequeue_task(rq, p, sleep);
977 dec_nr_running(p, rq);
981 * task_curr - is this task currently executing on a CPU?
982 * @p: the task in question.
984 inline int task_curr(const struct task_struct *p)
986 return cpu_curr(task_cpu(p)) == p;
989 /* Used instead of source_load when we know the type == 0 */
990 unsigned long weighted_cpuload(const int cpu)
992 return cpu_rq(cpu)->ls.load.weight;
995 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
998 task_thread_info(p)->cpu = cpu;
1005 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1007 int old_cpu = task_cpu(p);
1008 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1009 u64 clock_offset, fair_clock_offset;
1011 clock_offset = old_rq->clock - new_rq->clock;
1012 fair_clock_offset = old_rq->cfs.fair_clock - new_rq->cfs.fair_clock;
1014 if (p->se.wait_start_fair)
1015 p->se.wait_start_fair -= fair_clock_offset;
1016 if (p->se.sleep_start_fair)
1017 p->se.sleep_start_fair -= fair_clock_offset;
1019 #ifdef CONFIG_SCHEDSTATS
1020 if (p->se.wait_start)
1021 p->se.wait_start -= clock_offset;
1022 if (p->se.sleep_start)
1023 p->se.sleep_start -= clock_offset;
1024 if (p->se.block_start)
1025 p->se.block_start -= clock_offset;
1028 __set_task_cpu(p, new_cpu);
1031 struct migration_req {
1032 struct list_head list;
1034 struct task_struct *task;
1037 struct completion done;
1041 * The task's runqueue lock must be held.
1042 * Returns true if you have to wait for migration thread.
1045 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1047 struct rq *rq = task_rq(p);
1050 * If the task is not on a runqueue (and not running), then
1051 * it is sufficient to simply update the task's cpu field.
1053 if (!p->se.on_rq && !task_running(rq, p)) {
1054 set_task_cpu(p, dest_cpu);
1058 init_completion(&req->done);
1060 req->dest_cpu = dest_cpu;
1061 list_add(&req->list, &rq->migration_queue);
1067 * wait_task_inactive - wait for a thread to unschedule.
1069 * The caller must ensure that the task *will* unschedule sometime soon,
1070 * else this function might spin for a *long* time. This function can't
1071 * be called with interrupts off, or it may introduce deadlock with
1072 * smp_call_function() if an IPI is sent by the same process we are
1073 * waiting to become inactive.
1075 void wait_task_inactive(struct task_struct *p)
1077 unsigned long flags;
1083 * We do the initial early heuristics without holding
1084 * any task-queue locks at all. We'll only try to get
1085 * the runqueue lock when things look like they will
1091 * If the task is actively running on another CPU
1092 * still, just relax and busy-wait without holding
1095 * NOTE! Since we don't hold any locks, it's not
1096 * even sure that "rq" stays as the right runqueue!
1097 * But we don't care, since "task_running()" will
1098 * return false if the runqueue has changed and p
1099 * is actually now running somewhere else!
1101 while (task_running(rq, p))
1105 * Ok, time to look more closely! We need the rq
1106 * lock now, to be *sure*. If we're wrong, we'll
1107 * just go back and repeat.
1109 rq = task_rq_lock(p, &flags);
1110 running = task_running(rq, p);
1111 on_rq = p->se.on_rq;
1112 task_rq_unlock(rq, &flags);
1115 * Was it really running after all now that we
1116 * checked with the proper locks actually held?
1118 * Oops. Go back and try again..
1120 if (unlikely(running)) {
1126 * It's not enough that it's not actively running,
1127 * it must be off the runqueue _entirely_, and not
1130 * So if it wa still runnable (but just not actively
1131 * running right now), it's preempted, and we should
1132 * yield - it could be a while.
1134 if (unlikely(on_rq)) {
1140 * Ahh, all good. It wasn't running, and it wasn't
1141 * runnable, which means that it will never become
1142 * running in the future either. We're all done!
1147 * kick_process - kick a running thread to enter/exit the kernel
1148 * @p: the to-be-kicked thread
1150 * Cause a process which is running on another CPU to enter
1151 * kernel-mode, without any delay. (to get signals handled.)
1153 * NOTE: this function doesnt have to take the runqueue lock,
1154 * because all it wants to ensure is that the remote task enters
1155 * the kernel. If the IPI races and the task has been migrated
1156 * to another CPU then no harm is done and the purpose has been
1159 void kick_process(struct task_struct *p)
1165 if ((cpu != smp_processor_id()) && task_curr(p))
1166 smp_send_reschedule(cpu);
1171 * Return a low guess at the load of a migration-source cpu weighted
1172 * according to the scheduling class and "nice" value.
1174 * We want to under-estimate the load of migration sources, to
1175 * balance conservatively.
1177 static inline unsigned long source_load(int cpu, int type)
1179 struct rq *rq = cpu_rq(cpu);
1180 unsigned long total = weighted_cpuload(cpu);
1185 return min(rq->cpu_load[type-1], total);
1189 * Return a high guess at the load of a migration-target cpu weighted
1190 * according to the scheduling class and "nice" value.
1192 static inline unsigned long target_load(int cpu, int type)
1194 struct rq *rq = cpu_rq(cpu);
1195 unsigned long total = weighted_cpuload(cpu);
1200 return max(rq->cpu_load[type-1], total);
1204 * Return the average load per task on the cpu's run queue
1206 static inline unsigned long cpu_avg_load_per_task(int cpu)
1208 struct rq *rq = cpu_rq(cpu);
1209 unsigned long total = weighted_cpuload(cpu);
1210 unsigned long n = rq->nr_running;
1212 return n ? total / n : SCHED_LOAD_SCALE;
1216 * find_idlest_group finds and returns the least busy CPU group within the
1219 static struct sched_group *
1220 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1222 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1223 unsigned long min_load = ULONG_MAX, this_load = 0;
1224 int load_idx = sd->forkexec_idx;
1225 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1228 unsigned long load, avg_load;
1232 /* Skip over this group if it has no CPUs allowed */
1233 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1236 local_group = cpu_isset(this_cpu, group->cpumask);
1238 /* Tally up the load of all CPUs in the group */
1241 for_each_cpu_mask(i, group->cpumask) {
1242 /* Bias balancing toward cpus of our domain */
1244 load = source_load(i, load_idx);
1246 load = target_load(i, load_idx);
1251 /* Adjust by relative CPU power of the group */
1252 avg_load = sg_div_cpu_power(group,
1253 avg_load * SCHED_LOAD_SCALE);
1256 this_load = avg_load;
1258 } else if (avg_load < min_load) {
1259 min_load = avg_load;
1263 group = group->next;
1264 } while (group != sd->groups);
1266 if (!idlest || 100*this_load < imbalance*min_load)
1272 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1275 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1278 unsigned long load, min_load = ULONG_MAX;
1282 /* Traverse only the allowed CPUs */
1283 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1285 for_each_cpu_mask(i, tmp) {
1286 load = weighted_cpuload(i);
1288 if (load < min_load || (load == min_load && i == this_cpu)) {
1298 * sched_balance_self: balance the current task (running on cpu) in domains
1299 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1302 * Balance, ie. select the least loaded group.
1304 * Returns the target CPU number, or the same CPU if no balancing is needed.
1306 * preempt must be disabled.
1308 static int sched_balance_self(int cpu, int flag)
1310 struct task_struct *t = current;
1311 struct sched_domain *tmp, *sd = NULL;
1313 for_each_domain(cpu, tmp) {
1315 * If power savings logic is enabled for a domain, stop there.
1317 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1319 if (tmp->flags & flag)
1325 struct sched_group *group;
1326 int new_cpu, weight;
1328 if (!(sd->flags & flag)) {
1334 group = find_idlest_group(sd, t, cpu);
1340 new_cpu = find_idlest_cpu(group, t, cpu);
1341 if (new_cpu == -1 || new_cpu == cpu) {
1342 /* Now try balancing at a lower domain level of cpu */
1347 /* Now try balancing at a lower domain level of new_cpu */
1350 weight = cpus_weight(span);
1351 for_each_domain(cpu, tmp) {
1352 if (weight <= cpus_weight(tmp->span))
1354 if (tmp->flags & flag)
1357 /* while loop will break here if sd == NULL */
1363 #endif /* CONFIG_SMP */
1366 * wake_idle() will wake a task on an idle cpu if task->cpu is
1367 * not idle and an idle cpu is available. The span of cpus to
1368 * search starts with cpus closest then further out as needed,
1369 * so we always favor a closer, idle cpu.
1371 * Returns the CPU we should wake onto.
1373 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1374 static int wake_idle(int cpu, struct task_struct *p)
1377 struct sched_domain *sd;
1381 * If it is idle, then it is the best cpu to run this task.
1383 * This cpu is also the best, if it has more than one task already.
1384 * Siblings must be also busy(in most cases) as they didn't already
1385 * pickup the extra load from this cpu and hence we need not check
1386 * sibling runqueue info. This will avoid the checks and cache miss
1387 * penalities associated with that.
1389 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1392 for_each_domain(cpu, sd) {
1393 if (sd->flags & SD_WAKE_IDLE) {
1394 cpus_and(tmp, sd->span, p->cpus_allowed);
1395 for_each_cpu_mask(i, tmp) {
1406 static inline int wake_idle(int cpu, struct task_struct *p)
1413 * try_to_wake_up - wake up a thread
1414 * @p: the to-be-woken-up thread
1415 * @state: the mask of task states that can be woken
1416 * @sync: do a synchronous wakeup?
1418 * Put it on the run-queue if it's not already there. The "current"
1419 * thread is always on the run-queue (except when the actual
1420 * re-schedule is in progress), and as such you're allowed to do
1421 * the simpler "current->state = TASK_RUNNING" to mark yourself
1422 * runnable without the overhead of this.
1424 * returns failure only if the task is already active.
1426 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1428 int cpu, this_cpu, success = 0;
1429 unsigned long flags;
1433 struct sched_domain *sd, *this_sd = NULL;
1434 unsigned long load, this_load;
1438 rq = task_rq_lock(p, &flags);
1439 old_state = p->state;
1440 if (!(old_state & state))
1447 this_cpu = smp_processor_id();
1450 if (unlikely(task_running(rq, p)))
1455 schedstat_inc(rq, ttwu_cnt);
1456 if (cpu == this_cpu) {
1457 schedstat_inc(rq, ttwu_local);
1461 for_each_domain(this_cpu, sd) {
1462 if (cpu_isset(cpu, sd->span)) {
1463 schedstat_inc(sd, ttwu_wake_remote);
1469 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1473 * Check for affine wakeup and passive balancing possibilities.
1476 int idx = this_sd->wake_idx;
1477 unsigned int imbalance;
1479 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1481 load = source_load(cpu, idx);
1482 this_load = target_load(this_cpu, idx);
1484 new_cpu = this_cpu; /* Wake to this CPU if we can */
1486 if (this_sd->flags & SD_WAKE_AFFINE) {
1487 unsigned long tl = this_load;
1488 unsigned long tl_per_task;
1490 tl_per_task = cpu_avg_load_per_task(this_cpu);
1493 * If sync wakeup then subtract the (maximum possible)
1494 * effect of the currently running task from the load
1495 * of the current CPU:
1498 tl -= current->se.load.weight;
1501 tl + target_load(cpu, idx) <= tl_per_task) ||
1502 100*(tl + p->se.load.weight) <= imbalance*load) {
1504 * This domain has SD_WAKE_AFFINE and
1505 * p is cache cold in this domain, and
1506 * there is no bad imbalance.
1508 schedstat_inc(this_sd, ttwu_move_affine);
1514 * Start passive balancing when half the imbalance_pct
1517 if (this_sd->flags & SD_WAKE_BALANCE) {
1518 if (imbalance*this_load <= 100*load) {
1519 schedstat_inc(this_sd, ttwu_move_balance);
1525 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1527 new_cpu = wake_idle(new_cpu, p);
1528 if (new_cpu != cpu) {
1529 set_task_cpu(p, new_cpu);
1530 task_rq_unlock(rq, &flags);
1531 /* might preempt at this point */
1532 rq = task_rq_lock(p, &flags);
1533 old_state = p->state;
1534 if (!(old_state & state))
1539 this_cpu = smp_processor_id();
1544 #endif /* CONFIG_SMP */
1545 update_rq_clock(rq);
1546 activate_task(rq, p, 1);
1548 * Sync wakeups (i.e. those types of wakeups where the waker
1549 * has indicated that it will leave the CPU in short order)
1550 * don't trigger a preemption, if the woken up task will run on
1551 * this cpu. (in this case the 'I will reschedule' promise of
1552 * the waker guarantees that the freshly woken up task is going
1553 * to be considered on this CPU.)
1555 if (!sync || cpu != this_cpu)
1556 check_preempt_curr(rq, p);
1560 p->state = TASK_RUNNING;
1562 task_rq_unlock(rq, &flags);
1567 int fastcall wake_up_process(struct task_struct *p)
1569 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1570 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1572 EXPORT_SYMBOL(wake_up_process);
1574 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1576 return try_to_wake_up(p, state, 0);
1580 * Perform scheduler related setup for a newly forked process p.
1581 * p is forked by current.
1583 * __sched_fork() is basic setup used by init_idle() too:
1585 static void __sched_fork(struct task_struct *p)
1587 p->se.wait_start_fair = 0;
1588 p->se.exec_start = 0;
1589 p->se.sum_exec_runtime = 0;
1590 p->se.prev_sum_exec_runtime = 0;
1591 p->se.delta_exec = 0;
1592 p->se.delta_fair_run = 0;
1593 p->se.delta_fair_sleep = 0;
1594 p->se.wait_runtime = 0;
1595 p->se.sleep_start_fair = 0;
1597 #ifdef CONFIG_SCHEDSTATS
1598 p->se.wait_start = 0;
1599 p->se.sum_wait_runtime = 0;
1600 p->se.sum_sleep_runtime = 0;
1601 p->se.sleep_start = 0;
1602 p->se.block_start = 0;
1603 p->se.sleep_max = 0;
1604 p->se.block_max = 0;
1607 p->se.wait_runtime_overruns = 0;
1608 p->se.wait_runtime_underruns = 0;
1611 INIT_LIST_HEAD(&p->run_list);
1614 #ifdef CONFIG_PREEMPT_NOTIFIERS
1615 INIT_HLIST_HEAD(&p->preempt_notifiers);
1619 * We mark the process as running here, but have not actually
1620 * inserted it onto the runqueue yet. This guarantees that
1621 * nobody will actually run it, and a signal or other external
1622 * event cannot wake it up and insert it on the runqueue either.
1624 p->state = TASK_RUNNING;
1628 * fork()/clone()-time setup:
1630 void sched_fork(struct task_struct *p, int clone_flags)
1632 int cpu = get_cpu();
1637 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1639 __set_task_cpu(p, cpu);
1642 * Make sure we do not leak PI boosting priority to the child:
1644 p->prio = current->normal_prio;
1646 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1647 if (likely(sched_info_on()))
1648 memset(&p->sched_info, 0, sizeof(p->sched_info));
1650 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1653 #ifdef CONFIG_PREEMPT
1654 /* Want to start with kernel preemption disabled. */
1655 task_thread_info(p)->preempt_count = 1;
1661 * After fork, child runs first. (default) If set to 0 then
1662 * parent will (try to) run first.
1664 unsigned int __read_mostly sysctl_sched_child_runs_first = 1;
1667 * wake_up_new_task - wake up a newly created task for the first time.
1669 * This function will do some initial scheduler statistics housekeeping
1670 * that must be done for every newly created context, then puts the task
1671 * on the runqueue and wakes it.
1673 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1675 unsigned long flags;
1679 rq = task_rq_lock(p, &flags);
1680 BUG_ON(p->state != TASK_RUNNING);
1681 this_cpu = smp_processor_id(); /* parent's CPU */
1682 update_rq_clock(rq);
1684 p->prio = effective_prio(p);
1686 if (rt_prio(p->prio))
1687 p->sched_class = &rt_sched_class;
1689 p->sched_class = &fair_sched_class;
1691 if (task_cpu(p) != this_cpu || !p->sched_class->task_new ||
1692 !current->se.on_rq) {
1693 activate_task(rq, p, 0);
1696 * Let the scheduling class do new task startup
1697 * management (if any):
1699 p->sched_class->task_new(rq, p);
1700 inc_nr_running(p, rq);
1702 check_preempt_curr(rq, p);
1703 task_rq_unlock(rq, &flags);
1706 #ifdef CONFIG_PREEMPT_NOTIFIERS
1709 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1710 * @notifier: notifier struct to register
1712 void preempt_notifier_register(struct preempt_notifier *notifier)
1714 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1716 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1719 * preempt_notifier_unregister - no longer interested in preemption notifications
1720 * @notifier: notifier struct to unregister
1722 * This is safe to call from within a preemption notifier.
1724 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1726 hlist_del(¬ifier->link);
1728 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1730 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1732 struct preempt_notifier *notifier;
1733 struct hlist_node *node;
1735 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1736 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1740 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1741 struct task_struct *next)
1743 struct preempt_notifier *notifier;
1744 struct hlist_node *node;
1746 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1747 notifier->ops->sched_out(notifier, next);
1752 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1757 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1758 struct task_struct *next)
1765 * prepare_task_switch - prepare to switch tasks
1766 * @rq: the runqueue preparing to switch
1767 * @prev: the current task that is being switched out
1768 * @next: the task we are going to switch to.
1770 * This is called with the rq lock held and interrupts off. It must
1771 * be paired with a subsequent finish_task_switch after the context
1774 * prepare_task_switch sets up locking and calls architecture specific
1778 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1779 struct task_struct *next)
1781 fire_sched_out_preempt_notifiers(prev, next);
1782 prepare_lock_switch(rq, next);
1783 prepare_arch_switch(next);
1787 * finish_task_switch - clean up after a task-switch
1788 * @rq: runqueue associated with task-switch
1789 * @prev: the thread we just switched away from.
1791 * finish_task_switch must be called after the context switch, paired
1792 * with a prepare_task_switch call before the context switch.
1793 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1794 * and do any other architecture-specific cleanup actions.
1796 * Note that we may have delayed dropping an mm in context_switch(). If
1797 * so, we finish that here outside of the runqueue lock. (Doing it
1798 * with the lock held can cause deadlocks; see schedule() for
1801 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1802 __releases(rq->lock)
1804 struct mm_struct *mm = rq->prev_mm;
1810 * A task struct has one reference for the use as "current".
1811 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1812 * schedule one last time. The schedule call will never return, and
1813 * the scheduled task must drop that reference.
1814 * The test for TASK_DEAD must occur while the runqueue locks are
1815 * still held, otherwise prev could be scheduled on another cpu, die
1816 * there before we look at prev->state, and then the reference would
1818 * Manfred Spraul <manfred@colorfullife.com>
1820 prev_state = prev->state;
1821 finish_arch_switch(prev);
1822 finish_lock_switch(rq, prev);
1823 fire_sched_in_preempt_notifiers(current);
1826 if (unlikely(prev_state == TASK_DEAD)) {
1828 * Remove function-return probe instances associated with this
1829 * task and put them back on the free list.
1831 kprobe_flush_task(prev);
1832 put_task_struct(prev);
1837 * schedule_tail - first thing a freshly forked thread must call.
1838 * @prev: the thread we just switched away from.
1840 asmlinkage void schedule_tail(struct task_struct *prev)
1841 __releases(rq->lock)
1843 struct rq *rq = this_rq();
1845 finish_task_switch(rq, prev);
1846 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1847 /* In this case, finish_task_switch does not reenable preemption */
1850 if (current->set_child_tid)
1851 put_user(current->pid, current->set_child_tid);
1855 * context_switch - switch to the new MM and the new
1856 * thread's register state.
1859 context_switch(struct rq *rq, struct task_struct *prev,
1860 struct task_struct *next)
1862 struct mm_struct *mm, *oldmm;
1864 prepare_task_switch(rq, prev, next);
1866 oldmm = prev->active_mm;
1868 * For paravirt, this is coupled with an exit in switch_to to
1869 * combine the page table reload and the switch backend into
1872 arch_enter_lazy_cpu_mode();
1874 if (unlikely(!mm)) {
1875 next->active_mm = oldmm;
1876 atomic_inc(&oldmm->mm_count);
1877 enter_lazy_tlb(oldmm, next);
1879 switch_mm(oldmm, mm, next);
1881 if (unlikely(!prev->mm)) {
1882 prev->active_mm = NULL;
1883 rq->prev_mm = oldmm;
1886 * Since the runqueue lock will be released by the next
1887 * task (which is an invalid locking op but in the case
1888 * of the scheduler it's an obvious special-case), so we
1889 * do an early lockdep release here:
1891 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1892 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1895 /* Here we just switch the register state and the stack. */
1896 switch_to(prev, next, prev);
1900 * this_rq must be evaluated again because prev may have moved
1901 * CPUs since it called schedule(), thus the 'rq' on its stack
1902 * frame will be invalid.
1904 finish_task_switch(this_rq(), prev);
1908 * nr_running, nr_uninterruptible and nr_context_switches:
1910 * externally visible scheduler statistics: current number of runnable
1911 * threads, current number of uninterruptible-sleeping threads, total
1912 * number of context switches performed since bootup.
1914 unsigned long nr_running(void)
1916 unsigned long i, sum = 0;
1918 for_each_online_cpu(i)
1919 sum += cpu_rq(i)->nr_running;
1924 unsigned long nr_uninterruptible(void)
1926 unsigned long i, sum = 0;
1928 for_each_possible_cpu(i)
1929 sum += cpu_rq(i)->nr_uninterruptible;
1932 * Since we read the counters lockless, it might be slightly
1933 * inaccurate. Do not allow it to go below zero though:
1935 if (unlikely((long)sum < 0))
1941 unsigned long long nr_context_switches(void)
1944 unsigned long long sum = 0;
1946 for_each_possible_cpu(i)
1947 sum += cpu_rq(i)->nr_switches;
1952 unsigned long nr_iowait(void)
1954 unsigned long i, sum = 0;
1956 for_each_possible_cpu(i)
1957 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1962 unsigned long nr_active(void)
1964 unsigned long i, running = 0, uninterruptible = 0;
1966 for_each_online_cpu(i) {
1967 running += cpu_rq(i)->nr_running;
1968 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1971 if (unlikely((long)uninterruptible < 0))
1972 uninterruptible = 0;
1974 return running + uninterruptible;
1978 * Update rq->cpu_load[] statistics. This function is usually called every
1979 * scheduler tick (TICK_NSEC).
1981 static void update_cpu_load(struct rq *this_rq)
1983 u64 fair_delta64, exec_delta64, idle_delta64, sample_interval64, tmp64;
1984 unsigned long total_load = this_rq->ls.load.weight;
1985 unsigned long this_load = total_load;
1986 struct load_stat *ls = &this_rq->ls;
1989 this_rq->nr_load_updates++;
1990 if (unlikely(!(sysctl_sched_features & SCHED_FEAT_PRECISE_CPU_LOAD)))
1993 /* Update delta_fair/delta_exec fields first */
1994 update_curr_load(this_rq);
1996 fair_delta64 = ls->delta_fair + 1;
1999 exec_delta64 = ls->delta_exec + 1;
2002 sample_interval64 = this_rq->clock - ls->load_update_last;
2003 ls->load_update_last = this_rq->clock;
2005 if ((s64)sample_interval64 < (s64)TICK_NSEC)
2006 sample_interval64 = TICK_NSEC;
2008 if (exec_delta64 > sample_interval64)
2009 exec_delta64 = sample_interval64;
2011 idle_delta64 = sample_interval64 - exec_delta64;
2013 tmp64 = div64_64(SCHED_LOAD_SCALE * exec_delta64, fair_delta64);
2014 tmp64 = div64_64(tmp64 * exec_delta64, sample_interval64);
2016 this_load = (unsigned long)tmp64;
2020 /* Update our load: */
2021 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2022 unsigned long old_load, new_load;
2024 /* scale is effectively 1 << i now, and >> i divides by scale */
2026 old_load = this_rq->cpu_load[i];
2027 new_load = this_load;
2029 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2036 * double_rq_lock - safely lock two runqueues
2038 * Note this does not disable interrupts like task_rq_lock,
2039 * you need to do so manually before calling.
2041 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2042 __acquires(rq1->lock)
2043 __acquires(rq2->lock)
2045 BUG_ON(!irqs_disabled());
2047 spin_lock(&rq1->lock);
2048 __acquire(rq2->lock); /* Fake it out ;) */
2051 spin_lock(&rq1->lock);
2052 spin_lock(&rq2->lock);
2054 spin_lock(&rq2->lock);
2055 spin_lock(&rq1->lock);
2058 update_rq_clock(rq1);
2059 update_rq_clock(rq2);
2063 * double_rq_unlock - safely unlock two runqueues
2065 * Note this does not restore interrupts like task_rq_unlock,
2066 * you need to do so manually after calling.
2068 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2069 __releases(rq1->lock)
2070 __releases(rq2->lock)
2072 spin_unlock(&rq1->lock);
2074 spin_unlock(&rq2->lock);
2076 __release(rq2->lock);
2080 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2082 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2083 __releases(this_rq->lock)
2084 __acquires(busiest->lock)
2085 __acquires(this_rq->lock)
2087 if (unlikely(!irqs_disabled())) {
2088 /* printk() doesn't work good under rq->lock */
2089 spin_unlock(&this_rq->lock);
2092 if (unlikely(!spin_trylock(&busiest->lock))) {
2093 if (busiest < this_rq) {
2094 spin_unlock(&this_rq->lock);
2095 spin_lock(&busiest->lock);
2096 spin_lock(&this_rq->lock);
2098 spin_lock(&busiest->lock);
2103 * If dest_cpu is allowed for this process, migrate the task to it.
2104 * This is accomplished by forcing the cpu_allowed mask to only
2105 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2106 * the cpu_allowed mask is restored.
2108 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2110 struct migration_req req;
2111 unsigned long flags;
2114 rq = task_rq_lock(p, &flags);
2115 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2116 || unlikely(cpu_is_offline(dest_cpu)))
2119 /* force the process onto the specified CPU */
2120 if (migrate_task(p, dest_cpu, &req)) {
2121 /* Need to wait for migration thread (might exit: take ref). */
2122 struct task_struct *mt = rq->migration_thread;
2124 get_task_struct(mt);
2125 task_rq_unlock(rq, &flags);
2126 wake_up_process(mt);
2127 put_task_struct(mt);
2128 wait_for_completion(&req.done);
2133 task_rq_unlock(rq, &flags);
2137 * sched_exec - execve() is a valuable balancing opportunity, because at
2138 * this point the task has the smallest effective memory and cache footprint.
2140 void sched_exec(void)
2142 int new_cpu, this_cpu = get_cpu();
2143 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2145 if (new_cpu != this_cpu)
2146 sched_migrate_task(current, new_cpu);
2150 * pull_task - move a task from a remote runqueue to the local runqueue.
2151 * Both runqueues must be locked.
2153 static void pull_task(struct rq *src_rq, struct task_struct *p,
2154 struct rq *this_rq, int this_cpu)
2156 deactivate_task(src_rq, p, 0);
2157 set_task_cpu(p, this_cpu);
2158 activate_task(this_rq, p, 0);
2160 * Note that idle threads have a prio of MAX_PRIO, for this test
2161 * to be always true for them.
2163 check_preempt_curr(this_rq, p);
2167 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2170 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2171 struct sched_domain *sd, enum cpu_idle_type idle,
2175 * We do not migrate tasks that are:
2176 * 1) running (obviously), or
2177 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2178 * 3) are cache-hot on their current CPU.
2180 if (!cpu_isset(this_cpu, p->cpus_allowed))
2184 if (task_running(rq, p))
2190 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2191 unsigned long max_nr_move, unsigned long max_load_move,
2192 struct sched_domain *sd, enum cpu_idle_type idle,
2193 int *all_pinned, unsigned long *load_moved,
2194 int *this_best_prio, struct rq_iterator *iterator)
2196 int pulled = 0, pinned = 0, skip_for_load;
2197 struct task_struct *p;
2198 long rem_load_move = max_load_move;
2200 if (max_nr_move == 0 || max_load_move == 0)
2206 * Start the load-balancing iterator:
2208 p = iterator->start(iterator->arg);
2213 * To help distribute high priority tasks accross CPUs we don't
2214 * skip a task if it will be the highest priority task (i.e. smallest
2215 * prio value) on its new queue regardless of its load weight
2217 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2218 SCHED_LOAD_SCALE_FUZZ;
2219 if ((skip_for_load && p->prio >= *this_best_prio) ||
2220 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2221 p = iterator->next(iterator->arg);
2225 pull_task(busiest, p, this_rq, this_cpu);
2227 rem_load_move -= p->se.load.weight;
2230 * We only want to steal up to the prescribed number of tasks
2231 * and the prescribed amount of weighted load.
2233 if (pulled < max_nr_move && rem_load_move > 0) {
2234 if (p->prio < *this_best_prio)
2235 *this_best_prio = p->prio;
2236 p = iterator->next(iterator->arg);
2241 * Right now, this is the only place pull_task() is called,
2242 * so we can safely collect pull_task() stats here rather than
2243 * inside pull_task().
2245 schedstat_add(sd, lb_gained[idle], pulled);
2248 *all_pinned = pinned;
2249 *load_moved = max_load_move - rem_load_move;
2254 * move_tasks tries to move up to max_load_move weighted load from busiest to
2255 * this_rq, as part of a balancing operation within domain "sd".
2256 * Returns 1 if successful and 0 otherwise.
2258 * Called with both runqueues locked.
2260 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2261 unsigned long max_load_move,
2262 struct sched_domain *sd, enum cpu_idle_type idle,
2265 struct sched_class *class = sched_class_highest;
2266 unsigned long total_load_moved = 0;
2267 int this_best_prio = this_rq->curr->prio;
2271 class->load_balance(this_rq, this_cpu, busiest,
2272 ULONG_MAX, max_load_move - total_load_moved,
2273 sd, idle, all_pinned, &this_best_prio);
2274 class = class->next;
2275 } while (class && max_load_move > total_load_moved);
2277 return total_load_moved > 0;
2281 * move_one_task tries to move exactly one task from busiest to this_rq, as
2282 * part of active balancing operations within "domain".
2283 * Returns 1 if successful and 0 otherwise.
2285 * Called with both runqueues locked.
2287 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2288 struct sched_domain *sd, enum cpu_idle_type idle)
2290 struct sched_class *class;
2291 int this_best_prio = MAX_PRIO;
2293 for (class = sched_class_highest; class; class = class->next)
2294 if (class->load_balance(this_rq, this_cpu, busiest,
2295 1, ULONG_MAX, sd, idle, NULL,
2303 * find_busiest_group finds and returns the busiest CPU group within the
2304 * domain. It calculates and returns the amount of weighted load which
2305 * should be moved to restore balance via the imbalance parameter.
2307 static struct sched_group *
2308 find_busiest_group(struct sched_domain *sd, int this_cpu,
2309 unsigned long *imbalance, enum cpu_idle_type idle,
2310 int *sd_idle, cpumask_t *cpus, int *balance)
2312 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2313 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2314 unsigned long max_pull;
2315 unsigned long busiest_load_per_task, busiest_nr_running;
2316 unsigned long this_load_per_task, this_nr_running;
2318 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2319 int power_savings_balance = 1;
2320 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2321 unsigned long min_nr_running = ULONG_MAX;
2322 struct sched_group *group_min = NULL, *group_leader = NULL;
2325 max_load = this_load = total_load = total_pwr = 0;
2326 busiest_load_per_task = busiest_nr_running = 0;
2327 this_load_per_task = this_nr_running = 0;
2328 if (idle == CPU_NOT_IDLE)
2329 load_idx = sd->busy_idx;
2330 else if (idle == CPU_NEWLY_IDLE)
2331 load_idx = sd->newidle_idx;
2333 load_idx = sd->idle_idx;
2336 unsigned long load, group_capacity;
2339 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2340 unsigned long sum_nr_running, sum_weighted_load;
2342 local_group = cpu_isset(this_cpu, group->cpumask);
2345 balance_cpu = first_cpu(group->cpumask);
2347 /* Tally up the load of all CPUs in the group */
2348 sum_weighted_load = sum_nr_running = avg_load = 0;
2350 for_each_cpu_mask(i, group->cpumask) {
2353 if (!cpu_isset(i, *cpus))
2358 if (*sd_idle && rq->nr_running)
2361 /* Bias balancing toward cpus of our domain */
2363 if (idle_cpu(i) && !first_idle_cpu) {
2368 load = target_load(i, load_idx);
2370 load = source_load(i, load_idx);
2373 sum_nr_running += rq->nr_running;
2374 sum_weighted_load += weighted_cpuload(i);
2378 * First idle cpu or the first cpu(busiest) in this sched group
2379 * is eligible for doing load balancing at this and above
2380 * domains. In the newly idle case, we will allow all the cpu's
2381 * to do the newly idle load balance.
2383 if (idle != CPU_NEWLY_IDLE && local_group &&
2384 balance_cpu != this_cpu && balance) {
2389 total_load += avg_load;
2390 total_pwr += group->__cpu_power;
2392 /* Adjust by relative CPU power of the group */
2393 avg_load = sg_div_cpu_power(group,
2394 avg_load * SCHED_LOAD_SCALE);
2396 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2399 this_load = avg_load;
2401 this_nr_running = sum_nr_running;
2402 this_load_per_task = sum_weighted_load;
2403 } else if (avg_load > max_load &&
2404 sum_nr_running > group_capacity) {
2405 max_load = avg_load;
2407 busiest_nr_running = sum_nr_running;
2408 busiest_load_per_task = sum_weighted_load;
2411 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2413 * Busy processors will not participate in power savings
2416 if (idle == CPU_NOT_IDLE ||
2417 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2421 * If the local group is idle or completely loaded
2422 * no need to do power savings balance at this domain
2424 if (local_group && (this_nr_running >= group_capacity ||
2426 power_savings_balance = 0;
2429 * If a group is already running at full capacity or idle,
2430 * don't include that group in power savings calculations
2432 if (!power_savings_balance || sum_nr_running >= group_capacity
2437 * Calculate the group which has the least non-idle load.
2438 * This is the group from where we need to pick up the load
2441 if ((sum_nr_running < min_nr_running) ||
2442 (sum_nr_running == min_nr_running &&
2443 first_cpu(group->cpumask) <
2444 first_cpu(group_min->cpumask))) {
2446 min_nr_running = sum_nr_running;
2447 min_load_per_task = sum_weighted_load /
2452 * Calculate the group which is almost near its
2453 * capacity but still has some space to pick up some load
2454 * from other group and save more power
2456 if (sum_nr_running <= group_capacity - 1) {
2457 if (sum_nr_running > leader_nr_running ||
2458 (sum_nr_running == leader_nr_running &&
2459 first_cpu(group->cpumask) >
2460 first_cpu(group_leader->cpumask))) {
2461 group_leader = group;
2462 leader_nr_running = sum_nr_running;
2467 group = group->next;
2468 } while (group != sd->groups);
2470 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2473 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2475 if (this_load >= avg_load ||
2476 100*max_load <= sd->imbalance_pct*this_load)
2479 busiest_load_per_task /= busiest_nr_running;
2481 * We're trying to get all the cpus to the average_load, so we don't
2482 * want to push ourselves above the average load, nor do we wish to
2483 * reduce the max loaded cpu below the average load, as either of these
2484 * actions would just result in more rebalancing later, and ping-pong
2485 * tasks around. Thus we look for the minimum possible imbalance.
2486 * Negative imbalances (*we* are more loaded than anyone else) will
2487 * be counted as no imbalance for these purposes -- we can't fix that
2488 * by pulling tasks to us. Be careful of negative numbers as they'll
2489 * appear as very large values with unsigned longs.
2491 if (max_load <= busiest_load_per_task)
2495 * In the presence of smp nice balancing, certain scenarios can have
2496 * max load less than avg load(as we skip the groups at or below
2497 * its cpu_power, while calculating max_load..)
2499 if (max_load < avg_load) {
2501 goto small_imbalance;
2504 /* Don't want to pull so many tasks that a group would go idle */
2505 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2507 /* How much load to actually move to equalise the imbalance */
2508 *imbalance = min(max_pull * busiest->__cpu_power,
2509 (avg_load - this_load) * this->__cpu_power)
2513 * if *imbalance is less than the average load per runnable task
2514 * there is no gaurantee that any tasks will be moved so we'll have
2515 * a think about bumping its value to force at least one task to be
2518 if (*imbalance < busiest_load_per_task) {
2519 unsigned long tmp, pwr_now, pwr_move;
2523 pwr_move = pwr_now = 0;
2525 if (this_nr_running) {
2526 this_load_per_task /= this_nr_running;
2527 if (busiest_load_per_task > this_load_per_task)
2530 this_load_per_task = SCHED_LOAD_SCALE;
2532 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2533 busiest_load_per_task * imbn) {
2534 *imbalance = busiest_load_per_task;
2539 * OK, we don't have enough imbalance to justify moving tasks,
2540 * however we may be able to increase total CPU power used by
2544 pwr_now += busiest->__cpu_power *
2545 min(busiest_load_per_task, max_load);
2546 pwr_now += this->__cpu_power *
2547 min(this_load_per_task, this_load);
2548 pwr_now /= SCHED_LOAD_SCALE;
2550 /* Amount of load we'd subtract */
2551 tmp = sg_div_cpu_power(busiest,
2552 busiest_load_per_task * SCHED_LOAD_SCALE);
2554 pwr_move += busiest->__cpu_power *
2555 min(busiest_load_per_task, max_load - tmp);
2557 /* Amount of load we'd add */
2558 if (max_load * busiest->__cpu_power <
2559 busiest_load_per_task * SCHED_LOAD_SCALE)
2560 tmp = sg_div_cpu_power(this,
2561 max_load * busiest->__cpu_power);
2563 tmp = sg_div_cpu_power(this,
2564 busiest_load_per_task * SCHED_LOAD_SCALE);
2565 pwr_move += this->__cpu_power *
2566 min(this_load_per_task, this_load + tmp);
2567 pwr_move /= SCHED_LOAD_SCALE;
2569 /* Move if we gain throughput */
2570 if (pwr_move > pwr_now)
2571 *imbalance = busiest_load_per_task;
2577 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2578 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2581 if (this == group_leader && group_leader != group_min) {
2582 *imbalance = min_load_per_task;
2592 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2595 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2596 unsigned long imbalance, cpumask_t *cpus)
2598 struct rq *busiest = NULL, *rq;
2599 unsigned long max_load = 0;
2602 for_each_cpu_mask(i, group->cpumask) {
2605 if (!cpu_isset(i, *cpus))
2609 wl = weighted_cpuload(i);
2611 if (rq->nr_running == 1 && wl > imbalance)
2614 if (wl > max_load) {
2624 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2625 * so long as it is large enough.
2627 #define MAX_PINNED_INTERVAL 512
2630 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2631 * tasks if there is an imbalance.
2633 static int load_balance(int this_cpu, struct rq *this_rq,
2634 struct sched_domain *sd, enum cpu_idle_type idle,
2637 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2638 struct sched_group *group;
2639 unsigned long imbalance;
2641 cpumask_t cpus = CPU_MASK_ALL;
2642 unsigned long flags;
2645 * When power savings policy is enabled for the parent domain, idle
2646 * sibling can pick up load irrespective of busy siblings. In this case,
2647 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2648 * portraying it as CPU_NOT_IDLE.
2650 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2651 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2654 schedstat_inc(sd, lb_cnt[idle]);
2657 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2664 schedstat_inc(sd, lb_nobusyg[idle]);
2668 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2670 schedstat_inc(sd, lb_nobusyq[idle]);
2674 BUG_ON(busiest == this_rq);
2676 schedstat_add(sd, lb_imbalance[idle], imbalance);
2679 if (busiest->nr_running > 1) {
2681 * Attempt to move tasks. If find_busiest_group has found
2682 * an imbalance but busiest->nr_running <= 1, the group is
2683 * still unbalanced. ld_moved simply stays zero, so it is
2684 * correctly treated as an imbalance.
2686 local_irq_save(flags);
2687 double_rq_lock(this_rq, busiest);
2688 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2689 imbalance, sd, idle, &all_pinned);
2690 double_rq_unlock(this_rq, busiest);
2691 local_irq_restore(flags);
2694 * some other cpu did the load balance for us.
2696 if (ld_moved && this_cpu != smp_processor_id())
2697 resched_cpu(this_cpu);
2699 /* All tasks on this runqueue were pinned by CPU affinity */
2700 if (unlikely(all_pinned)) {
2701 cpu_clear(cpu_of(busiest), cpus);
2702 if (!cpus_empty(cpus))
2709 schedstat_inc(sd, lb_failed[idle]);
2710 sd->nr_balance_failed++;
2712 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2714 spin_lock_irqsave(&busiest->lock, flags);
2716 /* don't kick the migration_thread, if the curr
2717 * task on busiest cpu can't be moved to this_cpu
2719 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2720 spin_unlock_irqrestore(&busiest->lock, flags);
2722 goto out_one_pinned;
2725 if (!busiest->active_balance) {
2726 busiest->active_balance = 1;
2727 busiest->push_cpu = this_cpu;
2730 spin_unlock_irqrestore(&busiest->lock, flags);
2732 wake_up_process(busiest->migration_thread);
2735 * We've kicked active balancing, reset the failure
2738 sd->nr_balance_failed = sd->cache_nice_tries+1;
2741 sd->nr_balance_failed = 0;
2743 if (likely(!active_balance)) {
2744 /* We were unbalanced, so reset the balancing interval */
2745 sd->balance_interval = sd->min_interval;
2748 * If we've begun active balancing, start to back off. This
2749 * case may not be covered by the all_pinned logic if there
2750 * is only 1 task on the busy runqueue (because we don't call
2753 if (sd->balance_interval < sd->max_interval)
2754 sd->balance_interval *= 2;
2757 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2758 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2763 schedstat_inc(sd, lb_balanced[idle]);
2765 sd->nr_balance_failed = 0;
2768 /* tune up the balancing interval */
2769 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2770 (sd->balance_interval < sd->max_interval))
2771 sd->balance_interval *= 2;
2773 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2774 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2780 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2781 * tasks if there is an imbalance.
2783 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2784 * this_rq is locked.
2787 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2789 struct sched_group *group;
2790 struct rq *busiest = NULL;
2791 unsigned long imbalance;
2795 cpumask_t cpus = CPU_MASK_ALL;
2798 * When power savings policy is enabled for the parent domain, idle
2799 * sibling can pick up load irrespective of busy siblings. In this case,
2800 * let the state of idle sibling percolate up as IDLE, instead of
2801 * portraying it as CPU_NOT_IDLE.
2803 if (sd->flags & SD_SHARE_CPUPOWER &&
2804 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2807 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2809 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2810 &sd_idle, &cpus, NULL);
2812 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2816 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2819 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2823 BUG_ON(busiest == this_rq);
2825 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2828 if (busiest->nr_running > 1) {
2829 /* Attempt to move tasks */
2830 double_lock_balance(this_rq, busiest);
2831 /* this_rq->clock is already updated */
2832 update_rq_clock(busiest);
2833 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2834 imbalance, sd, CPU_NEWLY_IDLE,
2836 spin_unlock(&busiest->lock);
2838 if (unlikely(all_pinned)) {
2839 cpu_clear(cpu_of(busiest), cpus);
2840 if (!cpus_empty(cpus))
2846 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2847 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2848 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2851 sd->nr_balance_failed = 0;
2856 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2857 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2858 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2860 sd->nr_balance_failed = 0;
2866 * idle_balance is called by schedule() if this_cpu is about to become
2867 * idle. Attempts to pull tasks from other CPUs.
2869 static void idle_balance(int this_cpu, struct rq *this_rq)
2871 struct sched_domain *sd;
2872 int pulled_task = -1;
2873 unsigned long next_balance = jiffies + HZ;
2875 for_each_domain(this_cpu, sd) {
2876 unsigned long interval;
2878 if (!(sd->flags & SD_LOAD_BALANCE))
2881 if (sd->flags & SD_BALANCE_NEWIDLE)
2882 /* If we've pulled tasks over stop searching: */
2883 pulled_task = load_balance_newidle(this_cpu,
2886 interval = msecs_to_jiffies(sd->balance_interval);
2887 if (time_after(next_balance, sd->last_balance + interval))
2888 next_balance = sd->last_balance + interval;
2892 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2894 * We are going idle. next_balance may be set based on
2895 * a busy processor. So reset next_balance.
2897 this_rq->next_balance = next_balance;
2902 * active_load_balance is run by migration threads. It pushes running tasks
2903 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2904 * running on each physical CPU where possible, and avoids physical /
2905 * logical imbalances.
2907 * Called with busiest_rq locked.
2909 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2911 int target_cpu = busiest_rq->push_cpu;
2912 struct sched_domain *sd;
2913 struct rq *target_rq;
2915 /* Is there any task to move? */
2916 if (busiest_rq->nr_running <= 1)
2919 target_rq = cpu_rq(target_cpu);
2922 * This condition is "impossible", if it occurs
2923 * we need to fix it. Originally reported by
2924 * Bjorn Helgaas on a 128-cpu setup.
2926 BUG_ON(busiest_rq == target_rq);
2928 /* move a task from busiest_rq to target_rq */
2929 double_lock_balance(busiest_rq, target_rq);
2930 update_rq_clock(busiest_rq);
2931 update_rq_clock(target_rq);
2933 /* Search for an sd spanning us and the target CPU. */
2934 for_each_domain(target_cpu, sd) {
2935 if ((sd->flags & SD_LOAD_BALANCE) &&
2936 cpu_isset(busiest_cpu, sd->span))
2941 schedstat_inc(sd, alb_cnt);
2943 if (move_one_task(target_rq, target_cpu, busiest_rq,
2945 schedstat_inc(sd, alb_pushed);
2947 schedstat_inc(sd, alb_failed);
2949 spin_unlock(&target_rq->lock);
2954 atomic_t load_balancer;
2956 } nohz ____cacheline_aligned = {
2957 .load_balancer = ATOMIC_INIT(-1),
2958 .cpu_mask = CPU_MASK_NONE,
2962 * This routine will try to nominate the ilb (idle load balancing)
2963 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2964 * load balancing on behalf of all those cpus. If all the cpus in the system
2965 * go into this tickless mode, then there will be no ilb owner (as there is
2966 * no need for one) and all the cpus will sleep till the next wakeup event
2969 * For the ilb owner, tick is not stopped. And this tick will be used
2970 * for idle load balancing. ilb owner will still be part of
2973 * While stopping the tick, this cpu will become the ilb owner if there
2974 * is no other owner. And will be the owner till that cpu becomes busy
2975 * or if all cpus in the system stop their ticks at which point
2976 * there is no need for ilb owner.
2978 * When the ilb owner becomes busy, it nominates another owner, during the
2979 * next busy scheduler_tick()
2981 int select_nohz_load_balancer(int stop_tick)
2983 int cpu = smp_processor_id();
2986 cpu_set(cpu, nohz.cpu_mask);
2987 cpu_rq(cpu)->in_nohz_recently = 1;
2990 * If we are going offline and still the leader, give up!
2992 if (cpu_is_offline(cpu) &&
2993 atomic_read(&nohz.load_balancer) == cpu) {
2994 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2999 /* time for ilb owner also to sleep */
3000 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3001 if (atomic_read(&nohz.load_balancer) == cpu)
3002 atomic_set(&nohz.load_balancer, -1);
3006 if (atomic_read(&nohz.load_balancer) == -1) {
3007 /* make me the ilb owner */
3008 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3010 } else if (atomic_read(&nohz.load_balancer) == cpu)
3013 if (!cpu_isset(cpu, nohz.cpu_mask))
3016 cpu_clear(cpu, nohz.cpu_mask);
3018 if (atomic_read(&nohz.load_balancer) == cpu)
3019 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3026 static DEFINE_SPINLOCK(balancing);
3029 * It checks each scheduling domain to see if it is due to be balanced,
3030 * and initiates a balancing operation if so.
3032 * Balancing parameters are set up in arch_init_sched_domains.
3034 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
3037 struct rq *rq = cpu_rq(cpu);
3038 unsigned long interval;
3039 struct sched_domain *sd;
3040 /* Earliest time when we have to do rebalance again */
3041 unsigned long next_balance = jiffies + 60*HZ;
3042 int update_next_balance = 0;
3044 for_each_domain(cpu, sd) {
3045 if (!(sd->flags & SD_LOAD_BALANCE))
3048 interval = sd->balance_interval;
3049 if (idle != CPU_IDLE)
3050 interval *= sd->busy_factor;
3052 /* scale ms to jiffies */
3053 interval = msecs_to_jiffies(interval);
3054 if (unlikely(!interval))
3056 if (interval > HZ*NR_CPUS/10)
3057 interval = HZ*NR_CPUS/10;
3060 if (sd->flags & SD_SERIALIZE) {
3061 if (!spin_trylock(&balancing))
3065 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3066 if (load_balance(cpu, rq, sd, idle, &balance)) {
3068 * We've pulled tasks over so either we're no
3069 * longer idle, or one of our SMT siblings is
3072 idle = CPU_NOT_IDLE;
3074 sd->last_balance = jiffies;
3076 if (sd->flags & SD_SERIALIZE)
3077 spin_unlock(&balancing);
3079 if (time_after(next_balance, sd->last_balance + interval)) {
3080 next_balance = sd->last_balance + interval;
3081 update_next_balance = 1;
3085 * Stop the load balance at this level. There is another
3086 * CPU in our sched group which is doing load balancing more
3094 * next_balance will be updated only when there is a need.
3095 * When the cpu is attached to null domain for ex, it will not be
3098 if (likely(update_next_balance))
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 this_cpu = smp_processor_id();
3110 struct rq *this_rq = cpu_rq(this_cpu);
3111 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3112 CPU_IDLE : CPU_NOT_IDLE;
3114 rebalance_domains(this_cpu, idle);
3118 * If this cpu is the owner for idle load balancing, then do the
3119 * balancing on behalf of the other idle cpus whose ticks are
3122 if (this_rq->idle_at_tick &&
3123 atomic_read(&nohz.load_balancer) == this_cpu) {
3124 cpumask_t cpus = nohz.cpu_mask;
3128 cpu_clear(this_cpu, cpus);
3129 for_each_cpu_mask(balance_cpu, cpus) {
3131 * If this cpu gets work to do, stop the load balancing
3132 * work being done for other cpus. Next load
3133 * balancing owner will pick it up.
3138 rebalance_domains(balance_cpu, CPU_IDLE);
3140 rq = cpu_rq(balance_cpu);
3141 if (time_after(this_rq->next_balance, rq->next_balance))
3142 this_rq->next_balance = rq->next_balance;
3149 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3151 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3152 * idle load balancing owner or decide to stop the periodic load balancing,
3153 * if the whole system is idle.
3155 static inline void trigger_load_balance(struct rq *rq, int 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);
3209 #else /* CONFIG_SMP */
3212 * on UP we do not need to balance between CPUs:
3214 static inline void idle_balance(int cpu, struct rq *rq)
3218 /* Avoid "used but not defined" warning on UP */
3219 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3220 unsigned long max_nr_move, unsigned long max_load_move,
3221 struct sched_domain *sd, enum cpu_idle_type idle,
3222 int *all_pinned, unsigned long *load_moved,
3223 int *this_best_prio, struct rq_iterator *iterator)
3232 DEFINE_PER_CPU(struct kernel_stat, kstat);
3234 EXPORT_PER_CPU_SYMBOL(kstat);
3237 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3238 * that have not yet been banked in case the task is currently running.
3240 unsigned long long task_sched_runtime(struct task_struct *p)
3242 unsigned long flags;
3246 rq = task_rq_lock(p, &flags);
3247 ns = p->se.sum_exec_runtime;
3248 if (rq->curr == p) {
3249 update_rq_clock(rq);
3250 delta_exec = rq->clock - p->se.exec_start;
3251 if ((s64)delta_exec > 0)
3254 task_rq_unlock(rq, &flags);
3260 * Account user cpu time to a process.
3261 * @p: the process that the cpu time gets accounted to
3262 * @hardirq_offset: the offset to subtract from hardirq_count()
3263 * @cputime: the cpu time spent in user space since the last update
3265 void account_user_time(struct task_struct *p, cputime_t cputime)
3267 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3270 p->utime = cputime_add(p->utime, cputime);
3272 /* Add user time to cpustat. */
3273 tmp = cputime_to_cputime64(cputime);
3274 if (TASK_NICE(p) > 0)
3275 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3277 cpustat->user = cputime64_add(cpustat->user, tmp);
3281 * Account system cpu time to a process.
3282 * @p: the process that the cpu time gets accounted to
3283 * @hardirq_offset: the offset to subtract from hardirq_count()
3284 * @cputime: the cpu time spent in kernel space since the last update
3286 void account_system_time(struct task_struct *p, int hardirq_offset,
3289 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3290 struct rq *rq = this_rq();
3293 p->stime = cputime_add(p->stime, cputime);
3295 /* Add system time to cpustat. */
3296 tmp = cputime_to_cputime64(cputime);
3297 if (hardirq_count() - hardirq_offset)
3298 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3299 else if (softirq_count())
3300 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3301 else if (p != rq->idle)
3302 cpustat->system = cputime64_add(cpustat->system, tmp);
3303 else if (atomic_read(&rq->nr_iowait) > 0)
3304 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3306 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3307 /* Account for system time used */
3308 acct_update_integrals(p);
3312 * Account for involuntary wait time.
3313 * @p: the process from which the cpu time has been stolen
3314 * @steal: the cpu time spent in involuntary wait
3316 void account_steal_time(struct task_struct *p, cputime_t steal)
3318 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3319 cputime64_t tmp = cputime_to_cputime64(steal);
3320 struct rq *rq = this_rq();
3322 if (p == rq->idle) {
3323 p->stime = cputime_add(p->stime, steal);
3324 if (atomic_read(&rq->nr_iowait) > 0)
3325 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3327 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3329 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3333 * This function gets called by the timer code, with HZ frequency.
3334 * We call it with interrupts disabled.
3336 * It also gets called by the fork code, when changing the parent's
3339 void scheduler_tick(void)
3341 int cpu = smp_processor_id();
3342 struct rq *rq = cpu_rq(cpu);
3343 struct task_struct *curr = rq->curr;
3344 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3346 spin_lock(&rq->lock);
3347 __update_rq_clock(rq);
3349 * Let rq->clock advance by at least TICK_NSEC:
3351 if (unlikely(rq->clock < next_tick))
3352 rq->clock = next_tick;
3353 rq->tick_timestamp = rq->clock;
3354 update_cpu_load(rq);
3355 if (curr != rq->idle) /* FIXME: needed? */
3356 curr->sched_class->task_tick(rq, curr);
3357 spin_unlock(&rq->lock);
3360 rq->idle_at_tick = idle_cpu(cpu);
3361 trigger_load_balance(rq, cpu);
3365 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3367 void fastcall add_preempt_count(int val)
3372 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3374 preempt_count() += val;
3376 * Spinlock count overflowing soon?
3378 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3381 EXPORT_SYMBOL(add_preempt_count);
3383 void fastcall sub_preempt_count(int val)
3388 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3391 * Is the spinlock portion underflowing?
3393 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3394 !(preempt_count() & PREEMPT_MASK)))
3397 preempt_count() -= val;
3399 EXPORT_SYMBOL(sub_preempt_count);
3404 * Print scheduling while atomic bug:
3406 static noinline void __schedule_bug(struct task_struct *prev)
3408 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3409 prev->comm, preempt_count(), prev->pid);
3410 debug_show_held_locks(prev);
3411 if (irqs_disabled())
3412 print_irqtrace_events(prev);
3417 * Various schedule()-time debugging checks and statistics:
3419 static inline void schedule_debug(struct task_struct *prev)
3422 * Test if we are atomic. Since do_exit() needs to call into
3423 * schedule() atomically, we ignore that path for now.
3424 * Otherwise, whine if we are scheduling when we should not be.
3426 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3427 __schedule_bug(prev);
3429 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3431 schedstat_inc(this_rq(), sched_cnt);
3435 * Pick up the highest-prio task:
3437 static inline struct task_struct *
3438 pick_next_task(struct rq *rq, struct task_struct *prev)
3440 struct sched_class *class;
3441 struct task_struct *p;
3444 * Optimization: we know that if all tasks are in
3445 * the fair class we can call that function directly:
3447 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3448 p = fair_sched_class.pick_next_task(rq);
3453 class = sched_class_highest;
3455 p = class->pick_next_task(rq);
3459 * Will never be NULL as the idle class always
3460 * returns a non-NULL p:
3462 class = class->next;
3467 * schedule() is the main scheduler function.
3469 asmlinkage void __sched schedule(void)
3471 struct task_struct *prev, *next;
3478 cpu = smp_processor_id();
3482 switch_count = &prev->nivcsw;
3484 release_kernel_lock(prev);
3485 need_resched_nonpreemptible:
3487 schedule_debug(prev);
3489 spin_lock_irq(&rq->lock);
3490 clear_tsk_need_resched(prev);
3491 __update_rq_clock(rq);
3493 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3494 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3495 unlikely(signal_pending(prev)))) {
3496 prev->state = TASK_RUNNING;
3498 deactivate_task(rq, prev, 1);
3500 switch_count = &prev->nvcsw;
3503 if (unlikely(!rq->nr_running))
3504 idle_balance(cpu, rq);
3506 prev->sched_class->put_prev_task(rq, prev);
3507 next = pick_next_task(rq, prev);
3509 sched_info_switch(prev, next);
3511 if (likely(prev != next)) {
3516 context_switch(rq, prev, next); /* unlocks the rq */
3518 spin_unlock_irq(&rq->lock);
3520 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3521 cpu = smp_processor_id();
3523 goto need_resched_nonpreemptible;
3525 preempt_enable_no_resched();
3526 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3529 EXPORT_SYMBOL(schedule);
3531 #ifdef CONFIG_PREEMPT
3533 * this is the entry point to schedule() from in-kernel preemption
3534 * off of preempt_enable. Kernel preemptions off return from interrupt
3535 * occur there and call schedule directly.
3537 asmlinkage void __sched preempt_schedule(void)
3539 struct thread_info *ti = current_thread_info();
3540 #ifdef CONFIG_PREEMPT_BKL
3541 struct task_struct *task = current;
3542 int saved_lock_depth;
3545 * If there is a non-zero preempt_count or interrupts are disabled,
3546 * we do not want to preempt the current task. Just return..
3548 if (likely(ti->preempt_count || irqs_disabled()))
3552 add_preempt_count(PREEMPT_ACTIVE);
3554 * We keep the big kernel semaphore locked, but we
3555 * clear ->lock_depth so that schedule() doesnt
3556 * auto-release the semaphore:
3558 #ifdef CONFIG_PREEMPT_BKL
3559 saved_lock_depth = task->lock_depth;
3560 task->lock_depth = -1;
3563 #ifdef CONFIG_PREEMPT_BKL
3564 task->lock_depth = saved_lock_depth;
3566 sub_preempt_count(PREEMPT_ACTIVE);
3568 /* we could miss a preemption opportunity between schedule and now */
3570 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3573 EXPORT_SYMBOL(preempt_schedule);
3576 * this is the entry point to schedule() from kernel preemption
3577 * off of irq context.
3578 * Note, that this is called and return with irqs disabled. This will
3579 * protect us against recursive calling from irq.
3581 asmlinkage void __sched preempt_schedule_irq(void)
3583 struct thread_info *ti = current_thread_info();
3584 #ifdef CONFIG_PREEMPT_BKL
3585 struct task_struct *task = current;
3586 int saved_lock_depth;
3588 /* Catch callers which need to be fixed */
3589 BUG_ON(ti->preempt_count || !irqs_disabled());
3592 add_preempt_count(PREEMPT_ACTIVE);
3594 * We keep the big kernel semaphore locked, but we
3595 * clear ->lock_depth so that schedule() doesnt
3596 * auto-release the semaphore:
3598 #ifdef CONFIG_PREEMPT_BKL
3599 saved_lock_depth = task->lock_depth;
3600 task->lock_depth = -1;
3604 local_irq_disable();
3605 #ifdef CONFIG_PREEMPT_BKL
3606 task->lock_depth = saved_lock_depth;
3608 sub_preempt_count(PREEMPT_ACTIVE);
3610 /* we could miss a preemption opportunity between schedule and now */
3612 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3616 #endif /* CONFIG_PREEMPT */
3618 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3621 return try_to_wake_up(curr->private, mode, sync);
3623 EXPORT_SYMBOL(default_wake_function);
3626 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3627 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3628 * number) then we wake all the non-exclusive tasks and one exclusive task.
3630 * There are circumstances in which we can try to wake a task which has already
3631 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3632 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3634 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3635 int nr_exclusive, int sync, void *key)
3637 struct list_head *tmp, *next;
3639 list_for_each_safe(tmp, next, &q->task_list) {
3640 wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3641 unsigned flags = curr->flags;
3643 if (curr->func(curr, mode, sync, key) &&
3644 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3650 * __wake_up - wake up threads blocked on a waitqueue.
3652 * @mode: which threads
3653 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3654 * @key: is directly passed to the wakeup function
3656 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3657 int nr_exclusive, void *key)
3659 unsigned long flags;
3661 spin_lock_irqsave(&q->lock, flags);
3662 __wake_up_common(q, mode, nr_exclusive, 0, key);
3663 spin_unlock_irqrestore(&q->lock, flags);
3665 EXPORT_SYMBOL(__wake_up);
3668 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3670 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3672 __wake_up_common(q, mode, 1, 0, NULL);
3676 * __wake_up_sync - wake up threads blocked on a waitqueue.
3678 * @mode: which threads
3679 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3681 * The sync wakeup differs that the waker knows that it will schedule
3682 * away soon, so while the target thread will be woken up, it will not
3683 * be migrated to another CPU - ie. the two threads are 'synchronized'
3684 * with each other. This can prevent needless bouncing between CPUs.
3686 * On UP it can prevent extra preemption.
3689 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3691 unsigned long flags;
3697 if (unlikely(!nr_exclusive))
3700 spin_lock_irqsave(&q->lock, flags);
3701 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3702 spin_unlock_irqrestore(&q->lock, flags);
3704 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3706 void fastcall complete(struct completion *x)
3708 unsigned long flags;
3710 spin_lock_irqsave(&x->wait.lock, flags);
3712 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3714 spin_unlock_irqrestore(&x->wait.lock, flags);
3716 EXPORT_SYMBOL(complete);
3718 void fastcall complete_all(struct completion *x)
3720 unsigned long flags;
3722 spin_lock_irqsave(&x->wait.lock, flags);
3723 x->done += UINT_MAX/2;
3724 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3726 spin_unlock_irqrestore(&x->wait.lock, flags);
3728 EXPORT_SYMBOL(complete_all);
3730 void fastcall __sched wait_for_completion(struct completion *x)
3734 spin_lock_irq(&x->wait.lock);
3736 DECLARE_WAITQUEUE(wait, current);
3738 wait.flags |= WQ_FLAG_EXCLUSIVE;
3739 __add_wait_queue_tail(&x->wait, &wait);
3741 __set_current_state(TASK_UNINTERRUPTIBLE);
3742 spin_unlock_irq(&x->wait.lock);
3744 spin_lock_irq(&x->wait.lock);
3746 __remove_wait_queue(&x->wait, &wait);
3749 spin_unlock_irq(&x->wait.lock);
3751 EXPORT_SYMBOL(wait_for_completion);
3753 unsigned long fastcall __sched
3754 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3758 spin_lock_irq(&x->wait.lock);
3760 DECLARE_WAITQUEUE(wait, current);
3762 wait.flags |= WQ_FLAG_EXCLUSIVE;
3763 __add_wait_queue_tail(&x->wait, &wait);
3765 __set_current_state(TASK_UNINTERRUPTIBLE);
3766 spin_unlock_irq(&x->wait.lock);
3767 timeout = schedule_timeout(timeout);
3768 spin_lock_irq(&x->wait.lock);
3770 __remove_wait_queue(&x->wait, &wait);
3774 __remove_wait_queue(&x->wait, &wait);
3778 spin_unlock_irq(&x->wait.lock);
3781 EXPORT_SYMBOL(wait_for_completion_timeout);
3783 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3789 spin_lock_irq(&x->wait.lock);
3791 DECLARE_WAITQUEUE(wait, current);
3793 wait.flags |= WQ_FLAG_EXCLUSIVE;
3794 __add_wait_queue_tail(&x->wait, &wait);
3796 if (signal_pending(current)) {
3798 __remove_wait_queue(&x->wait, &wait);
3801 __set_current_state(TASK_INTERRUPTIBLE);
3802 spin_unlock_irq(&x->wait.lock);
3804 spin_lock_irq(&x->wait.lock);
3806 __remove_wait_queue(&x->wait, &wait);
3810 spin_unlock_irq(&x->wait.lock);
3814 EXPORT_SYMBOL(wait_for_completion_interruptible);
3816 unsigned long fastcall __sched
3817 wait_for_completion_interruptible_timeout(struct completion *x,
3818 unsigned long timeout)
3822 spin_lock_irq(&x->wait.lock);
3824 DECLARE_WAITQUEUE(wait, current);
3826 wait.flags |= WQ_FLAG_EXCLUSIVE;
3827 __add_wait_queue_tail(&x->wait, &wait);
3829 if (signal_pending(current)) {
3830 timeout = -ERESTARTSYS;
3831 __remove_wait_queue(&x->wait, &wait);
3834 __set_current_state(TASK_INTERRUPTIBLE);
3835 spin_unlock_irq(&x->wait.lock);
3836 timeout = schedule_timeout(timeout);
3837 spin_lock_irq(&x->wait.lock);
3839 __remove_wait_queue(&x->wait, &wait);
3843 __remove_wait_queue(&x->wait, &wait);
3847 spin_unlock_irq(&x->wait.lock);
3850 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3853 sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3855 spin_lock_irqsave(&q->lock, *flags);
3856 __add_wait_queue(q, wait);
3857 spin_unlock(&q->lock);
3861 sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3863 spin_lock_irq(&q->lock);
3864 __remove_wait_queue(q, wait);
3865 spin_unlock_irqrestore(&q->lock, *flags);
3868 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3870 unsigned long flags;
3873 init_waitqueue_entry(&wait, current);
3875 current->state = TASK_INTERRUPTIBLE;
3877 sleep_on_head(q, &wait, &flags);
3879 sleep_on_tail(q, &wait, &flags);
3881 EXPORT_SYMBOL(interruptible_sleep_on);
3884 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3886 unsigned long flags;
3889 init_waitqueue_entry(&wait, current);
3891 current->state = TASK_INTERRUPTIBLE;
3893 sleep_on_head(q, &wait, &flags);
3894 timeout = schedule_timeout(timeout);
3895 sleep_on_tail(q, &wait, &flags);
3899 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3901 void __sched sleep_on(wait_queue_head_t *q)
3903 unsigned long flags;
3906 init_waitqueue_entry(&wait, current);
3908 current->state = TASK_UNINTERRUPTIBLE;
3910 sleep_on_head(q, &wait, &flags);
3912 sleep_on_tail(q, &wait, &flags);
3914 EXPORT_SYMBOL(sleep_on);
3916 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3918 unsigned long flags;
3921 init_waitqueue_entry(&wait, current);
3923 current->state = TASK_UNINTERRUPTIBLE;
3925 sleep_on_head(q, &wait, &flags);
3926 timeout = schedule_timeout(timeout);
3927 sleep_on_tail(q, &wait, &flags);
3931 EXPORT_SYMBOL(sleep_on_timeout);
3933 #ifdef CONFIG_RT_MUTEXES
3936 * rt_mutex_setprio - set the current priority of a task
3938 * @prio: prio value (kernel-internal form)
3940 * This function changes the 'effective' priority of a task. It does
3941 * not touch ->normal_prio like __setscheduler().
3943 * Used by the rt_mutex code to implement priority inheritance logic.
3945 void rt_mutex_setprio(struct task_struct *p, int prio)
3947 unsigned long flags;
3951 BUG_ON(prio < 0 || prio > MAX_PRIO);
3953 rq = task_rq_lock(p, &flags);
3954 update_rq_clock(rq);
3957 on_rq = p->se.on_rq;
3959 dequeue_task(rq, p, 0);
3962 p->sched_class = &rt_sched_class;
3964 p->sched_class = &fair_sched_class;
3969 enqueue_task(rq, p, 0);
3971 * Reschedule if we are currently running on this runqueue and
3972 * our priority decreased, or if we are not currently running on
3973 * this runqueue and our priority is higher than the current's
3975 if (task_running(rq, p)) {
3976 if (p->prio > oldprio)
3977 resched_task(rq->curr);
3979 check_preempt_curr(rq, p);
3982 task_rq_unlock(rq, &flags);
3987 void set_user_nice(struct task_struct *p, long nice)
3989 int old_prio, delta, on_rq;
3990 unsigned long flags;
3993 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3996 * We have to be careful, if called from sys_setpriority(),
3997 * the task might be in the middle of scheduling on another CPU.
3999 rq = task_rq_lock(p, &flags);
4000 update_rq_clock(rq);
4002 * The RT priorities are set via sched_setscheduler(), but we still
4003 * allow the 'normal' nice value to be set - but as expected
4004 * it wont have any effect on scheduling until the task is
4005 * SCHED_FIFO/SCHED_RR:
4007 if (task_has_rt_policy(p)) {
4008 p->static_prio = NICE_TO_PRIO(nice);
4011 on_rq = p->se.on_rq;
4013 dequeue_task(rq, p, 0);
4017 p->static_prio = NICE_TO_PRIO(nice);
4020 p->prio = effective_prio(p);
4021 delta = p->prio - old_prio;
4024 enqueue_task(rq, p, 0);
4027 * If the task increased its priority or is running and
4028 * lowered its priority, then reschedule its CPU:
4030 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4031 resched_task(rq->curr);
4034 task_rq_unlock(rq, &flags);
4036 EXPORT_SYMBOL(set_user_nice);
4039 * can_nice - check if a task can reduce its nice value
4043 int can_nice(const struct task_struct *p, const int nice)
4045 /* convert nice value [19,-20] to rlimit style value [1,40] */
4046 int nice_rlim = 20 - nice;
4048 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4049 capable(CAP_SYS_NICE));
4052 #ifdef __ARCH_WANT_SYS_NICE
4055 * sys_nice - change the priority of the current process.
4056 * @increment: priority increment
4058 * sys_setpriority is a more generic, but much slower function that
4059 * does similar things.
4061 asmlinkage long sys_nice(int increment)
4066 * Setpriority might change our priority at the same moment.
4067 * We don't have to worry. Conceptually one call occurs first
4068 * and we have a single winner.
4070 if (increment < -40)
4075 nice = PRIO_TO_NICE(current->static_prio) + increment;
4081 if (increment < 0 && !can_nice(current, nice))
4084 retval = security_task_setnice(current, nice);
4088 set_user_nice(current, nice);
4095 * task_prio - return the priority value of a given task.
4096 * @p: the task in question.
4098 * This is the priority value as seen by users in /proc.
4099 * RT tasks are offset by -200. Normal tasks are centered
4100 * around 0, value goes from -16 to +15.
4102 int task_prio(const struct task_struct *p)
4104 return p->prio - MAX_RT_PRIO;
4108 * task_nice - return the nice value of a given task.
4109 * @p: the task in question.
4111 int task_nice(const struct task_struct *p)
4113 return TASK_NICE(p);
4115 EXPORT_SYMBOL_GPL(task_nice);
4118 * idle_cpu - is a given cpu idle currently?
4119 * @cpu: the processor in question.
4121 int idle_cpu(int cpu)
4123 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4127 * idle_task - return the idle task for a given cpu.
4128 * @cpu: the processor in question.
4130 struct task_struct *idle_task(int cpu)
4132 return cpu_rq(cpu)->idle;
4136 * find_process_by_pid - find a process with a matching PID value.
4137 * @pid: the pid in question.
4139 static inline struct task_struct *find_process_by_pid(pid_t pid)
4141 return pid ? find_task_by_pid(pid) : current;
4144 /* Actually do priority change: must hold rq lock. */
4146 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4148 BUG_ON(p->se.on_rq);
4151 switch (p->policy) {
4155 p->sched_class = &fair_sched_class;
4159 p->sched_class = &rt_sched_class;
4163 p->rt_priority = prio;
4164 p->normal_prio = normal_prio(p);
4165 /* we are holding p->pi_lock already */
4166 p->prio = rt_mutex_getprio(p);
4171 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4172 * @p: the task in question.
4173 * @policy: new policy.
4174 * @param: structure containing the new RT priority.
4176 * NOTE that the task may be already dead.
4178 int sched_setscheduler(struct task_struct *p, int policy,
4179 struct sched_param *param)
4181 int retval, oldprio, oldpolicy = -1, on_rq;
4182 unsigned long flags;
4185 /* may grab non-irq protected spin_locks */
4186 BUG_ON(in_interrupt());
4188 /* double check policy once rq lock held */
4190 policy = oldpolicy = p->policy;
4191 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4192 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4193 policy != SCHED_IDLE)
4196 * Valid priorities for SCHED_FIFO and SCHED_RR are
4197 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4198 * SCHED_BATCH and SCHED_IDLE is 0.
4200 if (param->sched_priority < 0 ||
4201 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4202 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4204 if (rt_policy(policy) != (param->sched_priority != 0))
4208 * Allow unprivileged RT tasks to decrease priority:
4210 if (!capable(CAP_SYS_NICE)) {
4211 if (rt_policy(policy)) {
4212 unsigned long rlim_rtprio;
4214 if (!lock_task_sighand(p, &flags))
4216 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4217 unlock_task_sighand(p, &flags);
4219 /* can't set/change the rt policy */
4220 if (policy != p->policy && !rlim_rtprio)
4223 /* can't increase priority */
4224 if (param->sched_priority > p->rt_priority &&
4225 param->sched_priority > rlim_rtprio)
4229 * Like positive nice levels, dont allow tasks to
4230 * move out of SCHED_IDLE either:
4232 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4235 /* can't change other user's priorities */
4236 if ((current->euid != p->euid) &&
4237 (current->euid != p->uid))
4241 retval = security_task_setscheduler(p, policy, param);
4245 * make sure no PI-waiters arrive (or leave) while we are
4246 * changing the priority of the task:
4248 spin_lock_irqsave(&p->pi_lock, flags);
4250 * To be able to change p->policy safely, the apropriate
4251 * runqueue lock must be held.
4253 rq = __task_rq_lock(p);
4254 /* recheck policy now with rq lock held */
4255 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4256 policy = oldpolicy = -1;
4257 __task_rq_unlock(rq);
4258 spin_unlock_irqrestore(&p->pi_lock, flags);
4261 update_rq_clock(rq);
4262 on_rq = p->se.on_rq;
4264 deactivate_task(rq, p, 0);
4266 __setscheduler(rq, p, policy, param->sched_priority);
4268 activate_task(rq, p, 0);
4270 * Reschedule if we are currently running on this runqueue and
4271 * our priority decreased, or if we are not currently running on
4272 * this runqueue and our priority is higher than the current's
4274 if (task_running(rq, p)) {
4275 if (p->prio > oldprio)
4276 resched_task(rq->curr);
4278 check_preempt_curr(rq, p);
4281 __task_rq_unlock(rq);
4282 spin_unlock_irqrestore(&p->pi_lock, flags);
4284 rt_mutex_adjust_pi(p);
4288 EXPORT_SYMBOL_GPL(sched_setscheduler);
4291 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4293 struct sched_param lparam;
4294 struct task_struct *p;
4297 if (!param || pid < 0)
4299 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4304 p = find_process_by_pid(pid);
4306 retval = sched_setscheduler(p, policy, &lparam);
4313 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4314 * @pid: the pid in question.
4315 * @policy: new policy.
4316 * @param: structure containing the new RT priority.
4318 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4319 struct sched_param __user *param)
4321 /* negative values for policy are not valid */
4325 return do_sched_setscheduler(pid, policy, param);
4329 * sys_sched_setparam - set/change the RT priority of a thread
4330 * @pid: the pid in question.
4331 * @param: structure containing the new RT priority.
4333 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4335 return do_sched_setscheduler(pid, -1, param);
4339 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4340 * @pid: the pid in question.
4342 asmlinkage long sys_sched_getscheduler(pid_t pid)
4344 struct task_struct *p;
4345 int retval = -EINVAL;
4351 read_lock(&tasklist_lock);
4352 p = find_process_by_pid(pid);
4354 retval = security_task_getscheduler(p);
4358 read_unlock(&tasklist_lock);
4365 * sys_sched_getscheduler - get the RT priority of a thread
4366 * @pid: the pid in question.
4367 * @param: structure containing the RT priority.
4369 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4371 struct sched_param lp;
4372 struct task_struct *p;
4373 int retval = -EINVAL;
4375 if (!param || pid < 0)
4378 read_lock(&tasklist_lock);
4379 p = find_process_by_pid(pid);
4384 retval = security_task_getscheduler(p);
4388 lp.sched_priority = p->rt_priority;
4389 read_unlock(&tasklist_lock);
4392 * This one might sleep, we cannot do it with a spinlock held ...
4394 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4400 read_unlock(&tasklist_lock);
4404 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4406 cpumask_t cpus_allowed;
4407 struct task_struct *p;
4410 mutex_lock(&sched_hotcpu_mutex);
4411 read_lock(&tasklist_lock);
4413 p = find_process_by_pid(pid);
4415 read_unlock(&tasklist_lock);
4416 mutex_unlock(&sched_hotcpu_mutex);
4421 * It is not safe to call set_cpus_allowed with the
4422 * tasklist_lock held. We will bump the task_struct's
4423 * usage count and then drop tasklist_lock.
4426 read_unlock(&tasklist_lock);
4429 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4430 !capable(CAP_SYS_NICE))
4433 retval = security_task_setscheduler(p, 0, NULL);
4437 cpus_allowed = cpuset_cpus_allowed(p);
4438 cpus_and(new_mask, new_mask, cpus_allowed);
4439 retval = set_cpus_allowed(p, new_mask);
4443 mutex_unlock(&sched_hotcpu_mutex);
4447 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4448 cpumask_t *new_mask)
4450 if (len < sizeof(cpumask_t)) {
4451 memset(new_mask, 0, sizeof(cpumask_t));
4452 } else if (len > sizeof(cpumask_t)) {
4453 len = sizeof(cpumask_t);
4455 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4459 * sys_sched_setaffinity - set the cpu affinity of a process
4460 * @pid: pid of the process
4461 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4462 * @user_mask_ptr: user-space pointer to the new cpu mask
4464 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4465 unsigned long __user *user_mask_ptr)
4470 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4474 return sched_setaffinity(pid, new_mask);
4478 * Represents all cpu's present in the system
4479 * In systems capable of hotplug, this map could dynamically grow
4480 * as new cpu's are detected in the system via any platform specific
4481 * method, such as ACPI for e.g.
4484 cpumask_t cpu_present_map __read_mostly;
4485 EXPORT_SYMBOL(cpu_present_map);
4488 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4489 EXPORT_SYMBOL(cpu_online_map);
4491 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4492 EXPORT_SYMBOL(cpu_possible_map);
4495 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4497 struct task_struct *p;
4500 mutex_lock(&sched_hotcpu_mutex);
4501 read_lock(&tasklist_lock);
4504 p = find_process_by_pid(pid);
4508 retval = security_task_getscheduler(p);
4512 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4515 read_unlock(&tasklist_lock);
4516 mutex_unlock(&sched_hotcpu_mutex);
4522 * sys_sched_getaffinity - get the cpu affinity of a process
4523 * @pid: pid of the process
4524 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4525 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4527 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4528 unsigned long __user *user_mask_ptr)
4533 if (len < sizeof(cpumask_t))
4536 ret = sched_getaffinity(pid, &mask);
4540 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4543 return sizeof(cpumask_t);
4547 * sys_sched_yield - yield the current processor to other threads.
4549 * This function yields the current CPU to other tasks. If there are no
4550 * other threads running on this CPU then this function will return.
4552 asmlinkage long sys_sched_yield(void)
4554 struct rq *rq = this_rq_lock();
4556 schedstat_inc(rq, yld_cnt);
4557 current->sched_class->yield_task(rq, current);
4560 * Since we are going to call schedule() anyway, there's
4561 * no need to preempt or enable interrupts:
4563 __release(rq->lock);
4564 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4565 _raw_spin_unlock(&rq->lock);
4566 preempt_enable_no_resched();
4573 static void __cond_resched(void)
4575 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4576 __might_sleep(__FILE__, __LINE__);
4579 * The BKS might be reacquired before we have dropped
4580 * PREEMPT_ACTIVE, which could trigger a second
4581 * cond_resched() call.
4584 add_preempt_count(PREEMPT_ACTIVE);
4586 sub_preempt_count(PREEMPT_ACTIVE);
4587 } while (need_resched());
4590 int __sched cond_resched(void)
4592 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4593 system_state == SYSTEM_RUNNING) {
4599 EXPORT_SYMBOL(cond_resched);
4602 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4603 * call schedule, and on return reacquire the lock.
4605 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4606 * operations here to prevent schedule() from being called twice (once via
4607 * spin_unlock(), once by hand).
4609 int cond_resched_lock(spinlock_t *lock)
4613 if (need_lockbreak(lock)) {
4619 if (need_resched() && system_state == SYSTEM_RUNNING) {
4620 spin_release(&lock->dep_map, 1, _THIS_IP_);
4621 _raw_spin_unlock(lock);
4622 preempt_enable_no_resched();
4629 EXPORT_SYMBOL(cond_resched_lock);
4631 int __sched cond_resched_softirq(void)
4633 BUG_ON(!in_softirq());
4635 if (need_resched() && system_state == SYSTEM_RUNNING) {
4643 EXPORT_SYMBOL(cond_resched_softirq);
4646 * yield - yield the current processor to other threads.
4648 * This is a shortcut for kernel-space yielding - it marks the
4649 * thread runnable and calls sys_sched_yield().
4651 void __sched yield(void)
4653 set_current_state(TASK_RUNNING);
4656 EXPORT_SYMBOL(yield);
4659 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4660 * that process accounting knows that this is a task in IO wait state.
4662 * But don't do that if it is a deliberate, throttling IO wait (this task
4663 * has set its backing_dev_info: the queue against which it should throttle)
4665 void __sched io_schedule(void)
4667 struct rq *rq = &__raw_get_cpu_var(runqueues);
4669 delayacct_blkio_start();
4670 atomic_inc(&rq->nr_iowait);
4672 atomic_dec(&rq->nr_iowait);
4673 delayacct_blkio_end();
4675 EXPORT_SYMBOL(io_schedule);
4677 long __sched io_schedule_timeout(long timeout)
4679 struct rq *rq = &__raw_get_cpu_var(runqueues);
4682 delayacct_blkio_start();
4683 atomic_inc(&rq->nr_iowait);
4684 ret = schedule_timeout(timeout);
4685 atomic_dec(&rq->nr_iowait);
4686 delayacct_blkio_end();
4691 * sys_sched_get_priority_max - return maximum RT priority.
4692 * @policy: scheduling class.
4694 * this syscall returns the maximum rt_priority that can be used
4695 * by a given scheduling class.
4697 asmlinkage long sys_sched_get_priority_max(int policy)
4704 ret = MAX_USER_RT_PRIO-1;
4716 * sys_sched_get_priority_min - return minimum RT priority.
4717 * @policy: scheduling class.
4719 * this syscall returns the minimum rt_priority that can be used
4720 * by a given scheduling class.
4722 asmlinkage long sys_sched_get_priority_min(int policy)
4740 * sys_sched_rr_get_interval - return the default timeslice of a process.
4741 * @pid: pid of the process.
4742 * @interval: userspace pointer to the timeslice value.
4744 * this syscall writes the default timeslice value of a given process
4745 * into the user-space timespec buffer. A value of '0' means infinity.
4748 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4750 struct task_struct *p;
4751 int retval = -EINVAL;
4758 read_lock(&tasklist_lock);
4759 p = find_process_by_pid(pid);
4763 retval = security_task_getscheduler(p);
4767 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4768 0 : static_prio_timeslice(p->static_prio), &t);
4769 read_unlock(&tasklist_lock);
4770 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4774 read_unlock(&tasklist_lock);
4778 static const char stat_nam[] = "RSDTtZX";
4780 static void show_task(struct task_struct *p)
4782 unsigned long free = 0;
4785 state = p->state ? __ffs(p->state) + 1 : 0;
4786 printk("%-13.13s %c", p->comm,
4787 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4788 #if BITS_PER_LONG == 32
4789 if (state == TASK_RUNNING)
4790 printk(" running ");
4792 printk(" %08lx ", thread_saved_pc(p));
4794 if (state == TASK_RUNNING)
4795 printk(" running task ");
4797 printk(" %016lx ", thread_saved_pc(p));
4799 #ifdef CONFIG_DEBUG_STACK_USAGE
4801 unsigned long *n = end_of_stack(p);
4804 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4807 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4809 if (state != TASK_RUNNING)
4810 show_stack(p, NULL);
4813 void show_state_filter(unsigned long state_filter)
4815 struct task_struct *g, *p;
4817 #if BITS_PER_LONG == 32
4819 " task PC stack pid father\n");
4822 " task PC stack pid father\n");
4824 read_lock(&tasklist_lock);
4825 do_each_thread(g, p) {
4827 * reset the NMI-timeout, listing all files on a slow
4828 * console might take alot of time:
4830 touch_nmi_watchdog();
4831 if (!state_filter || (p->state & state_filter))
4833 } while_each_thread(g, p);
4835 touch_all_softlockup_watchdogs();
4837 #ifdef CONFIG_SCHED_DEBUG
4838 sysrq_sched_debug_show();
4840 read_unlock(&tasklist_lock);
4842 * Only show locks if all tasks are dumped:
4844 if (state_filter == -1)
4845 debug_show_all_locks();
4848 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4850 idle->sched_class = &idle_sched_class;
4854 * init_idle - set up an idle thread for a given CPU
4855 * @idle: task in question
4856 * @cpu: cpu the idle task belongs to
4858 * NOTE: this function does not set the idle thread's NEED_RESCHED
4859 * flag, to make booting more robust.
4861 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4863 struct rq *rq = cpu_rq(cpu);
4864 unsigned long flags;
4867 idle->se.exec_start = sched_clock();
4869 idle->prio = idle->normal_prio = MAX_PRIO;
4870 idle->cpus_allowed = cpumask_of_cpu(cpu);
4871 __set_task_cpu(idle, cpu);
4873 spin_lock_irqsave(&rq->lock, flags);
4874 rq->curr = rq->idle = idle;
4875 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4878 spin_unlock_irqrestore(&rq->lock, flags);
4880 /* Set the preempt count _outside_ the spinlocks! */
4881 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4882 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4884 task_thread_info(idle)->preempt_count = 0;
4887 * The idle tasks have their own, simple scheduling class:
4889 idle->sched_class = &idle_sched_class;
4893 * In a system that switches off the HZ timer nohz_cpu_mask
4894 * indicates which cpus entered this state. This is used
4895 * in the rcu update to wait only for active cpus. For system
4896 * which do not switch off the HZ timer nohz_cpu_mask should
4897 * always be CPU_MASK_NONE.
4899 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4902 * Increase the granularity value when there are more CPUs,
4903 * because with more CPUs the 'effective latency' as visible
4904 * to users decreases. But the relationship is not linear,
4905 * so pick a second-best guess by going with the log2 of the
4908 * This idea comes from the SD scheduler of Con Kolivas:
4910 static inline void sched_init_granularity(void)
4912 unsigned int factor = 1 + ilog2(num_online_cpus());
4913 const unsigned long limit = 100000000;
4915 sysctl_sched_min_granularity *= factor;
4916 if (sysctl_sched_min_granularity > limit)
4917 sysctl_sched_min_granularity = limit;
4919 sysctl_sched_latency *= factor;
4920 if (sysctl_sched_latency > limit)
4921 sysctl_sched_latency = limit;
4923 sysctl_sched_runtime_limit = sysctl_sched_latency;
4924 sysctl_sched_wakeup_granularity = sysctl_sched_min_granularity / 2;
4929 * This is how migration works:
4931 * 1) we queue a struct migration_req structure in the source CPU's
4932 * runqueue and wake up that CPU's migration thread.
4933 * 2) we down() the locked semaphore => thread blocks.
4934 * 3) migration thread wakes up (implicitly it forces the migrated
4935 * thread off the CPU)
4936 * 4) it gets the migration request and checks whether the migrated
4937 * task is still in the wrong runqueue.
4938 * 5) if it's in the wrong runqueue then the migration thread removes
4939 * it and puts it into the right queue.
4940 * 6) migration thread up()s the semaphore.
4941 * 7) we wake up and the migration is done.
4945 * Change a given task's CPU affinity. Migrate the thread to a
4946 * proper CPU and schedule it away if the CPU it's executing on
4947 * is removed from the allowed bitmask.
4949 * NOTE: the caller must have a valid reference to the task, the
4950 * task must not exit() & deallocate itself prematurely. The
4951 * call is not atomic; no spinlocks may be held.
4953 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4955 struct migration_req req;
4956 unsigned long flags;
4960 rq = task_rq_lock(p, &flags);
4961 if (!cpus_intersects(new_mask, cpu_online_map)) {
4966 p->cpus_allowed = new_mask;
4967 /* Can the task run on the task's current CPU? If so, we're done */
4968 if (cpu_isset(task_cpu(p), new_mask))
4971 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4972 /* Need help from migration thread: drop lock and wait. */
4973 task_rq_unlock(rq, &flags);
4974 wake_up_process(rq->migration_thread);
4975 wait_for_completion(&req.done);
4976 tlb_migrate_finish(p->mm);
4980 task_rq_unlock(rq, &flags);
4984 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4987 * Move (not current) task off this cpu, onto dest cpu. We're doing
4988 * this because either it can't run here any more (set_cpus_allowed()
4989 * away from this CPU, or CPU going down), or because we're
4990 * attempting to rebalance this task on exec (sched_exec).
4992 * So we race with normal scheduler movements, but that's OK, as long
4993 * as the task is no longer on this CPU.
4995 * Returns non-zero if task was successfully migrated.
4997 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4999 struct rq *rq_dest, *rq_src;
5002 if (unlikely(cpu_is_offline(dest_cpu)))
5005 rq_src = cpu_rq(src_cpu);
5006 rq_dest = cpu_rq(dest_cpu);
5008 double_rq_lock(rq_src, rq_dest);
5009 /* Already moved. */
5010 if (task_cpu(p) != src_cpu)
5012 /* Affinity changed (again). */
5013 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5016 on_rq = p->se.on_rq;
5018 deactivate_task(rq_src, p, 0);
5020 set_task_cpu(p, dest_cpu);
5022 activate_task(rq_dest, p, 0);
5023 check_preempt_curr(rq_dest, p);
5027 double_rq_unlock(rq_src, rq_dest);
5032 * migration_thread - this is a highprio system thread that performs
5033 * thread migration by bumping thread off CPU then 'pushing' onto
5036 static int migration_thread(void *data)
5038 int cpu = (long)data;
5042 BUG_ON(rq->migration_thread != current);
5044 set_current_state(TASK_INTERRUPTIBLE);
5045 while (!kthread_should_stop()) {
5046 struct migration_req *req;
5047 struct list_head *head;
5049 spin_lock_irq(&rq->lock);
5051 if (cpu_is_offline(cpu)) {
5052 spin_unlock_irq(&rq->lock);
5056 if (rq->active_balance) {
5057 active_load_balance(rq, cpu);
5058 rq->active_balance = 0;
5061 head = &rq->migration_queue;
5063 if (list_empty(head)) {
5064 spin_unlock_irq(&rq->lock);
5066 set_current_state(TASK_INTERRUPTIBLE);
5069 req = list_entry(head->next, struct migration_req, list);
5070 list_del_init(head->next);
5072 spin_unlock(&rq->lock);
5073 __migrate_task(req->task, cpu, req->dest_cpu);
5076 complete(&req->done);
5078 __set_current_state(TASK_RUNNING);
5082 /* Wait for kthread_stop */
5083 set_current_state(TASK_INTERRUPTIBLE);
5084 while (!kthread_should_stop()) {
5086 set_current_state(TASK_INTERRUPTIBLE);
5088 __set_current_state(TASK_RUNNING);
5092 #ifdef CONFIG_HOTPLUG_CPU
5094 * Figure out where task on dead CPU should go, use force if neccessary.
5095 * NOTE: interrupts should be disabled by the caller
5097 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5099 unsigned long flags;
5106 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5107 cpus_and(mask, mask, p->cpus_allowed);
5108 dest_cpu = any_online_cpu(mask);
5110 /* On any allowed CPU? */
5111 if (dest_cpu == NR_CPUS)
5112 dest_cpu = any_online_cpu(p->cpus_allowed);
5114 /* No more Mr. Nice Guy. */
5115 if (dest_cpu == NR_CPUS) {
5116 rq = task_rq_lock(p, &flags);
5117 cpus_setall(p->cpus_allowed);
5118 dest_cpu = any_online_cpu(p->cpus_allowed);
5119 task_rq_unlock(rq, &flags);
5122 * Don't tell them about moving exiting tasks or
5123 * kernel threads (both mm NULL), since they never
5126 if (p->mm && printk_ratelimit())
5127 printk(KERN_INFO "process %d (%s) no "
5128 "longer affine to cpu%d\n",
5129 p->pid, p->comm, dead_cpu);
5131 if (!__migrate_task(p, dead_cpu, dest_cpu))
5136 * While a dead CPU has no uninterruptible tasks queued at this point,
5137 * it might still have a nonzero ->nr_uninterruptible counter, because
5138 * for performance reasons the counter is not stricly tracking tasks to
5139 * their home CPUs. So we just add the counter to another CPU's counter,
5140 * to keep the global sum constant after CPU-down:
5142 static void migrate_nr_uninterruptible(struct rq *rq_src)
5144 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5145 unsigned long flags;
5147 local_irq_save(flags);
5148 double_rq_lock(rq_src, rq_dest);
5149 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5150 rq_src->nr_uninterruptible = 0;
5151 double_rq_unlock(rq_src, rq_dest);
5152 local_irq_restore(flags);
5155 /* Run through task list and migrate tasks from the dead cpu. */
5156 static void migrate_live_tasks(int src_cpu)
5158 struct task_struct *p, *t;
5160 write_lock_irq(&tasklist_lock);
5162 do_each_thread(t, p) {
5166 if (task_cpu(p) == src_cpu)
5167 move_task_off_dead_cpu(src_cpu, p);
5168 } while_each_thread(t, p);
5170 write_unlock_irq(&tasklist_lock);
5174 * Schedules idle task to be the next runnable task on current CPU.
5175 * It does so by boosting its priority to highest possible and adding it to
5176 * the _front_ of the runqueue. Used by CPU offline code.
5178 void sched_idle_next(void)
5180 int this_cpu = smp_processor_id();
5181 struct rq *rq = cpu_rq(this_cpu);
5182 struct task_struct *p = rq->idle;
5183 unsigned long flags;
5185 /* cpu has to be offline */
5186 BUG_ON(cpu_online(this_cpu));
5189 * Strictly not necessary since rest of the CPUs are stopped by now
5190 * and interrupts disabled on the current cpu.
5192 spin_lock_irqsave(&rq->lock, flags);
5194 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5196 /* Add idle task to the _front_ of its priority queue: */
5197 activate_idle_task(p, rq);
5199 spin_unlock_irqrestore(&rq->lock, flags);
5203 * Ensures that the idle task is using init_mm right before its cpu goes
5206 void idle_task_exit(void)
5208 struct mm_struct *mm = current->active_mm;
5210 BUG_ON(cpu_online(smp_processor_id()));
5213 switch_mm(mm, &init_mm, current);
5217 /* called under rq->lock with disabled interrupts */
5218 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5220 struct rq *rq = cpu_rq(dead_cpu);
5222 /* Must be exiting, otherwise would be on tasklist. */
5223 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5225 /* Cannot have done final schedule yet: would have vanished. */
5226 BUG_ON(p->state == TASK_DEAD);
5231 * Drop lock around migration; if someone else moves it,
5232 * that's OK. No task can be added to this CPU, so iteration is
5234 * NOTE: interrupts should be left disabled --dev@
5236 spin_unlock(&rq->lock);
5237 move_task_off_dead_cpu(dead_cpu, p);
5238 spin_lock(&rq->lock);
5243 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5244 static void migrate_dead_tasks(unsigned int dead_cpu)
5246 struct rq *rq = cpu_rq(dead_cpu);
5247 struct task_struct *next;
5250 if (!rq->nr_running)
5252 update_rq_clock(rq);
5253 next = pick_next_task(rq, rq->curr);
5256 migrate_dead(dead_cpu, next);
5260 #endif /* CONFIG_HOTPLUG_CPU */
5262 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5264 static struct ctl_table sd_ctl_dir[] = {
5266 .procname = "sched_domain",
5272 static struct ctl_table sd_ctl_root[] = {
5274 .ctl_name = CTL_KERN,
5275 .procname = "kernel",
5277 .child = sd_ctl_dir,
5282 static struct ctl_table *sd_alloc_ctl_entry(int n)
5284 struct ctl_table *entry =
5285 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5288 memset(entry, 0, n * sizeof(struct ctl_table));
5294 set_table_entry(struct ctl_table *entry,
5295 const char *procname, void *data, int maxlen,
5296 mode_t mode, proc_handler *proc_handler)
5298 entry->procname = procname;
5300 entry->maxlen = maxlen;
5302 entry->proc_handler = proc_handler;
5305 static struct ctl_table *
5306 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5308 struct ctl_table *table = sd_alloc_ctl_entry(14);
5310 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5311 sizeof(long), 0644, proc_doulongvec_minmax);
5312 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5313 sizeof(long), 0644, proc_doulongvec_minmax);
5314 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5315 sizeof(int), 0644, proc_dointvec_minmax);
5316 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5317 sizeof(int), 0644, proc_dointvec_minmax);
5318 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5319 sizeof(int), 0644, proc_dointvec_minmax);
5320 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5321 sizeof(int), 0644, proc_dointvec_minmax);
5322 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5323 sizeof(int), 0644, proc_dointvec_minmax);
5324 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5325 sizeof(int), 0644, proc_dointvec_minmax);
5326 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5327 sizeof(int), 0644, proc_dointvec_minmax);
5328 set_table_entry(&table[10], "cache_nice_tries",
5329 &sd->cache_nice_tries,
5330 sizeof(int), 0644, proc_dointvec_minmax);
5331 set_table_entry(&table[12], "flags", &sd->flags,
5332 sizeof(int), 0644, proc_dointvec_minmax);
5337 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5339 struct ctl_table *entry, *table;
5340 struct sched_domain *sd;
5341 int domain_num = 0, i;
5344 for_each_domain(cpu, sd)
5346 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5349 for_each_domain(cpu, sd) {
5350 snprintf(buf, 32, "domain%d", i);
5351 entry->procname = kstrdup(buf, GFP_KERNEL);
5353 entry->child = sd_alloc_ctl_domain_table(sd);
5360 static struct ctl_table_header *sd_sysctl_header;
5361 static void init_sched_domain_sysctl(void)
5363 int i, cpu_num = num_online_cpus();
5364 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5367 sd_ctl_dir[0].child = entry;
5369 for (i = 0; i < cpu_num; i++, entry++) {
5370 snprintf(buf, 32, "cpu%d", i);
5371 entry->procname = kstrdup(buf, GFP_KERNEL);
5373 entry->child = sd_alloc_ctl_cpu_table(i);
5375 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5378 static void init_sched_domain_sysctl(void)
5384 * migration_call - callback that gets triggered when a CPU is added.
5385 * Here we can start up the necessary migration thread for the new CPU.
5387 static int __cpuinit
5388 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5390 struct task_struct *p;
5391 int cpu = (long)hcpu;
5392 unsigned long flags;
5396 case CPU_LOCK_ACQUIRE:
5397 mutex_lock(&sched_hotcpu_mutex);
5400 case CPU_UP_PREPARE:
5401 case CPU_UP_PREPARE_FROZEN:
5402 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5405 kthread_bind(p, cpu);
5406 /* Must be high prio: stop_machine expects to yield to it. */
5407 rq = task_rq_lock(p, &flags);
5408 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5409 task_rq_unlock(rq, &flags);
5410 cpu_rq(cpu)->migration_thread = p;
5414 case CPU_ONLINE_FROZEN:
5415 /* Strictly unneccessary, as first user will wake it. */
5416 wake_up_process(cpu_rq(cpu)->migration_thread);
5419 #ifdef CONFIG_HOTPLUG_CPU
5420 case CPU_UP_CANCELED:
5421 case CPU_UP_CANCELED_FROZEN:
5422 if (!cpu_rq(cpu)->migration_thread)
5424 /* Unbind it from offline cpu so it can run. Fall thru. */
5425 kthread_bind(cpu_rq(cpu)->migration_thread,
5426 any_online_cpu(cpu_online_map));
5427 kthread_stop(cpu_rq(cpu)->migration_thread);
5428 cpu_rq(cpu)->migration_thread = NULL;
5432 case CPU_DEAD_FROZEN:
5433 migrate_live_tasks(cpu);
5435 kthread_stop(rq->migration_thread);
5436 rq->migration_thread = NULL;
5437 /* Idle task back to normal (off runqueue, low prio) */
5438 rq = task_rq_lock(rq->idle, &flags);
5439 update_rq_clock(rq);
5440 deactivate_task(rq, rq->idle, 0);
5441 rq->idle->static_prio = MAX_PRIO;
5442 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5443 rq->idle->sched_class = &idle_sched_class;
5444 migrate_dead_tasks(cpu);
5445 task_rq_unlock(rq, &flags);
5446 migrate_nr_uninterruptible(rq);
5447 BUG_ON(rq->nr_running != 0);
5449 /* No need to migrate the tasks: it was best-effort if
5450 * they didn't take sched_hotcpu_mutex. Just wake up
5451 * the requestors. */
5452 spin_lock_irq(&rq->lock);
5453 while (!list_empty(&rq->migration_queue)) {
5454 struct migration_req *req;
5456 req = list_entry(rq->migration_queue.next,
5457 struct migration_req, list);
5458 list_del_init(&req->list);
5459 complete(&req->done);
5461 spin_unlock_irq(&rq->lock);
5464 case CPU_LOCK_RELEASE:
5465 mutex_unlock(&sched_hotcpu_mutex);
5471 /* Register at highest priority so that task migration (migrate_all_tasks)
5472 * happens before everything else.
5474 static struct notifier_block __cpuinitdata migration_notifier = {
5475 .notifier_call = migration_call,
5479 int __init migration_init(void)
5481 void *cpu = (void *)(long)smp_processor_id();
5484 /* Start one for the boot CPU: */
5485 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5486 BUG_ON(err == NOTIFY_BAD);
5487 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5488 register_cpu_notifier(&migration_notifier);
5496 /* Number of possible processor ids */
5497 int nr_cpu_ids __read_mostly = NR_CPUS;
5498 EXPORT_SYMBOL(nr_cpu_ids);
5500 #undef SCHED_DOMAIN_DEBUG
5501 #ifdef SCHED_DOMAIN_DEBUG
5502 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5507 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5511 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5516 struct sched_group *group = sd->groups;
5517 cpumask_t groupmask;
5519 cpumask_scnprintf(str, NR_CPUS, sd->span);
5520 cpus_clear(groupmask);
5523 for (i = 0; i < level + 1; i++)
5525 printk("domain %d: ", level);
5527 if (!(sd->flags & SD_LOAD_BALANCE)) {
5528 printk("does not load-balance\n");
5530 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5535 printk("span %s\n", str);
5537 if (!cpu_isset(cpu, sd->span))
5538 printk(KERN_ERR "ERROR: domain->span does not contain "
5540 if (!cpu_isset(cpu, group->cpumask))
5541 printk(KERN_ERR "ERROR: domain->groups does not contain"
5545 for (i = 0; i < level + 2; i++)
5551 printk(KERN_ERR "ERROR: group is NULL\n");
5555 if (!group->__cpu_power) {
5557 printk(KERN_ERR "ERROR: domain->cpu_power not "
5561 if (!cpus_weight(group->cpumask)) {
5563 printk(KERN_ERR "ERROR: empty group\n");
5566 if (cpus_intersects(groupmask, group->cpumask)) {
5568 printk(KERN_ERR "ERROR: repeated CPUs\n");
5571 cpus_or(groupmask, groupmask, group->cpumask);
5573 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5576 group = group->next;
5577 } while (group != sd->groups);
5580 if (!cpus_equal(sd->span, groupmask))
5581 printk(KERN_ERR "ERROR: groups don't span "
5589 if (!cpus_subset(groupmask, sd->span))
5590 printk(KERN_ERR "ERROR: parent span is not a superset "
5591 "of domain->span\n");
5596 # define sched_domain_debug(sd, cpu) do { } while (0)
5599 static int sd_degenerate(struct sched_domain *sd)
5601 if (cpus_weight(sd->span) == 1)
5604 /* Following flags need at least 2 groups */
5605 if (sd->flags & (SD_LOAD_BALANCE |
5606 SD_BALANCE_NEWIDLE |
5610 SD_SHARE_PKG_RESOURCES)) {
5611 if (sd->groups != sd->groups->next)
5615 /* Following flags don't use groups */
5616 if (sd->flags & (SD_WAKE_IDLE |
5625 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5627 unsigned long cflags = sd->flags, pflags = parent->flags;
5629 if (sd_degenerate(parent))
5632 if (!cpus_equal(sd->span, parent->span))
5635 /* Does parent contain flags not in child? */
5636 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5637 if (cflags & SD_WAKE_AFFINE)
5638 pflags &= ~SD_WAKE_BALANCE;
5639 /* Flags needing groups don't count if only 1 group in parent */
5640 if (parent->groups == parent->groups->next) {
5641 pflags &= ~(SD_LOAD_BALANCE |
5642 SD_BALANCE_NEWIDLE |
5646 SD_SHARE_PKG_RESOURCES);
5648 if (~cflags & pflags)
5655 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5656 * hold the hotplug lock.
5658 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5660 struct rq *rq = cpu_rq(cpu);
5661 struct sched_domain *tmp;
5663 /* Remove the sched domains which do not contribute to scheduling. */
5664 for (tmp = sd; tmp; tmp = tmp->parent) {
5665 struct sched_domain *parent = tmp->parent;
5668 if (sd_parent_degenerate(tmp, parent)) {
5669 tmp->parent = parent->parent;
5671 parent->parent->child = tmp;
5675 if (sd && sd_degenerate(sd)) {
5681 sched_domain_debug(sd, cpu);
5683 rcu_assign_pointer(rq->sd, sd);
5686 /* cpus with isolated domains */
5687 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5689 /* Setup the mask of cpus configured for isolated domains */
5690 static int __init isolated_cpu_setup(char *str)
5692 int ints[NR_CPUS], i;
5694 str = get_options(str, ARRAY_SIZE(ints), ints);
5695 cpus_clear(cpu_isolated_map);
5696 for (i = 1; i <= ints[0]; i++)
5697 if (ints[i] < NR_CPUS)
5698 cpu_set(ints[i], cpu_isolated_map);
5702 __setup ("isolcpus=", isolated_cpu_setup);
5705 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5706 * to a function which identifies what group(along with sched group) a CPU
5707 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5708 * (due to the fact that we keep track of groups covered with a cpumask_t).
5710 * init_sched_build_groups will build a circular linked list of the groups
5711 * covered by the given span, and will set each group's ->cpumask correctly,
5712 * and ->cpu_power to 0.
5715 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5716 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5717 struct sched_group **sg))
5719 struct sched_group *first = NULL, *last = NULL;
5720 cpumask_t covered = CPU_MASK_NONE;
5723 for_each_cpu_mask(i, span) {
5724 struct sched_group *sg;
5725 int group = group_fn(i, cpu_map, &sg);
5728 if (cpu_isset(i, covered))
5731 sg->cpumask = CPU_MASK_NONE;
5732 sg->__cpu_power = 0;
5734 for_each_cpu_mask(j, span) {
5735 if (group_fn(j, cpu_map, NULL) != group)
5738 cpu_set(j, covered);
5739 cpu_set(j, sg->cpumask);
5750 #define SD_NODES_PER_DOMAIN 16
5755 * find_next_best_node - find the next node to include in a sched_domain
5756 * @node: node whose sched_domain we're building
5757 * @used_nodes: nodes already in the sched_domain
5759 * Find the next node to include in a given scheduling domain. Simply
5760 * finds the closest node not already in the @used_nodes map.
5762 * Should use nodemask_t.
5764 static int find_next_best_node(int node, unsigned long *used_nodes)
5766 int i, n, val, min_val, best_node = 0;
5770 for (i = 0; i < MAX_NUMNODES; i++) {
5771 /* Start at @node */
5772 n = (node + i) % MAX_NUMNODES;
5774 if (!nr_cpus_node(n))
5777 /* Skip already used nodes */
5778 if (test_bit(n, used_nodes))
5781 /* Simple min distance search */
5782 val = node_distance(node, n);
5784 if (val < min_val) {
5790 set_bit(best_node, used_nodes);
5795 * sched_domain_node_span - get a cpumask for a node's sched_domain
5796 * @node: node whose cpumask we're constructing
5797 * @size: number of nodes to include in this span
5799 * Given a node, construct a good cpumask for its sched_domain to span. It
5800 * should be one that prevents unnecessary balancing, but also spreads tasks
5803 static cpumask_t sched_domain_node_span(int node)
5805 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5806 cpumask_t span, nodemask;
5810 bitmap_zero(used_nodes, MAX_NUMNODES);
5812 nodemask = node_to_cpumask(node);
5813 cpus_or(span, span, nodemask);
5814 set_bit(node, used_nodes);
5816 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5817 int next_node = find_next_best_node(node, used_nodes);
5819 nodemask = node_to_cpumask(next_node);
5820 cpus_or(span, span, nodemask);
5827 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5830 * SMT sched-domains:
5832 #ifdef CONFIG_SCHED_SMT
5833 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5834 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5836 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5837 struct sched_group **sg)
5840 *sg = &per_cpu(sched_group_cpus, cpu);
5846 * multi-core sched-domains:
5848 #ifdef CONFIG_SCHED_MC
5849 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5850 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5853 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5854 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5855 struct sched_group **sg)
5858 cpumask_t mask = cpu_sibling_map[cpu];
5859 cpus_and(mask, mask, *cpu_map);
5860 group = first_cpu(mask);
5862 *sg = &per_cpu(sched_group_core, group);
5865 #elif defined(CONFIG_SCHED_MC)
5866 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5867 struct sched_group **sg)
5870 *sg = &per_cpu(sched_group_core, cpu);
5875 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5876 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5878 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5879 struct sched_group **sg)
5882 #ifdef CONFIG_SCHED_MC
5883 cpumask_t mask = cpu_coregroup_map(cpu);
5884 cpus_and(mask, mask, *cpu_map);
5885 group = first_cpu(mask);
5886 #elif defined(CONFIG_SCHED_SMT)
5887 cpumask_t mask = cpu_sibling_map[cpu];
5888 cpus_and(mask, mask, *cpu_map);
5889 group = first_cpu(mask);
5894 *sg = &per_cpu(sched_group_phys, group);
5900 * The init_sched_build_groups can't handle what we want to do with node
5901 * groups, so roll our own. Now each node has its own list of groups which
5902 * gets dynamically allocated.
5904 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5905 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5907 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5908 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5910 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5911 struct sched_group **sg)
5913 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5916 cpus_and(nodemask, nodemask, *cpu_map);
5917 group = first_cpu(nodemask);
5920 *sg = &per_cpu(sched_group_allnodes, group);
5924 static void init_numa_sched_groups_power(struct sched_group *group_head)
5926 struct sched_group *sg = group_head;
5932 for_each_cpu_mask(j, sg->cpumask) {
5933 struct sched_domain *sd;
5935 sd = &per_cpu(phys_domains, j);
5936 if (j != first_cpu(sd->groups->cpumask)) {
5938 * Only add "power" once for each
5944 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5947 if (sg != group_head)
5953 /* Free memory allocated for various sched_group structures */
5954 static void free_sched_groups(const cpumask_t *cpu_map)
5958 for_each_cpu_mask(cpu, *cpu_map) {
5959 struct sched_group **sched_group_nodes
5960 = sched_group_nodes_bycpu[cpu];
5962 if (!sched_group_nodes)
5965 for (i = 0; i < MAX_NUMNODES; i++) {
5966 cpumask_t nodemask = node_to_cpumask(i);
5967 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5969 cpus_and(nodemask, nodemask, *cpu_map);
5970 if (cpus_empty(nodemask))
5980 if (oldsg != sched_group_nodes[i])
5983 kfree(sched_group_nodes);
5984 sched_group_nodes_bycpu[cpu] = NULL;
5988 static void free_sched_groups(const cpumask_t *cpu_map)
5994 * Initialize sched groups cpu_power.
5996 * cpu_power indicates the capacity of sched group, which is used while
5997 * distributing the load between different sched groups in a sched domain.
5998 * Typically cpu_power for all the groups in a sched domain will be same unless
5999 * there are asymmetries in the topology. If there are asymmetries, group
6000 * having more cpu_power will pickup more load compared to the group having
6003 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6004 * the maximum number of tasks a group can handle in the presence of other idle
6005 * or lightly loaded groups in the same sched domain.
6007 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6009 struct sched_domain *child;
6010 struct sched_group *group;
6012 WARN_ON(!sd || !sd->groups);
6014 if (cpu != first_cpu(sd->groups->cpumask))
6019 sd->groups->__cpu_power = 0;
6022 * For perf policy, if the groups in child domain share resources
6023 * (for example cores sharing some portions of the cache hierarchy
6024 * or SMT), then set this domain groups cpu_power such that each group
6025 * can handle only one task, when there are other idle groups in the
6026 * same sched domain.
6028 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6030 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6031 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6036 * add cpu_power of each child group to this groups cpu_power
6038 group = child->groups;
6040 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6041 group = group->next;
6042 } while (group != child->groups);
6046 * Build sched domains for a given set of cpus and attach the sched domains
6047 * to the individual cpus
6049 static int build_sched_domains(const cpumask_t *cpu_map)
6053 struct sched_group **sched_group_nodes = NULL;
6054 int sd_allnodes = 0;
6057 * Allocate the per-node list of sched groups
6059 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
6061 if (!sched_group_nodes) {
6062 printk(KERN_WARNING "Can not alloc sched group node list\n");
6065 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6069 * Set up domains for cpus specified by the cpu_map.
6071 for_each_cpu_mask(i, *cpu_map) {
6072 struct sched_domain *sd = NULL, *p;
6073 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6075 cpus_and(nodemask, nodemask, *cpu_map);
6078 if (cpus_weight(*cpu_map) >
6079 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6080 sd = &per_cpu(allnodes_domains, i);
6081 *sd = SD_ALLNODES_INIT;
6082 sd->span = *cpu_map;
6083 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6089 sd = &per_cpu(node_domains, i);
6091 sd->span = sched_domain_node_span(cpu_to_node(i));
6095 cpus_and(sd->span, sd->span, *cpu_map);
6099 sd = &per_cpu(phys_domains, i);
6101 sd->span = nodemask;
6105 cpu_to_phys_group(i, cpu_map, &sd->groups);
6107 #ifdef CONFIG_SCHED_MC
6109 sd = &per_cpu(core_domains, i);
6111 sd->span = cpu_coregroup_map(i);
6112 cpus_and(sd->span, sd->span, *cpu_map);
6115 cpu_to_core_group(i, cpu_map, &sd->groups);
6118 #ifdef CONFIG_SCHED_SMT
6120 sd = &per_cpu(cpu_domains, i);
6121 *sd = SD_SIBLING_INIT;
6122 sd->span = cpu_sibling_map[i];
6123 cpus_and(sd->span, sd->span, *cpu_map);
6126 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6130 #ifdef CONFIG_SCHED_SMT
6131 /* Set up CPU (sibling) groups */
6132 for_each_cpu_mask(i, *cpu_map) {
6133 cpumask_t this_sibling_map = cpu_sibling_map[i];
6134 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6135 if (i != first_cpu(this_sibling_map))
6138 init_sched_build_groups(this_sibling_map, cpu_map,
6143 #ifdef CONFIG_SCHED_MC
6144 /* Set up multi-core groups */
6145 for_each_cpu_mask(i, *cpu_map) {
6146 cpumask_t this_core_map = cpu_coregroup_map(i);
6147 cpus_and(this_core_map, this_core_map, *cpu_map);
6148 if (i != first_cpu(this_core_map))
6150 init_sched_build_groups(this_core_map, cpu_map,
6151 &cpu_to_core_group);
6155 /* Set up physical groups */
6156 for (i = 0; i < MAX_NUMNODES; i++) {
6157 cpumask_t nodemask = node_to_cpumask(i);
6159 cpus_and(nodemask, nodemask, *cpu_map);
6160 if (cpus_empty(nodemask))
6163 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6167 /* Set up node groups */
6169 init_sched_build_groups(*cpu_map, cpu_map,
6170 &cpu_to_allnodes_group);
6172 for (i = 0; i < MAX_NUMNODES; i++) {
6173 /* Set up node groups */
6174 struct sched_group *sg, *prev;
6175 cpumask_t nodemask = node_to_cpumask(i);
6176 cpumask_t domainspan;
6177 cpumask_t covered = CPU_MASK_NONE;
6180 cpus_and(nodemask, nodemask, *cpu_map);
6181 if (cpus_empty(nodemask)) {
6182 sched_group_nodes[i] = NULL;
6186 domainspan = sched_domain_node_span(i);
6187 cpus_and(domainspan, domainspan, *cpu_map);
6189 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6191 printk(KERN_WARNING "Can not alloc domain group for "
6195 sched_group_nodes[i] = sg;
6196 for_each_cpu_mask(j, nodemask) {
6197 struct sched_domain *sd;
6199 sd = &per_cpu(node_domains, j);
6202 sg->__cpu_power = 0;
6203 sg->cpumask = nodemask;
6205 cpus_or(covered, covered, nodemask);
6208 for (j = 0; j < MAX_NUMNODES; j++) {
6209 cpumask_t tmp, notcovered;
6210 int n = (i + j) % MAX_NUMNODES;
6212 cpus_complement(notcovered, covered);
6213 cpus_and(tmp, notcovered, *cpu_map);
6214 cpus_and(tmp, tmp, domainspan);
6215 if (cpus_empty(tmp))
6218 nodemask = node_to_cpumask(n);
6219 cpus_and(tmp, tmp, nodemask);
6220 if (cpus_empty(tmp))
6223 sg = kmalloc_node(sizeof(struct sched_group),
6227 "Can not alloc domain group for node %d\n", j);
6230 sg->__cpu_power = 0;
6232 sg->next = prev->next;
6233 cpus_or(covered, covered, tmp);
6240 /* Calculate CPU power for physical packages and nodes */
6241 #ifdef CONFIG_SCHED_SMT
6242 for_each_cpu_mask(i, *cpu_map) {
6243 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6245 init_sched_groups_power(i, sd);
6248 #ifdef CONFIG_SCHED_MC
6249 for_each_cpu_mask(i, *cpu_map) {
6250 struct sched_domain *sd = &per_cpu(core_domains, i);
6252 init_sched_groups_power(i, sd);
6256 for_each_cpu_mask(i, *cpu_map) {
6257 struct sched_domain *sd = &per_cpu(phys_domains, i);
6259 init_sched_groups_power(i, sd);
6263 for (i = 0; i < MAX_NUMNODES; i++)
6264 init_numa_sched_groups_power(sched_group_nodes[i]);
6267 struct sched_group *sg;
6269 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6270 init_numa_sched_groups_power(sg);
6274 /* Attach the domains */
6275 for_each_cpu_mask(i, *cpu_map) {
6276 struct sched_domain *sd;
6277 #ifdef CONFIG_SCHED_SMT
6278 sd = &per_cpu(cpu_domains, i);
6279 #elif defined(CONFIG_SCHED_MC)
6280 sd = &per_cpu(core_domains, i);
6282 sd = &per_cpu(phys_domains, i);
6284 cpu_attach_domain(sd, i);
6291 free_sched_groups(cpu_map);
6296 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6298 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6300 cpumask_t cpu_default_map;
6304 * Setup mask for cpus without special case scheduling requirements.
6305 * For now this just excludes isolated cpus, but could be used to
6306 * exclude other special cases in the future.
6308 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6310 err = build_sched_domains(&cpu_default_map);
6315 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6317 free_sched_groups(cpu_map);
6321 * Detach sched domains from a group of cpus specified in cpu_map
6322 * These cpus will now be attached to the NULL domain
6324 static void detach_destroy_domains(const cpumask_t *cpu_map)
6328 for_each_cpu_mask(i, *cpu_map)
6329 cpu_attach_domain(NULL, i);
6330 synchronize_sched();
6331 arch_destroy_sched_domains(cpu_map);
6335 * Partition sched domains as specified by the cpumasks below.
6336 * This attaches all cpus from the cpumasks to the NULL domain,
6337 * waits for a RCU quiescent period, recalculates sched
6338 * domain information and then attaches them back to the
6339 * correct sched domains
6340 * Call with hotplug lock held
6342 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6344 cpumask_t change_map;
6347 cpus_and(*partition1, *partition1, cpu_online_map);
6348 cpus_and(*partition2, *partition2, cpu_online_map);
6349 cpus_or(change_map, *partition1, *partition2);
6351 /* Detach sched domains from all of the affected cpus */
6352 detach_destroy_domains(&change_map);
6353 if (!cpus_empty(*partition1))
6354 err = build_sched_domains(partition1);
6355 if (!err && !cpus_empty(*partition2))
6356 err = build_sched_domains(partition2);
6361 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6362 static int arch_reinit_sched_domains(void)
6366 mutex_lock(&sched_hotcpu_mutex);
6367 detach_destroy_domains(&cpu_online_map);
6368 err = arch_init_sched_domains(&cpu_online_map);
6369 mutex_unlock(&sched_hotcpu_mutex);
6374 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6378 if (buf[0] != '0' && buf[0] != '1')
6382 sched_smt_power_savings = (buf[0] == '1');
6384 sched_mc_power_savings = (buf[0] == '1');
6386 ret = arch_reinit_sched_domains();
6388 return ret ? ret : count;
6391 #ifdef CONFIG_SCHED_MC
6392 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6394 return sprintf(page, "%u\n", sched_mc_power_savings);
6396 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6397 const char *buf, size_t count)
6399 return sched_power_savings_store(buf, count, 0);
6401 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6402 sched_mc_power_savings_store);
6405 #ifdef CONFIG_SCHED_SMT
6406 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6408 return sprintf(page, "%u\n", sched_smt_power_savings);
6410 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6411 const char *buf, size_t count)
6413 return sched_power_savings_store(buf, count, 1);
6415 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6416 sched_smt_power_savings_store);
6419 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6423 #ifdef CONFIG_SCHED_SMT
6425 err = sysfs_create_file(&cls->kset.kobj,
6426 &attr_sched_smt_power_savings.attr);
6428 #ifdef CONFIG_SCHED_MC
6429 if (!err && mc_capable())
6430 err = sysfs_create_file(&cls->kset.kobj,
6431 &attr_sched_mc_power_savings.attr);
6438 * Force a reinitialization of the sched domains hierarchy. The domains
6439 * and groups cannot be updated in place without racing with the balancing
6440 * code, so we temporarily attach all running cpus to the NULL domain
6441 * which will prevent rebalancing while the sched domains are recalculated.
6443 static int update_sched_domains(struct notifier_block *nfb,
6444 unsigned long action, void *hcpu)
6447 case CPU_UP_PREPARE:
6448 case CPU_UP_PREPARE_FROZEN:
6449 case CPU_DOWN_PREPARE:
6450 case CPU_DOWN_PREPARE_FROZEN:
6451 detach_destroy_domains(&cpu_online_map);
6454 case CPU_UP_CANCELED:
6455 case CPU_UP_CANCELED_FROZEN:
6456 case CPU_DOWN_FAILED:
6457 case CPU_DOWN_FAILED_FROZEN:
6459 case CPU_ONLINE_FROZEN:
6461 case CPU_DEAD_FROZEN:
6463 * Fall through and re-initialise the domains.
6470 /* The hotplug lock is already held by cpu_up/cpu_down */
6471 arch_init_sched_domains(&cpu_online_map);
6476 void __init sched_init_smp(void)
6478 cpumask_t non_isolated_cpus;
6480 mutex_lock(&sched_hotcpu_mutex);
6481 arch_init_sched_domains(&cpu_online_map);
6482 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6483 if (cpus_empty(non_isolated_cpus))
6484 cpu_set(smp_processor_id(), non_isolated_cpus);
6485 mutex_unlock(&sched_hotcpu_mutex);
6486 /* XXX: Theoretical race here - CPU may be hotplugged now */
6487 hotcpu_notifier(update_sched_domains, 0);
6489 init_sched_domain_sysctl();
6491 /* Move init over to a non-isolated CPU */
6492 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6494 sched_init_granularity();
6497 void __init sched_init_smp(void)
6499 sched_init_granularity();
6501 #endif /* CONFIG_SMP */
6503 int in_sched_functions(unsigned long addr)
6505 /* Linker adds these: start and end of __sched functions */
6506 extern char __sched_text_start[], __sched_text_end[];
6508 return in_lock_functions(addr) ||
6509 (addr >= (unsigned long)__sched_text_start
6510 && addr < (unsigned long)__sched_text_end);
6513 static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6515 cfs_rq->tasks_timeline = RB_ROOT;
6516 cfs_rq->fair_clock = 1;
6517 #ifdef CONFIG_FAIR_GROUP_SCHED
6522 void __init sched_init(void)
6524 u64 now = sched_clock();
6525 int highest_cpu = 0;
6529 * Link up the scheduling class hierarchy:
6531 rt_sched_class.next = &fair_sched_class;
6532 fair_sched_class.next = &idle_sched_class;
6533 idle_sched_class.next = NULL;
6535 for_each_possible_cpu(i) {
6536 struct rt_prio_array *array;
6540 spin_lock_init(&rq->lock);
6541 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6544 init_cfs_rq(&rq->cfs, rq);
6545 #ifdef CONFIG_FAIR_GROUP_SCHED
6546 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6547 list_add(&rq->cfs.leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6549 rq->ls.load_update_last = now;
6550 rq->ls.load_update_start = now;
6552 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6553 rq->cpu_load[j] = 0;
6556 rq->active_balance = 0;
6557 rq->next_balance = jiffies;
6560 rq->migration_thread = NULL;
6561 INIT_LIST_HEAD(&rq->migration_queue);
6563 atomic_set(&rq->nr_iowait, 0);
6565 array = &rq->rt.active;
6566 for (j = 0; j < MAX_RT_PRIO; j++) {
6567 INIT_LIST_HEAD(array->queue + j);
6568 __clear_bit(j, array->bitmap);
6571 /* delimiter for bitsearch: */
6572 __set_bit(MAX_RT_PRIO, array->bitmap);
6575 set_load_weight(&init_task);
6577 #ifdef CONFIG_PREEMPT_NOTIFIERS
6578 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6582 nr_cpu_ids = highest_cpu + 1;
6583 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6586 #ifdef CONFIG_RT_MUTEXES
6587 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6591 * The boot idle thread does lazy MMU switching as well:
6593 atomic_inc(&init_mm.mm_count);
6594 enter_lazy_tlb(&init_mm, current);
6597 * Make us the idle thread. Technically, schedule() should not be
6598 * called from this thread, however somewhere below it might be,
6599 * but because we are the idle thread, we just pick up running again
6600 * when this runqueue becomes "idle".
6602 init_idle(current, smp_processor_id());
6604 * During early bootup we pretend to be a normal task:
6606 current->sched_class = &fair_sched_class;
6609 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6610 void __might_sleep(char *file, int line)
6613 static unsigned long prev_jiffy; /* ratelimiting */
6615 if ((in_atomic() || irqs_disabled()) &&
6616 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6617 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6619 prev_jiffy = jiffies;
6620 printk(KERN_ERR "BUG: sleeping function called from invalid"
6621 " context at %s:%d\n", file, line);
6622 printk("in_atomic():%d, irqs_disabled():%d\n",
6623 in_atomic(), irqs_disabled());
6624 debug_show_held_locks(current);
6625 if (irqs_disabled())
6626 print_irqtrace_events(current);
6631 EXPORT_SYMBOL(__might_sleep);
6634 #ifdef CONFIG_MAGIC_SYSRQ
6635 void normalize_rt_tasks(void)
6637 struct task_struct *g, *p;
6638 unsigned long flags;
6642 read_lock_irq(&tasklist_lock);
6643 do_each_thread(g, p) {
6645 p->se.wait_runtime = 0;
6646 p->se.exec_start = 0;
6647 p->se.wait_start_fair = 0;
6648 p->se.sleep_start_fair = 0;
6649 #ifdef CONFIG_SCHEDSTATS
6650 p->se.wait_start = 0;
6651 p->se.sleep_start = 0;
6652 p->se.block_start = 0;
6654 task_rq(p)->cfs.fair_clock = 0;
6655 task_rq(p)->clock = 0;
6659 * Renice negative nice level userspace
6662 if (TASK_NICE(p) < 0 && p->mm)
6663 set_user_nice(p, 0);
6667 spin_lock_irqsave(&p->pi_lock, flags);
6668 rq = __task_rq_lock(p);
6671 * Do not touch the migration thread:
6673 if (p == rq->migration_thread)
6677 update_rq_clock(rq);
6678 on_rq = p->se.on_rq;
6680 deactivate_task(rq, p, 0);
6681 __setscheduler(rq, p, SCHED_NORMAL, 0);
6683 activate_task(rq, p, 0);
6684 resched_task(rq->curr);
6689 __task_rq_unlock(rq);
6690 spin_unlock_irqrestore(&p->pi_lock, flags);
6691 } while_each_thread(g, p);
6693 read_unlock_irq(&tasklist_lock);
6696 #endif /* CONFIG_MAGIC_SYSRQ */
6700 * These functions are only useful for the IA64 MCA handling.
6702 * They can only be called when the whole system has been
6703 * stopped - every CPU needs to be quiescent, and no scheduling
6704 * activity can take place. Using them for anything else would
6705 * be a serious bug, and as a result, they aren't even visible
6706 * under any other configuration.
6710 * curr_task - return the current task for a given cpu.
6711 * @cpu: the processor in question.
6713 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6715 struct task_struct *curr_task(int cpu)
6717 return cpu_curr(cpu);
6721 * set_curr_task - set the current task for a given cpu.
6722 * @cpu: the processor in question.
6723 * @p: the task pointer to set.
6725 * Description: This function must only be used when non-maskable interrupts
6726 * are serviced on a separate stack. It allows the architecture to switch the
6727 * notion of the current task on a cpu in a non-blocking manner. This function
6728 * must be called with all CPU's synchronized, and interrupts disabled, the
6729 * and caller must save the original value of the current task (see
6730 * curr_task() above) and restore that value before reenabling interrupts and
6731 * re-starting the system.
6733 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6735 void set_curr_task(int cpu, struct task_struct *p)