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
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.wait_runtime = 0;
1592 p->se.sleep_start_fair = 0;
1594 #ifdef CONFIG_SCHEDSTATS
1595 p->se.wait_start = 0;
1596 p->se.sum_wait_runtime = 0;
1597 p->se.sum_sleep_runtime = 0;
1598 p->se.sleep_start = 0;
1599 p->se.block_start = 0;
1600 p->se.sleep_max = 0;
1601 p->se.block_max = 0;
1603 p->se.slice_max = 0;
1605 p->se.wait_runtime_overruns = 0;
1606 p->se.wait_runtime_underruns = 0;
1609 INIT_LIST_HEAD(&p->run_list);
1612 #ifdef CONFIG_PREEMPT_NOTIFIERS
1613 INIT_HLIST_HEAD(&p->preempt_notifiers);
1617 * We mark the process as running here, but have not actually
1618 * inserted it onto the runqueue yet. This guarantees that
1619 * nobody will actually run it, and a signal or other external
1620 * event cannot wake it up and insert it on the runqueue either.
1622 p->state = TASK_RUNNING;
1626 * fork()/clone()-time setup:
1628 void sched_fork(struct task_struct *p, int clone_flags)
1630 int cpu = get_cpu();
1635 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1637 __set_task_cpu(p, cpu);
1640 * Make sure we do not leak PI boosting priority to the child:
1642 p->prio = current->normal_prio;
1644 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1645 if (likely(sched_info_on()))
1646 memset(&p->sched_info, 0, sizeof(p->sched_info));
1648 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1651 #ifdef CONFIG_PREEMPT
1652 /* Want to start with kernel preemption disabled. */
1653 task_thread_info(p)->preempt_count = 1;
1659 * wake_up_new_task - wake up a newly created task for the first time.
1661 * This function will do some initial scheduler statistics housekeeping
1662 * that must be done for every newly created context, then puts the task
1663 * on the runqueue and wakes it.
1665 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1667 unsigned long flags;
1671 rq = task_rq_lock(p, &flags);
1672 BUG_ON(p->state != TASK_RUNNING);
1673 this_cpu = smp_processor_id(); /* parent's CPU */
1674 update_rq_clock(rq);
1676 p->prio = effective_prio(p);
1678 if (rt_prio(p->prio))
1679 p->sched_class = &rt_sched_class;
1681 p->sched_class = &fair_sched_class;
1683 if (task_cpu(p) != this_cpu || !p->sched_class->task_new ||
1684 !current->se.on_rq) {
1685 activate_task(rq, p, 0);
1688 * Let the scheduling class do new task startup
1689 * management (if any):
1691 p->sched_class->task_new(rq, p);
1692 inc_nr_running(p, rq);
1694 check_preempt_curr(rq, p);
1695 task_rq_unlock(rq, &flags);
1698 #ifdef CONFIG_PREEMPT_NOTIFIERS
1701 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1702 * @notifier: notifier struct to register
1704 void preempt_notifier_register(struct preempt_notifier *notifier)
1706 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1708 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1711 * preempt_notifier_unregister - no longer interested in preemption notifications
1712 * @notifier: notifier struct to unregister
1714 * This is safe to call from within a preemption notifier.
1716 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1718 hlist_del(¬ifier->link);
1720 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1722 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1724 struct preempt_notifier *notifier;
1725 struct hlist_node *node;
1727 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1728 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1732 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1733 struct task_struct *next)
1735 struct preempt_notifier *notifier;
1736 struct hlist_node *node;
1738 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1739 notifier->ops->sched_out(notifier, next);
1744 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1749 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1750 struct task_struct *next)
1757 * prepare_task_switch - prepare to switch tasks
1758 * @rq: the runqueue preparing to switch
1759 * @prev: the current task that is being switched out
1760 * @next: the task we are going to switch to.
1762 * This is called with the rq lock held and interrupts off. It must
1763 * be paired with a subsequent finish_task_switch after the context
1766 * prepare_task_switch sets up locking and calls architecture specific
1770 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1771 struct task_struct *next)
1773 fire_sched_out_preempt_notifiers(prev, next);
1774 prepare_lock_switch(rq, next);
1775 prepare_arch_switch(next);
1779 * finish_task_switch - clean up after a task-switch
1780 * @rq: runqueue associated with task-switch
1781 * @prev: the thread we just switched away from.
1783 * finish_task_switch must be called after the context switch, paired
1784 * with a prepare_task_switch call before the context switch.
1785 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1786 * and do any other architecture-specific cleanup actions.
1788 * Note that we may have delayed dropping an mm in context_switch(). If
1789 * so, we finish that here outside of the runqueue lock. (Doing it
1790 * with the lock held can cause deadlocks; see schedule() for
1793 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1794 __releases(rq->lock)
1796 struct mm_struct *mm = rq->prev_mm;
1802 * A task struct has one reference for the use as "current".
1803 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1804 * schedule one last time. The schedule call will never return, and
1805 * the scheduled task must drop that reference.
1806 * The test for TASK_DEAD must occur while the runqueue locks are
1807 * still held, otherwise prev could be scheduled on another cpu, die
1808 * there before we look at prev->state, and then the reference would
1810 * Manfred Spraul <manfred@colorfullife.com>
1812 prev_state = prev->state;
1813 finish_arch_switch(prev);
1814 finish_lock_switch(rq, prev);
1815 fire_sched_in_preempt_notifiers(current);
1818 if (unlikely(prev_state == TASK_DEAD)) {
1820 * Remove function-return probe instances associated with this
1821 * task and put them back on the free list.
1823 kprobe_flush_task(prev);
1824 put_task_struct(prev);
1829 * schedule_tail - first thing a freshly forked thread must call.
1830 * @prev: the thread we just switched away from.
1832 asmlinkage void schedule_tail(struct task_struct *prev)
1833 __releases(rq->lock)
1835 struct rq *rq = this_rq();
1837 finish_task_switch(rq, prev);
1838 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1839 /* In this case, finish_task_switch does not reenable preemption */
1842 if (current->set_child_tid)
1843 put_user(current->pid, current->set_child_tid);
1847 * context_switch - switch to the new MM and the new
1848 * thread's register state.
1851 context_switch(struct rq *rq, struct task_struct *prev,
1852 struct task_struct *next)
1854 struct mm_struct *mm, *oldmm;
1856 prepare_task_switch(rq, prev, next);
1858 oldmm = prev->active_mm;
1860 * For paravirt, this is coupled with an exit in switch_to to
1861 * combine the page table reload and the switch backend into
1864 arch_enter_lazy_cpu_mode();
1866 if (unlikely(!mm)) {
1867 next->active_mm = oldmm;
1868 atomic_inc(&oldmm->mm_count);
1869 enter_lazy_tlb(oldmm, next);
1871 switch_mm(oldmm, mm, next);
1873 if (unlikely(!prev->mm)) {
1874 prev->active_mm = NULL;
1875 rq->prev_mm = oldmm;
1878 * Since the runqueue lock will be released by the next
1879 * task (which is an invalid locking op but in the case
1880 * of the scheduler it's an obvious special-case), so we
1881 * do an early lockdep release here:
1883 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1884 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1887 /* Here we just switch the register state and the stack. */
1888 switch_to(prev, next, prev);
1892 * this_rq must be evaluated again because prev may have moved
1893 * CPUs since it called schedule(), thus the 'rq' on its stack
1894 * frame will be invalid.
1896 finish_task_switch(this_rq(), prev);
1900 * nr_running, nr_uninterruptible and nr_context_switches:
1902 * externally visible scheduler statistics: current number of runnable
1903 * threads, current number of uninterruptible-sleeping threads, total
1904 * number of context switches performed since bootup.
1906 unsigned long nr_running(void)
1908 unsigned long i, sum = 0;
1910 for_each_online_cpu(i)
1911 sum += cpu_rq(i)->nr_running;
1916 unsigned long nr_uninterruptible(void)
1918 unsigned long i, sum = 0;
1920 for_each_possible_cpu(i)
1921 sum += cpu_rq(i)->nr_uninterruptible;
1924 * Since we read the counters lockless, it might be slightly
1925 * inaccurate. Do not allow it to go below zero though:
1927 if (unlikely((long)sum < 0))
1933 unsigned long long nr_context_switches(void)
1936 unsigned long long sum = 0;
1938 for_each_possible_cpu(i)
1939 sum += cpu_rq(i)->nr_switches;
1944 unsigned long nr_iowait(void)
1946 unsigned long i, sum = 0;
1948 for_each_possible_cpu(i)
1949 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1954 unsigned long nr_active(void)
1956 unsigned long i, running = 0, uninterruptible = 0;
1958 for_each_online_cpu(i) {
1959 running += cpu_rq(i)->nr_running;
1960 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1963 if (unlikely((long)uninterruptible < 0))
1964 uninterruptible = 0;
1966 return running + uninterruptible;
1970 * Update rq->cpu_load[] statistics. This function is usually called every
1971 * scheduler tick (TICK_NSEC).
1973 static void update_cpu_load(struct rq *this_rq)
1975 unsigned long total_load = this_rq->ls.load.weight;
1976 unsigned long this_load = total_load;
1979 this_rq->nr_load_updates++;
1981 /* Update our load: */
1982 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
1983 unsigned long old_load, new_load;
1985 /* scale is effectively 1 << i now, and >> i divides by scale */
1987 old_load = this_rq->cpu_load[i];
1988 new_load = this_load;
1990 * Round up the averaging division if load is increasing. This
1991 * prevents us from getting stuck on 9 if the load is 10, for
1994 if (new_load > old_load)
1995 new_load += scale-1;
1996 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2003 * double_rq_lock - safely lock two runqueues
2005 * Note this does not disable interrupts like task_rq_lock,
2006 * you need to do so manually before calling.
2008 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2009 __acquires(rq1->lock)
2010 __acquires(rq2->lock)
2012 BUG_ON(!irqs_disabled());
2014 spin_lock(&rq1->lock);
2015 __acquire(rq2->lock); /* Fake it out ;) */
2018 spin_lock(&rq1->lock);
2019 spin_lock(&rq2->lock);
2021 spin_lock(&rq2->lock);
2022 spin_lock(&rq1->lock);
2025 update_rq_clock(rq1);
2026 update_rq_clock(rq2);
2030 * double_rq_unlock - safely unlock two runqueues
2032 * Note this does not restore interrupts like task_rq_unlock,
2033 * you need to do so manually after calling.
2035 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2036 __releases(rq1->lock)
2037 __releases(rq2->lock)
2039 spin_unlock(&rq1->lock);
2041 spin_unlock(&rq2->lock);
2043 __release(rq2->lock);
2047 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2049 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2050 __releases(this_rq->lock)
2051 __acquires(busiest->lock)
2052 __acquires(this_rq->lock)
2054 if (unlikely(!irqs_disabled())) {
2055 /* printk() doesn't work good under rq->lock */
2056 spin_unlock(&this_rq->lock);
2059 if (unlikely(!spin_trylock(&busiest->lock))) {
2060 if (busiest < this_rq) {
2061 spin_unlock(&this_rq->lock);
2062 spin_lock(&busiest->lock);
2063 spin_lock(&this_rq->lock);
2065 spin_lock(&busiest->lock);
2070 * If dest_cpu is allowed for this process, migrate the task to it.
2071 * This is accomplished by forcing the cpu_allowed mask to only
2072 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2073 * the cpu_allowed mask is restored.
2075 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2077 struct migration_req req;
2078 unsigned long flags;
2081 rq = task_rq_lock(p, &flags);
2082 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2083 || unlikely(cpu_is_offline(dest_cpu)))
2086 /* force the process onto the specified CPU */
2087 if (migrate_task(p, dest_cpu, &req)) {
2088 /* Need to wait for migration thread (might exit: take ref). */
2089 struct task_struct *mt = rq->migration_thread;
2091 get_task_struct(mt);
2092 task_rq_unlock(rq, &flags);
2093 wake_up_process(mt);
2094 put_task_struct(mt);
2095 wait_for_completion(&req.done);
2100 task_rq_unlock(rq, &flags);
2104 * sched_exec - execve() is a valuable balancing opportunity, because at
2105 * this point the task has the smallest effective memory and cache footprint.
2107 void sched_exec(void)
2109 int new_cpu, this_cpu = get_cpu();
2110 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2112 if (new_cpu != this_cpu)
2113 sched_migrate_task(current, new_cpu);
2117 * pull_task - move a task from a remote runqueue to the local runqueue.
2118 * Both runqueues must be locked.
2120 static void pull_task(struct rq *src_rq, struct task_struct *p,
2121 struct rq *this_rq, int this_cpu)
2123 deactivate_task(src_rq, p, 0);
2124 set_task_cpu(p, this_cpu);
2125 activate_task(this_rq, p, 0);
2127 * Note that idle threads have a prio of MAX_PRIO, for this test
2128 * to be always true for them.
2130 check_preempt_curr(this_rq, p);
2134 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2137 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2138 struct sched_domain *sd, enum cpu_idle_type idle,
2142 * We do not migrate tasks that are:
2143 * 1) running (obviously), or
2144 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2145 * 3) are cache-hot on their current CPU.
2147 if (!cpu_isset(this_cpu, p->cpus_allowed))
2151 if (task_running(rq, p))
2157 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2158 unsigned long max_nr_move, unsigned long max_load_move,
2159 struct sched_domain *sd, enum cpu_idle_type idle,
2160 int *all_pinned, unsigned long *load_moved,
2161 int *this_best_prio, struct rq_iterator *iterator)
2163 int pulled = 0, pinned = 0, skip_for_load;
2164 struct task_struct *p;
2165 long rem_load_move = max_load_move;
2167 if (max_nr_move == 0 || max_load_move == 0)
2173 * Start the load-balancing iterator:
2175 p = iterator->start(iterator->arg);
2180 * To help distribute high priority tasks accross CPUs we don't
2181 * skip a task if it will be the highest priority task (i.e. smallest
2182 * prio value) on its new queue regardless of its load weight
2184 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2185 SCHED_LOAD_SCALE_FUZZ;
2186 if ((skip_for_load && p->prio >= *this_best_prio) ||
2187 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2188 p = iterator->next(iterator->arg);
2192 pull_task(busiest, p, this_rq, this_cpu);
2194 rem_load_move -= p->se.load.weight;
2197 * We only want to steal up to the prescribed number of tasks
2198 * and the prescribed amount of weighted load.
2200 if (pulled < max_nr_move && rem_load_move > 0) {
2201 if (p->prio < *this_best_prio)
2202 *this_best_prio = p->prio;
2203 p = iterator->next(iterator->arg);
2208 * Right now, this is the only place pull_task() is called,
2209 * so we can safely collect pull_task() stats here rather than
2210 * inside pull_task().
2212 schedstat_add(sd, lb_gained[idle], pulled);
2215 *all_pinned = pinned;
2216 *load_moved = max_load_move - rem_load_move;
2221 * move_tasks tries to move up to max_load_move weighted load from busiest to
2222 * this_rq, as part of a balancing operation within domain "sd".
2223 * Returns 1 if successful and 0 otherwise.
2225 * Called with both runqueues locked.
2227 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2228 unsigned long max_load_move,
2229 struct sched_domain *sd, enum cpu_idle_type idle,
2232 struct sched_class *class = sched_class_highest;
2233 unsigned long total_load_moved = 0;
2234 int this_best_prio = this_rq->curr->prio;
2238 class->load_balance(this_rq, this_cpu, busiest,
2239 ULONG_MAX, max_load_move - total_load_moved,
2240 sd, idle, all_pinned, &this_best_prio);
2241 class = class->next;
2242 } while (class && max_load_move > total_load_moved);
2244 return total_load_moved > 0;
2248 * move_one_task tries to move exactly one task from busiest to this_rq, as
2249 * part of active balancing operations within "domain".
2250 * Returns 1 if successful and 0 otherwise.
2252 * Called with both runqueues locked.
2254 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2255 struct sched_domain *sd, enum cpu_idle_type idle)
2257 struct sched_class *class;
2258 int this_best_prio = MAX_PRIO;
2260 for (class = sched_class_highest; class; class = class->next)
2261 if (class->load_balance(this_rq, this_cpu, busiest,
2262 1, ULONG_MAX, sd, idle, NULL,
2270 * find_busiest_group finds and returns the busiest CPU group within the
2271 * domain. It calculates and returns the amount of weighted load which
2272 * should be moved to restore balance via the imbalance parameter.
2274 static struct sched_group *
2275 find_busiest_group(struct sched_domain *sd, int this_cpu,
2276 unsigned long *imbalance, enum cpu_idle_type idle,
2277 int *sd_idle, cpumask_t *cpus, int *balance)
2279 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2280 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2281 unsigned long max_pull;
2282 unsigned long busiest_load_per_task, busiest_nr_running;
2283 unsigned long this_load_per_task, this_nr_running;
2285 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2286 int power_savings_balance = 1;
2287 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2288 unsigned long min_nr_running = ULONG_MAX;
2289 struct sched_group *group_min = NULL, *group_leader = NULL;
2292 max_load = this_load = total_load = total_pwr = 0;
2293 busiest_load_per_task = busiest_nr_running = 0;
2294 this_load_per_task = this_nr_running = 0;
2295 if (idle == CPU_NOT_IDLE)
2296 load_idx = sd->busy_idx;
2297 else if (idle == CPU_NEWLY_IDLE)
2298 load_idx = sd->newidle_idx;
2300 load_idx = sd->idle_idx;
2303 unsigned long load, group_capacity;
2306 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2307 unsigned long sum_nr_running, sum_weighted_load;
2309 local_group = cpu_isset(this_cpu, group->cpumask);
2312 balance_cpu = first_cpu(group->cpumask);
2314 /* Tally up the load of all CPUs in the group */
2315 sum_weighted_load = sum_nr_running = avg_load = 0;
2317 for_each_cpu_mask(i, group->cpumask) {
2320 if (!cpu_isset(i, *cpus))
2325 if (*sd_idle && rq->nr_running)
2328 /* Bias balancing toward cpus of our domain */
2330 if (idle_cpu(i) && !first_idle_cpu) {
2335 load = target_load(i, load_idx);
2337 load = source_load(i, load_idx);
2340 sum_nr_running += rq->nr_running;
2341 sum_weighted_load += weighted_cpuload(i);
2345 * First idle cpu or the first cpu(busiest) in this sched group
2346 * is eligible for doing load balancing at this and above
2347 * domains. In the newly idle case, we will allow all the cpu's
2348 * to do the newly idle load balance.
2350 if (idle != CPU_NEWLY_IDLE && local_group &&
2351 balance_cpu != this_cpu && balance) {
2356 total_load += avg_load;
2357 total_pwr += group->__cpu_power;
2359 /* Adjust by relative CPU power of the group */
2360 avg_load = sg_div_cpu_power(group,
2361 avg_load * SCHED_LOAD_SCALE);
2363 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2366 this_load = avg_load;
2368 this_nr_running = sum_nr_running;
2369 this_load_per_task = sum_weighted_load;
2370 } else if (avg_load > max_load &&
2371 sum_nr_running > group_capacity) {
2372 max_load = avg_load;
2374 busiest_nr_running = sum_nr_running;
2375 busiest_load_per_task = sum_weighted_load;
2378 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2380 * Busy processors will not participate in power savings
2383 if (idle == CPU_NOT_IDLE ||
2384 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2388 * If the local group is idle or completely loaded
2389 * no need to do power savings balance at this domain
2391 if (local_group && (this_nr_running >= group_capacity ||
2393 power_savings_balance = 0;
2396 * If a group is already running at full capacity or idle,
2397 * don't include that group in power savings calculations
2399 if (!power_savings_balance || sum_nr_running >= group_capacity
2404 * Calculate the group which has the least non-idle load.
2405 * This is the group from where we need to pick up the load
2408 if ((sum_nr_running < min_nr_running) ||
2409 (sum_nr_running == min_nr_running &&
2410 first_cpu(group->cpumask) <
2411 first_cpu(group_min->cpumask))) {
2413 min_nr_running = sum_nr_running;
2414 min_load_per_task = sum_weighted_load /
2419 * Calculate the group which is almost near its
2420 * capacity but still has some space to pick up some load
2421 * from other group and save more power
2423 if (sum_nr_running <= group_capacity - 1) {
2424 if (sum_nr_running > leader_nr_running ||
2425 (sum_nr_running == leader_nr_running &&
2426 first_cpu(group->cpumask) >
2427 first_cpu(group_leader->cpumask))) {
2428 group_leader = group;
2429 leader_nr_running = sum_nr_running;
2434 group = group->next;
2435 } while (group != sd->groups);
2437 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2440 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2442 if (this_load >= avg_load ||
2443 100*max_load <= sd->imbalance_pct*this_load)
2446 busiest_load_per_task /= busiest_nr_running;
2448 * We're trying to get all the cpus to the average_load, so we don't
2449 * want to push ourselves above the average load, nor do we wish to
2450 * reduce the max loaded cpu below the average load, as either of these
2451 * actions would just result in more rebalancing later, and ping-pong
2452 * tasks around. Thus we look for the minimum possible imbalance.
2453 * Negative imbalances (*we* are more loaded than anyone else) will
2454 * be counted as no imbalance for these purposes -- we can't fix that
2455 * by pulling tasks to us. Be careful of negative numbers as they'll
2456 * appear as very large values with unsigned longs.
2458 if (max_load <= busiest_load_per_task)
2462 * In the presence of smp nice balancing, certain scenarios can have
2463 * max load less than avg load(as we skip the groups at or below
2464 * its cpu_power, while calculating max_load..)
2466 if (max_load < avg_load) {
2468 goto small_imbalance;
2471 /* Don't want to pull so many tasks that a group would go idle */
2472 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2474 /* How much load to actually move to equalise the imbalance */
2475 *imbalance = min(max_pull * busiest->__cpu_power,
2476 (avg_load - this_load) * this->__cpu_power)
2480 * if *imbalance is less than the average load per runnable task
2481 * there is no gaurantee that any tasks will be moved so we'll have
2482 * a think about bumping its value to force at least one task to be
2485 if (*imbalance < busiest_load_per_task) {
2486 unsigned long tmp, pwr_now, pwr_move;
2490 pwr_move = pwr_now = 0;
2492 if (this_nr_running) {
2493 this_load_per_task /= this_nr_running;
2494 if (busiest_load_per_task > this_load_per_task)
2497 this_load_per_task = SCHED_LOAD_SCALE;
2499 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2500 busiest_load_per_task * imbn) {
2501 *imbalance = busiest_load_per_task;
2506 * OK, we don't have enough imbalance to justify moving tasks,
2507 * however we may be able to increase total CPU power used by
2511 pwr_now += busiest->__cpu_power *
2512 min(busiest_load_per_task, max_load);
2513 pwr_now += this->__cpu_power *
2514 min(this_load_per_task, this_load);
2515 pwr_now /= SCHED_LOAD_SCALE;
2517 /* Amount of load we'd subtract */
2518 tmp = sg_div_cpu_power(busiest,
2519 busiest_load_per_task * SCHED_LOAD_SCALE);
2521 pwr_move += busiest->__cpu_power *
2522 min(busiest_load_per_task, max_load - tmp);
2524 /* Amount of load we'd add */
2525 if (max_load * busiest->__cpu_power <
2526 busiest_load_per_task * SCHED_LOAD_SCALE)
2527 tmp = sg_div_cpu_power(this,
2528 max_load * busiest->__cpu_power);
2530 tmp = sg_div_cpu_power(this,
2531 busiest_load_per_task * SCHED_LOAD_SCALE);
2532 pwr_move += this->__cpu_power *
2533 min(this_load_per_task, this_load + tmp);
2534 pwr_move /= SCHED_LOAD_SCALE;
2536 /* Move if we gain throughput */
2537 if (pwr_move > pwr_now)
2538 *imbalance = busiest_load_per_task;
2544 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2545 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2548 if (this == group_leader && group_leader != group_min) {
2549 *imbalance = min_load_per_task;
2559 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2562 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2563 unsigned long imbalance, cpumask_t *cpus)
2565 struct rq *busiest = NULL, *rq;
2566 unsigned long max_load = 0;
2569 for_each_cpu_mask(i, group->cpumask) {
2572 if (!cpu_isset(i, *cpus))
2576 wl = weighted_cpuload(i);
2578 if (rq->nr_running == 1 && wl > imbalance)
2581 if (wl > max_load) {
2591 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2592 * so long as it is large enough.
2594 #define MAX_PINNED_INTERVAL 512
2597 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2598 * tasks if there is an imbalance.
2600 static int load_balance(int this_cpu, struct rq *this_rq,
2601 struct sched_domain *sd, enum cpu_idle_type idle,
2604 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2605 struct sched_group *group;
2606 unsigned long imbalance;
2608 cpumask_t cpus = CPU_MASK_ALL;
2609 unsigned long flags;
2612 * When power savings policy is enabled for the parent domain, idle
2613 * sibling can pick up load irrespective of busy siblings. In this case,
2614 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2615 * portraying it as CPU_NOT_IDLE.
2617 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2618 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2621 schedstat_inc(sd, lb_cnt[idle]);
2624 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2631 schedstat_inc(sd, lb_nobusyg[idle]);
2635 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2637 schedstat_inc(sd, lb_nobusyq[idle]);
2641 BUG_ON(busiest == this_rq);
2643 schedstat_add(sd, lb_imbalance[idle], imbalance);
2646 if (busiest->nr_running > 1) {
2648 * Attempt to move tasks. If find_busiest_group has found
2649 * an imbalance but busiest->nr_running <= 1, the group is
2650 * still unbalanced. ld_moved simply stays zero, so it is
2651 * correctly treated as an imbalance.
2653 local_irq_save(flags);
2654 double_rq_lock(this_rq, busiest);
2655 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2656 imbalance, sd, idle, &all_pinned);
2657 double_rq_unlock(this_rq, busiest);
2658 local_irq_restore(flags);
2661 * some other cpu did the load balance for us.
2663 if (ld_moved && this_cpu != smp_processor_id())
2664 resched_cpu(this_cpu);
2666 /* All tasks on this runqueue were pinned by CPU affinity */
2667 if (unlikely(all_pinned)) {
2668 cpu_clear(cpu_of(busiest), cpus);
2669 if (!cpus_empty(cpus))
2676 schedstat_inc(sd, lb_failed[idle]);
2677 sd->nr_balance_failed++;
2679 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2681 spin_lock_irqsave(&busiest->lock, flags);
2683 /* don't kick the migration_thread, if the curr
2684 * task on busiest cpu can't be moved to this_cpu
2686 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2687 spin_unlock_irqrestore(&busiest->lock, flags);
2689 goto out_one_pinned;
2692 if (!busiest->active_balance) {
2693 busiest->active_balance = 1;
2694 busiest->push_cpu = this_cpu;
2697 spin_unlock_irqrestore(&busiest->lock, flags);
2699 wake_up_process(busiest->migration_thread);
2702 * We've kicked active balancing, reset the failure
2705 sd->nr_balance_failed = sd->cache_nice_tries+1;
2708 sd->nr_balance_failed = 0;
2710 if (likely(!active_balance)) {
2711 /* We were unbalanced, so reset the balancing interval */
2712 sd->balance_interval = sd->min_interval;
2715 * If we've begun active balancing, start to back off. This
2716 * case may not be covered by the all_pinned logic if there
2717 * is only 1 task on the busy runqueue (because we don't call
2720 if (sd->balance_interval < sd->max_interval)
2721 sd->balance_interval *= 2;
2724 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2725 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2730 schedstat_inc(sd, lb_balanced[idle]);
2732 sd->nr_balance_failed = 0;
2735 /* tune up the balancing interval */
2736 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2737 (sd->balance_interval < sd->max_interval))
2738 sd->balance_interval *= 2;
2740 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2741 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2747 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2748 * tasks if there is an imbalance.
2750 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2751 * this_rq is locked.
2754 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2756 struct sched_group *group;
2757 struct rq *busiest = NULL;
2758 unsigned long imbalance;
2762 cpumask_t cpus = CPU_MASK_ALL;
2765 * When power savings policy is enabled for the parent domain, idle
2766 * sibling can pick up load irrespective of busy siblings. In this case,
2767 * let the state of idle sibling percolate up as IDLE, instead of
2768 * portraying it as CPU_NOT_IDLE.
2770 if (sd->flags & SD_SHARE_CPUPOWER &&
2771 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2774 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2776 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2777 &sd_idle, &cpus, NULL);
2779 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2783 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2786 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2790 BUG_ON(busiest == this_rq);
2792 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2795 if (busiest->nr_running > 1) {
2796 /* Attempt to move tasks */
2797 double_lock_balance(this_rq, busiest);
2798 /* this_rq->clock is already updated */
2799 update_rq_clock(busiest);
2800 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2801 imbalance, sd, CPU_NEWLY_IDLE,
2803 spin_unlock(&busiest->lock);
2805 if (unlikely(all_pinned)) {
2806 cpu_clear(cpu_of(busiest), cpus);
2807 if (!cpus_empty(cpus))
2813 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2814 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2815 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2818 sd->nr_balance_failed = 0;
2823 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2824 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2825 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2827 sd->nr_balance_failed = 0;
2833 * idle_balance is called by schedule() if this_cpu is about to become
2834 * idle. Attempts to pull tasks from other CPUs.
2836 static void idle_balance(int this_cpu, struct rq *this_rq)
2838 struct sched_domain *sd;
2839 int pulled_task = -1;
2840 unsigned long next_balance = jiffies + HZ;
2842 for_each_domain(this_cpu, sd) {
2843 unsigned long interval;
2845 if (!(sd->flags & SD_LOAD_BALANCE))
2848 if (sd->flags & SD_BALANCE_NEWIDLE)
2849 /* If we've pulled tasks over stop searching: */
2850 pulled_task = load_balance_newidle(this_cpu,
2853 interval = msecs_to_jiffies(sd->balance_interval);
2854 if (time_after(next_balance, sd->last_balance + interval))
2855 next_balance = sd->last_balance + interval;
2859 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2861 * We are going idle. next_balance may be set based on
2862 * a busy processor. So reset next_balance.
2864 this_rq->next_balance = next_balance;
2869 * active_load_balance is run by migration threads. It pushes running tasks
2870 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2871 * running on each physical CPU where possible, and avoids physical /
2872 * logical imbalances.
2874 * Called with busiest_rq locked.
2876 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2878 int target_cpu = busiest_rq->push_cpu;
2879 struct sched_domain *sd;
2880 struct rq *target_rq;
2882 /* Is there any task to move? */
2883 if (busiest_rq->nr_running <= 1)
2886 target_rq = cpu_rq(target_cpu);
2889 * This condition is "impossible", if it occurs
2890 * we need to fix it. Originally reported by
2891 * Bjorn Helgaas on a 128-cpu setup.
2893 BUG_ON(busiest_rq == target_rq);
2895 /* move a task from busiest_rq to target_rq */
2896 double_lock_balance(busiest_rq, target_rq);
2897 update_rq_clock(busiest_rq);
2898 update_rq_clock(target_rq);
2900 /* Search for an sd spanning us and the target CPU. */
2901 for_each_domain(target_cpu, sd) {
2902 if ((sd->flags & SD_LOAD_BALANCE) &&
2903 cpu_isset(busiest_cpu, sd->span))
2908 schedstat_inc(sd, alb_cnt);
2910 if (move_one_task(target_rq, target_cpu, busiest_rq,
2912 schedstat_inc(sd, alb_pushed);
2914 schedstat_inc(sd, alb_failed);
2916 spin_unlock(&target_rq->lock);
2921 atomic_t load_balancer;
2923 } nohz ____cacheline_aligned = {
2924 .load_balancer = ATOMIC_INIT(-1),
2925 .cpu_mask = CPU_MASK_NONE,
2929 * This routine will try to nominate the ilb (idle load balancing)
2930 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2931 * load balancing on behalf of all those cpus. If all the cpus in the system
2932 * go into this tickless mode, then there will be no ilb owner (as there is
2933 * no need for one) and all the cpus will sleep till the next wakeup event
2936 * For the ilb owner, tick is not stopped. And this tick will be used
2937 * for idle load balancing. ilb owner will still be part of
2940 * While stopping the tick, this cpu will become the ilb owner if there
2941 * is no other owner. And will be the owner till that cpu becomes busy
2942 * or if all cpus in the system stop their ticks at which point
2943 * there is no need for ilb owner.
2945 * When the ilb owner becomes busy, it nominates another owner, during the
2946 * next busy scheduler_tick()
2948 int select_nohz_load_balancer(int stop_tick)
2950 int cpu = smp_processor_id();
2953 cpu_set(cpu, nohz.cpu_mask);
2954 cpu_rq(cpu)->in_nohz_recently = 1;
2957 * If we are going offline and still the leader, give up!
2959 if (cpu_is_offline(cpu) &&
2960 atomic_read(&nohz.load_balancer) == cpu) {
2961 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2966 /* time for ilb owner also to sleep */
2967 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2968 if (atomic_read(&nohz.load_balancer) == cpu)
2969 atomic_set(&nohz.load_balancer, -1);
2973 if (atomic_read(&nohz.load_balancer) == -1) {
2974 /* make me the ilb owner */
2975 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2977 } else if (atomic_read(&nohz.load_balancer) == cpu)
2980 if (!cpu_isset(cpu, nohz.cpu_mask))
2983 cpu_clear(cpu, nohz.cpu_mask);
2985 if (atomic_read(&nohz.load_balancer) == cpu)
2986 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2993 static DEFINE_SPINLOCK(balancing);
2996 * It checks each scheduling domain to see if it is due to be balanced,
2997 * and initiates a balancing operation if so.
2999 * Balancing parameters are set up in arch_init_sched_domains.
3001 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
3004 struct rq *rq = cpu_rq(cpu);
3005 unsigned long interval;
3006 struct sched_domain *sd;
3007 /* Earliest time when we have to do rebalance again */
3008 unsigned long next_balance = jiffies + 60*HZ;
3009 int update_next_balance = 0;
3011 for_each_domain(cpu, sd) {
3012 if (!(sd->flags & SD_LOAD_BALANCE))
3015 interval = sd->balance_interval;
3016 if (idle != CPU_IDLE)
3017 interval *= sd->busy_factor;
3019 /* scale ms to jiffies */
3020 interval = msecs_to_jiffies(interval);
3021 if (unlikely(!interval))
3023 if (interval > HZ*NR_CPUS/10)
3024 interval = HZ*NR_CPUS/10;
3027 if (sd->flags & SD_SERIALIZE) {
3028 if (!spin_trylock(&balancing))
3032 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3033 if (load_balance(cpu, rq, sd, idle, &balance)) {
3035 * We've pulled tasks over so either we're no
3036 * longer idle, or one of our SMT siblings is
3039 idle = CPU_NOT_IDLE;
3041 sd->last_balance = jiffies;
3043 if (sd->flags & SD_SERIALIZE)
3044 spin_unlock(&balancing);
3046 if (time_after(next_balance, sd->last_balance + interval)) {
3047 next_balance = sd->last_balance + interval;
3048 update_next_balance = 1;
3052 * Stop the load balance at this level. There is another
3053 * CPU in our sched group which is doing load balancing more
3061 * next_balance will be updated only when there is a need.
3062 * When the cpu is attached to null domain for ex, it will not be
3065 if (likely(update_next_balance))
3066 rq->next_balance = next_balance;
3070 * run_rebalance_domains is triggered when needed from the scheduler tick.
3071 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3072 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3074 static void run_rebalance_domains(struct softirq_action *h)
3076 int this_cpu = smp_processor_id();
3077 struct rq *this_rq = cpu_rq(this_cpu);
3078 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3079 CPU_IDLE : CPU_NOT_IDLE;
3081 rebalance_domains(this_cpu, idle);
3085 * If this cpu is the owner for idle load balancing, then do the
3086 * balancing on behalf of the other idle cpus whose ticks are
3089 if (this_rq->idle_at_tick &&
3090 atomic_read(&nohz.load_balancer) == this_cpu) {
3091 cpumask_t cpus = nohz.cpu_mask;
3095 cpu_clear(this_cpu, cpus);
3096 for_each_cpu_mask(balance_cpu, cpus) {
3098 * If this cpu gets work to do, stop the load balancing
3099 * work being done for other cpus. Next load
3100 * balancing owner will pick it up.
3105 rebalance_domains(balance_cpu, CPU_IDLE);
3107 rq = cpu_rq(balance_cpu);
3108 if (time_after(this_rq->next_balance, rq->next_balance))
3109 this_rq->next_balance = rq->next_balance;
3116 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3118 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3119 * idle load balancing owner or decide to stop the periodic load balancing,
3120 * if the whole system is idle.
3122 static inline void trigger_load_balance(struct rq *rq, int cpu)
3126 * If we were in the nohz mode recently and busy at the current
3127 * scheduler tick, then check if we need to nominate new idle
3130 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3131 rq->in_nohz_recently = 0;
3133 if (atomic_read(&nohz.load_balancer) == cpu) {
3134 cpu_clear(cpu, nohz.cpu_mask);
3135 atomic_set(&nohz.load_balancer, -1);
3138 if (atomic_read(&nohz.load_balancer) == -1) {
3140 * simple selection for now: Nominate the
3141 * first cpu in the nohz list to be the next
3144 * TBD: Traverse the sched domains and nominate
3145 * the nearest cpu in the nohz.cpu_mask.
3147 int ilb = first_cpu(nohz.cpu_mask);
3155 * If this cpu is idle and doing idle load balancing for all the
3156 * cpus with ticks stopped, is it time for that to stop?
3158 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3159 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3165 * If this cpu is idle and the idle load balancing is done by
3166 * someone else, then no need raise the SCHED_SOFTIRQ
3168 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3169 cpu_isset(cpu, nohz.cpu_mask))
3172 if (time_after_eq(jiffies, rq->next_balance))
3173 raise_softirq(SCHED_SOFTIRQ);
3176 #else /* CONFIG_SMP */
3179 * on UP we do not need to balance between CPUs:
3181 static inline void idle_balance(int cpu, struct rq *rq)
3185 /* Avoid "used but not defined" warning on UP */
3186 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3187 unsigned long max_nr_move, unsigned long max_load_move,
3188 struct sched_domain *sd, enum cpu_idle_type idle,
3189 int *all_pinned, unsigned long *load_moved,
3190 int *this_best_prio, struct rq_iterator *iterator)
3199 DEFINE_PER_CPU(struct kernel_stat, kstat);
3201 EXPORT_PER_CPU_SYMBOL(kstat);
3204 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3205 * that have not yet been banked in case the task is currently running.
3207 unsigned long long task_sched_runtime(struct task_struct *p)
3209 unsigned long flags;
3213 rq = task_rq_lock(p, &flags);
3214 ns = p->se.sum_exec_runtime;
3215 if (rq->curr == p) {
3216 update_rq_clock(rq);
3217 delta_exec = rq->clock - p->se.exec_start;
3218 if ((s64)delta_exec > 0)
3221 task_rq_unlock(rq, &flags);
3227 * Account user cpu time to a process.
3228 * @p: the process that the cpu time gets accounted to
3229 * @hardirq_offset: the offset to subtract from hardirq_count()
3230 * @cputime: the cpu time spent in user space since the last update
3232 void account_user_time(struct task_struct *p, cputime_t cputime)
3234 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3237 p->utime = cputime_add(p->utime, cputime);
3239 /* Add user time to cpustat. */
3240 tmp = cputime_to_cputime64(cputime);
3241 if (TASK_NICE(p) > 0)
3242 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3244 cpustat->user = cputime64_add(cpustat->user, tmp);
3248 * Account system cpu time to a process.
3249 * @p: the process that the cpu time gets accounted to
3250 * @hardirq_offset: the offset to subtract from hardirq_count()
3251 * @cputime: the cpu time spent in kernel space since the last update
3253 void account_system_time(struct task_struct *p, int hardirq_offset,
3256 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3257 struct rq *rq = this_rq();
3260 p->stime = cputime_add(p->stime, cputime);
3262 /* Add system time to cpustat. */
3263 tmp = cputime_to_cputime64(cputime);
3264 if (hardirq_count() - hardirq_offset)
3265 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3266 else if (softirq_count())
3267 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3268 else if (p != rq->idle)
3269 cpustat->system = cputime64_add(cpustat->system, tmp);
3270 else if (atomic_read(&rq->nr_iowait) > 0)
3271 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3273 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3274 /* Account for system time used */
3275 acct_update_integrals(p);
3279 * Account for involuntary wait time.
3280 * @p: the process from which the cpu time has been stolen
3281 * @steal: the cpu time spent in involuntary wait
3283 void account_steal_time(struct task_struct *p, cputime_t steal)
3285 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3286 cputime64_t tmp = cputime_to_cputime64(steal);
3287 struct rq *rq = this_rq();
3289 if (p == rq->idle) {
3290 p->stime = cputime_add(p->stime, steal);
3291 if (atomic_read(&rq->nr_iowait) > 0)
3292 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3294 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3296 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3300 * This function gets called by the timer code, with HZ frequency.
3301 * We call it with interrupts disabled.
3303 * It also gets called by the fork code, when changing the parent's
3306 void scheduler_tick(void)
3308 int cpu = smp_processor_id();
3309 struct rq *rq = cpu_rq(cpu);
3310 struct task_struct *curr = rq->curr;
3311 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3313 spin_lock(&rq->lock);
3314 __update_rq_clock(rq);
3316 * Let rq->clock advance by at least TICK_NSEC:
3318 if (unlikely(rq->clock < next_tick))
3319 rq->clock = next_tick;
3320 rq->tick_timestamp = rq->clock;
3321 update_cpu_load(rq);
3322 if (curr != rq->idle) /* FIXME: needed? */
3323 curr->sched_class->task_tick(rq, curr);
3324 spin_unlock(&rq->lock);
3327 rq->idle_at_tick = idle_cpu(cpu);
3328 trigger_load_balance(rq, cpu);
3332 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3334 void fastcall add_preempt_count(int val)
3339 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3341 preempt_count() += val;
3343 * Spinlock count overflowing soon?
3345 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3348 EXPORT_SYMBOL(add_preempt_count);
3350 void fastcall sub_preempt_count(int val)
3355 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3358 * Is the spinlock portion underflowing?
3360 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3361 !(preempt_count() & PREEMPT_MASK)))
3364 preempt_count() -= val;
3366 EXPORT_SYMBOL(sub_preempt_count);
3371 * Print scheduling while atomic bug:
3373 static noinline void __schedule_bug(struct task_struct *prev)
3375 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3376 prev->comm, preempt_count(), prev->pid);
3377 debug_show_held_locks(prev);
3378 if (irqs_disabled())
3379 print_irqtrace_events(prev);
3384 * Various schedule()-time debugging checks and statistics:
3386 static inline void schedule_debug(struct task_struct *prev)
3389 * Test if we are atomic. Since do_exit() needs to call into
3390 * schedule() atomically, we ignore that path for now.
3391 * Otherwise, whine if we are scheduling when we should not be.
3393 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3394 __schedule_bug(prev);
3396 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3398 schedstat_inc(this_rq(), sched_cnt);
3402 * Pick up the highest-prio task:
3404 static inline struct task_struct *
3405 pick_next_task(struct rq *rq, struct task_struct *prev)
3407 struct sched_class *class;
3408 struct task_struct *p;
3411 * Optimization: we know that if all tasks are in
3412 * the fair class we can call that function directly:
3414 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3415 p = fair_sched_class.pick_next_task(rq);
3420 class = sched_class_highest;
3422 p = class->pick_next_task(rq);
3426 * Will never be NULL as the idle class always
3427 * returns a non-NULL p:
3429 class = class->next;
3434 * schedule() is the main scheduler function.
3436 asmlinkage void __sched schedule(void)
3438 struct task_struct *prev, *next;
3445 cpu = smp_processor_id();
3449 switch_count = &prev->nivcsw;
3451 release_kernel_lock(prev);
3452 need_resched_nonpreemptible:
3454 schedule_debug(prev);
3456 spin_lock_irq(&rq->lock);
3457 clear_tsk_need_resched(prev);
3458 __update_rq_clock(rq);
3460 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3461 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3462 unlikely(signal_pending(prev)))) {
3463 prev->state = TASK_RUNNING;
3465 deactivate_task(rq, prev, 1);
3467 switch_count = &prev->nvcsw;
3470 if (unlikely(!rq->nr_running))
3471 idle_balance(cpu, rq);
3473 prev->sched_class->put_prev_task(rq, prev);
3474 next = pick_next_task(rq, prev);
3476 sched_info_switch(prev, next);
3478 if (likely(prev != next)) {
3483 context_switch(rq, prev, next); /* unlocks the rq */
3485 spin_unlock_irq(&rq->lock);
3487 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3488 cpu = smp_processor_id();
3490 goto need_resched_nonpreemptible;
3492 preempt_enable_no_resched();
3493 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3496 EXPORT_SYMBOL(schedule);
3498 #ifdef CONFIG_PREEMPT
3500 * this is the entry point to schedule() from in-kernel preemption
3501 * off of preempt_enable. Kernel preemptions off return from interrupt
3502 * occur there and call schedule directly.
3504 asmlinkage void __sched preempt_schedule(void)
3506 struct thread_info *ti = current_thread_info();
3507 #ifdef CONFIG_PREEMPT_BKL
3508 struct task_struct *task = current;
3509 int saved_lock_depth;
3512 * If there is a non-zero preempt_count or interrupts are disabled,
3513 * we do not want to preempt the current task. Just return..
3515 if (likely(ti->preempt_count || irqs_disabled()))
3519 add_preempt_count(PREEMPT_ACTIVE);
3521 * We keep the big kernel semaphore locked, but we
3522 * clear ->lock_depth so that schedule() doesnt
3523 * auto-release the semaphore:
3525 #ifdef CONFIG_PREEMPT_BKL
3526 saved_lock_depth = task->lock_depth;
3527 task->lock_depth = -1;
3530 #ifdef CONFIG_PREEMPT_BKL
3531 task->lock_depth = saved_lock_depth;
3533 sub_preempt_count(PREEMPT_ACTIVE);
3535 /* we could miss a preemption opportunity between schedule and now */
3537 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3540 EXPORT_SYMBOL(preempt_schedule);
3543 * this is the entry point to schedule() from kernel preemption
3544 * off of irq context.
3545 * Note, that this is called and return with irqs disabled. This will
3546 * protect us against recursive calling from irq.
3548 asmlinkage void __sched preempt_schedule_irq(void)
3550 struct thread_info *ti = current_thread_info();
3551 #ifdef CONFIG_PREEMPT_BKL
3552 struct task_struct *task = current;
3553 int saved_lock_depth;
3555 /* Catch callers which need to be fixed */
3556 BUG_ON(ti->preempt_count || !irqs_disabled());
3559 add_preempt_count(PREEMPT_ACTIVE);
3561 * We keep the big kernel semaphore locked, but we
3562 * clear ->lock_depth so that schedule() doesnt
3563 * auto-release the semaphore:
3565 #ifdef CONFIG_PREEMPT_BKL
3566 saved_lock_depth = task->lock_depth;
3567 task->lock_depth = -1;
3571 local_irq_disable();
3572 #ifdef CONFIG_PREEMPT_BKL
3573 task->lock_depth = saved_lock_depth;
3575 sub_preempt_count(PREEMPT_ACTIVE);
3577 /* we could miss a preemption opportunity between schedule and now */
3579 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3583 #endif /* CONFIG_PREEMPT */
3585 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3588 return try_to_wake_up(curr->private, mode, sync);
3590 EXPORT_SYMBOL(default_wake_function);
3593 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3594 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3595 * number) then we wake all the non-exclusive tasks and one exclusive task.
3597 * There are circumstances in which we can try to wake a task which has already
3598 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3599 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3601 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3602 int nr_exclusive, int sync, void *key)
3604 wait_queue_t *curr, *next;
3606 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3607 unsigned flags = curr->flags;
3609 if (curr->func(curr, mode, sync, key) &&
3610 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3616 * __wake_up - wake up threads blocked on a waitqueue.
3618 * @mode: which threads
3619 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3620 * @key: is directly passed to the wakeup function
3622 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3623 int nr_exclusive, void *key)
3625 unsigned long flags;
3627 spin_lock_irqsave(&q->lock, flags);
3628 __wake_up_common(q, mode, nr_exclusive, 0, key);
3629 spin_unlock_irqrestore(&q->lock, flags);
3631 EXPORT_SYMBOL(__wake_up);
3634 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3636 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3638 __wake_up_common(q, mode, 1, 0, NULL);
3642 * __wake_up_sync - wake up threads blocked on a waitqueue.
3644 * @mode: which threads
3645 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3647 * The sync wakeup differs that the waker knows that it will schedule
3648 * away soon, so while the target thread will be woken up, it will not
3649 * be migrated to another CPU - ie. the two threads are 'synchronized'
3650 * with each other. This can prevent needless bouncing between CPUs.
3652 * On UP it can prevent extra preemption.
3655 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3657 unsigned long flags;
3663 if (unlikely(!nr_exclusive))
3666 spin_lock_irqsave(&q->lock, flags);
3667 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3668 spin_unlock_irqrestore(&q->lock, flags);
3670 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3672 void fastcall complete(struct completion *x)
3674 unsigned long flags;
3676 spin_lock_irqsave(&x->wait.lock, flags);
3678 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3680 spin_unlock_irqrestore(&x->wait.lock, flags);
3682 EXPORT_SYMBOL(complete);
3684 void fastcall complete_all(struct completion *x)
3686 unsigned long flags;
3688 spin_lock_irqsave(&x->wait.lock, flags);
3689 x->done += UINT_MAX/2;
3690 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3692 spin_unlock_irqrestore(&x->wait.lock, flags);
3694 EXPORT_SYMBOL(complete_all);
3696 void fastcall __sched wait_for_completion(struct completion *x)
3700 spin_lock_irq(&x->wait.lock);
3702 DECLARE_WAITQUEUE(wait, current);
3704 wait.flags |= WQ_FLAG_EXCLUSIVE;
3705 __add_wait_queue_tail(&x->wait, &wait);
3707 __set_current_state(TASK_UNINTERRUPTIBLE);
3708 spin_unlock_irq(&x->wait.lock);
3710 spin_lock_irq(&x->wait.lock);
3712 __remove_wait_queue(&x->wait, &wait);
3715 spin_unlock_irq(&x->wait.lock);
3717 EXPORT_SYMBOL(wait_for_completion);
3719 unsigned long fastcall __sched
3720 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3724 spin_lock_irq(&x->wait.lock);
3726 DECLARE_WAITQUEUE(wait, current);
3728 wait.flags |= WQ_FLAG_EXCLUSIVE;
3729 __add_wait_queue_tail(&x->wait, &wait);
3731 __set_current_state(TASK_UNINTERRUPTIBLE);
3732 spin_unlock_irq(&x->wait.lock);
3733 timeout = schedule_timeout(timeout);
3734 spin_lock_irq(&x->wait.lock);
3736 __remove_wait_queue(&x->wait, &wait);
3740 __remove_wait_queue(&x->wait, &wait);
3744 spin_unlock_irq(&x->wait.lock);
3747 EXPORT_SYMBOL(wait_for_completion_timeout);
3749 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3755 spin_lock_irq(&x->wait.lock);
3757 DECLARE_WAITQUEUE(wait, current);
3759 wait.flags |= WQ_FLAG_EXCLUSIVE;
3760 __add_wait_queue_tail(&x->wait, &wait);
3762 if (signal_pending(current)) {
3764 __remove_wait_queue(&x->wait, &wait);
3767 __set_current_state(TASK_INTERRUPTIBLE);
3768 spin_unlock_irq(&x->wait.lock);
3770 spin_lock_irq(&x->wait.lock);
3772 __remove_wait_queue(&x->wait, &wait);
3776 spin_unlock_irq(&x->wait.lock);
3780 EXPORT_SYMBOL(wait_for_completion_interruptible);
3782 unsigned long fastcall __sched
3783 wait_for_completion_interruptible_timeout(struct completion *x,
3784 unsigned long timeout)
3788 spin_lock_irq(&x->wait.lock);
3790 DECLARE_WAITQUEUE(wait, current);
3792 wait.flags |= WQ_FLAG_EXCLUSIVE;
3793 __add_wait_queue_tail(&x->wait, &wait);
3795 if (signal_pending(current)) {
3796 timeout = -ERESTARTSYS;
3797 __remove_wait_queue(&x->wait, &wait);
3800 __set_current_state(TASK_INTERRUPTIBLE);
3801 spin_unlock_irq(&x->wait.lock);
3802 timeout = schedule_timeout(timeout);
3803 spin_lock_irq(&x->wait.lock);
3805 __remove_wait_queue(&x->wait, &wait);
3809 __remove_wait_queue(&x->wait, &wait);
3813 spin_unlock_irq(&x->wait.lock);
3816 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3819 sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3821 spin_lock_irqsave(&q->lock, *flags);
3822 __add_wait_queue(q, wait);
3823 spin_unlock(&q->lock);
3827 sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3829 spin_lock_irq(&q->lock);
3830 __remove_wait_queue(q, wait);
3831 spin_unlock_irqrestore(&q->lock, *flags);
3834 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3836 unsigned long flags;
3839 init_waitqueue_entry(&wait, current);
3841 current->state = TASK_INTERRUPTIBLE;
3843 sleep_on_head(q, &wait, &flags);
3845 sleep_on_tail(q, &wait, &flags);
3847 EXPORT_SYMBOL(interruptible_sleep_on);
3850 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3852 unsigned long flags;
3855 init_waitqueue_entry(&wait, current);
3857 current->state = TASK_INTERRUPTIBLE;
3859 sleep_on_head(q, &wait, &flags);
3860 timeout = schedule_timeout(timeout);
3861 sleep_on_tail(q, &wait, &flags);
3865 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3867 void __sched sleep_on(wait_queue_head_t *q)
3869 unsigned long flags;
3872 init_waitqueue_entry(&wait, current);
3874 current->state = TASK_UNINTERRUPTIBLE;
3876 sleep_on_head(q, &wait, &flags);
3878 sleep_on_tail(q, &wait, &flags);
3880 EXPORT_SYMBOL(sleep_on);
3882 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3884 unsigned long flags;
3887 init_waitqueue_entry(&wait, current);
3889 current->state = TASK_UNINTERRUPTIBLE;
3891 sleep_on_head(q, &wait, &flags);
3892 timeout = schedule_timeout(timeout);
3893 sleep_on_tail(q, &wait, &flags);
3897 EXPORT_SYMBOL(sleep_on_timeout);
3899 #ifdef CONFIG_RT_MUTEXES
3902 * rt_mutex_setprio - set the current priority of a task
3904 * @prio: prio value (kernel-internal form)
3906 * This function changes the 'effective' priority of a task. It does
3907 * not touch ->normal_prio like __setscheduler().
3909 * Used by the rt_mutex code to implement priority inheritance logic.
3911 void rt_mutex_setprio(struct task_struct *p, int prio)
3913 unsigned long flags;
3917 BUG_ON(prio < 0 || prio > MAX_PRIO);
3919 rq = task_rq_lock(p, &flags);
3920 update_rq_clock(rq);
3923 on_rq = p->se.on_rq;
3925 dequeue_task(rq, p, 0);
3928 p->sched_class = &rt_sched_class;
3930 p->sched_class = &fair_sched_class;
3935 enqueue_task(rq, p, 0);
3937 * Reschedule if we are currently running on this runqueue and
3938 * our priority decreased, or if we are not currently running on
3939 * this runqueue and our priority is higher than the current's
3941 if (task_running(rq, p)) {
3942 if (p->prio > oldprio)
3943 resched_task(rq->curr);
3945 check_preempt_curr(rq, p);
3948 task_rq_unlock(rq, &flags);
3953 void set_user_nice(struct task_struct *p, long nice)
3955 int old_prio, delta, on_rq;
3956 unsigned long flags;
3959 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3962 * We have to be careful, if called from sys_setpriority(),
3963 * the task might be in the middle of scheduling on another CPU.
3965 rq = task_rq_lock(p, &flags);
3966 update_rq_clock(rq);
3968 * The RT priorities are set via sched_setscheduler(), but we still
3969 * allow the 'normal' nice value to be set - but as expected
3970 * it wont have any effect on scheduling until the task is
3971 * SCHED_FIFO/SCHED_RR:
3973 if (task_has_rt_policy(p)) {
3974 p->static_prio = NICE_TO_PRIO(nice);
3977 on_rq = p->se.on_rq;
3979 dequeue_task(rq, p, 0);
3983 p->static_prio = NICE_TO_PRIO(nice);
3986 p->prio = effective_prio(p);
3987 delta = p->prio - old_prio;
3990 enqueue_task(rq, p, 0);
3993 * If the task increased its priority or is running and
3994 * lowered its priority, then reschedule its CPU:
3996 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3997 resched_task(rq->curr);
4000 task_rq_unlock(rq, &flags);
4002 EXPORT_SYMBOL(set_user_nice);
4005 * can_nice - check if a task can reduce its nice value
4009 int can_nice(const struct task_struct *p, const int nice)
4011 /* convert nice value [19,-20] to rlimit style value [1,40] */
4012 int nice_rlim = 20 - nice;
4014 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4015 capable(CAP_SYS_NICE));
4018 #ifdef __ARCH_WANT_SYS_NICE
4021 * sys_nice - change the priority of the current process.
4022 * @increment: priority increment
4024 * sys_setpriority is a more generic, but much slower function that
4025 * does similar things.
4027 asmlinkage long sys_nice(int increment)
4032 * Setpriority might change our priority at the same moment.
4033 * We don't have to worry. Conceptually one call occurs first
4034 * and we have a single winner.
4036 if (increment < -40)
4041 nice = PRIO_TO_NICE(current->static_prio) + increment;
4047 if (increment < 0 && !can_nice(current, nice))
4050 retval = security_task_setnice(current, nice);
4054 set_user_nice(current, nice);
4061 * task_prio - return the priority value of a given task.
4062 * @p: the task in question.
4064 * This is the priority value as seen by users in /proc.
4065 * RT tasks are offset by -200. Normal tasks are centered
4066 * around 0, value goes from -16 to +15.
4068 int task_prio(const struct task_struct *p)
4070 return p->prio - MAX_RT_PRIO;
4074 * task_nice - return the nice value of a given task.
4075 * @p: the task in question.
4077 int task_nice(const struct task_struct *p)
4079 return TASK_NICE(p);
4081 EXPORT_SYMBOL_GPL(task_nice);
4084 * idle_cpu - is a given cpu idle currently?
4085 * @cpu: the processor in question.
4087 int idle_cpu(int cpu)
4089 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4093 * idle_task - return the idle task for a given cpu.
4094 * @cpu: the processor in question.
4096 struct task_struct *idle_task(int cpu)
4098 return cpu_rq(cpu)->idle;
4102 * find_process_by_pid - find a process with a matching PID value.
4103 * @pid: the pid in question.
4105 static inline struct task_struct *find_process_by_pid(pid_t pid)
4107 return pid ? find_task_by_pid(pid) : current;
4110 /* Actually do priority change: must hold rq lock. */
4112 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4114 BUG_ON(p->se.on_rq);
4117 switch (p->policy) {
4121 p->sched_class = &fair_sched_class;
4125 p->sched_class = &rt_sched_class;
4129 p->rt_priority = prio;
4130 p->normal_prio = normal_prio(p);
4131 /* we are holding p->pi_lock already */
4132 p->prio = rt_mutex_getprio(p);
4137 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4138 * @p: the task in question.
4139 * @policy: new policy.
4140 * @param: structure containing the new RT priority.
4142 * NOTE that the task may be already dead.
4144 int sched_setscheduler(struct task_struct *p, int policy,
4145 struct sched_param *param)
4147 int retval, oldprio, oldpolicy = -1, on_rq;
4148 unsigned long flags;
4151 /* may grab non-irq protected spin_locks */
4152 BUG_ON(in_interrupt());
4154 /* double check policy once rq lock held */
4156 policy = oldpolicy = p->policy;
4157 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4158 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4159 policy != SCHED_IDLE)
4162 * Valid priorities for SCHED_FIFO and SCHED_RR are
4163 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4164 * SCHED_BATCH and SCHED_IDLE is 0.
4166 if (param->sched_priority < 0 ||
4167 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4168 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4170 if (rt_policy(policy) != (param->sched_priority != 0))
4174 * Allow unprivileged RT tasks to decrease priority:
4176 if (!capable(CAP_SYS_NICE)) {
4177 if (rt_policy(policy)) {
4178 unsigned long rlim_rtprio;
4180 if (!lock_task_sighand(p, &flags))
4182 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4183 unlock_task_sighand(p, &flags);
4185 /* can't set/change the rt policy */
4186 if (policy != p->policy && !rlim_rtprio)
4189 /* can't increase priority */
4190 if (param->sched_priority > p->rt_priority &&
4191 param->sched_priority > rlim_rtprio)
4195 * Like positive nice levels, dont allow tasks to
4196 * move out of SCHED_IDLE either:
4198 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4201 /* can't change other user's priorities */
4202 if ((current->euid != p->euid) &&
4203 (current->euid != p->uid))
4207 retval = security_task_setscheduler(p, policy, param);
4211 * make sure no PI-waiters arrive (or leave) while we are
4212 * changing the priority of the task:
4214 spin_lock_irqsave(&p->pi_lock, flags);
4216 * To be able to change p->policy safely, the apropriate
4217 * runqueue lock must be held.
4219 rq = __task_rq_lock(p);
4220 /* recheck policy now with rq lock held */
4221 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4222 policy = oldpolicy = -1;
4223 __task_rq_unlock(rq);
4224 spin_unlock_irqrestore(&p->pi_lock, flags);
4227 update_rq_clock(rq);
4228 on_rq = p->se.on_rq;
4230 deactivate_task(rq, p, 0);
4232 __setscheduler(rq, p, policy, param->sched_priority);
4234 activate_task(rq, p, 0);
4236 * Reschedule if we are currently running on this runqueue and
4237 * our priority decreased, or if we are not currently running on
4238 * this runqueue and our priority is higher than the current's
4240 if (task_running(rq, p)) {
4241 if (p->prio > oldprio)
4242 resched_task(rq->curr);
4244 check_preempt_curr(rq, p);
4247 __task_rq_unlock(rq);
4248 spin_unlock_irqrestore(&p->pi_lock, flags);
4250 rt_mutex_adjust_pi(p);
4254 EXPORT_SYMBOL_GPL(sched_setscheduler);
4257 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4259 struct sched_param lparam;
4260 struct task_struct *p;
4263 if (!param || pid < 0)
4265 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4270 p = find_process_by_pid(pid);
4272 retval = sched_setscheduler(p, policy, &lparam);
4279 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4280 * @pid: the pid in question.
4281 * @policy: new policy.
4282 * @param: structure containing the new RT priority.
4284 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4285 struct sched_param __user *param)
4287 /* negative values for policy are not valid */
4291 return do_sched_setscheduler(pid, policy, param);
4295 * sys_sched_setparam - set/change the RT priority of a thread
4296 * @pid: the pid in question.
4297 * @param: structure containing the new RT priority.
4299 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4301 return do_sched_setscheduler(pid, -1, param);
4305 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4306 * @pid: the pid in question.
4308 asmlinkage long sys_sched_getscheduler(pid_t pid)
4310 struct task_struct *p;
4311 int retval = -EINVAL;
4317 read_lock(&tasklist_lock);
4318 p = find_process_by_pid(pid);
4320 retval = security_task_getscheduler(p);
4324 read_unlock(&tasklist_lock);
4331 * sys_sched_getscheduler - get the RT priority of a thread
4332 * @pid: the pid in question.
4333 * @param: structure containing the RT priority.
4335 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4337 struct sched_param lp;
4338 struct task_struct *p;
4339 int retval = -EINVAL;
4341 if (!param || pid < 0)
4344 read_lock(&tasklist_lock);
4345 p = find_process_by_pid(pid);
4350 retval = security_task_getscheduler(p);
4354 lp.sched_priority = p->rt_priority;
4355 read_unlock(&tasklist_lock);
4358 * This one might sleep, we cannot do it with a spinlock held ...
4360 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4366 read_unlock(&tasklist_lock);
4370 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4372 cpumask_t cpus_allowed;
4373 struct task_struct *p;
4376 mutex_lock(&sched_hotcpu_mutex);
4377 read_lock(&tasklist_lock);
4379 p = find_process_by_pid(pid);
4381 read_unlock(&tasklist_lock);
4382 mutex_unlock(&sched_hotcpu_mutex);
4387 * It is not safe to call set_cpus_allowed with the
4388 * tasklist_lock held. We will bump the task_struct's
4389 * usage count and then drop tasklist_lock.
4392 read_unlock(&tasklist_lock);
4395 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4396 !capable(CAP_SYS_NICE))
4399 retval = security_task_setscheduler(p, 0, NULL);
4403 cpus_allowed = cpuset_cpus_allowed(p);
4404 cpus_and(new_mask, new_mask, cpus_allowed);
4405 retval = set_cpus_allowed(p, new_mask);
4409 mutex_unlock(&sched_hotcpu_mutex);
4413 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4414 cpumask_t *new_mask)
4416 if (len < sizeof(cpumask_t)) {
4417 memset(new_mask, 0, sizeof(cpumask_t));
4418 } else if (len > sizeof(cpumask_t)) {
4419 len = sizeof(cpumask_t);
4421 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4425 * sys_sched_setaffinity - set the cpu affinity of a process
4426 * @pid: pid of the process
4427 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4428 * @user_mask_ptr: user-space pointer to the new cpu mask
4430 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4431 unsigned long __user *user_mask_ptr)
4436 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4440 return sched_setaffinity(pid, new_mask);
4444 * Represents all cpu's present in the system
4445 * In systems capable of hotplug, this map could dynamically grow
4446 * as new cpu's are detected in the system via any platform specific
4447 * method, such as ACPI for e.g.
4450 cpumask_t cpu_present_map __read_mostly;
4451 EXPORT_SYMBOL(cpu_present_map);
4454 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4455 EXPORT_SYMBOL(cpu_online_map);
4457 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4458 EXPORT_SYMBOL(cpu_possible_map);
4461 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4463 struct task_struct *p;
4466 mutex_lock(&sched_hotcpu_mutex);
4467 read_lock(&tasklist_lock);
4470 p = find_process_by_pid(pid);
4474 retval = security_task_getscheduler(p);
4478 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4481 read_unlock(&tasklist_lock);
4482 mutex_unlock(&sched_hotcpu_mutex);
4488 * sys_sched_getaffinity - get the cpu affinity of a process
4489 * @pid: pid of the process
4490 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4491 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4493 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4494 unsigned long __user *user_mask_ptr)
4499 if (len < sizeof(cpumask_t))
4502 ret = sched_getaffinity(pid, &mask);
4506 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4509 return sizeof(cpumask_t);
4513 * sys_sched_yield - yield the current processor to other threads.
4515 * This function yields the current CPU to other tasks. If there are no
4516 * other threads running on this CPU then this function will return.
4518 asmlinkage long sys_sched_yield(void)
4520 struct rq *rq = this_rq_lock();
4522 schedstat_inc(rq, yld_cnt);
4523 current->sched_class->yield_task(rq, current);
4526 * Since we are going to call schedule() anyway, there's
4527 * no need to preempt or enable interrupts:
4529 __release(rq->lock);
4530 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4531 _raw_spin_unlock(&rq->lock);
4532 preempt_enable_no_resched();
4539 static void __cond_resched(void)
4541 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4542 __might_sleep(__FILE__, __LINE__);
4545 * The BKS might be reacquired before we have dropped
4546 * PREEMPT_ACTIVE, which could trigger a second
4547 * cond_resched() call.
4550 add_preempt_count(PREEMPT_ACTIVE);
4552 sub_preempt_count(PREEMPT_ACTIVE);
4553 } while (need_resched());
4556 int __sched cond_resched(void)
4558 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4559 system_state == SYSTEM_RUNNING) {
4565 EXPORT_SYMBOL(cond_resched);
4568 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4569 * call schedule, and on return reacquire the lock.
4571 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4572 * operations here to prevent schedule() from being called twice (once via
4573 * spin_unlock(), once by hand).
4575 int cond_resched_lock(spinlock_t *lock)
4579 if (need_lockbreak(lock)) {
4585 if (need_resched() && system_state == SYSTEM_RUNNING) {
4586 spin_release(&lock->dep_map, 1, _THIS_IP_);
4587 _raw_spin_unlock(lock);
4588 preempt_enable_no_resched();
4595 EXPORT_SYMBOL(cond_resched_lock);
4597 int __sched cond_resched_softirq(void)
4599 BUG_ON(!in_softirq());
4601 if (need_resched() && system_state == SYSTEM_RUNNING) {
4609 EXPORT_SYMBOL(cond_resched_softirq);
4612 * yield - yield the current processor to other threads.
4614 * This is a shortcut for kernel-space yielding - it marks the
4615 * thread runnable and calls sys_sched_yield().
4617 void __sched yield(void)
4619 set_current_state(TASK_RUNNING);
4622 EXPORT_SYMBOL(yield);
4625 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4626 * that process accounting knows that this is a task in IO wait state.
4628 * But don't do that if it is a deliberate, throttling IO wait (this task
4629 * has set its backing_dev_info: the queue against which it should throttle)
4631 void __sched io_schedule(void)
4633 struct rq *rq = &__raw_get_cpu_var(runqueues);
4635 delayacct_blkio_start();
4636 atomic_inc(&rq->nr_iowait);
4638 atomic_dec(&rq->nr_iowait);
4639 delayacct_blkio_end();
4641 EXPORT_SYMBOL(io_schedule);
4643 long __sched io_schedule_timeout(long timeout)
4645 struct rq *rq = &__raw_get_cpu_var(runqueues);
4648 delayacct_blkio_start();
4649 atomic_inc(&rq->nr_iowait);
4650 ret = schedule_timeout(timeout);
4651 atomic_dec(&rq->nr_iowait);
4652 delayacct_blkio_end();
4657 * sys_sched_get_priority_max - return maximum RT priority.
4658 * @policy: scheduling class.
4660 * this syscall returns the maximum rt_priority that can be used
4661 * by a given scheduling class.
4663 asmlinkage long sys_sched_get_priority_max(int policy)
4670 ret = MAX_USER_RT_PRIO-1;
4682 * sys_sched_get_priority_min - return minimum RT priority.
4683 * @policy: scheduling class.
4685 * this syscall returns the minimum rt_priority that can be used
4686 * by a given scheduling class.
4688 asmlinkage long sys_sched_get_priority_min(int policy)
4706 * sys_sched_rr_get_interval - return the default timeslice of a process.
4707 * @pid: pid of the process.
4708 * @interval: userspace pointer to the timeslice value.
4710 * this syscall writes the default timeslice value of a given process
4711 * into the user-space timespec buffer. A value of '0' means infinity.
4714 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4716 struct task_struct *p;
4717 int retval = -EINVAL;
4724 read_lock(&tasklist_lock);
4725 p = find_process_by_pid(pid);
4729 retval = security_task_getscheduler(p);
4733 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4734 0 : static_prio_timeslice(p->static_prio), &t);
4735 read_unlock(&tasklist_lock);
4736 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4740 read_unlock(&tasklist_lock);
4744 static const char stat_nam[] = "RSDTtZX";
4746 static void show_task(struct task_struct *p)
4748 unsigned long free = 0;
4751 state = p->state ? __ffs(p->state) + 1 : 0;
4752 printk("%-13.13s %c", p->comm,
4753 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4754 #if BITS_PER_LONG == 32
4755 if (state == TASK_RUNNING)
4756 printk(" running ");
4758 printk(" %08lx ", thread_saved_pc(p));
4760 if (state == TASK_RUNNING)
4761 printk(" running task ");
4763 printk(" %016lx ", thread_saved_pc(p));
4765 #ifdef CONFIG_DEBUG_STACK_USAGE
4767 unsigned long *n = end_of_stack(p);
4770 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4773 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4775 if (state != TASK_RUNNING)
4776 show_stack(p, NULL);
4779 void show_state_filter(unsigned long state_filter)
4781 struct task_struct *g, *p;
4783 #if BITS_PER_LONG == 32
4785 " task PC stack pid father\n");
4788 " task PC stack pid father\n");
4790 read_lock(&tasklist_lock);
4791 do_each_thread(g, p) {
4793 * reset the NMI-timeout, listing all files on a slow
4794 * console might take alot of time:
4796 touch_nmi_watchdog();
4797 if (!state_filter || (p->state & state_filter))
4799 } while_each_thread(g, p);
4801 touch_all_softlockup_watchdogs();
4803 #ifdef CONFIG_SCHED_DEBUG
4804 sysrq_sched_debug_show();
4806 read_unlock(&tasklist_lock);
4808 * Only show locks if all tasks are dumped:
4810 if (state_filter == -1)
4811 debug_show_all_locks();
4814 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4816 idle->sched_class = &idle_sched_class;
4820 * init_idle - set up an idle thread for a given CPU
4821 * @idle: task in question
4822 * @cpu: cpu the idle task belongs to
4824 * NOTE: this function does not set the idle thread's NEED_RESCHED
4825 * flag, to make booting more robust.
4827 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4829 struct rq *rq = cpu_rq(cpu);
4830 unsigned long flags;
4833 idle->se.exec_start = sched_clock();
4835 idle->prio = idle->normal_prio = MAX_PRIO;
4836 idle->cpus_allowed = cpumask_of_cpu(cpu);
4837 __set_task_cpu(idle, cpu);
4839 spin_lock_irqsave(&rq->lock, flags);
4840 rq->curr = rq->idle = idle;
4841 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4844 spin_unlock_irqrestore(&rq->lock, flags);
4846 /* Set the preempt count _outside_ the spinlocks! */
4847 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4848 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4850 task_thread_info(idle)->preempt_count = 0;
4853 * The idle tasks have their own, simple scheduling class:
4855 idle->sched_class = &idle_sched_class;
4859 * In a system that switches off the HZ timer nohz_cpu_mask
4860 * indicates which cpus entered this state. This is used
4861 * in the rcu update to wait only for active cpus. For system
4862 * which do not switch off the HZ timer nohz_cpu_mask should
4863 * always be CPU_MASK_NONE.
4865 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4869 * This is how migration works:
4871 * 1) we queue a struct migration_req structure in the source CPU's
4872 * runqueue and wake up that CPU's migration thread.
4873 * 2) we down() the locked semaphore => thread blocks.
4874 * 3) migration thread wakes up (implicitly it forces the migrated
4875 * thread off the CPU)
4876 * 4) it gets the migration request and checks whether the migrated
4877 * task is still in the wrong runqueue.
4878 * 5) if it's in the wrong runqueue then the migration thread removes
4879 * it and puts it into the right queue.
4880 * 6) migration thread up()s the semaphore.
4881 * 7) we wake up and the migration is done.
4885 * Change a given task's CPU affinity. Migrate the thread to a
4886 * proper CPU and schedule it away if the CPU it's executing on
4887 * is removed from the allowed bitmask.
4889 * NOTE: the caller must have a valid reference to the task, the
4890 * task must not exit() & deallocate itself prematurely. The
4891 * call is not atomic; no spinlocks may be held.
4893 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4895 struct migration_req req;
4896 unsigned long flags;
4900 rq = task_rq_lock(p, &flags);
4901 if (!cpus_intersects(new_mask, cpu_online_map)) {
4906 p->cpus_allowed = new_mask;
4907 /* Can the task run on the task's current CPU? If so, we're done */
4908 if (cpu_isset(task_cpu(p), new_mask))
4911 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4912 /* Need help from migration thread: drop lock and wait. */
4913 task_rq_unlock(rq, &flags);
4914 wake_up_process(rq->migration_thread);
4915 wait_for_completion(&req.done);
4916 tlb_migrate_finish(p->mm);
4920 task_rq_unlock(rq, &flags);
4924 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4927 * Move (not current) task off this cpu, onto dest cpu. We're doing
4928 * this because either it can't run here any more (set_cpus_allowed()
4929 * away from this CPU, or CPU going down), or because we're
4930 * attempting to rebalance this task on exec (sched_exec).
4932 * So we race with normal scheduler movements, but that's OK, as long
4933 * as the task is no longer on this CPU.
4935 * Returns non-zero if task was successfully migrated.
4937 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4939 struct rq *rq_dest, *rq_src;
4942 if (unlikely(cpu_is_offline(dest_cpu)))
4945 rq_src = cpu_rq(src_cpu);
4946 rq_dest = cpu_rq(dest_cpu);
4948 double_rq_lock(rq_src, rq_dest);
4949 /* Already moved. */
4950 if (task_cpu(p) != src_cpu)
4952 /* Affinity changed (again). */
4953 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4956 on_rq = p->se.on_rq;
4958 deactivate_task(rq_src, p, 0);
4960 set_task_cpu(p, dest_cpu);
4962 activate_task(rq_dest, p, 0);
4963 check_preempt_curr(rq_dest, p);
4967 double_rq_unlock(rq_src, rq_dest);
4972 * migration_thread - this is a highprio system thread that performs
4973 * thread migration by bumping thread off CPU then 'pushing' onto
4976 static int migration_thread(void *data)
4978 int cpu = (long)data;
4982 BUG_ON(rq->migration_thread != current);
4984 set_current_state(TASK_INTERRUPTIBLE);
4985 while (!kthread_should_stop()) {
4986 struct migration_req *req;
4987 struct list_head *head;
4989 spin_lock_irq(&rq->lock);
4991 if (cpu_is_offline(cpu)) {
4992 spin_unlock_irq(&rq->lock);
4996 if (rq->active_balance) {
4997 active_load_balance(rq, cpu);
4998 rq->active_balance = 0;
5001 head = &rq->migration_queue;
5003 if (list_empty(head)) {
5004 spin_unlock_irq(&rq->lock);
5006 set_current_state(TASK_INTERRUPTIBLE);
5009 req = list_entry(head->next, struct migration_req, list);
5010 list_del_init(head->next);
5012 spin_unlock(&rq->lock);
5013 __migrate_task(req->task, cpu, req->dest_cpu);
5016 complete(&req->done);
5018 __set_current_state(TASK_RUNNING);
5022 /* Wait for kthread_stop */
5023 set_current_state(TASK_INTERRUPTIBLE);
5024 while (!kthread_should_stop()) {
5026 set_current_state(TASK_INTERRUPTIBLE);
5028 __set_current_state(TASK_RUNNING);
5032 #ifdef CONFIG_HOTPLUG_CPU
5034 * Figure out where task on dead CPU should go, use force if neccessary.
5035 * NOTE: interrupts should be disabled by the caller
5037 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5039 unsigned long flags;
5046 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5047 cpus_and(mask, mask, p->cpus_allowed);
5048 dest_cpu = any_online_cpu(mask);
5050 /* On any allowed CPU? */
5051 if (dest_cpu == NR_CPUS)
5052 dest_cpu = any_online_cpu(p->cpus_allowed);
5054 /* No more Mr. Nice Guy. */
5055 if (dest_cpu == NR_CPUS) {
5056 rq = task_rq_lock(p, &flags);
5057 cpus_setall(p->cpus_allowed);
5058 dest_cpu = any_online_cpu(p->cpus_allowed);
5059 task_rq_unlock(rq, &flags);
5062 * Don't tell them about moving exiting tasks or
5063 * kernel threads (both mm NULL), since they never
5066 if (p->mm && printk_ratelimit())
5067 printk(KERN_INFO "process %d (%s) no "
5068 "longer affine to cpu%d\n",
5069 p->pid, p->comm, dead_cpu);
5071 if (!__migrate_task(p, dead_cpu, dest_cpu))
5076 * While a dead CPU has no uninterruptible tasks queued at this point,
5077 * it might still have a nonzero ->nr_uninterruptible counter, because
5078 * for performance reasons the counter is not stricly tracking tasks to
5079 * their home CPUs. So we just add the counter to another CPU's counter,
5080 * to keep the global sum constant after CPU-down:
5082 static void migrate_nr_uninterruptible(struct rq *rq_src)
5084 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5085 unsigned long flags;
5087 local_irq_save(flags);
5088 double_rq_lock(rq_src, rq_dest);
5089 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5090 rq_src->nr_uninterruptible = 0;
5091 double_rq_unlock(rq_src, rq_dest);
5092 local_irq_restore(flags);
5095 /* Run through task list and migrate tasks from the dead cpu. */
5096 static void migrate_live_tasks(int src_cpu)
5098 struct task_struct *p, *t;
5100 write_lock_irq(&tasklist_lock);
5102 do_each_thread(t, p) {
5106 if (task_cpu(p) == src_cpu)
5107 move_task_off_dead_cpu(src_cpu, p);
5108 } while_each_thread(t, p);
5110 write_unlock_irq(&tasklist_lock);
5114 * Schedules idle task to be the next runnable task on current CPU.
5115 * It does so by boosting its priority to highest possible and adding it to
5116 * the _front_ of the runqueue. Used by CPU offline code.
5118 void sched_idle_next(void)
5120 int this_cpu = smp_processor_id();
5121 struct rq *rq = cpu_rq(this_cpu);
5122 struct task_struct *p = rq->idle;
5123 unsigned long flags;
5125 /* cpu has to be offline */
5126 BUG_ON(cpu_online(this_cpu));
5129 * Strictly not necessary since rest of the CPUs are stopped by now
5130 * and interrupts disabled on the current cpu.
5132 spin_lock_irqsave(&rq->lock, flags);
5134 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5136 /* Add idle task to the _front_ of its priority queue: */
5137 activate_idle_task(p, rq);
5139 spin_unlock_irqrestore(&rq->lock, flags);
5143 * Ensures that the idle task is using init_mm right before its cpu goes
5146 void idle_task_exit(void)
5148 struct mm_struct *mm = current->active_mm;
5150 BUG_ON(cpu_online(smp_processor_id()));
5153 switch_mm(mm, &init_mm, current);
5157 /* called under rq->lock with disabled interrupts */
5158 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5160 struct rq *rq = cpu_rq(dead_cpu);
5162 /* Must be exiting, otherwise would be on tasklist. */
5163 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5165 /* Cannot have done final schedule yet: would have vanished. */
5166 BUG_ON(p->state == TASK_DEAD);
5171 * Drop lock around migration; if someone else moves it,
5172 * that's OK. No task can be added to this CPU, so iteration is
5174 * NOTE: interrupts should be left disabled --dev@
5176 spin_unlock(&rq->lock);
5177 move_task_off_dead_cpu(dead_cpu, p);
5178 spin_lock(&rq->lock);
5183 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5184 static void migrate_dead_tasks(unsigned int dead_cpu)
5186 struct rq *rq = cpu_rq(dead_cpu);
5187 struct task_struct *next;
5190 if (!rq->nr_running)
5192 update_rq_clock(rq);
5193 next = pick_next_task(rq, rq->curr);
5196 migrate_dead(dead_cpu, next);
5200 #endif /* CONFIG_HOTPLUG_CPU */
5202 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5204 static struct ctl_table sd_ctl_dir[] = {
5206 .procname = "sched_domain",
5212 static struct ctl_table sd_ctl_root[] = {
5214 .ctl_name = CTL_KERN,
5215 .procname = "kernel",
5217 .child = sd_ctl_dir,
5222 static struct ctl_table *sd_alloc_ctl_entry(int n)
5224 struct ctl_table *entry =
5225 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5228 memset(entry, 0, n * sizeof(struct ctl_table));
5234 set_table_entry(struct ctl_table *entry,
5235 const char *procname, void *data, int maxlen,
5236 mode_t mode, proc_handler *proc_handler)
5238 entry->procname = procname;
5240 entry->maxlen = maxlen;
5242 entry->proc_handler = proc_handler;
5245 static struct ctl_table *
5246 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5248 struct ctl_table *table = sd_alloc_ctl_entry(14);
5250 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5251 sizeof(long), 0644, proc_doulongvec_minmax);
5252 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5253 sizeof(long), 0644, proc_doulongvec_minmax);
5254 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5255 sizeof(int), 0644, proc_dointvec_minmax);
5256 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5257 sizeof(int), 0644, proc_dointvec_minmax);
5258 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5259 sizeof(int), 0644, proc_dointvec_minmax);
5260 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5261 sizeof(int), 0644, proc_dointvec_minmax);
5262 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5263 sizeof(int), 0644, proc_dointvec_minmax);
5264 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5265 sizeof(int), 0644, proc_dointvec_minmax);
5266 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5267 sizeof(int), 0644, proc_dointvec_minmax);
5268 set_table_entry(&table[10], "cache_nice_tries",
5269 &sd->cache_nice_tries,
5270 sizeof(int), 0644, proc_dointvec_minmax);
5271 set_table_entry(&table[12], "flags", &sd->flags,
5272 sizeof(int), 0644, proc_dointvec_minmax);
5277 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5279 struct ctl_table *entry, *table;
5280 struct sched_domain *sd;
5281 int domain_num = 0, i;
5284 for_each_domain(cpu, sd)
5286 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5289 for_each_domain(cpu, sd) {
5290 snprintf(buf, 32, "domain%d", i);
5291 entry->procname = kstrdup(buf, GFP_KERNEL);
5293 entry->child = sd_alloc_ctl_domain_table(sd);
5300 static struct ctl_table_header *sd_sysctl_header;
5301 static void init_sched_domain_sysctl(void)
5303 int i, cpu_num = num_online_cpus();
5304 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5307 sd_ctl_dir[0].child = entry;
5309 for (i = 0; i < cpu_num; i++, entry++) {
5310 snprintf(buf, 32, "cpu%d", i);
5311 entry->procname = kstrdup(buf, GFP_KERNEL);
5313 entry->child = sd_alloc_ctl_cpu_table(i);
5315 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5318 static void init_sched_domain_sysctl(void)
5324 * migration_call - callback that gets triggered when a CPU is added.
5325 * Here we can start up the necessary migration thread for the new CPU.
5327 static int __cpuinit
5328 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5330 struct task_struct *p;
5331 int cpu = (long)hcpu;
5332 unsigned long flags;
5336 case CPU_LOCK_ACQUIRE:
5337 mutex_lock(&sched_hotcpu_mutex);
5340 case CPU_UP_PREPARE:
5341 case CPU_UP_PREPARE_FROZEN:
5342 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5345 kthread_bind(p, cpu);
5346 /* Must be high prio: stop_machine expects to yield to it. */
5347 rq = task_rq_lock(p, &flags);
5348 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5349 task_rq_unlock(rq, &flags);
5350 cpu_rq(cpu)->migration_thread = p;
5354 case CPU_ONLINE_FROZEN:
5355 /* Strictly unneccessary, as first user will wake it. */
5356 wake_up_process(cpu_rq(cpu)->migration_thread);
5359 #ifdef CONFIG_HOTPLUG_CPU
5360 case CPU_UP_CANCELED:
5361 case CPU_UP_CANCELED_FROZEN:
5362 if (!cpu_rq(cpu)->migration_thread)
5364 /* Unbind it from offline cpu so it can run. Fall thru. */
5365 kthread_bind(cpu_rq(cpu)->migration_thread,
5366 any_online_cpu(cpu_online_map));
5367 kthread_stop(cpu_rq(cpu)->migration_thread);
5368 cpu_rq(cpu)->migration_thread = NULL;
5372 case CPU_DEAD_FROZEN:
5373 migrate_live_tasks(cpu);
5375 kthread_stop(rq->migration_thread);
5376 rq->migration_thread = NULL;
5377 /* Idle task back to normal (off runqueue, low prio) */
5378 rq = task_rq_lock(rq->idle, &flags);
5379 update_rq_clock(rq);
5380 deactivate_task(rq, rq->idle, 0);
5381 rq->idle->static_prio = MAX_PRIO;
5382 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5383 rq->idle->sched_class = &idle_sched_class;
5384 migrate_dead_tasks(cpu);
5385 task_rq_unlock(rq, &flags);
5386 migrate_nr_uninterruptible(rq);
5387 BUG_ON(rq->nr_running != 0);
5389 /* No need to migrate the tasks: it was best-effort if
5390 * they didn't take sched_hotcpu_mutex. Just wake up
5391 * the requestors. */
5392 spin_lock_irq(&rq->lock);
5393 while (!list_empty(&rq->migration_queue)) {
5394 struct migration_req *req;
5396 req = list_entry(rq->migration_queue.next,
5397 struct migration_req, list);
5398 list_del_init(&req->list);
5399 complete(&req->done);
5401 spin_unlock_irq(&rq->lock);
5404 case CPU_LOCK_RELEASE:
5405 mutex_unlock(&sched_hotcpu_mutex);
5411 /* Register at highest priority so that task migration (migrate_all_tasks)
5412 * happens before everything else.
5414 static struct notifier_block __cpuinitdata migration_notifier = {
5415 .notifier_call = migration_call,
5419 int __init migration_init(void)
5421 void *cpu = (void *)(long)smp_processor_id();
5424 /* Start one for the boot CPU: */
5425 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5426 BUG_ON(err == NOTIFY_BAD);
5427 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5428 register_cpu_notifier(&migration_notifier);
5436 /* Number of possible processor ids */
5437 int nr_cpu_ids __read_mostly = NR_CPUS;
5438 EXPORT_SYMBOL(nr_cpu_ids);
5440 #undef SCHED_DOMAIN_DEBUG
5441 #ifdef SCHED_DOMAIN_DEBUG
5442 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5447 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5451 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5456 struct sched_group *group = sd->groups;
5457 cpumask_t groupmask;
5459 cpumask_scnprintf(str, NR_CPUS, sd->span);
5460 cpus_clear(groupmask);
5463 for (i = 0; i < level + 1; i++)
5465 printk("domain %d: ", level);
5467 if (!(sd->flags & SD_LOAD_BALANCE)) {
5468 printk("does not load-balance\n");
5470 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5475 printk("span %s\n", str);
5477 if (!cpu_isset(cpu, sd->span))
5478 printk(KERN_ERR "ERROR: domain->span does not contain "
5480 if (!cpu_isset(cpu, group->cpumask))
5481 printk(KERN_ERR "ERROR: domain->groups does not contain"
5485 for (i = 0; i < level + 2; i++)
5491 printk(KERN_ERR "ERROR: group is NULL\n");
5495 if (!group->__cpu_power) {
5497 printk(KERN_ERR "ERROR: domain->cpu_power not "
5501 if (!cpus_weight(group->cpumask)) {
5503 printk(KERN_ERR "ERROR: empty group\n");
5506 if (cpus_intersects(groupmask, group->cpumask)) {
5508 printk(KERN_ERR "ERROR: repeated CPUs\n");
5511 cpus_or(groupmask, groupmask, group->cpumask);
5513 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5516 group = group->next;
5517 } while (group != sd->groups);
5520 if (!cpus_equal(sd->span, groupmask))
5521 printk(KERN_ERR "ERROR: groups don't span "
5529 if (!cpus_subset(groupmask, sd->span))
5530 printk(KERN_ERR "ERROR: parent span is not a superset "
5531 "of domain->span\n");
5536 # define sched_domain_debug(sd, cpu) do { } while (0)
5539 static int sd_degenerate(struct sched_domain *sd)
5541 if (cpus_weight(sd->span) == 1)
5544 /* Following flags need at least 2 groups */
5545 if (sd->flags & (SD_LOAD_BALANCE |
5546 SD_BALANCE_NEWIDLE |
5550 SD_SHARE_PKG_RESOURCES)) {
5551 if (sd->groups != sd->groups->next)
5555 /* Following flags don't use groups */
5556 if (sd->flags & (SD_WAKE_IDLE |
5565 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5567 unsigned long cflags = sd->flags, pflags = parent->flags;
5569 if (sd_degenerate(parent))
5572 if (!cpus_equal(sd->span, parent->span))
5575 /* Does parent contain flags not in child? */
5576 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5577 if (cflags & SD_WAKE_AFFINE)
5578 pflags &= ~SD_WAKE_BALANCE;
5579 /* Flags needing groups don't count if only 1 group in parent */
5580 if (parent->groups == parent->groups->next) {
5581 pflags &= ~(SD_LOAD_BALANCE |
5582 SD_BALANCE_NEWIDLE |
5586 SD_SHARE_PKG_RESOURCES);
5588 if (~cflags & pflags)
5595 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5596 * hold the hotplug lock.
5598 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5600 struct rq *rq = cpu_rq(cpu);
5601 struct sched_domain *tmp;
5603 /* Remove the sched domains which do not contribute to scheduling. */
5604 for (tmp = sd; tmp; tmp = tmp->parent) {
5605 struct sched_domain *parent = tmp->parent;
5608 if (sd_parent_degenerate(tmp, parent)) {
5609 tmp->parent = parent->parent;
5611 parent->parent->child = tmp;
5615 if (sd && sd_degenerate(sd)) {
5621 sched_domain_debug(sd, cpu);
5623 rcu_assign_pointer(rq->sd, sd);
5626 /* cpus with isolated domains */
5627 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5629 /* Setup the mask of cpus configured for isolated domains */
5630 static int __init isolated_cpu_setup(char *str)
5632 int ints[NR_CPUS], i;
5634 str = get_options(str, ARRAY_SIZE(ints), ints);
5635 cpus_clear(cpu_isolated_map);
5636 for (i = 1; i <= ints[0]; i++)
5637 if (ints[i] < NR_CPUS)
5638 cpu_set(ints[i], cpu_isolated_map);
5642 __setup ("isolcpus=", isolated_cpu_setup);
5645 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5646 * to a function which identifies what group(along with sched group) a CPU
5647 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5648 * (due to the fact that we keep track of groups covered with a cpumask_t).
5650 * init_sched_build_groups will build a circular linked list of the groups
5651 * covered by the given span, and will set each group's ->cpumask correctly,
5652 * and ->cpu_power to 0.
5655 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5656 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5657 struct sched_group **sg))
5659 struct sched_group *first = NULL, *last = NULL;
5660 cpumask_t covered = CPU_MASK_NONE;
5663 for_each_cpu_mask(i, span) {
5664 struct sched_group *sg;
5665 int group = group_fn(i, cpu_map, &sg);
5668 if (cpu_isset(i, covered))
5671 sg->cpumask = CPU_MASK_NONE;
5672 sg->__cpu_power = 0;
5674 for_each_cpu_mask(j, span) {
5675 if (group_fn(j, cpu_map, NULL) != group)
5678 cpu_set(j, covered);
5679 cpu_set(j, sg->cpumask);
5690 #define SD_NODES_PER_DOMAIN 16
5695 * find_next_best_node - find the next node to include in a sched_domain
5696 * @node: node whose sched_domain we're building
5697 * @used_nodes: nodes already in the sched_domain
5699 * Find the next node to include in a given scheduling domain. Simply
5700 * finds the closest node not already in the @used_nodes map.
5702 * Should use nodemask_t.
5704 static int find_next_best_node(int node, unsigned long *used_nodes)
5706 int i, n, val, min_val, best_node = 0;
5710 for (i = 0; i < MAX_NUMNODES; i++) {
5711 /* Start at @node */
5712 n = (node + i) % MAX_NUMNODES;
5714 if (!nr_cpus_node(n))
5717 /* Skip already used nodes */
5718 if (test_bit(n, used_nodes))
5721 /* Simple min distance search */
5722 val = node_distance(node, n);
5724 if (val < min_val) {
5730 set_bit(best_node, used_nodes);
5735 * sched_domain_node_span - get a cpumask for a node's sched_domain
5736 * @node: node whose cpumask we're constructing
5737 * @size: number of nodes to include in this span
5739 * Given a node, construct a good cpumask for its sched_domain to span. It
5740 * should be one that prevents unnecessary balancing, but also spreads tasks
5743 static cpumask_t sched_domain_node_span(int node)
5745 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5746 cpumask_t span, nodemask;
5750 bitmap_zero(used_nodes, MAX_NUMNODES);
5752 nodemask = node_to_cpumask(node);
5753 cpus_or(span, span, nodemask);
5754 set_bit(node, used_nodes);
5756 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5757 int next_node = find_next_best_node(node, used_nodes);
5759 nodemask = node_to_cpumask(next_node);
5760 cpus_or(span, span, nodemask);
5767 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5770 * SMT sched-domains:
5772 #ifdef CONFIG_SCHED_SMT
5773 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5774 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5776 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5777 struct sched_group **sg)
5780 *sg = &per_cpu(sched_group_cpus, cpu);
5786 * multi-core sched-domains:
5788 #ifdef CONFIG_SCHED_MC
5789 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5790 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5793 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5794 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5795 struct sched_group **sg)
5798 cpumask_t mask = cpu_sibling_map[cpu];
5799 cpus_and(mask, mask, *cpu_map);
5800 group = first_cpu(mask);
5802 *sg = &per_cpu(sched_group_core, group);
5805 #elif defined(CONFIG_SCHED_MC)
5806 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5807 struct sched_group **sg)
5810 *sg = &per_cpu(sched_group_core, cpu);
5815 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5816 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5818 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5819 struct sched_group **sg)
5822 #ifdef CONFIG_SCHED_MC
5823 cpumask_t mask = cpu_coregroup_map(cpu);
5824 cpus_and(mask, mask, *cpu_map);
5825 group = first_cpu(mask);
5826 #elif defined(CONFIG_SCHED_SMT)
5827 cpumask_t mask = cpu_sibling_map[cpu];
5828 cpus_and(mask, mask, *cpu_map);
5829 group = first_cpu(mask);
5834 *sg = &per_cpu(sched_group_phys, group);
5840 * The init_sched_build_groups can't handle what we want to do with node
5841 * groups, so roll our own. Now each node has its own list of groups which
5842 * gets dynamically allocated.
5844 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5845 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5847 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5848 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5850 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5851 struct sched_group **sg)
5853 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5856 cpus_and(nodemask, nodemask, *cpu_map);
5857 group = first_cpu(nodemask);
5860 *sg = &per_cpu(sched_group_allnodes, group);
5864 static void init_numa_sched_groups_power(struct sched_group *group_head)
5866 struct sched_group *sg = group_head;
5872 for_each_cpu_mask(j, sg->cpumask) {
5873 struct sched_domain *sd;
5875 sd = &per_cpu(phys_domains, j);
5876 if (j != first_cpu(sd->groups->cpumask)) {
5878 * Only add "power" once for each
5884 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5887 if (sg != group_head)
5893 /* Free memory allocated for various sched_group structures */
5894 static void free_sched_groups(const cpumask_t *cpu_map)
5898 for_each_cpu_mask(cpu, *cpu_map) {
5899 struct sched_group **sched_group_nodes
5900 = sched_group_nodes_bycpu[cpu];
5902 if (!sched_group_nodes)
5905 for (i = 0; i < MAX_NUMNODES; i++) {
5906 cpumask_t nodemask = node_to_cpumask(i);
5907 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5909 cpus_and(nodemask, nodemask, *cpu_map);
5910 if (cpus_empty(nodemask))
5920 if (oldsg != sched_group_nodes[i])
5923 kfree(sched_group_nodes);
5924 sched_group_nodes_bycpu[cpu] = NULL;
5928 static void free_sched_groups(const cpumask_t *cpu_map)
5934 * Initialize sched groups cpu_power.
5936 * cpu_power indicates the capacity of sched group, which is used while
5937 * distributing the load between different sched groups in a sched domain.
5938 * Typically cpu_power for all the groups in a sched domain will be same unless
5939 * there are asymmetries in the topology. If there are asymmetries, group
5940 * having more cpu_power will pickup more load compared to the group having
5943 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5944 * the maximum number of tasks a group can handle in the presence of other idle
5945 * or lightly loaded groups in the same sched domain.
5947 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5949 struct sched_domain *child;
5950 struct sched_group *group;
5952 WARN_ON(!sd || !sd->groups);
5954 if (cpu != first_cpu(sd->groups->cpumask))
5959 sd->groups->__cpu_power = 0;
5962 * For perf policy, if the groups in child domain share resources
5963 * (for example cores sharing some portions of the cache hierarchy
5964 * or SMT), then set this domain groups cpu_power such that each group
5965 * can handle only one task, when there are other idle groups in the
5966 * same sched domain.
5968 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5970 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5971 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5976 * add cpu_power of each child group to this groups cpu_power
5978 group = child->groups;
5980 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5981 group = group->next;
5982 } while (group != child->groups);
5986 * Build sched domains for a given set of cpus and attach the sched domains
5987 * to the individual cpus
5989 static int build_sched_domains(const cpumask_t *cpu_map)
5993 struct sched_group **sched_group_nodes = NULL;
5994 int sd_allnodes = 0;
5997 * Allocate the per-node list of sched groups
5999 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
6001 if (!sched_group_nodes) {
6002 printk(KERN_WARNING "Can not alloc sched group node list\n");
6005 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6009 * Set up domains for cpus specified by the cpu_map.
6011 for_each_cpu_mask(i, *cpu_map) {
6012 struct sched_domain *sd = NULL, *p;
6013 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6015 cpus_and(nodemask, nodemask, *cpu_map);
6018 if (cpus_weight(*cpu_map) >
6019 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6020 sd = &per_cpu(allnodes_domains, i);
6021 *sd = SD_ALLNODES_INIT;
6022 sd->span = *cpu_map;
6023 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6029 sd = &per_cpu(node_domains, i);
6031 sd->span = sched_domain_node_span(cpu_to_node(i));
6035 cpus_and(sd->span, sd->span, *cpu_map);
6039 sd = &per_cpu(phys_domains, i);
6041 sd->span = nodemask;
6045 cpu_to_phys_group(i, cpu_map, &sd->groups);
6047 #ifdef CONFIG_SCHED_MC
6049 sd = &per_cpu(core_domains, i);
6051 sd->span = cpu_coregroup_map(i);
6052 cpus_and(sd->span, sd->span, *cpu_map);
6055 cpu_to_core_group(i, cpu_map, &sd->groups);
6058 #ifdef CONFIG_SCHED_SMT
6060 sd = &per_cpu(cpu_domains, i);
6061 *sd = SD_SIBLING_INIT;
6062 sd->span = cpu_sibling_map[i];
6063 cpus_and(sd->span, sd->span, *cpu_map);
6066 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6070 #ifdef CONFIG_SCHED_SMT
6071 /* Set up CPU (sibling) groups */
6072 for_each_cpu_mask(i, *cpu_map) {
6073 cpumask_t this_sibling_map = cpu_sibling_map[i];
6074 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6075 if (i != first_cpu(this_sibling_map))
6078 init_sched_build_groups(this_sibling_map, cpu_map,
6083 #ifdef CONFIG_SCHED_MC
6084 /* Set up multi-core groups */
6085 for_each_cpu_mask(i, *cpu_map) {
6086 cpumask_t this_core_map = cpu_coregroup_map(i);
6087 cpus_and(this_core_map, this_core_map, *cpu_map);
6088 if (i != first_cpu(this_core_map))
6090 init_sched_build_groups(this_core_map, cpu_map,
6091 &cpu_to_core_group);
6095 /* Set up physical groups */
6096 for (i = 0; i < MAX_NUMNODES; i++) {
6097 cpumask_t nodemask = node_to_cpumask(i);
6099 cpus_and(nodemask, nodemask, *cpu_map);
6100 if (cpus_empty(nodemask))
6103 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6107 /* Set up node groups */
6109 init_sched_build_groups(*cpu_map, cpu_map,
6110 &cpu_to_allnodes_group);
6112 for (i = 0; i < MAX_NUMNODES; i++) {
6113 /* Set up node groups */
6114 struct sched_group *sg, *prev;
6115 cpumask_t nodemask = node_to_cpumask(i);
6116 cpumask_t domainspan;
6117 cpumask_t covered = CPU_MASK_NONE;
6120 cpus_and(nodemask, nodemask, *cpu_map);
6121 if (cpus_empty(nodemask)) {
6122 sched_group_nodes[i] = NULL;
6126 domainspan = sched_domain_node_span(i);
6127 cpus_and(domainspan, domainspan, *cpu_map);
6129 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6131 printk(KERN_WARNING "Can not alloc domain group for "
6135 sched_group_nodes[i] = sg;
6136 for_each_cpu_mask(j, nodemask) {
6137 struct sched_domain *sd;
6139 sd = &per_cpu(node_domains, j);
6142 sg->__cpu_power = 0;
6143 sg->cpumask = nodemask;
6145 cpus_or(covered, covered, nodemask);
6148 for (j = 0; j < MAX_NUMNODES; j++) {
6149 cpumask_t tmp, notcovered;
6150 int n = (i + j) % MAX_NUMNODES;
6152 cpus_complement(notcovered, covered);
6153 cpus_and(tmp, notcovered, *cpu_map);
6154 cpus_and(tmp, tmp, domainspan);
6155 if (cpus_empty(tmp))
6158 nodemask = node_to_cpumask(n);
6159 cpus_and(tmp, tmp, nodemask);
6160 if (cpus_empty(tmp))
6163 sg = kmalloc_node(sizeof(struct sched_group),
6167 "Can not alloc domain group for node %d\n", j);
6170 sg->__cpu_power = 0;
6172 sg->next = prev->next;
6173 cpus_or(covered, covered, tmp);
6180 /* Calculate CPU power for physical packages and nodes */
6181 #ifdef CONFIG_SCHED_SMT
6182 for_each_cpu_mask(i, *cpu_map) {
6183 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6185 init_sched_groups_power(i, sd);
6188 #ifdef CONFIG_SCHED_MC
6189 for_each_cpu_mask(i, *cpu_map) {
6190 struct sched_domain *sd = &per_cpu(core_domains, i);
6192 init_sched_groups_power(i, sd);
6196 for_each_cpu_mask(i, *cpu_map) {
6197 struct sched_domain *sd = &per_cpu(phys_domains, i);
6199 init_sched_groups_power(i, sd);
6203 for (i = 0; i < MAX_NUMNODES; i++)
6204 init_numa_sched_groups_power(sched_group_nodes[i]);
6207 struct sched_group *sg;
6209 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6210 init_numa_sched_groups_power(sg);
6214 /* Attach the domains */
6215 for_each_cpu_mask(i, *cpu_map) {
6216 struct sched_domain *sd;
6217 #ifdef CONFIG_SCHED_SMT
6218 sd = &per_cpu(cpu_domains, i);
6219 #elif defined(CONFIG_SCHED_MC)
6220 sd = &per_cpu(core_domains, i);
6222 sd = &per_cpu(phys_domains, i);
6224 cpu_attach_domain(sd, i);
6231 free_sched_groups(cpu_map);
6236 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6238 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6240 cpumask_t cpu_default_map;
6244 * Setup mask for cpus without special case scheduling requirements.
6245 * For now this just excludes isolated cpus, but could be used to
6246 * exclude other special cases in the future.
6248 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6250 err = build_sched_domains(&cpu_default_map);
6255 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6257 free_sched_groups(cpu_map);
6261 * Detach sched domains from a group of cpus specified in cpu_map
6262 * These cpus will now be attached to the NULL domain
6264 static void detach_destroy_domains(const cpumask_t *cpu_map)
6268 for_each_cpu_mask(i, *cpu_map)
6269 cpu_attach_domain(NULL, i);
6270 synchronize_sched();
6271 arch_destroy_sched_domains(cpu_map);
6275 * Partition sched domains as specified by the cpumasks below.
6276 * This attaches all cpus from the cpumasks to the NULL domain,
6277 * waits for a RCU quiescent period, recalculates sched
6278 * domain information and then attaches them back to the
6279 * correct sched domains
6280 * Call with hotplug lock held
6282 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6284 cpumask_t change_map;
6287 cpus_and(*partition1, *partition1, cpu_online_map);
6288 cpus_and(*partition2, *partition2, cpu_online_map);
6289 cpus_or(change_map, *partition1, *partition2);
6291 /* Detach sched domains from all of the affected cpus */
6292 detach_destroy_domains(&change_map);
6293 if (!cpus_empty(*partition1))
6294 err = build_sched_domains(partition1);
6295 if (!err && !cpus_empty(*partition2))
6296 err = build_sched_domains(partition2);
6301 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6302 static int arch_reinit_sched_domains(void)
6306 mutex_lock(&sched_hotcpu_mutex);
6307 detach_destroy_domains(&cpu_online_map);
6308 err = arch_init_sched_domains(&cpu_online_map);
6309 mutex_unlock(&sched_hotcpu_mutex);
6314 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6318 if (buf[0] != '0' && buf[0] != '1')
6322 sched_smt_power_savings = (buf[0] == '1');
6324 sched_mc_power_savings = (buf[0] == '1');
6326 ret = arch_reinit_sched_domains();
6328 return ret ? ret : count;
6331 #ifdef CONFIG_SCHED_MC
6332 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6334 return sprintf(page, "%u\n", sched_mc_power_savings);
6336 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6337 const char *buf, size_t count)
6339 return sched_power_savings_store(buf, count, 0);
6341 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6342 sched_mc_power_savings_store);
6345 #ifdef CONFIG_SCHED_SMT
6346 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6348 return sprintf(page, "%u\n", sched_smt_power_savings);
6350 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6351 const char *buf, size_t count)
6353 return sched_power_savings_store(buf, count, 1);
6355 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6356 sched_smt_power_savings_store);
6359 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6363 #ifdef CONFIG_SCHED_SMT
6365 err = sysfs_create_file(&cls->kset.kobj,
6366 &attr_sched_smt_power_savings.attr);
6368 #ifdef CONFIG_SCHED_MC
6369 if (!err && mc_capable())
6370 err = sysfs_create_file(&cls->kset.kobj,
6371 &attr_sched_mc_power_savings.attr);
6378 * Force a reinitialization of the sched domains hierarchy. The domains
6379 * and groups cannot be updated in place without racing with the balancing
6380 * code, so we temporarily attach all running cpus to the NULL domain
6381 * which will prevent rebalancing while the sched domains are recalculated.
6383 static int update_sched_domains(struct notifier_block *nfb,
6384 unsigned long action, void *hcpu)
6387 case CPU_UP_PREPARE:
6388 case CPU_UP_PREPARE_FROZEN:
6389 case CPU_DOWN_PREPARE:
6390 case CPU_DOWN_PREPARE_FROZEN:
6391 detach_destroy_domains(&cpu_online_map);
6394 case CPU_UP_CANCELED:
6395 case CPU_UP_CANCELED_FROZEN:
6396 case CPU_DOWN_FAILED:
6397 case CPU_DOWN_FAILED_FROZEN:
6399 case CPU_ONLINE_FROZEN:
6401 case CPU_DEAD_FROZEN:
6403 * Fall through and re-initialise the domains.
6410 /* The hotplug lock is already held by cpu_up/cpu_down */
6411 arch_init_sched_domains(&cpu_online_map);
6416 void __init sched_init_smp(void)
6418 cpumask_t non_isolated_cpus;
6420 mutex_lock(&sched_hotcpu_mutex);
6421 arch_init_sched_domains(&cpu_online_map);
6422 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6423 if (cpus_empty(non_isolated_cpus))
6424 cpu_set(smp_processor_id(), non_isolated_cpus);
6425 mutex_unlock(&sched_hotcpu_mutex);
6426 /* XXX: Theoretical race here - CPU may be hotplugged now */
6427 hotcpu_notifier(update_sched_domains, 0);
6429 init_sched_domain_sysctl();
6431 /* Move init over to a non-isolated CPU */
6432 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6436 void __init sched_init_smp(void)
6439 #endif /* CONFIG_SMP */
6441 int in_sched_functions(unsigned long addr)
6443 /* Linker adds these: start and end of __sched functions */
6444 extern char __sched_text_start[], __sched_text_end[];
6446 return in_lock_functions(addr) ||
6447 (addr >= (unsigned long)__sched_text_start
6448 && addr < (unsigned long)__sched_text_end);
6451 static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6453 cfs_rq->tasks_timeline = RB_ROOT;
6454 cfs_rq->fair_clock = 1;
6455 #ifdef CONFIG_FAIR_GROUP_SCHED
6460 void __init sched_init(void)
6462 int highest_cpu = 0;
6466 * Link up the scheduling class hierarchy:
6468 rt_sched_class.next = &fair_sched_class;
6469 fair_sched_class.next = &idle_sched_class;
6470 idle_sched_class.next = NULL;
6472 for_each_possible_cpu(i) {
6473 struct rt_prio_array *array;
6477 spin_lock_init(&rq->lock);
6478 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6481 init_cfs_rq(&rq->cfs, rq);
6482 #ifdef CONFIG_FAIR_GROUP_SCHED
6483 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6484 list_add(&rq->cfs.leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6487 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6488 rq->cpu_load[j] = 0;
6491 rq->active_balance = 0;
6492 rq->next_balance = jiffies;
6495 rq->migration_thread = NULL;
6496 INIT_LIST_HEAD(&rq->migration_queue);
6498 atomic_set(&rq->nr_iowait, 0);
6500 array = &rq->rt.active;
6501 for (j = 0; j < MAX_RT_PRIO; j++) {
6502 INIT_LIST_HEAD(array->queue + j);
6503 __clear_bit(j, array->bitmap);
6506 /* delimiter for bitsearch: */
6507 __set_bit(MAX_RT_PRIO, array->bitmap);
6510 set_load_weight(&init_task);
6512 #ifdef CONFIG_PREEMPT_NOTIFIERS
6513 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6517 nr_cpu_ids = highest_cpu + 1;
6518 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6521 #ifdef CONFIG_RT_MUTEXES
6522 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6526 * The boot idle thread does lazy MMU switching as well:
6528 atomic_inc(&init_mm.mm_count);
6529 enter_lazy_tlb(&init_mm, current);
6532 * Make us the idle thread. Technically, schedule() should not be
6533 * called from this thread, however somewhere below it might be,
6534 * but because we are the idle thread, we just pick up running again
6535 * when this runqueue becomes "idle".
6537 init_idle(current, smp_processor_id());
6539 * During early bootup we pretend to be a normal task:
6541 current->sched_class = &fair_sched_class;
6544 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6545 void __might_sleep(char *file, int line)
6548 static unsigned long prev_jiffy; /* ratelimiting */
6550 if ((in_atomic() || irqs_disabled()) &&
6551 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6552 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6554 prev_jiffy = jiffies;
6555 printk(KERN_ERR "BUG: sleeping function called from invalid"
6556 " context at %s:%d\n", file, line);
6557 printk("in_atomic():%d, irqs_disabled():%d\n",
6558 in_atomic(), irqs_disabled());
6559 debug_show_held_locks(current);
6560 if (irqs_disabled())
6561 print_irqtrace_events(current);
6566 EXPORT_SYMBOL(__might_sleep);
6569 #ifdef CONFIG_MAGIC_SYSRQ
6570 void normalize_rt_tasks(void)
6572 struct task_struct *g, *p;
6573 unsigned long flags;
6577 read_lock_irq(&tasklist_lock);
6578 do_each_thread(g, p) {
6580 p->se.wait_runtime = 0;
6581 p->se.exec_start = 0;
6582 p->se.wait_start_fair = 0;
6583 p->se.sleep_start_fair = 0;
6584 #ifdef CONFIG_SCHEDSTATS
6585 p->se.wait_start = 0;
6586 p->se.sleep_start = 0;
6587 p->se.block_start = 0;
6589 task_rq(p)->cfs.fair_clock = 0;
6590 task_rq(p)->clock = 0;
6594 * Renice negative nice level userspace
6597 if (TASK_NICE(p) < 0 && p->mm)
6598 set_user_nice(p, 0);
6602 spin_lock_irqsave(&p->pi_lock, flags);
6603 rq = __task_rq_lock(p);
6606 * Do not touch the migration thread:
6608 if (p == rq->migration_thread)
6612 update_rq_clock(rq);
6613 on_rq = p->se.on_rq;
6615 deactivate_task(rq, p, 0);
6616 __setscheduler(rq, p, SCHED_NORMAL, 0);
6618 activate_task(rq, p, 0);
6619 resched_task(rq->curr);
6624 __task_rq_unlock(rq);
6625 spin_unlock_irqrestore(&p->pi_lock, flags);
6626 } while_each_thread(g, p);
6628 read_unlock_irq(&tasklist_lock);
6631 #endif /* CONFIG_MAGIC_SYSRQ */
6635 * These functions are only useful for the IA64 MCA handling.
6637 * They can only be called when the whole system has been
6638 * stopped - every CPU needs to be quiescent, and no scheduling
6639 * activity can take place. Using them for anything else would
6640 * be a serious bug, and as a result, they aren't even visible
6641 * under any other configuration.
6645 * curr_task - return the current task for a given cpu.
6646 * @cpu: the processor in question.
6648 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6650 struct task_struct *curr_task(int cpu)
6652 return cpu_curr(cpu);
6656 * set_curr_task - set the current task for a given cpu.
6657 * @cpu: the processor in question.
6658 * @p: the task pointer to set.
6660 * Description: This function must only be used when non-maskable interrupts
6661 * are serviced on a separate stack. It allows the architecture to switch the
6662 * notion of the current task on a cpu in a non-blocking manner. This function
6663 * must be called with all CPU's synchronized, and interrupts disabled, the
6664 * and caller must save the original value of the current task (see
6665 * curr_task() above) and restore that value before reenabling interrupts and
6666 * re-starting the system.
6668 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6670 void set_curr_task(int cpu, struct task_struct *p)