4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
64 #include <linux/pagemap.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __attribute__((weak)) sched_clock(void)
75 return (unsigned long long)jiffies * (1000000000 / HZ);
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Some helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
100 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
109 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
110 * Timeslices get refilled after they expire.
112 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
122 return reciprocal_divide(load, sg->reciprocal_cpu_power);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
131 sg->__cpu_power += val;
132 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
136 #define SCALE_PRIO(x, prio) \
137 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
140 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
141 * to time slice values: [800ms ... 100ms ... 5ms]
143 static unsigned int static_prio_timeslice(int static_prio)
145 if (static_prio == NICE_TO_PRIO(19))
148 if (static_prio < NICE_TO_PRIO(0))
149 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
151 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
154 static inline int rt_policy(int policy)
156 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
161 static inline int task_has_rt_policy(struct task_struct *p)
163 return rt_policy(p->policy);
167 * This is the priority-queue data structure of the RT scheduling class:
169 struct rt_prio_array {
170 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
171 struct list_head queue[MAX_RT_PRIO];
175 struct load_weight load;
178 /* CFS-related fields in a runqueue */
180 struct load_weight load;
181 unsigned long nr_running;
188 unsigned long wait_runtime_overruns, wait_runtime_underruns;
190 struct rb_root tasks_timeline;
191 struct rb_node *rb_leftmost;
192 struct rb_node *rb_load_balance_curr;
193 /* 'curr' points to currently running entity on this cfs_rq.
194 * It is set to NULL otherwise (i.e when none are currently running).
196 struct sched_entity *curr;
197 #ifdef CONFIG_FAIR_GROUP_SCHED
198 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
200 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
201 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
202 * (like users, containers etc.)
204 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
205 * list is used during load balance.
207 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
211 /* Real-Time classes' related field in a runqueue: */
213 struct rt_prio_array active;
214 int rt_load_balance_idx;
215 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
219 * This is the main, per-CPU runqueue data structure.
221 * Locking rule: those places that want to lock multiple runqueues
222 * (such as the load balancing or the thread migration code), lock
223 * acquire operations must be ordered by ascending &runqueue.
226 spinlock_t lock; /* runqueue lock */
229 * nr_running and cpu_load should be in the same cacheline because
230 * remote CPUs use both these fields when doing load calculation.
232 unsigned long nr_running;
233 #define CPU_LOAD_IDX_MAX 5
234 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
235 unsigned char idle_at_tick;
237 unsigned char in_nohz_recently;
239 struct load_stat ls; /* capture load from *all* tasks on this cpu */
240 unsigned long nr_load_updates;
244 #ifdef CONFIG_FAIR_GROUP_SCHED
245 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
250 * This is part of a global counter where only the total sum
251 * over all CPUs matters. A task can increase this counter on
252 * one CPU and if it got migrated afterwards it may decrease
253 * it on another CPU. Always updated under the runqueue lock:
255 unsigned long nr_uninterruptible;
257 struct task_struct *curr, *idle;
258 unsigned long next_balance;
259 struct mm_struct *prev_mm;
261 u64 clock, prev_clock_raw;
264 unsigned int clock_warps, clock_overflows;
266 unsigned int clock_deep_idle_events;
272 struct sched_domain *sd;
274 /* For active balancing */
277 int cpu; /* cpu of this runqueue */
279 struct task_struct *migration_thread;
280 struct list_head migration_queue;
283 #ifdef CONFIG_SCHEDSTATS
285 struct sched_info rq_sched_info;
287 /* sys_sched_yield() stats */
288 unsigned long yld_exp_empty;
289 unsigned long yld_act_empty;
290 unsigned long yld_both_empty;
291 unsigned long yld_cnt;
293 /* schedule() stats */
294 unsigned long sched_switch;
295 unsigned long sched_cnt;
296 unsigned long sched_goidle;
298 /* try_to_wake_up() stats */
299 unsigned long ttwu_cnt;
300 unsigned long ttwu_local;
302 struct lock_class_key rq_lock_key;
305 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
306 static DEFINE_MUTEX(sched_hotcpu_mutex);
308 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
310 rq->curr->sched_class->check_preempt_curr(rq, p);
313 static inline int cpu_of(struct rq *rq)
323 * Update the per-runqueue clock, as finegrained as the platform can give
324 * us, but without assuming monotonicity, etc.:
326 static void __update_rq_clock(struct rq *rq)
328 u64 prev_raw = rq->prev_clock_raw;
329 u64 now = sched_clock();
330 s64 delta = now - prev_raw;
331 u64 clock = rq->clock;
333 #ifdef CONFIG_SCHED_DEBUG
334 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
337 * Protect against sched_clock() occasionally going backwards:
339 if (unlikely(delta < 0)) {
344 * Catch too large forward jumps too:
346 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
347 if (clock < rq->tick_timestamp + TICK_NSEC)
348 clock = rq->tick_timestamp + TICK_NSEC;
351 rq->clock_overflows++;
353 if (unlikely(delta > rq->clock_max_delta))
354 rq->clock_max_delta = delta;
359 rq->prev_clock_raw = now;
363 static void update_rq_clock(struct rq *rq)
365 if (likely(smp_processor_id() == cpu_of(rq)))
366 __update_rq_clock(rq);
370 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
371 * See detach_destroy_domains: synchronize_sched for details.
373 * The domain tree of any CPU may only be accessed from within
374 * preempt-disabled sections.
376 #define for_each_domain(cpu, __sd) \
377 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
379 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
380 #define this_rq() (&__get_cpu_var(runqueues))
381 #define task_rq(p) cpu_rq(task_cpu(p))
382 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
385 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
387 #ifdef CONFIG_SCHED_DEBUG
388 # define const_debug __read_mostly
390 # define const_debug static const
394 * Debugging: various feature bits
397 SCHED_FEAT_FAIR_SLEEPERS = 1,
398 SCHED_FEAT_NEW_FAIR_SLEEPERS = 2,
399 SCHED_FEAT_SLEEPER_AVG = 4,
400 SCHED_FEAT_SLEEPER_LOAD_AVG = 8,
401 SCHED_FEAT_START_DEBIT = 16,
402 SCHED_FEAT_USE_TREE_AVG = 32,
403 SCHED_FEAT_APPROX_AVG = 64,
406 const_debug unsigned int sysctl_sched_features =
407 SCHED_FEAT_FAIR_SLEEPERS *0 |
408 SCHED_FEAT_NEW_FAIR_SLEEPERS *1 |
409 SCHED_FEAT_SLEEPER_AVG *0 |
410 SCHED_FEAT_SLEEPER_LOAD_AVG *1 |
411 SCHED_FEAT_START_DEBIT *1 |
412 SCHED_FEAT_USE_TREE_AVG *0 |
413 SCHED_FEAT_APPROX_AVG *0;
415 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
418 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
419 * clock constructed from sched_clock():
421 unsigned long long cpu_clock(int cpu)
423 unsigned long long now;
427 local_irq_save(flags);
431 local_irq_restore(flags);
436 #ifdef CONFIG_FAIR_GROUP_SCHED
437 /* Change a task's ->cfs_rq if it moves across CPUs */
438 static inline void set_task_cfs_rq(struct task_struct *p)
440 p->se.cfs_rq = &task_rq(p)->cfs;
443 static inline void set_task_cfs_rq(struct task_struct *p)
448 #ifndef prepare_arch_switch
449 # define prepare_arch_switch(next) do { } while (0)
451 #ifndef finish_arch_switch
452 # define finish_arch_switch(prev) do { } while (0)
455 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
456 static inline int task_running(struct rq *rq, struct task_struct *p)
458 return rq->curr == p;
461 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
465 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
467 #ifdef CONFIG_DEBUG_SPINLOCK
468 /* this is a valid case when another task releases the spinlock */
469 rq->lock.owner = current;
472 * If we are tracking spinlock dependencies then we have to
473 * fix up the runqueue lock - which gets 'carried over' from
476 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
478 spin_unlock_irq(&rq->lock);
481 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
482 static inline int task_running(struct rq *rq, struct task_struct *p)
487 return rq->curr == p;
491 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
495 * We can optimise this out completely for !SMP, because the
496 * SMP rebalancing from interrupt is the only thing that cares
501 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
502 spin_unlock_irq(&rq->lock);
504 spin_unlock(&rq->lock);
508 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
512 * After ->oncpu is cleared, the task can be moved to a different CPU.
513 * We must ensure this doesn't happen until the switch is completely
519 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
523 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
526 * __task_rq_lock - lock the runqueue a given task resides on.
527 * Must be called interrupts disabled.
529 static inline struct rq *__task_rq_lock(struct task_struct *p)
536 spin_lock(&rq->lock);
537 if (unlikely(rq != task_rq(p))) {
538 spin_unlock(&rq->lock);
539 goto repeat_lock_task;
545 * task_rq_lock - lock the runqueue a given task resides on and disable
546 * interrupts. Note the ordering: we can safely lookup the task_rq without
547 * explicitly disabling preemption.
549 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
555 local_irq_save(*flags);
557 spin_lock(&rq->lock);
558 if (unlikely(rq != task_rq(p))) {
559 spin_unlock_irqrestore(&rq->lock, *flags);
560 goto repeat_lock_task;
565 static inline void __task_rq_unlock(struct rq *rq)
568 spin_unlock(&rq->lock);
571 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
574 spin_unlock_irqrestore(&rq->lock, *flags);
578 * this_rq_lock - lock this runqueue and disable interrupts.
580 static inline struct rq *this_rq_lock(void)
587 spin_lock(&rq->lock);
593 * We are going deep-idle (irqs are disabled):
595 void sched_clock_idle_sleep_event(void)
597 struct rq *rq = cpu_rq(smp_processor_id());
599 spin_lock(&rq->lock);
600 __update_rq_clock(rq);
601 spin_unlock(&rq->lock);
602 rq->clock_deep_idle_events++;
604 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
607 * We just idled delta nanoseconds (called with irqs disabled):
609 void sched_clock_idle_wakeup_event(u64 delta_ns)
611 struct rq *rq = cpu_rq(smp_processor_id());
612 u64 now = sched_clock();
614 rq->idle_clock += delta_ns;
616 * Override the previous timestamp and ignore all
617 * sched_clock() deltas that occured while we idled,
618 * and use the PM-provided delta_ns to advance the
621 spin_lock(&rq->lock);
622 rq->prev_clock_raw = now;
623 rq->clock += delta_ns;
624 spin_unlock(&rq->lock);
626 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
629 * resched_task - mark a task 'to be rescheduled now'.
631 * On UP this means the setting of the need_resched flag, on SMP it
632 * might also involve a cross-CPU call to trigger the scheduler on
637 #ifndef tsk_is_polling
638 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
641 static void resched_task(struct task_struct *p)
645 assert_spin_locked(&task_rq(p)->lock);
647 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
650 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
653 if (cpu == smp_processor_id())
656 /* NEED_RESCHED must be visible before we test polling */
658 if (!tsk_is_polling(p))
659 smp_send_reschedule(cpu);
662 static void resched_cpu(int cpu)
664 struct rq *rq = cpu_rq(cpu);
667 if (!spin_trylock_irqsave(&rq->lock, flags))
669 resched_task(cpu_curr(cpu));
670 spin_unlock_irqrestore(&rq->lock, flags);
673 static inline void resched_task(struct task_struct *p)
675 assert_spin_locked(&task_rq(p)->lock);
676 set_tsk_need_resched(p);
680 static u64 div64_likely32(u64 divident, unsigned long divisor)
682 #if BITS_PER_LONG == 32
683 if (likely(divident <= 0xffffffffULL))
684 return (u32)divident / divisor;
685 do_div(divident, divisor);
689 return divident / divisor;
693 #if BITS_PER_LONG == 32
694 # define WMULT_CONST (~0UL)
696 # define WMULT_CONST (1UL << 32)
699 #define WMULT_SHIFT 32
702 * Shift right and round:
704 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
707 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
708 struct load_weight *lw)
712 if (unlikely(!lw->inv_weight))
713 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
715 tmp = (u64)delta_exec * weight;
717 * Check whether we'd overflow the 64-bit multiplication:
719 if (unlikely(tmp > WMULT_CONST))
720 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
723 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
725 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
728 static inline unsigned long
729 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
731 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
734 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
737 if (sched_feat(FAIR_SLEEPERS))
738 lw->inv_weight = WMULT_CONST / lw->weight;
741 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
744 if (sched_feat(FAIR_SLEEPERS) && likely(lw->weight))
745 lw->inv_weight = WMULT_CONST / lw->weight;
749 * To aid in avoiding the subversion of "niceness" due to uneven distribution
750 * of tasks with abnormal "nice" values across CPUs the contribution that
751 * each task makes to its run queue's load is weighted according to its
752 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
753 * scaled version of the new time slice allocation that they receive on time
757 #define WEIGHT_IDLEPRIO 2
758 #define WMULT_IDLEPRIO (1 << 31)
761 * Nice levels are multiplicative, with a gentle 10% change for every
762 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
763 * nice 1, it will get ~10% less CPU time than another CPU-bound task
764 * that remained on nice 0.
766 * The "10% effect" is relative and cumulative: from _any_ nice level,
767 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
768 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
769 * If a task goes up by ~10% and another task goes down by ~10% then
770 * the relative distance between them is ~25%.)
772 static const int prio_to_weight[40] = {
773 /* -20 */ 88761, 71755, 56483, 46273, 36291,
774 /* -15 */ 29154, 23254, 18705, 14949, 11916,
775 /* -10 */ 9548, 7620, 6100, 4904, 3906,
776 /* -5 */ 3121, 2501, 1991, 1586, 1277,
777 /* 0 */ 1024, 820, 655, 526, 423,
778 /* 5 */ 335, 272, 215, 172, 137,
779 /* 10 */ 110, 87, 70, 56, 45,
780 /* 15 */ 36, 29, 23, 18, 15,
784 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
786 * In cases where the weight does not change often, we can use the
787 * precalculated inverse to speed up arithmetics by turning divisions
788 * into multiplications:
790 static const u32 prio_to_wmult[40] = {
791 /* -20 */ 48388, 59856, 76040, 92818, 118348,
792 /* -15 */ 147320, 184698, 229616, 287308, 360437,
793 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
794 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
795 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
796 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
797 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
798 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
801 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
804 * runqueue iterator, to support SMP load-balancing between different
805 * scheduling classes, without having to expose their internal data
806 * structures to the load-balancing proper:
810 struct task_struct *(*start)(void *);
811 struct task_struct *(*next)(void *);
814 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
815 unsigned long max_nr_move, unsigned long max_load_move,
816 struct sched_domain *sd, enum cpu_idle_type idle,
817 int *all_pinned, unsigned long *load_moved,
818 int *this_best_prio, struct rq_iterator *iterator);
820 #include "sched_stats.h"
821 #include "sched_rt.c"
822 #include "sched_fair.c"
823 #include "sched_idletask.c"
824 #ifdef CONFIG_SCHED_DEBUG
825 # include "sched_debug.c"
828 #define sched_class_highest (&rt_sched_class)
831 * Update delta_exec, delta_fair fields for rq.
833 * delta_fair clock advances at a rate inversely proportional to
834 * total load (rq->ls.load.weight) on the runqueue, while
835 * delta_exec advances at the same rate as wall-clock (provided
838 * delta_exec / delta_fair is a measure of the (smoothened) load on this
839 * runqueue over any given interval. This (smoothened) load is used
840 * during load balance.
842 * This function is called /before/ updating rq->ls.load
843 * and when switching tasks.
845 static inline void inc_load(struct rq *rq, const struct task_struct *p)
847 update_load_add(&rq->ls.load, p->se.load.weight);
850 static inline void dec_load(struct rq *rq, const struct task_struct *p)
852 update_load_sub(&rq->ls.load, p->se.load.weight);
855 static void inc_nr_running(struct task_struct *p, struct rq *rq)
861 static void dec_nr_running(struct task_struct *p, struct rq *rq)
867 static void set_load_weight(struct task_struct *p)
869 p->se.wait_runtime = 0;
871 if (task_has_rt_policy(p)) {
872 p->se.load.weight = prio_to_weight[0] * 2;
873 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
878 * SCHED_IDLE tasks get minimal weight:
880 if (p->policy == SCHED_IDLE) {
881 p->se.load.weight = WEIGHT_IDLEPRIO;
882 p->se.load.inv_weight = WMULT_IDLEPRIO;
886 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
887 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
890 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
892 sched_info_queued(p);
893 p->sched_class->enqueue_task(rq, p, wakeup);
897 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
899 p->sched_class->dequeue_task(rq, p, sleep);
904 * __normal_prio - return the priority that is based on the static prio
906 static inline int __normal_prio(struct task_struct *p)
908 return p->static_prio;
912 * Calculate the expected normal priority: i.e. priority
913 * without taking RT-inheritance into account. Might be
914 * boosted by interactivity modifiers. Changes upon fork,
915 * setprio syscalls, and whenever the interactivity
916 * estimator recalculates.
918 static inline int normal_prio(struct task_struct *p)
922 if (task_has_rt_policy(p))
923 prio = MAX_RT_PRIO-1 - p->rt_priority;
925 prio = __normal_prio(p);
930 * Calculate the current priority, i.e. the priority
931 * taken into account by the scheduler. This value might
932 * be boosted by RT tasks, or might be boosted by
933 * interactivity modifiers. Will be RT if the task got
934 * RT-boosted. If not then it returns p->normal_prio.
936 static int effective_prio(struct task_struct *p)
938 p->normal_prio = normal_prio(p);
940 * If we are RT tasks or we were boosted to RT priority,
941 * keep the priority unchanged. Otherwise, update priority
942 * to the normal priority:
944 if (!rt_prio(p->prio))
945 return p->normal_prio;
950 * activate_task - move a task to the runqueue.
952 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
954 if (p->state == TASK_UNINTERRUPTIBLE)
955 rq->nr_uninterruptible--;
957 enqueue_task(rq, p, wakeup);
958 inc_nr_running(p, rq);
962 * activate_idle_task - move idle task to the _front_ of runqueue.
964 static inline void activate_idle_task(struct task_struct *p, struct rq *rq)
968 if (p->state == TASK_UNINTERRUPTIBLE)
969 rq->nr_uninterruptible--;
971 enqueue_task(rq, p, 0);
972 inc_nr_running(p, rq);
976 * deactivate_task - remove a task from the runqueue.
978 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
980 if (p->state == TASK_UNINTERRUPTIBLE)
981 rq->nr_uninterruptible++;
983 dequeue_task(rq, p, sleep);
984 dec_nr_running(p, rq);
988 * task_curr - is this task currently executing on a CPU?
989 * @p: the task in question.
991 inline int task_curr(const struct task_struct *p)
993 return cpu_curr(task_cpu(p)) == p;
996 /* Used instead of source_load when we know the type == 0 */
997 unsigned long weighted_cpuload(const int cpu)
999 return cpu_rq(cpu)->ls.load.weight;
1002 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1005 task_thread_info(p)->cpu = cpu;
1012 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1014 int old_cpu = task_cpu(p);
1015 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1016 u64 clock_offset, fair_clock_offset;
1018 clock_offset = old_rq->clock - new_rq->clock;
1019 fair_clock_offset = old_rq->cfs.fair_clock - new_rq->cfs.fair_clock;
1021 if (p->se.wait_start_fair)
1022 p->se.wait_start_fair -= fair_clock_offset;
1023 if (p->se.sleep_start_fair)
1024 p->se.sleep_start_fair -= fair_clock_offset;
1026 #ifdef CONFIG_SCHEDSTATS
1027 if (p->se.wait_start)
1028 p->se.wait_start -= clock_offset;
1029 if (p->se.sleep_start)
1030 p->se.sleep_start -= clock_offset;
1031 if (p->se.block_start)
1032 p->se.block_start -= clock_offset;
1035 __set_task_cpu(p, new_cpu);
1038 struct migration_req {
1039 struct list_head list;
1041 struct task_struct *task;
1044 struct completion done;
1048 * The task's runqueue lock must be held.
1049 * Returns true if you have to wait for migration thread.
1052 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1054 struct rq *rq = task_rq(p);
1057 * If the task is not on a runqueue (and not running), then
1058 * it is sufficient to simply update the task's cpu field.
1060 if (!p->se.on_rq && !task_running(rq, p)) {
1061 set_task_cpu(p, dest_cpu);
1065 init_completion(&req->done);
1067 req->dest_cpu = dest_cpu;
1068 list_add(&req->list, &rq->migration_queue);
1074 * wait_task_inactive - wait for a thread to unschedule.
1076 * The caller must ensure that the task *will* unschedule sometime soon,
1077 * else this function might spin for a *long* time. This function can't
1078 * be called with interrupts off, or it may introduce deadlock with
1079 * smp_call_function() if an IPI is sent by the same process we are
1080 * waiting to become inactive.
1082 void wait_task_inactive(struct task_struct *p)
1084 unsigned long flags;
1090 * We do the initial early heuristics without holding
1091 * any task-queue locks at all. We'll only try to get
1092 * the runqueue lock when things look like they will
1098 * If the task is actively running on another CPU
1099 * still, just relax and busy-wait without holding
1102 * NOTE! Since we don't hold any locks, it's not
1103 * even sure that "rq" stays as the right runqueue!
1104 * But we don't care, since "task_running()" will
1105 * return false if the runqueue has changed and p
1106 * is actually now running somewhere else!
1108 while (task_running(rq, p))
1112 * Ok, time to look more closely! We need the rq
1113 * lock now, to be *sure*. If we're wrong, we'll
1114 * just go back and repeat.
1116 rq = task_rq_lock(p, &flags);
1117 running = task_running(rq, p);
1118 on_rq = p->se.on_rq;
1119 task_rq_unlock(rq, &flags);
1122 * Was it really running after all now that we
1123 * checked with the proper locks actually held?
1125 * Oops. Go back and try again..
1127 if (unlikely(running)) {
1133 * It's not enough that it's not actively running,
1134 * it must be off the runqueue _entirely_, and not
1137 * So if it wa still runnable (but just not actively
1138 * running right now), it's preempted, and we should
1139 * yield - it could be a while.
1141 if (unlikely(on_rq)) {
1147 * Ahh, all good. It wasn't running, and it wasn't
1148 * runnable, which means that it will never become
1149 * running in the future either. We're all done!
1154 * kick_process - kick a running thread to enter/exit the kernel
1155 * @p: the to-be-kicked thread
1157 * Cause a process which is running on another CPU to enter
1158 * kernel-mode, without any delay. (to get signals handled.)
1160 * NOTE: this function doesnt have to take the runqueue lock,
1161 * because all it wants to ensure is that the remote task enters
1162 * the kernel. If the IPI races and the task has been migrated
1163 * to another CPU then no harm is done and the purpose has been
1166 void kick_process(struct task_struct *p)
1172 if ((cpu != smp_processor_id()) && task_curr(p))
1173 smp_send_reschedule(cpu);
1178 * Return a low guess at the load of a migration-source cpu weighted
1179 * according to the scheduling class and "nice" value.
1181 * We want to under-estimate the load of migration sources, to
1182 * balance conservatively.
1184 static inline unsigned long source_load(int cpu, int type)
1186 struct rq *rq = cpu_rq(cpu);
1187 unsigned long total = weighted_cpuload(cpu);
1192 return min(rq->cpu_load[type-1], total);
1196 * Return a high guess at the load of a migration-target cpu weighted
1197 * according to the scheduling class and "nice" value.
1199 static inline unsigned long target_load(int cpu, int type)
1201 struct rq *rq = cpu_rq(cpu);
1202 unsigned long total = weighted_cpuload(cpu);
1207 return max(rq->cpu_load[type-1], total);
1211 * Return the average load per task on the cpu's run queue
1213 static inline unsigned long cpu_avg_load_per_task(int cpu)
1215 struct rq *rq = cpu_rq(cpu);
1216 unsigned long total = weighted_cpuload(cpu);
1217 unsigned long n = rq->nr_running;
1219 return n ? total / n : SCHED_LOAD_SCALE;
1223 * find_idlest_group finds and returns the least busy CPU group within the
1226 static struct sched_group *
1227 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1229 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1230 unsigned long min_load = ULONG_MAX, this_load = 0;
1231 int load_idx = sd->forkexec_idx;
1232 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1235 unsigned long load, avg_load;
1239 /* Skip over this group if it has no CPUs allowed */
1240 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1243 local_group = cpu_isset(this_cpu, group->cpumask);
1245 /* Tally up the load of all CPUs in the group */
1248 for_each_cpu_mask(i, group->cpumask) {
1249 /* Bias balancing toward cpus of our domain */
1251 load = source_load(i, load_idx);
1253 load = target_load(i, load_idx);
1258 /* Adjust by relative CPU power of the group */
1259 avg_load = sg_div_cpu_power(group,
1260 avg_load * SCHED_LOAD_SCALE);
1263 this_load = avg_load;
1265 } else if (avg_load < min_load) {
1266 min_load = avg_load;
1270 group = group->next;
1271 } while (group != sd->groups);
1273 if (!idlest || 100*this_load < imbalance*min_load)
1279 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1282 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1285 unsigned long load, min_load = ULONG_MAX;
1289 /* Traverse only the allowed CPUs */
1290 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1292 for_each_cpu_mask(i, tmp) {
1293 load = weighted_cpuload(i);
1295 if (load < min_load || (load == min_load && i == this_cpu)) {
1305 * sched_balance_self: balance the current task (running on cpu) in domains
1306 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1309 * Balance, ie. select the least loaded group.
1311 * Returns the target CPU number, or the same CPU if no balancing is needed.
1313 * preempt must be disabled.
1315 static int sched_balance_self(int cpu, int flag)
1317 struct task_struct *t = current;
1318 struct sched_domain *tmp, *sd = NULL;
1320 for_each_domain(cpu, tmp) {
1322 * If power savings logic is enabled for a domain, stop there.
1324 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1326 if (tmp->flags & flag)
1332 struct sched_group *group;
1333 int new_cpu, weight;
1335 if (!(sd->flags & flag)) {
1341 group = find_idlest_group(sd, t, cpu);
1347 new_cpu = find_idlest_cpu(group, t, cpu);
1348 if (new_cpu == -1 || new_cpu == cpu) {
1349 /* Now try balancing at a lower domain level of cpu */
1354 /* Now try balancing at a lower domain level of new_cpu */
1357 weight = cpus_weight(span);
1358 for_each_domain(cpu, tmp) {
1359 if (weight <= cpus_weight(tmp->span))
1361 if (tmp->flags & flag)
1364 /* while loop will break here if sd == NULL */
1370 #endif /* CONFIG_SMP */
1373 * wake_idle() will wake a task on an idle cpu if task->cpu is
1374 * not idle and an idle cpu is available. The span of cpus to
1375 * search starts with cpus closest then further out as needed,
1376 * so we always favor a closer, idle cpu.
1378 * Returns the CPU we should wake onto.
1380 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1381 static int wake_idle(int cpu, struct task_struct *p)
1384 struct sched_domain *sd;
1388 * If it is idle, then it is the best cpu to run this task.
1390 * This cpu is also the best, if it has more than one task already.
1391 * Siblings must be also busy(in most cases) as they didn't already
1392 * pickup the extra load from this cpu and hence we need not check
1393 * sibling runqueue info. This will avoid the checks and cache miss
1394 * penalities associated with that.
1396 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1399 for_each_domain(cpu, sd) {
1400 if (sd->flags & SD_WAKE_IDLE) {
1401 cpus_and(tmp, sd->span, p->cpus_allowed);
1402 for_each_cpu_mask(i, tmp) {
1413 static inline int wake_idle(int cpu, struct task_struct *p)
1420 * try_to_wake_up - wake up a thread
1421 * @p: the to-be-woken-up thread
1422 * @state: the mask of task states that can be woken
1423 * @sync: do a synchronous wakeup?
1425 * Put it on the run-queue if it's not already there. The "current"
1426 * thread is always on the run-queue (except when the actual
1427 * re-schedule is in progress), and as such you're allowed to do
1428 * the simpler "current->state = TASK_RUNNING" to mark yourself
1429 * runnable without the overhead of this.
1431 * returns failure only if the task is already active.
1433 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1435 int cpu, this_cpu, success = 0;
1436 unsigned long flags;
1440 struct sched_domain *sd, *this_sd = NULL;
1441 unsigned long load, this_load;
1445 rq = task_rq_lock(p, &flags);
1446 old_state = p->state;
1447 if (!(old_state & state))
1454 this_cpu = smp_processor_id();
1457 if (unlikely(task_running(rq, p)))
1462 schedstat_inc(rq, ttwu_cnt);
1463 if (cpu == this_cpu) {
1464 schedstat_inc(rq, ttwu_local);
1468 for_each_domain(this_cpu, sd) {
1469 if (cpu_isset(cpu, sd->span)) {
1470 schedstat_inc(sd, ttwu_wake_remote);
1476 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1480 * Check for affine wakeup and passive balancing possibilities.
1483 int idx = this_sd->wake_idx;
1484 unsigned int imbalance;
1486 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1488 load = source_load(cpu, idx);
1489 this_load = target_load(this_cpu, idx);
1491 new_cpu = this_cpu; /* Wake to this CPU if we can */
1493 if (this_sd->flags & SD_WAKE_AFFINE) {
1494 unsigned long tl = this_load;
1495 unsigned long tl_per_task;
1497 tl_per_task = cpu_avg_load_per_task(this_cpu);
1500 * If sync wakeup then subtract the (maximum possible)
1501 * effect of the currently running task from the load
1502 * of the current CPU:
1505 tl -= current->se.load.weight;
1508 tl + target_load(cpu, idx) <= tl_per_task) ||
1509 100*(tl + p->se.load.weight) <= imbalance*load) {
1511 * This domain has SD_WAKE_AFFINE and
1512 * p is cache cold in this domain, and
1513 * there is no bad imbalance.
1515 schedstat_inc(this_sd, ttwu_move_affine);
1521 * Start passive balancing when half the imbalance_pct
1524 if (this_sd->flags & SD_WAKE_BALANCE) {
1525 if (imbalance*this_load <= 100*load) {
1526 schedstat_inc(this_sd, ttwu_move_balance);
1532 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1534 new_cpu = wake_idle(new_cpu, p);
1535 if (new_cpu != cpu) {
1536 set_task_cpu(p, new_cpu);
1537 task_rq_unlock(rq, &flags);
1538 /* might preempt at this point */
1539 rq = task_rq_lock(p, &flags);
1540 old_state = p->state;
1541 if (!(old_state & state))
1546 this_cpu = smp_processor_id();
1551 #endif /* CONFIG_SMP */
1552 update_rq_clock(rq);
1553 activate_task(rq, p, 1);
1555 * Sync wakeups (i.e. those types of wakeups where the waker
1556 * has indicated that it will leave the CPU in short order)
1557 * don't trigger a preemption, if the woken up task will run on
1558 * this cpu. (in this case the 'I will reschedule' promise of
1559 * the waker guarantees that the freshly woken up task is going
1560 * to be considered on this CPU.)
1562 if (!sync || cpu != this_cpu)
1563 check_preempt_curr(rq, p);
1567 p->state = TASK_RUNNING;
1569 task_rq_unlock(rq, &flags);
1574 int fastcall wake_up_process(struct task_struct *p)
1576 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1577 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1579 EXPORT_SYMBOL(wake_up_process);
1581 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1583 return try_to_wake_up(p, state, 0);
1587 * Perform scheduler related setup for a newly forked process p.
1588 * p is forked by current.
1590 * __sched_fork() is basic setup used by init_idle() too:
1592 static void __sched_fork(struct task_struct *p)
1594 p->se.wait_start_fair = 0;
1595 p->se.exec_start = 0;
1596 p->se.sum_exec_runtime = 0;
1597 p->se.prev_sum_exec_runtime = 0;
1598 p->se.wait_runtime = 0;
1599 p->se.sleep_start_fair = 0;
1601 #ifdef CONFIG_SCHEDSTATS
1602 p->se.wait_start = 0;
1603 p->se.sum_wait_runtime = 0;
1604 p->se.sum_sleep_runtime = 0;
1605 p->se.sleep_start = 0;
1606 p->se.block_start = 0;
1607 p->se.sleep_max = 0;
1608 p->se.block_max = 0;
1610 p->se.slice_max = 0;
1612 p->se.wait_runtime_overruns = 0;
1613 p->se.wait_runtime_underruns = 0;
1616 INIT_LIST_HEAD(&p->run_list);
1619 #ifdef CONFIG_PREEMPT_NOTIFIERS
1620 INIT_HLIST_HEAD(&p->preempt_notifiers);
1624 * We mark the process as running here, but have not actually
1625 * inserted it onto the runqueue yet. This guarantees that
1626 * nobody will actually run it, and a signal or other external
1627 * event cannot wake it up and insert it on the runqueue either.
1629 p->state = TASK_RUNNING;
1633 * fork()/clone()-time setup:
1635 void sched_fork(struct task_struct *p, int clone_flags)
1637 int cpu = get_cpu();
1642 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1644 __set_task_cpu(p, cpu);
1647 * Make sure we do not leak PI boosting priority to the child:
1649 p->prio = current->normal_prio;
1651 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1652 if (likely(sched_info_on()))
1653 memset(&p->sched_info, 0, sizeof(p->sched_info));
1655 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1658 #ifdef CONFIG_PREEMPT
1659 /* Want to start with kernel preemption disabled. */
1660 task_thread_info(p)->preempt_count = 1;
1666 * wake_up_new_task - wake up a newly created task for the first time.
1668 * This function will do some initial scheduler statistics housekeeping
1669 * that must be done for every newly created context, then puts the task
1670 * on the runqueue and wakes it.
1672 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1674 unsigned long flags;
1678 rq = task_rq_lock(p, &flags);
1679 BUG_ON(p->state != TASK_RUNNING);
1680 this_cpu = smp_processor_id(); /* parent's CPU */
1681 update_rq_clock(rq);
1683 p->prio = effective_prio(p);
1685 if (rt_prio(p->prio))
1686 p->sched_class = &rt_sched_class;
1688 p->sched_class = &fair_sched_class;
1690 if (task_cpu(p) != this_cpu || !p->sched_class->task_new ||
1691 !current->se.on_rq) {
1692 activate_task(rq, p, 0);
1695 * Let the scheduling class do new task startup
1696 * management (if any):
1698 p->sched_class->task_new(rq, p);
1699 inc_nr_running(p, rq);
1701 check_preempt_curr(rq, p);
1702 task_rq_unlock(rq, &flags);
1705 #ifdef CONFIG_PREEMPT_NOTIFIERS
1708 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1709 * @notifier: notifier struct to register
1711 void preempt_notifier_register(struct preempt_notifier *notifier)
1713 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1715 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1718 * preempt_notifier_unregister - no longer interested in preemption notifications
1719 * @notifier: notifier struct to unregister
1721 * This is safe to call from within a preemption notifier.
1723 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1725 hlist_del(¬ifier->link);
1727 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1729 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1731 struct preempt_notifier *notifier;
1732 struct hlist_node *node;
1734 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1735 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1739 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1740 struct task_struct *next)
1742 struct preempt_notifier *notifier;
1743 struct hlist_node *node;
1745 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1746 notifier->ops->sched_out(notifier, next);
1751 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1756 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1757 struct task_struct *next)
1764 * prepare_task_switch - prepare to switch tasks
1765 * @rq: the runqueue preparing to switch
1766 * @prev: the current task that is being switched out
1767 * @next: the task we are going to switch to.
1769 * This is called with the rq lock held and interrupts off. It must
1770 * be paired with a subsequent finish_task_switch after the context
1773 * prepare_task_switch sets up locking and calls architecture specific
1777 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1778 struct task_struct *next)
1780 fire_sched_out_preempt_notifiers(prev, next);
1781 prepare_lock_switch(rq, next);
1782 prepare_arch_switch(next);
1786 * finish_task_switch - clean up after a task-switch
1787 * @rq: runqueue associated with task-switch
1788 * @prev: the thread we just switched away from.
1790 * finish_task_switch must be called after the context switch, paired
1791 * with a prepare_task_switch call before the context switch.
1792 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1793 * and do any other architecture-specific cleanup actions.
1795 * Note that we may have delayed dropping an mm in context_switch(). If
1796 * so, we finish that here outside of the runqueue lock. (Doing it
1797 * with the lock held can cause deadlocks; see schedule() for
1800 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1801 __releases(rq->lock)
1803 struct mm_struct *mm = rq->prev_mm;
1809 * A task struct has one reference for the use as "current".
1810 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1811 * schedule one last time. The schedule call will never return, and
1812 * the scheduled task must drop that reference.
1813 * The test for TASK_DEAD must occur while the runqueue locks are
1814 * still held, otherwise prev could be scheduled on another cpu, die
1815 * there before we look at prev->state, and then the reference would
1817 * Manfred Spraul <manfred@colorfullife.com>
1819 prev_state = prev->state;
1820 finish_arch_switch(prev);
1821 finish_lock_switch(rq, prev);
1822 fire_sched_in_preempt_notifiers(current);
1825 if (unlikely(prev_state == TASK_DEAD)) {
1827 * Remove function-return probe instances associated with this
1828 * task and put them back on the free list.
1830 kprobe_flush_task(prev);
1831 put_task_struct(prev);
1836 * schedule_tail - first thing a freshly forked thread must call.
1837 * @prev: the thread we just switched away from.
1839 asmlinkage void schedule_tail(struct task_struct *prev)
1840 __releases(rq->lock)
1842 struct rq *rq = this_rq();
1844 finish_task_switch(rq, prev);
1845 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1846 /* In this case, finish_task_switch does not reenable preemption */
1849 if (current->set_child_tid)
1850 put_user(current->pid, current->set_child_tid);
1854 * context_switch - switch to the new MM and the new
1855 * thread's register state.
1858 context_switch(struct rq *rq, struct task_struct *prev,
1859 struct task_struct *next)
1861 struct mm_struct *mm, *oldmm;
1863 prepare_task_switch(rq, prev, next);
1865 oldmm = prev->active_mm;
1867 * For paravirt, this is coupled with an exit in switch_to to
1868 * combine the page table reload and the switch backend into
1871 arch_enter_lazy_cpu_mode();
1873 if (unlikely(!mm)) {
1874 next->active_mm = oldmm;
1875 atomic_inc(&oldmm->mm_count);
1876 enter_lazy_tlb(oldmm, next);
1878 switch_mm(oldmm, mm, next);
1880 if (unlikely(!prev->mm)) {
1881 prev->active_mm = NULL;
1882 rq->prev_mm = oldmm;
1885 * Since the runqueue lock will be released by the next
1886 * task (which is an invalid locking op but in the case
1887 * of the scheduler it's an obvious special-case), so we
1888 * do an early lockdep release here:
1890 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1891 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1894 /* Here we just switch the register state and the stack. */
1895 switch_to(prev, next, prev);
1899 * this_rq must be evaluated again because prev may have moved
1900 * CPUs since it called schedule(), thus the 'rq' on its stack
1901 * frame will be invalid.
1903 finish_task_switch(this_rq(), prev);
1907 * nr_running, nr_uninterruptible and nr_context_switches:
1909 * externally visible scheduler statistics: current number of runnable
1910 * threads, current number of uninterruptible-sleeping threads, total
1911 * number of context switches performed since bootup.
1913 unsigned long nr_running(void)
1915 unsigned long i, sum = 0;
1917 for_each_online_cpu(i)
1918 sum += cpu_rq(i)->nr_running;
1923 unsigned long nr_uninterruptible(void)
1925 unsigned long i, sum = 0;
1927 for_each_possible_cpu(i)
1928 sum += cpu_rq(i)->nr_uninterruptible;
1931 * Since we read the counters lockless, it might be slightly
1932 * inaccurate. Do not allow it to go below zero though:
1934 if (unlikely((long)sum < 0))
1940 unsigned long long nr_context_switches(void)
1943 unsigned long long sum = 0;
1945 for_each_possible_cpu(i)
1946 sum += cpu_rq(i)->nr_switches;
1951 unsigned long nr_iowait(void)
1953 unsigned long i, sum = 0;
1955 for_each_possible_cpu(i)
1956 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1961 unsigned long nr_active(void)
1963 unsigned long i, running = 0, uninterruptible = 0;
1965 for_each_online_cpu(i) {
1966 running += cpu_rq(i)->nr_running;
1967 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1970 if (unlikely((long)uninterruptible < 0))
1971 uninterruptible = 0;
1973 return running + uninterruptible;
1977 * Update rq->cpu_load[] statistics. This function is usually called every
1978 * scheduler tick (TICK_NSEC).
1980 static void update_cpu_load(struct rq *this_rq)
1982 unsigned long this_load = this_rq->ls.load.weight;
1985 this_rq->nr_load_updates++;
1987 /* Update our load: */
1988 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
1989 unsigned long old_load, new_load;
1991 /* scale is effectively 1 << i now, and >> i divides by scale */
1993 old_load = this_rq->cpu_load[i];
1994 new_load = this_load;
1996 * Round up the averaging division if load is increasing. This
1997 * prevents us from getting stuck on 9 if the load is 10, for
2000 if (new_load > old_load)
2001 new_load += scale-1;
2002 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2009 * double_rq_lock - safely lock two runqueues
2011 * Note this does not disable interrupts like task_rq_lock,
2012 * you need to do so manually before calling.
2014 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2015 __acquires(rq1->lock)
2016 __acquires(rq2->lock)
2018 BUG_ON(!irqs_disabled());
2020 spin_lock(&rq1->lock);
2021 __acquire(rq2->lock); /* Fake it out ;) */
2024 spin_lock(&rq1->lock);
2025 spin_lock(&rq2->lock);
2027 spin_lock(&rq2->lock);
2028 spin_lock(&rq1->lock);
2031 update_rq_clock(rq1);
2032 update_rq_clock(rq2);
2036 * double_rq_unlock - safely unlock two runqueues
2038 * Note this does not restore interrupts like task_rq_unlock,
2039 * you need to do so manually after calling.
2041 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2042 __releases(rq1->lock)
2043 __releases(rq2->lock)
2045 spin_unlock(&rq1->lock);
2047 spin_unlock(&rq2->lock);
2049 __release(rq2->lock);
2053 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2055 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2056 __releases(this_rq->lock)
2057 __acquires(busiest->lock)
2058 __acquires(this_rq->lock)
2060 if (unlikely(!irqs_disabled())) {
2061 /* printk() doesn't work good under rq->lock */
2062 spin_unlock(&this_rq->lock);
2065 if (unlikely(!spin_trylock(&busiest->lock))) {
2066 if (busiest < this_rq) {
2067 spin_unlock(&this_rq->lock);
2068 spin_lock(&busiest->lock);
2069 spin_lock(&this_rq->lock);
2071 spin_lock(&busiest->lock);
2076 * If dest_cpu is allowed for this process, migrate the task to it.
2077 * This is accomplished by forcing the cpu_allowed mask to only
2078 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2079 * the cpu_allowed mask is restored.
2081 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2083 struct migration_req req;
2084 unsigned long flags;
2087 rq = task_rq_lock(p, &flags);
2088 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2089 || unlikely(cpu_is_offline(dest_cpu)))
2092 /* force the process onto the specified CPU */
2093 if (migrate_task(p, dest_cpu, &req)) {
2094 /* Need to wait for migration thread (might exit: take ref). */
2095 struct task_struct *mt = rq->migration_thread;
2097 get_task_struct(mt);
2098 task_rq_unlock(rq, &flags);
2099 wake_up_process(mt);
2100 put_task_struct(mt);
2101 wait_for_completion(&req.done);
2106 task_rq_unlock(rq, &flags);
2110 * sched_exec - execve() is a valuable balancing opportunity, because at
2111 * this point the task has the smallest effective memory and cache footprint.
2113 void sched_exec(void)
2115 int new_cpu, this_cpu = get_cpu();
2116 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2118 if (new_cpu != this_cpu)
2119 sched_migrate_task(current, new_cpu);
2123 * pull_task - move a task from a remote runqueue to the local runqueue.
2124 * Both runqueues must be locked.
2126 static void pull_task(struct rq *src_rq, struct task_struct *p,
2127 struct rq *this_rq, int this_cpu)
2129 deactivate_task(src_rq, p, 0);
2130 set_task_cpu(p, this_cpu);
2131 activate_task(this_rq, p, 0);
2133 * Note that idle threads have a prio of MAX_PRIO, for this test
2134 * to be always true for them.
2136 check_preempt_curr(this_rq, p);
2140 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2143 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2144 struct sched_domain *sd, enum cpu_idle_type idle,
2148 * We do not migrate tasks that are:
2149 * 1) running (obviously), or
2150 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2151 * 3) are cache-hot on their current CPU.
2153 if (!cpu_isset(this_cpu, p->cpus_allowed))
2157 if (task_running(rq, p))
2163 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2164 unsigned long max_nr_move, unsigned long max_load_move,
2165 struct sched_domain *sd, enum cpu_idle_type idle,
2166 int *all_pinned, unsigned long *load_moved,
2167 int *this_best_prio, struct rq_iterator *iterator)
2169 int pulled = 0, pinned = 0, skip_for_load;
2170 struct task_struct *p;
2171 long rem_load_move = max_load_move;
2173 if (max_nr_move == 0 || max_load_move == 0)
2179 * Start the load-balancing iterator:
2181 p = iterator->start(iterator->arg);
2186 * To help distribute high priority tasks accross CPUs we don't
2187 * skip a task if it will be the highest priority task (i.e. smallest
2188 * prio value) on its new queue regardless of its load weight
2190 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2191 SCHED_LOAD_SCALE_FUZZ;
2192 if ((skip_for_load && p->prio >= *this_best_prio) ||
2193 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2194 p = iterator->next(iterator->arg);
2198 pull_task(busiest, p, this_rq, this_cpu);
2200 rem_load_move -= p->se.load.weight;
2203 * We only want to steal up to the prescribed number of tasks
2204 * and the prescribed amount of weighted load.
2206 if (pulled < max_nr_move && rem_load_move > 0) {
2207 if (p->prio < *this_best_prio)
2208 *this_best_prio = p->prio;
2209 p = iterator->next(iterator->arg);
2214 * Right now, this is the only place pull_task() is called,
2215 * so we can safely collect pull_task() stats here rather than
2216 * inside pull_task().
2218 schedstat_add(sd, lb_gained[idle], pulled);
2221 *all_pinned = pinned;
2222 *load_moved = max_load_move - rem_load_move;
2227 * move_tasks tries to move up to max_load_move weighted load from busiest to
2228 * this_rq, as part of a balancing operation within domain "sd".
2229 * Returns 1 if successful and 0 otherwise.
2231 * Called with both runqueues locked.
2233 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2234 unsigned long max_load_move,
2235 struct sched_domain *sd, enum cpu_idle_type idle,
2238 struct sched_class *class = sched_class_highest;
2239 unsigned long total_load_moved = 0;
2240 int this_best_prio = this_rq->curr->prio;
2244 class->load_balance(this_rq, this_cpu, busiest,
2245 ULONG_MAX, max_load_move - total_load_moved,
2246 sd, idle, all_pinned, &this_best_prio);
2247 class = class->next;
2248 } while (class && max_load_move > total_load_moved);
2250 return total_load_moved > 0;
2254 * move_one_task tries to move exactly one task from busiest to this_rq, as
2255 * part of active balancing operations within "domain".
2256 * Returns 1 if successful and 0 otherwise.
2258 * Called with both runqueues locked.
2260 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2261 struct sched_domain *sd, enum cpu_idle_type idle)
2263 struct sched_class *class;
2264 int this_best_prio = MAX_PRIO;
2266 for (class = sched_class_highest; class; class = class->next)
2267 if (class->load_balance(this_rq, this_cpu, busiest,
2268 1, ULONG_MAX, sd, idle, NULL,
2276 * find_busiest_group finds and returns the busiest CPU group within the
2277 * domain. It calculates and returns the amount of weighted load which
2278 * should be moved to restore balance via the imbalance parameter.
2280 static struct sched_group *
2281 find_busiest_group(struct sched_domain *sd, int this_cpu,
2282 unsigned long *imbalance, enum cpu_idle_type idle,
2283 int *sd_idle, cpumask_t *cpus, int *balance)
2285 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2286 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2287 unsigned long max_pull;
2288 unsigned long busiest_load_per_task, busiest_nr_running;
2289 unsigned long this_load_per_task, this_nr_running;
2291 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2292 int power_savings_balance = 1;
2293 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2294 unsigned long min_nr_running = ULONG_MAX;
2295 struct sched_group *group_min = NULL, *group_leader = NULL;
2298 max_load = this_load = total_load = total_pwr = 0;
2299 busiest_load_per_task = busiest_nr_running = 0;
2300 this_load_per_task = this_nr_running = 0;
2301 if (idle == CPU_NOT_IDLE)
2302 load_idx = sd->busy_idx;
2303 else if (idle == CPU_NEWLY_IDLE)
2304 load_idx = sd->newidle_idx;
2306 load_idx = sd->idle_idx;
2309 unsigned long load, group_capacity;
2312 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2313 unsigned long sum_nr_running, sum_weighted_load;
2315 local_group = cpu_isset(this_cpu, group->cpumask);
2318 balance_cpu = first_cpu(group->cpumask);
2320 /* Tally up the load of all CPUs in the group */
2321 sum_weighted_load = sum_nr_running = avg_load = 0;
2323 for_each_cpu_mask(i, group->cpumask) {
2326 if (!cpu_isset(i, *cpus))
2331 if (*sd_idle && rq->nr_running)
2334 /* Bias balancing toward cpus of our domain */
2336 if (idle_cpu(i) && !first_idle_cpu) {
2341 load = target_load(i, load_idx);
2343 load = source_load(i, load_idx);
2346 sum_nr_running += rq->nr_running;
2347 sum_weighted_load += weighted_cpuload(i);
2351 * First idle cpu or the first cpu(busiest) in this sched group
2352 * is eligible for doing load balancing at this and above
2353 * domains. In the newly idle case, we will allow all the cpu's
2354 * to do the newly idle load balance.
2356 if (idle != CPU_NEWLY_IDLE && local_group &&
2357 balance_cpu != this_cpu && balance) {
2362 total_load += avg_load;
2363 total_pwr += group->__cpu_power;
2365 /* Adjust by relative CPU power of the group */
2366 avg_load = sg_div_cpu_power(group,
2367 avg_load * SCHED_LOAD_SCALE);
2369 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2372 this_load = avg_load;
2374 this_nr_running = sum_nr_running;
2375 this_load_per_task = sum_weighted_load;
2376 } else if (avg_load > max_load &&
2377 sum_nr_running > group_capacity) {
2378 max_load = avg_load;
2380 busiest_nr_running = sum_nr_running;
2381 busiest_load_per_task = sum_weighted_load;
2384 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2386 * Busy processors will not participate in power savings
2389 if (idle == CPU_NOT_IDLE ||
2390 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2394 * If the local group is idle or completely loaded
2395 * no need to do power savings balance at this domain
2397 if (local_group && (this_nr_running >= group_capacity ||
2399 power_savings_balance = 0;
2402 * If a group is already running at full capacity or idle,
2403 * don't include that group in power savings calculations
2405 if (!power_savings_balance || sum_nr_running >= group_capacity
2410 * Calculate the group which has the least non-idle load.
2411 * This is the group from where we need to pick up the load
2414 if ((sum_nr_running < min_nr_running) ||
2415 (sum_nr_running == min_nr_running &&
2416 first_cpu(group->cpumask) <
2417 first_cpu(group_min->cpumask))) {
2419 min_nr_running = sum_nr_running;
2420 min_load_per_task = sum_weighted_load /
2425 * Calculate the group which is almost near its
2426 * capacity but still has some space to pick up some load
2427 * from other group and save more power
2429 if (sum_nr_running <= group_capacity - 1) {
2430 if (sum_nr_running > leader_nr_running ||
2431 (sum_nr_running == leader_nr_running &&
2432 first_cpu(group->cpumask) >
2433 first_cpu(group_leader->cpumask))) {
2434 group_leader = group;
2435 leader_nr_running = sum_nr_running;
2440 group = group->next;
2441 } while (group != sd->groups);
2443 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2446 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2448 if (this_load >= avg_load ||
2449 100*max_load <= sd->imbalance_pct*this_load)
2452 busiest_load_per_task /= busiest_nr_running;
2454 * We're trying to get all the cpus to the average_load, so we don't
2455 * want to push ourselves above the average load, nor do we wish to
2456 * reduce the max loaded cpu below the average load, as either of these
2457 * actions would just result in more rebalancing later, and ping-pong
2458 * tasks around. Thus we look for the minimum possible imbalance.
2459 * Negative imbalances (*we* are more loaded than anyone else) will
2460 * be counted as no imbalance for these purposes -- we can't fix that
2461 * by pulling tasks to us. Be careful of negative numbers as they'll
2462 * appear as very large values with unsigned longs.
2464 if (max_load <= busiest_load_per_task)
2468 * In the presence of smp nice balancing, certain scenarios can have
2469 * max load less than avg load(as we skip the groups at or below
2470 * its cpu_power, while calculating max_load..)
2472 if (max_load < avg_load) {
2474 goto small_imbalance;
2477 /* Don't want to pull so many tasks that a group would go idle */
2478 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2480 /* How much load to actually move to equalise the imbalance */
2481 *imbalance = min(max_pull * busiest->__cpu_power,
2482 (avg_load - this_load) * this->__cpu_power)
2486 * if *imbalance is less than the average load per runnable task
2487 * there is no gaurantee that any tasks will be moved so we'll have
2488 * a think about bumping its value to force at least one task to be
2491 if (*imbalance < busiest_load_per_task) {
2492 unsigned long tmp, pwr_now, pwr_move;
2496 pwr_move = pwr_now = 0;
2498 if (this_nr_running) {
2499 this_load_per_task /= this_nr_running;
2500 if (busiest_load_per_task > this_load_per_task)
2503 this_load_per_task = SCHED_LOAD_SCALE;
2505 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2506 busiest_load_per_task * imbn) {
2507 *imbalance = busiest_load_per_task;
2512 * OK, we don't have enough imbalance to justify moving tasks,
2513 * however we may be able to increase total CPU power used by
2517 pwr_now += busiest->__cpu_power *
2518 min(busiest_load_per_task, max_load);
2519 pwr_now += this->__cpu_power *
2520 min(this_load_per_task, this_load);
2521 pwr_now /= SCHED_LOAD_SCALE;
2523 /* Amount of load we'd subtract */
2524 tmp = sg_div_cpu_power(busiest,
2525 busiest_load_per_task * SCHED_LOAD_SCALE);
2527 pwr_move += busiest->__cpu_power *
2528 min(busiest_load_per_task, max_load - tmp);
2530 /* Amount of load we'd add */
2531 if (max_load * busiest->__cpu_power <
2532 busiest_load_per_task * SCHED_LOAD_SCALE)
2533 tmp = sg_div_cpu_power(this,
2534 max_load * busiest->__cpu_power);
2536 tmp = sg_div_cpu_power(this,
2537 busiest_load_per_task * SCHED_LOAD_SCALE);
2538 pwr_move += this->__cpu_power *
2539 min(this_load_per_task, this_load + tmp);
2540 pwr_move /= SCHED_LOAD_SCALE;
2542 /* Move if we gain throughput */
2543 if (pwr_move > pwr_now)
2544 *imbalance = busiest_load_per_task;
2550 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2551 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2554 if (this == group_leader && group_leader != group_min) {
2555 *imbalance = min_load_per_task;
2565 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2568 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2569 unsigned long imbalance, cpumask_t *cpus)
2571 struct rq *busiest = NULL, *rq;
2572 unsigned long max_load = 0;
2575 for_each_cpu_mask(i, group->cpumask) {
2578 if (!cpu_isset(i, *cpus))
2582 wl = weighted_cpuload(i);
2584 if (rq->nr_running == 1 && wl > imbalance)
2587 if (wl > max_load) {
2597 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2598 * so long as it is large enough.
2600 #define MAX_PINNED_INTERVAL 512
2603 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2604 * tasks if there is an imbalance.
2606 static int load_balance(int this_cpu, struct rq *this_rq,
2607 struct sched_domain *sd, enum cpu_idle_type idle,
2610 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2611 struct sched_group *group;
2612 unsigned long imbalance;
2614 cpumask_t cpus = CPU_MASK_ALL;
2615 unsigned long flags;
2618 * When power savings policy is enabled for the parent domain, idle
2619 * sibling can pick up load irrespective of busy siblings. In this case,
2620 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2621 * portraying it as CPU_NOT_IDLE.
2623 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2624 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2627 schedstat_inc(sd, lb_cnt[idle]);
2630 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2637 schedstat_inc(sd, lb_nobusyg[idle]);
2641 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2643 schedstat_inc(sd, lb_nobusyq[idle]);
2647 BUG_ON(busiest == this_rq);
2649 schedstat_add(sd, lb_imbalance[idle], imbalance);
2652 if (busiest->nr_running > 1) {
2654 * Attempt to move tasks. If find_busiest_group has found
2655 * an imbalance but busiest->nr_running <= 1, the group is
2656 * still unbalanced. ld_moved simply stays zero, so it is
2657 * correctly treated as an imbalance.
2659 local_irq_save(flags);
2660 double_rq_lock(this_rq, busiest);
2661 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2662 imbalance, sd, idle, &all_pinned);
2663 double_rq_unlock(this_rq, busiest);
2664 local_irq_restore(flags);
2667 * some other cpu did the load balance for us.
2669 if (ld_moved && this_cpu != smp_processor_id())
2670 resched_cpu(this_cpu);
2672 /* All tasks on this runqueue were pinned by CPU affinity */
2673 if (unlikely(all_pinned)) {
2674 cpu_clear(cpu_of(busiest), cpus);
2675 if (!cpus_empty(cpus))
2682 schedstat_inc(sd, lb_failed[idle]);
2683 sd->nr_balance_failed++;
2685 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2687 spin_lock_irqsave(&busiest->lock, flags);
2689 /* don't kick the migration_thread, if the curr
2690 * task on busiest cpu can't be moved to this_cpu
2692 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2693 spin_unlock_irqrestore(&busiest->lock, flags);
2695 goto out_one_pinned;
2698 if (!busiest->active_balance) {
2699 busiest->active_balance = 1;
2700 busiest->push_cpu = this_cpu;
2703 spin_unlock_irqrestore(&busiest->lock, flags);
2705 wake_up_process(busiest->migration_thread);
2708 * We've kicked active balancing, reset the failure
2711 sd->nr_balance_failed = sd->cache_nice_tries+1;
2714 sd->nr_balance_failed = 0;
2716 if (likely(!active_balance)) {
2717 /* We were unbalanced, so reset the balancing interval */
2718 sd->balance_interval = sd->min_interval;
2721 * If we've begun active balancing, start to back off. This
2722 * case may not be covered by the all_pinned logic if there
2723 * is only 1 task on the busy runqueue (because we don't call
2726 if (sd->balance_interval < sd->max_interval)
2727 sd->balance_interval *= 2;
2730 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2731 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2736 schedstat_inc(sd, lb_balanced[idle]);
2738 sd->nr_balance_failed = 0;
2741 /* tune up the balancing interval */
2742 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2743 (sd->balance_interval < sd->max_interval))
2744 sd->balance_interval *= 2;
2746 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2747 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2753 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2754 * tasks if there is an imbalance.
2756 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2757 * this_rq is locked.
2760 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2762 struct sched_group *group;
2763 struct rq *busiest = NULL;
2764 unsigned long imbalance;
2768 cpumask_t cpus = CPU_MASK_ALL;
2771 * When power savings policy is enabled for the parent domain, idle
2772 * sibling can pick up load irrespective of busy siblings. In this case,
2773 * let the state of idle sibling percolate up as IDLE, instead of
2774 * portraying it as CPU_NOT_IDLE.
2776 if (sd->flags & SD_SHARE_CPUPOWER &&
2777 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2780 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2782 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2783 &sd_idle, &cpus, NULL);
2785 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2789 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2792 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2796 BUG_ON(busiest == this_rq);
2798 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2801 if (busiest->nr_running > 1) {
2802 /* Attempt to move tasks */
2803 double_lock_balance(this_rq, busiest);
2804 /* this_rq->clock is already updated */
2805 update_rq_clock(busiest);
2806 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2807 imbalance, sd, CPU_NEWLY_IDLE,
2809 spin_unlock(&busiest->lock);
2811 if (unlikely(all_pinned)) {
2812 cpu_clear(cpu_of(busiest), cpus);
2813 if (!cpus_empty(cpus))
2819 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2820 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2821 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2824 sd->nr_balance_failed = 0;
2829 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2830 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2831 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2833 sd->nr_balance_failed = 0;
2839 * idle_balance is called by schedule() if this_cpu is about to become
2840 * idle. Attempts to pull tasks from other CPUs.
2842 static void idle_balance(int this_cpu, struct rq *this_rq)
2844 struct sched_domain *sd;
2845 int pulled_task = -1;
2846 unsigned long next_balance = jiffies + HZ;
2848 for_each_domain(this_cpu, sd) {
2849 unsigned long interval;
2851 if (!(sd->flags & SD_LOAD_BALANCE))
2854 if (sd->flags & SD_BALANCE_NEWIDLE)
2855 /* If we've pulled tasks over stop searching: */
2856 pulled_task = load_balance_newidle(this_cpu,
2859 interval = msecs_to_jiffies(sd->balance_interval);
2860 if (time_after(next_balance, sd->last_balance + interval))
2861 next_balance = sd->last_balance + interval;
2865 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2867 * We are going idle. next_balance may be set based on
2868 * a busy processor. So reset next_balance.
2870 this_rq->next_balance = next_balance;
2875 * active_load_balance is run by migration threads. It pushes running tasks
2876 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2877 * running on each physical CPU where possible, and avoids physical /
2878 * logical imbalances.
2880 * Called with busiest_rq locked.
2882 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2884 int target_cpu = busiest_rq->push_cpu;
2885 struct sched_domain *sd;
2886 struct rq *target_rq;
2888 /* Is there any task to move? */
2889 if (busiest_rq->nr_running <= 1)
2892 target_rq = cpu_rq(target_cpu);
2895 * This condition is "impossible", if it occurs
2896 * we need to fix it. Originally reported by
2897 * Bjorn Helgaas on a 128-cpu setup.
2899 BUG_ON(busiest_rq == target_rq);
2901 /* move a task from busiest_rq to target_rq */
2902 double_lock_balance(busiest_rq, target_rq);
2903 update_rq_clock(busiest_rq);
2904 update_rq_clock(target_rq);
2906 /* Search for an sd spanning us and the target CPU. */
2907 for_each_domain(target_cpu, sd) {
2908 if ((sd->flags & SD_LOAD_BALANCE) &&
2909 cpu_isset(busiest_cpu, sd->span))
2914 schedstat_inc(sd, alb_cnt);
2916 if (move_one_task(target_rq, target_cpu, busiest_rq,
2918 schedstat_inc(sd, alb_pushed);
2920 schedstat_inc(sd, alb_failed);
2922 spin_unlock(&target_rq->lock);
2927 atomic_t load_balancer;
2929 } nohz ____cacheline_aligned = {
2930 .load_balancer = ATOMIC_INIT(-1),
2931 .cpu_mask = CPU_MASK_NONE,
2935 * This routine will try to nominate the ilb (idle load balancing)
2936 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2937 * load balancing on behalf of all those cpus. If all the cpus in the system
2938 * go into this tickless mode, then there will be no ilb owner (as there is
2939 * no need for one) and all the cpus will sleep till the next wakeup event
2942 * For the ilb owner, tick is not stopped. And this tick will be used
2943 * for idle load balancing. ilb owner will still be part of
2946 * While stopping the tick, this cpu will become the ilb owner if there
2947 * is no other owner. And will be the owner till that cpu becomes busy
2948 * or if all cpus in the system stop their ticks at which point
2949 * there is no need for ilb owner.
2951 * When the ilb owner becomes busy, it nominates another owner, during the
2952 * next busy scheduler_tick()
2954 int select_nohz_load_balancer(int stop_tick)
2956 int cpu = smp_processor_id();
2959 cpu_set(cpu, nohz.cpu_mask);
2960 cpu_rq(cpu)->in_nohz_recently = 1;
2963 * If we are going offline and still the leader, give up!
2965 if (cpu_is_offline(cpu) &&
2966 atomic_read(&nohz.load_balancer) == cpu) {
2967 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2972 /* time for ilb owner also to sleep */
2973 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2974 if (atomic_read(&nohz.load_balancer) == cpu)
2975 atomic_set(&nohz.load_balancer, -1);
2979 if (atomic_read(&nohz.load_balancer) == -1) {
2980 /* make me the ilb owner */
2981 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
2983 } else if (atomic_read(&nohz.load_balancer) == cpu)
2986 if (!cpu_isset(cpu, nohz.cpu_mask))
2989 cpu_clear(cpu, nohz.cpu_mask);
2991 if (atomic_read(&nohz.load_balancer) == cpu)
2992 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2999 static DEFINE_SPINLOCK(balancing);
3002 * It checks each scheduling domain to see if it is due to be balanced,
3003 * and initiates a balancing operation if so.
3005 * Balancing parameters are set up in arch_init_sched_domains.
3007 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
3010 struct rq *rq = cpu_rq(cpu);
3011 unsigned long interval;
3012 struct sched_domain *sd;
3013 /* Earliest time when we have to do rebalance again */
3014 unsigned long next_balance = jiffies + 60*HZ;
3015 int update_next_balance = 0;
3017 for_each_domain(cpu, sd) {
3018 if (!(sd->flags & SD_LOAD_BALANCE))
3021 interval = sd->balance_interval;
3022 if (idle != CPU_IDLE)
3023 interval *= sd->busy_factor;
3025 /* scale ms to jiffies */
3026 interval = msecs_to_jiffies(interval);
3027 if (unlikely(!interval))
3029 if (interval > HZ*NR_CPUS/10)
3030 interval = HZ*NR_CPUS/10;
3033 if (sd->flags & SD_SERIALIZE) {
3034 if (!spin_trylock(&balancing))
3038 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3039 if (load_balance(cpu, rq, sd, idle, &balance)) {
3041 * We've pulled tasks over so either we're no
3042 * longer idle, or one of our SMT siblings is
3045 idle = CPU_NOT_IDLE;
3047 sd->last_balance = jiffies;
3049 if (sd->flags & SD_SERIALIZE)
3050 spin_unlock(&balancing);
3052 if (time_after(next_balance, sd->last_balance + interval)) {
3053 next_balance = sd->last_balance + interval;
3054 update_next_balance = 1;
3058 * Stop the load balance at this level. There is another
3059 * CPU in our sched group which is doing load balancing more
3067 * next_balance will be updated only when there is a need.
3068 * When the cpu is attached to null domain for ex, it will not be
3071 if (likely(update_next_balance))
3072 rq->next_balance = next_balance;
3076 * run_rebalance_domains is triggered when needed from the scheduler tick.
3077 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3078 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3080 static void run_rebalance_domains(struct softirq_action *h)
3082 int this_cpu = smp_processor_id();
3083 struct rq *this_rq = cpu_rq(this_cpu);
3084 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3085 CPU_IDLE : CPU_NOT_IDLE;
3087 rebalance_domains(this_cpu, idle);
3091 * If this cpu is the owner for idle load balancing, then do the
3092 * balancing on behalf of the other idle cpus whose ticks are
3095 if (this_rq->idle_at_tick &&
3096 atomic_read(&nohz.load_balancer) == this_cpu) {
3097 cpumask_t cpus = nohz.cpu_mask;
3101 cpu_clear(this_cpu, cpus);
3102 for_each_cpu_mask(balance_cpu, cpus) {
3104 * If this cpu gets work to do, stop the load balancing
3105 * work being done for other cpus. Next load
3106 * balancing owner will pick it up.
3111 rebalance_domains(balance_cpu, CPU_IDLE);
3113 rq = cpu_rq(balance_cpu);
3114 if (time_after(this_rq->next_balance, rq->next_balance))
3115 this_rq->next_balance = rq->next_balance;
3122 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3124 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3125 * idle load balancing owner or decide to stop the periodic load balancing,
3126 * if the whole system is idle.
3128 static inline void trigger_load_balance(struct rq *rq, int cpu)
3132 * If we were in the nohz mode recently and busy at the current
3133 * scheduler tick, then check if we need to nominate new idle
3136 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3137 rq->in_nohz_recently = 0;
3139 if (atomic_read(&nohz.load_balancer) == cpu) {
3140 cpu_clear(cpu, nohz.cpu_mask);
3141 atomic_set(&nohz.load_balancer, -1);
3144 if (atomic_read(&nohz.load_balancer) == -1) {
3146 * simple selection for now: Nominate the
3147 * first cpu in the nohz list to be the next
3150 * TBD: Traverse the sched domains and nominate
3151 * the nearest cpu in the nohz.cpu_mask.
3153 int ilb = first_cpu(nohz.cpu_mask);
3161 * If this cpu is idle and doing idle load balancing for all the
3162 * cpus with ticks stopped, is it time for that to stop?
3164 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3165 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3171 * If this cpu is idle and the idle load balancing is done by
3172 * someone else, then no need raise the SCHED_SOFTIRQ
3174 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3175 cpu_isset(cpu, nohz.cpu_mask))
3178 if (time_after_eq(jiffies, rq->next_balance))
3179 raise_softirq(SCHED_SOFTIRQ);
3182 #else /* CONFIG_SMP */
3185 * on UP we do not need to balance between CPUs:
3187 static inline void idle_balance(int cpu, struct rq *rq)
3191 /* Avoid "used but not defined" warning on UP */
3192 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3193 unsigned long max_nr_move, unsigned long max_load_move,
3194 struct sched_domain *sd, enum cpu_idle_type idle,
3195 int *all_pinned, unsigned long *load_moved,
3196 int *this_best_prio, struct rq_iterator *iterator)
3205 DEFINE_PER_CPU(struct kernel_stat, kstat);
3207 EXPORT_PER_CPU_SYMBOL(kstat);
3210 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3211 * that have not yet been banked in case the task is currently running.
3213 unsigned long long task_sched_runtime(struct task_struct *p)
3215 unsigned long flags;
3219 rq = task_rq_lock(p, &flags);
3220 ns = p->se.sum_exec_runtime;
3221 if (rq->curr == p) {
3222 update_rq_clock(rq);
3223 delta_exec = rq->clock - p->se.exec_start;
3224 if ((s64)delta_exec > 0)
3227 task_rq_unlock(rq, &flags);
3233 * Account user cpu time to a process.
3234 * @p: the process that the cpu time gets accounted to
3235 * @hardirq_offset: the offset to subtract from hardirq_count()
3236 * @cputime: the cpu time spent in user space since the last update
3238 void account_user_time(struct task_struct *p, cputime_t cputime)
3240 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3243 p->utime = cputime_add(p->utime, cputime);
3245 /* Add user time to cpustat. */
3246 tmp = cputime_to_cputime64(cputime);
3247 if (TASK_NICE(p) > 0)
3248 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3250 cpustat->user = cputime64_add(cpustat->user, tmp);
3254 * Account system cpu time to a process.
3255 * @p: the process that the cpu time gets accounted to
3256 * @hardirq_offset: the offset to subtract from hardirq_count()
3257 * @cputime: the cpu time spent in kernel space since the last update
3259 void account_system_time(struct task_struct *p, int hardirq_offset,
3262 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3263 struct rq *rq = this_rq();
3266 p->stime = cputime_add(p->stime, cputime);
3268 /* Add system time to cpustat. */
3269 tmp = cputime_to_cputime64(cputime);
3270 if (hardirq_count() - hardirq_offset)
3271 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3272 else if (softirq_count())
3273 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3274 else if (p != rq->idle)
3275 cpustat->system = cputime64_add(cpustat->system, tmp);
3276 else if (atomic_read(&rq->nr_iowait) > 0)
3277 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3279 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3280 /* Account for system time used */
3281 acct_update_integrals(p);
3285 * Account for involuntary wait time.
3286 * @p: the process from which the cpu time has been stolen
3287 * @steal: the cpu time spent in involuntary wait
3289 void account_steal_time(struct task_struct *p, cputime_t steal)
3291 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3292 cputime64_t tmp = cputime_to_cputime64(steal);
3293 struct rq *rq = this_rq();
3295 if (p == rq->idle) {
3296 p->stime = cputime_add(p->stime, steal);
3297 if (atomic_read(&rq->nr_iowait) > 0)
3298 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3300 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3302 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3306 * This function gets called by the timer code, with HZ frequency.
3307 * We call it with interrupts disabled.
3309 * It also gets called by the fork code, when changing the parent's
3312 void scheduler_tick(void)
3314 int cpu = smp_processor_id();
3315 struct rq *rq = cpu_rq(cpu);
3316 struct task_struct *curr = rq->curr;
3317 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3319 spin_lock(&rq->lock);
3320 __update_rq_clock(rq);
3322 * Let rq->clock advance by at least TICK_NSEC:
3324 if (unlikely(rq->clock < next_tick))
3325 rq->clock = next_tick;
3326 rq->tick_timestamp = rq->clock;
3327 update_cpu_load(rq);
3328 if (curr != rq->idle) /* FIXME: needed? */
3329 curr->sched_class->task_tick(rq, curr);
3330 spin_unlock(&rq->lock);
3333 rq->idle_at_tick = idle_cpu(cpu);
3334 trigger_load_balance(rq, cpu);
3338 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3340 void fastcall add_preempt_count(int val)
3345 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3347 preempt_count() += val;
3349 * Spinlock count overflowing soon?
3351 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3354 EXPORT_SYMBOL(add_preempt_count);
3356 void fastcall sub_preempt_count(int val)
3361 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3364 * Is the spinlock portion underflowing?
3366 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3367 !(preempt_count() & PREEMPT_MASK)))
3370 preempt_count() -= val;
3372 EXPORT_SYMBOL(sub_preempt_count);
3377 * Print scheduling while atomic bug:
3379 static noinline void __schedule_bug(struct task_struct *prev)
3381 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3382 prev->comm, preempt_count(), prev->pid);
3383 debug_show_held_locks(prev);
3384 if (irqs_disabled())
3385 print_irqtrace_events(prev);
3390 * Various schedule()-time debugging checks and statistics:
3392 static inline void schedule_debug(struct task_struct *prev)
3395 * Test if we are atomic. Since do_exit() needs to call into
3396 * schedule() atomically, we ignore that path for now.
3397 * Otherwise, whine if we are scheduling when we should not be.
3399 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3400 __schedule_bug(prev);
3402 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3404 schedstat_inc(this_rq(), sched_cnt);
3408 * Pick up the highest-prio task:
3410 static inline struct task_struct *
3411 pick_next_task(struct rq *rq, struct task_struct *prev)
3413 struct sched_class *class;
3414 struct task_struct *p;
3417 * Optimization: we know that if all tasks are in
3418 * the fair class we can call that function directly:
3420 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3421 p = fair_sched_class.pick_next_task(rq);
3426 class = sched_class_highest;
3428 p = class->pick_next_task(rq);
3432 * Will never be NULL as the idle class always
3433 * returns a non-NULL p:
3435 class = class->next;
3440 * schedule() is the main scheduler function.
3442 asmlinkage void __sched schedule(void)
3444 struct task_struct *prev, *next;
3451 cpu = smp_processor_id();
3455 switch_count = &prev->nivcsw;
3457 release_kernel_lock(prev);
3458 need_resched_nonpreemptible:
3460 schedule_debug(prev);
3462 spin_lock_irq(&rq->lock);
3463 clear_tsk_need_resched(prev);
3464 __update_rq_clock(rq);
3466 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3467 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3468 unlikely(signal_pending(prev)))) {
3469 prev->state = TASK_RUNNING;
3471 deactivate_task(rq, prev, 1);
3473 switch_count = &prev->nvcsw;
3476 if (unlikely(!rq->nr_running))
3477 idle_balance(cpu, rq);
3479 prev->sched_class->put_prev_task(rq, prev);
3480 next = pick_next_task(rq, prev);
3482 sched_info_switch(prev, next);
3484 if (likely(prev != next)) {
3489 context_switch(rq, prev, next); /* unlocks the rq */
3491 spin_unlock_irq(&rq->lock);
3493 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3494 cpu = smp_processor_id();
3496 goto need_resched_nonpreemptible;
3498 preempt_enable_no_resched();
3499 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3502 EXPORT_SYMBOL(schedule);
3504 #ifdef CONFIG_PREEMPT
3506 * this is the entry point to schedule() from in-kernel preemption
3507 * off of preempt_enable. Kernel preemptions off return from interrupt
3508 * occur there and call schedule directly.
3510 asmlinkage void __sched preempt_schedule(void)
3512 struct thread_info *ti = current_thread_info();
3513 #ifdef CONFIG_PREEMPT_BKL
3514 struct task_struct *task = current;
3515 int saved_lock_depth;
3518 * If there is a non-zero preempt_count or interrupts are disabled,
3519 * we do not want to preempt the current task. Just return..
3521 if (likely(ti->preempt_count || irqs_disabled()))
3525 add_preempt_count(PREEMPT_ACTIVE);
3527 * We keep the big kernel semaphore locked, but we
3528 * clear ->lock_depth so that schedule() doesnt
3529 * auto-release the semaphore:
3531 #ifdef CONFIG_PREEMPT_BKL
3532 saved_lock_depth = task->lock_depth;
3533 task->lock_depth = -1;
3536 #ifdef CONFIG_PREEMPT_BKL
3537 task->lock_depth = saved_lock_depth;
3539 sub_preempt_count(PREEMPT_ACTIVE);
3541 /* we could miss a preemption opportunity between schedule and now */
3543 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3546 EXPORT_SYMBOL(preempt_schedule);
3549 * this is the entry point to schedule() from kernel preemption
3550 * off of irq context.
3551 * Note, that this is called and return with irqs disabled. This will
3552 * protect us against recursive calling from irq.
3554 asmlinkage void __sched preempt_schedule_irq(void)
3556 struct thread_info *ti = current_thread_info();
3557 #ifdef CONFIG_PREEMPT_BKL
3558 struct task_struct *task = current;
3559 int saved_lock_depth;
3561 /* Catch callers which need to be fixed */
3562 BUG_ON(ti->preempt_count || !irqs_disabled());
3565 add_preempt_count(PREEMPT_ACTIVE);
3567 * We keep the big kernel semaphore locked, but we
3568 * clear ->lock_depth so that schedule() doesnt
3569 * auto-release the semaphore:
3571 #ifdef CONFIG_PREEMPT_BKL
3572 saved_lock_depth = task->lock_depth;
3573 task->lock_depth = -1;
3577 local_irq_disable();
3578 #ifdef CONFIG_PREEMPT_BKL
3579 task->lock_depth = saved_lock_depth;
3581 sub_preempt_count(PREEMPT_ACTIVE);
3583 /* we could miss a preemption opportunity between schedule and now */
3585 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3589 #endif /* CONFIG_PREEMPT */
3591 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3594 return try_to_wake_up(curr->private, mode, sync);
3596 EXPORT_SYMBOL(default_wake_function);
3599 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3600 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3601 * number) then we wake all the non-exclusive tasks and one exclusive task.
3603 * There are circumstances in which we can try to wake a task which has already
3604 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3605 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3607 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3608 int nr_exclusive, int sync, void *key)
3610 wait_queue_t *curr, *next;
3612 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3613 unsigned flags = curr->flags;
3615 if (curr->func(curr, mode, sync, key) &&
3616 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3622 * __wake_up - wake up threads blocked on a waitqueue.
3624 * @mode: which threads
3625 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3626 * @key: is directly passed to the wakeup function
3628 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3629 int nr_exclusive, void *key)
3631 unsigned long flags;
3633 spin_lock_irqsave(&q->lock, flags);
3634 __wake_up_common(q, mode, nr_exclusive, 0, key);
3635 spin_unlock_irqrestore(&q->lock, flags);
3637 EXPORT_SYMBOL(__wake_up);
3640 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3642 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3644 __wake_up_common(q, mode, 1, 0, NULL);
3648 * __wake_up_sync - wake up threads blocked on a waitqueue.
3650 * @mode: which threads
3651 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3653 * The sync wakeup differs that the waker knows that it will schedule
3654 * away soon, so while the target thread will be woken up, it will not
3655 * be migrated to another CPU - ie. the two threads are 'synchronized'
3656 * with each other. This can prevent needless bouncing between CPUs.
3658 * On UP it can prevent extra preemption.
3661 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3663 unsigned long flags;
3669 if (unlikely(!nr_exclusive))
3672 spin_lock_irqsave(&q->lock, flags);
3673 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3674 spin_unlock_irqrestore(&q->lock, flags);
3676 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3678 void fastcall complete(struct completion *x)
3680 unsigned long flags;
3682 spin_lock_irqsave(&x->wait.lock, flags);
3684 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3686 spin_unlock_irqrestore(&x->wait.lock, flags);
3688 EXPORT_SYMBOL(complete);
3690 void fastcall complete_all(struct completion *x)
3692 unsigned long flags;
3694 spin_lock_irqsave(&x->wait.lock, flags);
3695 x->done += UINT_MAX/2;
3696 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3698 spin_unlock_irqrestore(&x->wait.lock, flags);
3700 EXPORT_SYMBOL(complete_all);
3702 void fastcall __sched wait_for_completion(struct completion *x)
3706 spin_lock_irq(&x->wait.lock);
3708 DECLARE_WAITQUEUE(wait, current);
3710 wait.flags |= WQ_FLAG_EXCLUSIVE;
3711 __add_wait_queue_tail(&x->wait, &wait);
3713 __set_current_state(TASK_UNINTERRUPTIBLE);
3714 spin_unlock_irq(&x->wait.lock);
3716 spin_lock_irq(&x->wait.lock);
3718 __remove_wait_queue(&x->wait, &wait);
3721 spin_unlock_irq(&x->wait.lock);
3723 EXPORT_SYMBOL(wait_for_completion);
3725 unsigned long fastcall __sched
3726 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3730 spin_lock_irq(&x->wait.lock);
3732 DECLARE_WAITQUEUE(wait, current);
3734 wait.flags |= WQ_FLAG_EXCLUSIVE;
3735 __add_wait_queue_tail(&x->wait, &wait);
3737 __set_current_state(TASK_UNINTERRUPTIBLE);
3738 spin_unlock_irq(&x->wait.lock);
3739 timeout = schedule_timeout(timeout);
3740 spin_lock_irq(&x->wait.lock);
3742 __remove_wait_queue(&x->wait, &wait);
3746 __remove_wait_queue(&x->wait, &wait);
3750 spin_unlock_irq(&x->wait.lock);
3753 EXPORT_SYMBOL(wait_for_completion_timeout);
3755 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3761 spin_lock_irq(&x->wait.lock);
3763 DECLARE_WAITQUEUE(wait, current);
3765 wait.flags |= WQ_FLAG_EXCLUSIVE;
3766 __add_wait_queue_tail(&x->wait, &wait);
3768 if (signal_pending(current)) {
3770 __remove_wait_queue(&x->wait, &wait);
3773 __set_current_state(TASK_INTERRUPTIBLE);
3774 spin_unlock_irq(&x->wait.lock);
3776 spin_lock_irq(&x->wait.lock);
3778 __remove_wait_queue(&x->wait, &wait);
3782 spin_unlock_irq(&x->wait.lock);
3786 EXPORT_SYMBOL(wait_for_completion_interruptible);
3788 unsigned long fastcall __sched
3789 wait_for_completion_interruptible_timeout(struct completion *x,
3790 unsigned long timeout)
3794 spin_lock_irq(&x->wait.lock);
3796 DECLARE_WAITQUEUE(wait, current);
3798 wait.flags |= WQ_FLAG_EXCLUSIVE;
3799 __add_wait_queue_tail(&x->wait, &wait);
3801 if (signal_pending(current)) {
3802 timeout = -ERESTARTSYS;
3803 __remove_wait_queue(&x->wait, &wait);
3806 __set_current_state(TASK_INTERRUPTIBLE);
3807 spin_unlock_irq(&x->wait.lock);
3808 timeout = schedule_timeout(timeout);
3809 spin_lock_irq(&x->wait.lock);
3811 __remove_wait_queue(&x->wait, &wait);
3815 __remove_wait_queue(&x->wait, &wait);
3819 spin_unlock_irq(&x->wait.lock);
3822 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3825 sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3827 spin_lock_irqsave(&q->lock, *flags);
3828 __add_wait_queue(q, wait);
3829 spin_unlock(&q->lock);
3833 sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3835 spin_lock_irq(&q->lock);
3836 __remove_wait_queue(q, wait);
3837 spin_unlock_irqrestore(&q->lock, *flags);
3840 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3842 unsigned long flags;
3845 init_waitqueue_entry(&wait, current);
3847 current->state = TASK_INTERRUPTIBLE;
3849 sleep_on_head(q, &wait, &flags);
3851 sleep_on_tail(q, &wait, &flags);
3853 EXPORT_SYMBOL(interruptible_sleep_on);
3856 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3858 unsigned long flags;
3861 init_waitqueue_entry(&wait, current);
3863 current->state = TASK_INTERRUPTIBLE;
3865 sleep_on_head(q, &wait, &flags);
3866 timeout = schedule_timeout(timeout);
3867 sleep_on_tail(q, &wait, &flags);
3871 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3873 void __sched sleep_on(wait_queue_head_t *q)
3875 unsigned long flags;
3878 init_waitqueue_entry(&wait, current);
3880 current->state = TASK_UNINTERRUPTIBLE;
3882 sleep_on_head(q, &wait, &flags);
3884 sleep_on_tail(q, &wait, &flags);
3886 EXPORT_SYMBOL(sleep_on);
3888 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3890 unsigned long flags;
3893 init_waitqueue_entry(&wait, current);
3895 current->state = TASK_UNINTERRUPTIBLE;
3897 sleep_on_head(q, &wait, &flags);
3898 timeout = schedule_timeout(timeout);
3899 sleep_on_tail(q, &wait, &flags);
3903 EXPORT_SYMBOL(sleep_on_timeout);
3905 #ifdef CONFIG_RT_MUTEXES
3908 * rt_mutex_setprio - set the current priority of a task
3910 * @prio: prio value (kernel-internal form)
3912 * This function changes the 'effective' priority of a task. It does
3913 * not touch ->normal_prio like __setscheduler().
3915 * Used by the rt_mutex code to implement priority inheritance logic.
3917 void rt_mutex_setprio(struct task_struct *p, int prio)
3919 unsigned long flags;
3923 BUG_ON(prio < 0 || prio > MAX_PRIO);
3925 rq = task_rq_lock(p, &flags);
3926 update_rq_clock(rq);
3929 on_rq = p->se.on_rq;
3931 dequeue_task(rq, p, 0);
3934 p->sched_class = &rt_sched_class;
3936 p->sched_class = &fair_sched_class;
3941 enqueue_task(rq, p, 0);
3943 * Reschedule if we are currently running on this runqueue and
3944 * our priority decreased, or if we are not currently running on
3945 * this runqueue and our priority is higher than the current's
3947 if (task_running(rq, p)) {
3948 if (p->prio > oldprio)
3949 resched_task(rq->curr);
3951 check_preempt_curr(rq, p);
3954 task_rq_unlock(rq, &flags);
3959 void set_user_nice(struct task_struct *p, long nice)
3961 int old_prio, delta, on_rq;
3962 unsigned long flags;
3965 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3968 * We have to be careful, if called from sys_setpriority(),
3969 * the task might be in the middle of scheduling on another CPU.
3971 rq = task_rq_lock(p, &flags);
3972 update_rq_clock(rq);
3974 * The RT priorities are set via sched_setscheduler(), but we still
3975 * allow the 'normal' nice value to be set - but as expected
3976 * it wont have any effect on scheduling until the task is
3977 * SCHED_FIFO/SCHED_RR:
3979 if (task_has_rt_policy(p)) {
3980 p->static_prio = NICE_TO_PRIO(nice);
3983 on_rq = p->se.on_rq;
3985 dequeue_task(rq, p, 0);
3989 p->static_prio = NICE_TO_PRIO(nice);
3992 p->prio = effective_prio(p);
3993 delta = p->prio - old_prio;
3996 enqueue_task(rq, p, 0);
3999 * If the task increased its priority or is running and
4000 * lowered its priority, then reschedule its CPU:
4002 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4003 resched_task(rq->curr);
4006 task_rq_unlock(rq, &flags);
4008 EXPORT_SYMBOL(set_user_nice);
4011 * can_nice - check if a task can reduce its nice value
4015 int can_nice(const struct task_struct *p, const int nice)
4017 /* convert nice value [19,-20] to rlimit style value [1,40] */
4018 int nice_rlim = 20 - nice;
4020 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4021 capable(CAP_SYS_NICE));
4024 #ifdef __ARCH_WANT_SYS_NICE
4027 * sys_nice - change the priority of the current process.
4028 * @increment: priority increment
4030 * sys_setpriority is a more generic, but much slower function that
4031 * does similar things.
4033 asmlinkage long sys_nice(int increment)
4038 * Setpriority might change our priority at the same moment.
4039 * We don't have to worry. Conceptually one call occurs first
4040 * and we have a single winner.
4042 if (increment < -40)
4047 nice = PRIO_TO_NICE(current->static_prio) + increment;
4053 if (increment < 0 && !can_nice(current, nice))
4056 retval = security_task_setnice(current, nice);
4060 set_user_nice(current, nice);
4067 * task_prio - return the priority value of a given task.
4068 * @p: the task in question.
4070 * This is the priority value as seen by users in /proc.
4071 * RT tasks are offset by -200. Normal tasks are centered
4072 * around 0, value goes from -16 to +15.
4074 int task_prio(const struct task_struct *p)
4076 return p->prio - MAX_RT_PRIO;
4080 * task_nice - return the nice value of a given task.
4081 * @p: the task in question.
4083 int task_nice(const struct task_struct *p)
4085 return TASK_NICE(p);
4087 EXPORT_SYMBOL_GPL(task_nice);
4090 * idle_cpu - is a given cpu idle currently?
4091 * @cpu: the processor in question.
4093 int idle_cpu(int cpu)
4095 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4099 * idle_task - return the idle task for a given cpu.
4100 * @cpu: the processor in question.
4102 struct task_struct *idle_task(int cpu)
4104 return cpu_rq(cpu)->idle;
4108 * find_process_by_pid - find a process with a matching PID value.
4109 * @pid: the pid in question.
4111 static inline struct task_struct *find_process_by_pid(pid_t pid)
4113 return pid ? find_task_by_pid(pid) : current;
4116 /* Actually do priority change: must hold rq lock. */
4118 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4120 BUG_ON(p->se.on_rq);
4123 switch (p->policy) {
4127 p->sched_class = &fair_sched_class;
4131 p->sched_class = &rt_sched_class;
4135 p->rt_priority = prio;
4136 p->normal_prio = normal_prio(p);
4137 /* we are holding p->pi_lock already */
4138 p->prio = rt_mutex_getprio(p);
4143 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4144 * @p: the task in question.
4145 * @policy: new policy.
4146 * @param: structure containing the new RT priority.
4148 * NOTE that the task may be already dead.
4150 int sched_setscheduler(struct task_struct *p, int policy,
4151 struct sched_param *param)
4153 int retval, oldprio, oldpolicy = -1, on_rq;
4154 unsigned long flags;
4157 /* may grab non-irq protected spin_locks */
4158 BUG_ON(in_interrupt());
4160 /* double check policy once rq lock held */
4162 policy = oldpolicy = p->policy;
4163 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4164 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4165 policy != SCHED_IDLE)
4168 * Valid priorities for SCHED_FIFO and SCHED_RR are
4169 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4170 * SCHED_BATCH and SCHED_IDLE is 0.
4172 if (param->sched_priority < 0 ||
4173 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4174 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4176 if (rt_policy(policy) != (param->sched_priority != 0))
4180 * Allow unprivileged RT tasks to decrease priority:
4182 if (!capable(CAP_SYS_NICE)) {
4183 if (rt_policy(policy)) {
4184 unsigned long rlim_rtprio;
4186 if (!lock_task_sighand(p, &flags))
4188 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4189 unlock_task_sighand(p, &flags);
4191 /* can't set/change the rt policy */
4192 if (policy != p->policy && !rlim_rtprio)
4195 /* can't increase priority */
4196 if (param->sched_priority > p->rt_priority &&
4197 param->sched_priority > rlim_rtprio)
4201 * Like positive nice levels, dont allow tasks to
4202 * move out of SCHED_IDLE either:
4204 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4207 /* can't change other user's priorities */
4208 if ((current->euid != p->euid) &&
4209 (current->euid != p->uid))
4213 retval = security_task_setscheduler(p, policy, param);
4217 * make sure no PI-waiters arrive (or leave) while we are
4218 * changing the priority of the task:
4220 spin_lock_irqsave(&p->pi_lock, flags);
4222 * To be able to change p->policy safely, the apropriate
4223 * runqueue lock must be held.
4225 rq = __task_rq_lock(p);
4226 /* recheck policy now with rq lock held */
4227 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4228 policy = oldpolicy = -1;
4229 __task_rq_unlock(rq);
4230 spin_unlock_irqrestore(&p->pi_lock, flags);
4233 update_rq_clock(rq);
4234 on_rq = p->se.on_rq;
4236 deactivate_task(rq, p, 0);
4238 __setscheduler(rq, p, policy, param->sched_priority);
4240 activate_task(rq, p, 0);
4242 * Reschedule if we are currently running on this runqueue and
4243 * our priority decreased, or if we are not currently running on
4244 * this runqueue and our priority is higher than the current's
4246 if (task_running(rq, p)) {
4247 if (p->prio > oldprio)
4248 resched_task(rq->curr);
4250 check_preempt_curr(rq, p);
4253 __task_rq_unlock(rq);
4254 spin_unlock_irqrestore(&p->pi_lock, flags);
4256 rt_mutex_adjust_pi(p);
4260 EXPORT_SYMBOL_GPL(sched_setscheduler);
4263 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4265 struct sched_param lparam;
4266 struct task_struct *p;
4269 if (!param || pid < 0)
4271 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4276 p = find_process_by_pid(pid);
4278 retval = sched_setscheduler(p, policy, &lparam);
4285 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4286 * @pid: the pid in question.
4287 * @policy: new policy.
4288 * @param: structure containing the new RT priority.
4290 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4291 struct sched_param __user *param)
4293 /* negative values for policy are not valid */
4297 return do_sched_setscheduler(pid, policy, param);
4301 * sys_sched_setparam - set/change the RT priority of a thread
4302 * @pid: the pid in question.
4303 * @param: structure containing the new RT priority.
4305 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4307 return do_sched_setscheduler(pid, -1, param);
4311 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4312 * @pid: the pid in question.
4314 asmlinkage long sys_sched_getscheduler(pid_t pid)
4316 struct task_struct *p;
4317 int retval = -EINVAL;
4323 read_lock(&tasklist_lock);
4324 p = find_process_by_pid(pid);
4326 retval = security_task_getscheduler(p);
4330 read_unlock(&tasklist_lock);
4337 * sys_sched_getscheduler - get the RT priority of a thread
4338 * @pid: the pid in question.
4339 * @param: structure containing the RT priority.
4341 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4343 struct sched_param lp;
4344 struct task_struct *p;
4345 int retval = -EINVAL;
4347 if (!param || pid < 0)
4350 read_lock(&tasklist_lock);
4351 p = find_process_by_pid(pid);
4356 retval = security_task_getscheduler(p);
4360 lp.sched_priority = p->rt_priority;
4361 read_unlock(&tasklist_lock);
4364 * This one might sleep, we cannot do it with a spinlock held ...
4366 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4372 read_unlock(&tasklist_lock);
4376 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4378 cpumask_t cpus_allowed;
4379 struct task_struct *p;
4382 mutex_lock(&sched_hotcpu_mutex);
4383 read_lock(&tasklist_lock);
4385 p = find_process_by_pid(pid);
4387 read_unlock(&tasklist_lock);
4388 mutex_unlock(&sched_hotcpu_mutex);
4393 * It is not safe to call set_cpus_allowed with the
4394 * tasklist_lock held. We will bump the task_struct's
4395 * usage count and then drop tasklist_lock.
4398 read_unlock(&tasklist_lock);
4401 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4402 !capable(CAP_SYS_NICE))
4405 retval = security_task_setscheduler(p, 0, NULL);
4409 cpus_allowed = cpuset_cpus_allowed(p);
4410 cpus_and(new_mask, new_mask, cpus_allowed);
4411 retval = set_cpus_allowed(p, new_mask);
4415 mutex_unlock(&sched_hotcpu_mutex);
4419 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4420 cpumask_t *new_mask)
4422 if (len < sizeof(cpumask_t)) {
4423 memset(new_mask, 0, sizeof(cpumask_t));
4424 } else if (len > sizeof(cpumask_t)) {
4425 len = sizeof(cpumask_t);
4427 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4431 * sys_sched_setaffinity - set the cpu affinity of a process
4432 * @pid: pid of the process
4433 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4434 * @user_mask_ptr: user-space pointer to the new cpu mask
4436 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4437 unsigned long __user *user_mask_ptr)
4442 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4446 return sched_setaffinity(pid, new_mask);
4450 * Represents all cpu's present in the system
4451 * In systems capable of hotplug, this map could dynamically grow
4452 * as new cpu's are detected in the system via any platform specific
4453 * method, such as ACPI for e.g.
4456 cpumask_t cpu_present_map __read_mostly;
4457 EXPORT_SYMBOL(cpu_present_map);
4460 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4461 EXPORT_SYMBOL(cpu_online_map);
4463 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4464 EXPORT_SYMBOL(cpu_possible_map);
4467 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4469 struct task_struct *p;
4472 mutex_lock(&sched_hotcpu_mutex);
4473 read_lock(&tasklist_lock);
4476 p = find_process_by_pid(pid);
4480 retval = security_task_getscheduler(p);
4484 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4487 read_unlock(&tasklist_lock);
4488 mutex_unlock(&sched_hotcpu_mutex);
4494 * sys_sched_getaffinity - get the cpu affinity of a process
4495 * @pid: pid of the process
4496 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4497 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4499 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4500 unsigned long __user *user_mask_ptr)
4505 if (len < sizeof(cpumask_t))
4508 ret = sched_getaffinity(pid, &mask);
4512 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4515 return sizeof(cpumask_t);
4519 * sys_sched_yield - yield the current processor to other threads.
4521 * This function yields the current CPU to other tasks. If there are no
4522 * other threads running on this CPU then this function will return.
4524 asmlinkage long sys_sched_yield(void)
4526 struct rq *rq = this_rq_lock();
4528 schedstat_inc(rq, yld_cnt);
4529 current->sched_class->yield_task(rq, current);
4532 * Since we are going to call schedule() anyway, there's
4533 * no need to preempt or enable interrupts:
4535 __release(rq->lock);
4536 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4537 _raw_spin_unlock(&rq->lock);
4538 preempt_enable_no_resched();
4545 static void __cond_resched(void)
4547 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4548 __might_sleep(__FILE__, __LINE__);
4551 * The BKS might be reacquired before we have dropped
4552 * PREEMPT_ACTIVE, which could trigger a second
4553 * cond_resched() call.
4556 add_preempt_count(PREEMPT_ACTIVE);
4558 sub_preempt_count(PREEMPT_ACTIVE);
4559 } while (need_resched());
4562 int __sched cond_resched(void)
4564 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4565 system_state == SYSTEM_RUNNING) {
4571 EXPORT_SYMBOL(cond_resched);
4574 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4575 * call schedule, and on return reacquire the lock.
4577 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4578 * operations here to prevent schedule() from being called twice (once via
4579 * spin_unlock(), once by hand).
4581 int cond_resched_lock(spinlock_t *lock)
4585 if (need_lockbreak(lock)) {
4591 if (need_resched() && system_state == SYSTEM_RUNNING) {
4592 spin_release(&lock->dep_map, 1, _THIS_IP_);
4593 _raw_spin_unlock(lock);
4594 preempt_enable_no_resched();
4601 EXPORT_SYMBOL(cond_resched_lock);
4603 int __sched cond_resched_softirq(void)
4605 BUG_ON(!in_softirq());
4607 if (need_resched() && system_state == SYSTEM_RUNNING) {
4615 EXPORT_SYMBOL(cond_resched_softirq);
4618 * yield - yield the current processor to other threads.
4620 * This is a shortcut for kernel-space yielding - it marks the
4621 * thread runnable and calls sys_sched_yield().
4623 void __sched yield(void)
4625 set_current_state(TASK_RUNNING);
4628 EXPORT_SYMBOL(yield);
4631 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4632 * that process accounting knows that this is a task in IO wait state.
4634 * But don't do that if it is a deliberate, throttling IO wait (this task
4635 * has set its backing_dev_info: the queue against which it should throttle)
4637 void __sched io_schedule(void)
4639 struct rq *rq = &__raw_get_cpu_var(runqueues);
4641 delayacct_blkio_start();
4642 atomic_inc(&rq->nr_iowait);
4644 atomic_dec(&rq->nr_iowait);
4645 delayacct_blkio_end();
4647 EXPORT_SYMBOL(io_schedule);
4649 long __sched io_schedule_timeout(long timeout)
4651 struct rq *rq = &__raw_get_cpu_var(runqueues);
4654 delayacct_blkio_start();
4655 atomic_inc(&rq->nr_iowait);
4656 ret = schedule_timeout(timeout);
4657 atomic_dec(&rq->nr_iowait);
4658 delayacct_blkio_end();
4663 * sys_sched_get_priority_max - return maximum RT priority.
4664 * @policy: scheduling class.
4666 * this syscall returns the maximum rt_priority that can be used
4667 * by a given scheduling class.
4669 asmlinkage long sys_sched_get_priority_max(int policy)
4676 ret = MAX_USER_RT_PRIO-1;
4688 * sys_sched_get_priority_min - return minimum RT priority.
4689 * @policy: scheduling class.
4691 * this syscall returns the minimum rt_priority that can be used
4692 * by a given scheduling class.
4694 asmlinkage long sys_sched_get_priority_min(int policy)
4712 * sys_sched_rr_get_interval - return the default timeslice of a process.
4713 * @pid: pid of the process.
4714 * @interval: userspace pointer to the timeslice value.
4716 * this syscall writes the default timeslice value of a given process
4717 * into the user-space timespec buffer. A value of '0' means infinity.
4720 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4722 struct task_struct *p;
4723 int retval = -EINVAL;
4730 read_lock(&tasklist_lock);
4731 p = find_process_by_pid(pid);
4735 retval = security_task_getscheduler(p);
4739 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4740 0 : static_prio_timeslice(p->static_prio), &t);
4741 read_unlock(&tasklist_lock);
4742 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4746 read_unlock(&tasklist_lock);
4750 static const char stat_nam[] = "RSDTtZX";
4752 static void show_task(struct task_struct *p)
4754 unsigned long free = 0;
4757 state = p->state ? __ffs(p->state) + 1 : 0;
4758 printk("%-13.13s %c", p->comm,
4759 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4760 #if BITS_PER_LONG == 32
4761 if (state == TASK_RUNNING)
4762 printk(" running ");
4764 printk(" %08lx ", thread_saved_pc(p));
4766 if (state == TASK_RUNNING)
4767 printk(" running task ");
4769 printk(" %016lx ", thread_saved_pc(p));
4771 #ifdef CONFIG_DEBUG_STACK_USAGE
4773 unsigned long *n = end_of_stack(p);
4776 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4779 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4781 if (state != TASK_RUNNING)
4782 show_stack(p, NULL);
4785 void show_state_filter(unsigned long state_filter)
4787 struct task_struct *g, *p;
4789 #if BITS_PER_LONG == 32
4791 " task PC stack pid father\n");
4794 " task PC stack pid father\n");
4796 read_lock(&tasklist_lock);
4797 do_each_thread(g, p) {
4799 * reset the NMI-timeout, listing all files on a slow
4800 * console might take alot of time:
4802 touch_nmi_watchdog();
4803 if (!state_filter || (p->state & state_filter))
4805 } while_each_thread(g, p);
4807 touch_all_softlockup_watchdogs();
4809 #ifdef CONFIG_SCHED_DEBUG
4810 sysrq_sched_debug_show();
4812 read_unlock(&tasklist_lock);
4814 * Only show locks if all tasks are dumped:
4816 if (state_filter == -1)
4817 debug_show_all_locks();
4820 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4822 idle->sched_class = &idle_sched_class;
4826 * init_idle - set up an idle thread for a given CPU
4827 * @idle: task in question
4828 * @cpu: cpu the idle task belongs to
4830 * NOTE: this function does not set the idle thread's NEED_RESCHED
4831 * flag, to make booting more robust.
4833 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4835 struct rq *rq = cpu_rq(cpu);
4836 unsigned long flags;
4839 idle->se.exec_start = sched_clock();
4841 idle->prio = idle->normal_prio = MAX_PRIO;
4842 idle->cpus_allowed = cpumask_of_cpu(cpu);
4843 __set_task_cpu(idle, cpu);
4845 spin_lock_irqsave(&rq->lock, flags);
4846 rq->curr = rq->idle = idle;
4847 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4850 spin_unlock_irqrestore(&rq->lock, flags);
4852 /* Set the preempt count _outside_ the spinlocks! */
4853 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4854 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4856 task_thread_info(idle)->preempt_count = 0;
4859 * The idle tasks have their own, simple scheduling class:
4861 idle->sched_class = &idle_sched_class;
4865 * In a system that switches off the HZ timer nohz_cpu_mask
4866 * indicates which cpus entered this state. This is used
4867 * in the rcu update to wait only for active cpus. For system
4868 * which do not switch off the HZ timer nohz_cpu_mask should
4869 * always be CPU_MASK_NONE.
4871 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4875 * This is how migration works:
4877 * 1) we queue a struct migration_req structure in the source CPU's
4878 * runqueue and wake up that CPU's migration thread.
4879 * 2) we down() the locked semaphore => thread blocks.
4880 * 3) migration thread wakes up (implicitly it forces the migrated
4881 * thread off the CPU)
4882 * 4) it gets the migration request and checks whether the migrated
4883 * task is still in the wrong runqueue.
4884 * 5) if it's in the wrong runqueue then the migration thread removes
4885 * it and puts it into the right queue.
4886 * 6) migration thread up()s the semaphore.
4887 * 7) we wake up and the migration is done.
4891 * Change a given task's CPU affinity. Migrate the thread to a
4892 * proper CPU and schedule it away if the CPU it's executing on
4893 * is removed from the allowed bitmask.
4895 * NOTE: the caller must have a valid reference to the task, the
4896 * task must not exit() & deallocate itself prematurely. The
4897 * call is not atomic; no spinlocks may be held.
4899 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4901 struct migration_req req;
4902 unsigned long flags;
4906 rq = task_rq_lock(p, &flags);
4907 if (!cpus_intersects(new_mask, cpu_online_map)) {
4912 p->cpus_allowed = new_mask;
4913 /* Can the task run on the task's current CPU? If so, we're done */
4914 if (cpu_isset(task_cpu(p), new_mask))
4917 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4918 /* Need help from migration thread: drop lock and wait. */
4919 task_rq_unlock(rq, &flags);
4920 wake_up_process(rq->migration_thread);
4921 wait_for_completion(&req.done);
4922 tlb_migrate_finish(p->mm);
4926 task_rq_unlock(rq, &flags);
4930 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4933 * Move (not current) task off this cpu, onto dest cpu. We're doing
4934 * this because either it can't run here any more (set_cpus_allowed()
4935 * away from this CPU, or CPU going down), or because we're
4936 * attempting to rebalance this task on exec (sched_exec).
4938 * So we race with normal scheduler movements, but that's OK, as long
4939 * as the task is no longer on this CPU.
4941 * Returns non-zero if task was successfully migrated.
4943 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4945 struct rq *rq_dest, *rq_src;
4948 if (unlikely(cpu_is_offline(dest_cpu)))
4951 rq_src = cpu_rq(src_cpu);
4952 rq_dest = cpu_rq(dest_cpu);
4954 double_rq_lock(rq_src, rq_dest);
4955 /* Already moved. */
4956 if (task_cpu(p) != src_cpu)
4958 /* Affinity changed (again). */
4959 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4962 on_rq = p->se.on_rq;
4964 deactivate_task(rq_src, p, 0);
4966 set_task_cpu(p, dest_cpu);
4968 activate_task(rq_dest, p, 0);
4969 check_preempt_curr(rq_dest, p);
4973 double_rq_unlock(rq_src, rq_dest);
4978 * migration_thread - this is a highprio system thread that performs
4979 * thread migration by bumping thread off CPU then 'pushing' onto
4982 static int migration_thread(void *data)
4984 int cpu = (long)data;
4988 BUG_ON(rq->migration_thread != current);
4990 set_current_state(TASK_INTERRUPTIBLE);
4991 while (!kthread_should_stop()) {
4992 struct migration_req *req;
4993 struct list_head *head;
4995 spin_lock_irq(&rq->lock);
4997 if (cpu_is_offline(cpu)) {
4998 spin_unlock_irq(&rq->lock);
5002 if (rq->active_balance) {
5003 active_load_balance(rq, cpu);
5004 rq->active_balance = 0;
5007 head = &rq->migration_queue;
5009 if (list_empty(head)) {
5010 spin_unlock_irq(&rq->lock);
5012 set_current_state(TASK_INTERRUPTIBLE);
5015 req = list_entry(head->next, struct migration_req, list);
5016 list_del_init(head->next);
5018 spin_unlock(&rq->lock);
5019 __migrate_task(req->task, cpu, req->dest_cpu);
5022 complete(&req->done);
5024 __set_current_state(TASK_RUNNING);
5028 /* Wait for kthread_stop */
5029 set_current_state(TASK_INTERRUPTIBLE);
5030 while (!kthread_should_stop()) {
5032 set_current_state(TASK_INTERRUPTIBLE);
5034 __set_current_state(TASK_RUNNING);
5038 #ifdef CONFIG_HOTPLUG_CPU
5040 * Figure out where task on dead CPU should go, use force if neccessary.
5041 * NOTE: interrupts should be disabled by the caller
5043 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5045 unsigned long flags;
5052 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5053 cpus_and(mask, mask, p->cpus_allowed);
5054 dest_cpu = any_online_cpu(mask);
5056 /* On any allowed CPU? */
5057 if (dest_cpu == NR_CPUS)
5058 dest_cpu = any_online_cpu(p->cpus_allowed);
5060 /* No more Mr. Nice Guy. */
5061 if (dest_cpu == NR_CPUS) {
5062 rq = task_rq_lock(p, &flags);
5063 cpus_setall(p->cpus_allowed);
5064 dest_cpu = any_online_cpu(p->cpus_allowed);
5065 task_rq_unlock(rq, &flags);
5068 * Don't tell them about moving exiting tasks or
5069 * kernel threads (both mm NULL), since they never
5072 if (p->mm && printk_ratelimit())
5073 printk(KERN_INFO "process %d (%s) no "
5074 "longer affine to cpu%d\n",
5075 p->pid, p->comm, dead_cpu);
5077 if (!__migrate_task(p, dead_cpu, dest_cpu))
5082 * While a dead CPU has no uninterruptible tasks queued at this point,
5083 * it might still have a nonzero ->nr_uninterruptible counter, because
5084 * for performance reasons the counter is not stricly tracking tasks to
5085 * their home CPUs. So we just add the counter to another CPU's counter,
5086 * to keep the global sum constant after CPU-down:
5088 static void migrate_nr_uninterruptible(struct rq *rq_src)
5090 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5091 unsigned long flags;
5093 local_irq_save(flags);
5094 double_rq_lock(rq_src, rq_dest);
5095 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5096 rq_src->nr_uninterruptible = 0;
5097 double_rq_unlock(rq_src, rq_dest);
5098 local_irq_restore(flags);
5101 /* Run through task list and migrate tasks from the dead cpu. */
5102 static void migrate_live_tasks(int src_cpu)
5104 struct task_struct *p, *t;
5106 write_lock_irq(&tasklist_lock);
5108 do_each_thread(t, p) {
5112 if (task_cpu(p) == src_cpu)
5113 move_task_off_dead_cpu(src_cpu, p);
5114 } while_each_thread(t, p);
5116 write_unlock_irq(&tasklist_lock);
5120 * Schedules idle task to be the next runnable task on current CPU.
5121 * It does so by boosting its priority to highest possible and adding it to
5122 * the _front_ of the runqueue. Used by CPU offline code.
5124 void sched_idle_next(void)
5126 int this_cpu = smp_processor_id();
5127 struct rq *rq = cpu_rq(this_cpu);
5128 struct task_struct *p = rq->idle;
5129 unsigned long flags;
5131 /* cpu has to be offline */
5132 BUG_ON(cpu_online(this_cpu));
5135 * Strictly not necessary since rest of the CPUs are stopped by now
5136 * and interrupts disabled on the current cpu.
5138 spin_lock_irqsave(&rq->lock, flags);
5140 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5142 /* Add idle task to the _front_ of its priority queue: */
5143 activate_idle_task(p, rq);
5145 spin_unlock_irqrestore(&rq->lock, flags);
5149 * Ensures that the idle task is using init_mm right before its cpu goes
5152 void idle_task_exit(void)
5154 struct mm_struct *mm = current->active_mm;
5156 BUG_ON(cpu_online(smp_processor_id()));
5159 switch_mm(mm, &init_mm, current);
5163 /* called under rq->lock with disabled interrupts */
5164 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5166 struct rq *rq = cpu_rq(dead_cpu);
5168 /* Must be exiting, otherwise would be on tasklist. */
5169 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5171 /* Cannot have done final schedule yet: would have vanished. */
5172 BUG_ON(p->state == TASK_DEAD);
5177 * Drop lock around migration; if someone else moves it,
5178 * that's OK. No task can be added to this CPU, so iteration is
5180 * NOTE: interrupts should be left disabled --dev@
5182 spin_unlock(&rq->lock);
5183 move_task_off_dead_cpu(dead_cpu, p);
5184 spin_lock(&rq->lock);
5189 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5190 static void migrate_dead_tasks(unsigned int dead_cpu)
5192 struct rq *rq = cpu_rq(dead_cpu);
5193 struct task_struct *next;
5196 if (!rq->nr_running)
5198 update_rq_clock(rq);
5199 next = pick_next_task(rq, rq->curr);
5202 migrate_dead(dead_cpu, next);
5206 #endif /* CONFIG_HOTPLUG_CPU */
5208 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5210 static struct ctl_table sd_ctl_dir[] = {
5212 .procname = "sched_domain",
5218 static struct ctl_table sd_ctl_root[] = {
5220 .ctl_name = CTL_KERN,
5221 .procname = "kernel",
5223 .child = sd_ctl_dir,
5228 static struct ctl_table *sd_alloc_ctl_entry(int n)
5230 struct ctl_table *entry =
5231 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5234 memset(entry, 0, n * sizeof(struct ctl_table));
5240 set_table_entry(struct ctl_table *entry,
5241 const char *procname, void *data, int maxlen,
5242 mode_t mode, proc_handler *proc_handler)
5244 entry->procname = procname;
5246 entry->maxlen = maxlen;
5248 entry->proc_handler = proc_handler;
5251 static struct ctl_table *
5252 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5254 struct ctl_table *table = sd_alloc_ctl_entry(14);
5256 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5257 sizeof(long), 0644, proc_doulongvec_minmax);
5258 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5259 sizeof(long), 0644, proc_doulongvec_minmax);
5260 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5261 sizeof(int), 0644, proc_dointvec_minmax);
5262 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5263 sizeof(int), 0644, proc_dointvec_minmax);
5264 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5265 sizeof(int), 0644, proc_dointvec_minmax);
5266 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5267 sizeof(int), 0644, proc_dointvec_minmax);
5268 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5269 sizeof(int), 0644, proc_dointvec_minmax);
5270 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5271 sizeof(int), 0644, proc_dointvec_minmax);
5272 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5273 sizeof(int), 0644, proc_dointvec_minmax);
5274 set_table_entry(&table[10], "cache_nice_tries",
5275 &sd->cache_nice_tries,
5276 sizeof(int), 0644, proc_dointvec_minmax);
5277 set_table_entry(&table[12], "flags", &sd->flags,
5278 sizeof(int), 0644, proc_dointvec_minmax);
5283 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5285 struct ctl_table *entry, *table;
5286 struct sched_domain *sd;
5287 int domain_num = 0, i;
5290 for_each_domain(cpu, sd)
5292 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5295 for_each_domain(cpu, sd) {
5296 snprintf(buf, 32, "domain%d", i);
5297 entry->procname = kstrdup(buf, GFP_KERNEL);
5299 entry->child = sd_alloc_ctl_domain_table(sd);
5306 static struct ctl_table_header *sd_sysctl_header;
5307 static void init_sched_domain_sysctl(void)
5309 int i, cpu_num = num_online_cpus();
5310 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5313 sd_ctl_dir[0].child = entry;
5315 for (i = 0; i < cpu_num; i++, entry++) {
5316 snprintf(buf, 32, "cpu%d", i);
5317 entry->procname = kstrdup(buf, GFP_KERNEL);
5319 entry->child = sd_alloc_ctl_cpu_table(i);
5321 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5324 static void init_sched_domain_sysctl(void)
5330 * migration_call - callback that gets triggered when a CPU is added.
5331 * Here we can start up the necessary migration thread for the new CPU.
5333 static int __cpuinit
5334 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5336 struct task_struct *p;
5337 int cpu = (long)hcpu;
5338 unsigned long flags;
5342 case CPU_LOCK_ACQUIRE:
5343 mutex_lock(&sched_hotcpu_mutex);
5346 case CPU_UP_PREPARE:
5347 case CPU_UP_PREPARE_FROZEN:
5348 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5351 kthread_bind(p, cpu);
5352 /* Must be high prio: stop_machine expects to yield to it. */
5353 rq = task_rq_lock(p, &flags);
5354 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5355 task_rq_unlock(rq, &flags);
5356 cpu_rq(cpu)->migration_thread = p;
5360 case CPU_ONLINE_FROZEN:
5361 /* Strictly unneccessary, as first user will wake it. */
5362 wake_up_process(cpu_rq(cpu)->migration_thread);
5365 #ifdef CONFIG_HOTPLUG_CPU
5366 case CPU_UP_CANCELED:
5367 case CPU_UP_CANCELED_FROZEN:
5368 if (!cpu_rq(cpu)->migration_thread)
5370 /* Unbind it from offline cpu so it can run. Fall thru. */
5371 kthread_bind(cpu_rq(cpu)->migration_thread,
5372 any_online_cpu(cpu_online_map));
5373 kthread_stop(cpu_rq(cpu)->migration_thread);
5374 cpu_rq(cpu)->migration_thread = NULL;
5378 case CPU_DEAD_FROZEN:
5379 migrate_live_tasks(cpu);
5381 kthread_stop(rq->migration_thread);
5382 rq->migration_thread = NULL;
5383 /* Idle task back to normal (off runqueue, low prio) */
5384 rq = task_rq_lock(rq->idle, &flags);
5385 update_rq_clock(rq);
5386 deactivate_task(rq, rq->idle, 0);
5387 rq->idle->static_prio = MAX_PRIO;
5388 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5389 rq->idle->sched_class = &idle_sched_class;
5390 migrate_dead_tasks(cpu);
5391 task_rq_unlock(rq, &flags);
5392 migrate_nr_uninterruptible(rq);
5393 BUG_ON(rq->nr_running != 0);
5395 /* No need to migrate the tasks: it was best-effort if
5396 * they didn't take sched_hotcpu_mutex. Just wake up
5397 * the requestors. */
5398 spin_lock_irq(&rq->lock);
5399 while (!list_empty(&rq->migration_queue)) {
5400 struct migration_req *req;
5402 req = list_entry(rq->migration_queue.next,
5403 struct migration_req, list);
5404 list_del_init(&req->list);
5405 complete(&req->done);
5407 spin_unlock_irq(&rq->lock);
5410 case CPU_LOCK_RELEASE:
5411 mutex_unlock(&sched_hotcpu_mutex);
5417 /* Register at highest priority so that task migration (migrate_all_tasks)
5418 * happens before everything else.
5420 static struct notifier_block __cpuinitdata migration_notifier = {
5421 .notifier_call = migration_call,
5425 int __init migration_init(void)
5427 void *cpu = (void *)(long)smp_processor_id();
5430 /* Start one for the boot CPU: */
5431 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5432 BUG_ON(err == NOTIFY_BAD);
5433 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5434 register_cpu_notifier(&migration_notifier);
5442 /* Number of possible processor ids */
5443 int nr_cpu_ids __read_mostly = NR_CPUS;
5444 EXPORT_SYMBOL(nr_cpu_ids);
5446 #undef SCHED_DOMAIN_DEBUG
5447 #ifdef SCHED_DOMAIN_DEBUG
5448 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5453 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5457 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5462 struct sched_group *group = sd->groups;
5463 cpumask_t groupmask;
5465 cpumask_scnprintf(str, NR_CPUS, sd->span);
5466 cpus_clear(groupmask);
5469 for (i = 0; i < level + 1; i++)
5471 printk("domain %d: ", level);
5473 if (!(sd->flags & SD_LOAD_BALANCE)) {
5474 printk("does not load-balance\n");
5476 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5481 printk("span %s\n", str);
5483 if (!cpu_isset(cpu, sd->span))
5484 printk(KERN_ERR "ERROR: domain->span does not contain "
5486 if (!cpu_isset(cpu, group->cpumask))
5487 printk(KERN_ERR "ERROR: domain->groups does not contain"
5491 for (i = 0; i < level + 2; i++)
5497 printk(KERN_ERR "ERROR: group is NULL\n");
5501 if (!group->__cpu_power) {
5503 printk(KERN_ERR "ERROR: domain->cpu_power not "
5507 if (!cpus_weight(group->cpumask)) {
5509 printk(KERN_ERR "ERROR: empty group\n");
5512 if (cpus_intersects(groupmask, group->cpumask)) {
5514 printk(KERN_ERR "ERROR: repeated CPUs\n");
5517 cpus_or(groupmask, groupmask, group->cpumask);
5519 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5522 group = group->next;
5523 } while (group != sd->groups);
5526 if (!cpus_equal(sd->span, groupmask))
5527 printk(KERN_ERR "ERROR: groups don't span "
5535 if (!cpus_subset(groupmask, sd->span))
5536 printk(KERN_ERR "ERROR: parent span is not a superset "
5537 "of domain->span\n");
5542 # define sched_domain_debug(sd, cpu) do { } while (0)
5545 static int sd_degenerate(struct sched_domain *sd)
5547 if (cpus_weight(sd->span) == 1)
5550 /* Following flags need at least 2 groups */
5551 if (sd->flags & (SD_LOAD_BALANCE |
5552 SD_BALANCE_NEWIDLE |
5556 SD_SHARE_PKG_RESOURCES)) {
5557 if (sd->groups != sd->groups->next)
5561 /* Following flags don't use groups */
5562 if (sd->flags & (SD_WAKE_IDLE |
5571 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5573 unsigned long cflags = sd->flags, pflags = parent->flags;
5575 if (sd_degenerate(parent))
5578 if (!cpus_equal(sd->span, parent->span))
5581 /* Does parent contain flags not in child? */
5582 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5583 if (cflags & SD_WAKE_AFFINE)
5584 pflags &= ~SD_WAKE_BALANCE;
5585 /* Flags needing groups don't count if only 1 group in parent */
5586 if (parent->groups == parent->groups->next) {
5587 pflags &= ~(SD_LOAD_BALANCE |
5588 SD_BALANCE_NEWIDLE |
5592 SD_SHARE_PKG_RESOURCES);
5594 if (~cflags & pflags)
5601 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5602 * hold the hotplug lock.
5604 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5606 struct rq *rq = cpu_rq(cpu);
5607 struct sched_domain *tmp;
5609 /* Remove the sched domains which do not contribute to scheduling. */
5610 for (tmp = sd; tmp; tmp = tmp->parent) {
5611 struct sched_domain *parent = tmp->parent;
5614 if (sd_parent_degenerate(tmp, parent)) {
5615 tmp->parent = parent->parent;
5617 parent->parent->child = tmp;
5621 if (sd && sd_degenerate(sd)) {
5627 sched_domain_debug(sd, cpu);
5629 rcu_assign_pointer(rq->sd, sd);
5632 /* cpus with isolated domains */
5633 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5635 /* Setup the mask of cpus configured for isolated domains */
5636 static int __init isolated_cpu_setup(char *str)
5638 int ints[NR_CPUS], i;
5640 str = get_options(str, ARRAY_SIZE(ints), ints);
5641 cpus_clear(cpu_isolated_map);
5642 for (i = 1; i <= ints[0]; i++)
5643 if (ints[i] < NR_CPUS)
5644 cpu_set(ints[i], cpu_isolated_map);
5648 __setup ("isolcpus=", isolated_cpu_setup);
5651 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5652 * to a function which identifies what group(along with sched group) a CPU
5653 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5654 * (due to the fact that we keep track of groups covered with a cpumask_t).
5656 * init_sched_build_groups will build a circular linked list of the groups
5657 * covered by the given span, and will set each group's ->cpumask correctly,
5658 * and ->cpu_power to 0.
5661 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5662 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5663 struct sched_group **sg))
5665 struct sched_group *first = NULL, *last = NULL;
5666 cpumask_t covered = CPU_MASK_NONE;
5669 for_each_cpu_mask(i, span) {
5670 struct sched_group *sg;
5671 int group = group_fn(i, cpu_map, &sg);
5674 if (cpu_isset(i, covered))
5677 sg->cpumask = CPU_MASK_NONE;
5678 sg->__cpu_power = 0;
5680 for_each_cpu_mask(j, span) {
5681 if (group_fn(j, cpu_map, NULL) != group)
5684 cpu_set(j, covered);
5685 cpu_set(j, sg->cpumask);
5696 #define SD_NODES_PER_DOMAIN 16
5701 * find_next_best_node - find the next node to include in a sched_domain
5702 * @node: node whose sched_domain we're building
5703 * @used_nodes: nodes already in the sched_domain
5705 * Find the next node to include in a given scheduling domain. Simply
5706 * finds the closest node not already in the @used_nodes map.
5708 * Should use nodemask_t.
5710 static int find_next_best_node(int node, unsigned long *used_nodes)
5712 int i, n, val, min_val, best_node = 0;
5716 for (i = 0; i < MAX_NUMNODES; i++) {
5717 /* Start at @node */
5718 n = (node + i) % MAX_NUMNODES;
5720 if (!nr_cpus_node(n))
5723 /* Skip already used nodes */
5724 if (test_bit(n, used_nodes))
5727 /* Simple min distance search */
5728 val = node_distance(node, n);
5730 if (val < min_val) {
5736 set_bit(best_node, used_nodes);
5741 * sched_domain_node_span - get a cpumask for a node's sched_domain
5742 * @node: node whose cpumask we're constructing
5743 * @size: number of nodes to include in this span
5745 * Given a node, construct a good cpumask for its sched_domain to span. It
5746 * should be one that prevents unnecessary balancing, but also spreads tasks
5749 static cpumask_t sched_domain_node_span(int node)
5751 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5752 cpumask_t span, nodemask;
5756 bitmap_zero(used_nodes, MAX_NUMNODES);
5758 nodemask = node_to_cpumask(node);
5759 cpus_or(span, span, nodemask);
5760 set_bit(node, used_nodes);
5762 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5763 int next_node = find_next_best_node(node, used_nodes);
5765 nodemask = node_to_cpumask(next_node);
5766 cpus_or(span, span, nodemask);
5773 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5776 * SMT sched-domains:
5778 #ifdef CONFIG_SCHED_SMT
5779 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5780 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5782 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5783 struct sched_group **sg)
5786 *sg = &per_cpu(sched_group_cpus, cpu);
5792 * multi-core sched-domains:
5794 #ifdef CONFIG_SCHED_MC
5795 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5796 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5799 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5800 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5801 struct sched_group **sg)
5804 cpumask_t mask = cpu_sibling_map[cpu];
5805 cpus_and(mask, mask, *cpu_map);
5806 group = first_cpu(mask);
5808 *sg = &per_cpu(sched_group_core, group);
5811 #elif defined(CONFIG_SCHED_MC)
5812 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5813 struct sched_group **sg)
5816 *sg = &per_cpu(sched_group_core, cpu);
5821 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5822 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5824 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5825 struct sched_group **sg)
5828 #ifdef CONFIG_SCHED_MC
5829 cpumask_t mask = cpu_coregroup_map(cpu);
5830 cpus_and(mask, mask, *cpu_map);
5831 group = first_cpu(mask);
5832 #elif defined(CONFIG_SCHED_SMT)
5833 cpumask_t mask = cpu_sibling_map[cpu];
5834 cpus_and(mask, mask, *cpu_map);
5835 group = first_cpu(mask);
5840 *sg = &per_cpu(sched_group_phys, group);
5846 * The init_sched_build_groups can't handle what we want to do with node
5847 * groups, so roll our own. Now each node has its own list of groups which
5848 * gets dynamically allocated.
5850 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5851 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5853 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5854 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5856 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5857 struct sched_group **sg)
5859 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5862 cpus_and(nodemask, nodemask, *cpu_map);
5863 group = first_cpu(nodemask);
5866 *sg = &per_cpu(sched_group_allnodes, group);
5870 static void init_numa_sched_groups_power(struct sched_group *group_head)
5872 struct sched_group *sg = group_head;
5878 for_each_cpu_mask(j, sg->cpumask) {
5879 struct sched_domain *sd;
5881 sd = &per_cpu(phys_domains, j);
5882 if (j != first_cpu(sd->groups->cpumask)) {
5884 * Only add "power" once for each
5890 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5893 if (sg != group_head)
5899 /* Free memory allocated for various sched_group structures */
5900 static void free_sched_groups(const cpumask_t *cpu_map)
5904 for_each_cpu_mask(cpu, *cpu_map) {
5905 struct sched_group **sched_group_nodes
5906 = sched_group_nodes_bycpu[cpu];
5908 if (!sched_group_nodes)
5911 for (i = 0; i < MAX_NUMNODES; i++) {
5912 cpumask_t nodemask = node_to_cpumask(i);
5913 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5915 cpus_and(nodemask, nodemask, *cpu_map);
5916 if (cpus_empty(nodemask))
5926 if (oldsg != sched_group_nodes[i])
5929 kfree(sched_group_nodes);
5930 sched_group_nodes_bycpu[cpu] = NULL;
5934 static void free_sched_groups(const cpumask_t *cpu_map)
5940 * Initialize sched groups cpu_power.
5942 * cpu_power indicates the capacity of sched group, which is used while
5943 * distributing the load between different sched groups in a sched domain.
5944 * Typically cpu_power for all the groups in a sched domain will be same unless
5945 * there are asymmetries in the topology. If there are asymmetries, group
5946 * having more cpu_power will pickup more load compared to the group having
5949 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5950 * the maximum number of tasks a group can handle in the presence of other idle
5951 * or lightly loaded groups in the same sched domain.
5953 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5955 struct sched_domain *child;
5956 struct sched_group *group;
5958 WARN_ON(!sd || !sd->groups);
5960 if (cpu != first_cpu(sd->groups->cpumask))
5965 sd->groups->__cpu_power = 0;
5968 * For perf policy, if the groups in child domain share resources
5969 * (for example cores sharing some portions of the cache hierarchy
5970 * or SMT), then set this domain groups cpu_power such that each group
5971 * can handle only one task, when there are other idle groups in the
5972 * same sched domain.
5974 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
5976 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5977 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
5982 * add cpu_power of each child group to this groups cpu_power
5984 group = child->groups;
5986 sg_inc_cpu_power(sd->groups, group->__cpu_power);
5987 group = group->next;
5988 } while (group != child->groups);
5992 * Build sched domains for a given set of cpus and attach the sched domains
5993 * to the individual cpus
5995 static int build_sched_domains(const cpumask_t *cpu_map)
5999 struct sched_group **sched_group_nodes = NULL;
6000 int sd_allnodes = 0;
6003 * Allocate the per-node list of sched groups
6005 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
6007 if (!sched_group_nodes) {
6008 printk(KERN_WARNING "Can not alloc sched group node list\n");
6011 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6015 * Set up domains for cpus specified by the cpu_map.
6017 for_each_cpu_mask(i, *cpu_map) {
6018 struct sched_domain *sd = NULL, *p;
6019 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6021 cpus_and(nodemask, nodemask, *cpu_map);
6024 if (cpus_weight(*cpu_map) >
6025 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6026 sd = &per_cpu(allnodes_domains, i);
6027 *sd = SD_ALLNODES_INIT;
6028 sd->span = *cpu_map;
6029 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6035 sd = &per_cpu(node_domains, i);
6037 sd->span = sched_domain_node_span(cpu_to_node(i));
6041 cpus_and(sd->span, sd->span, *cpu_map);
6045 sd = &per_cpu(phys_domains, i);
6047 sd->span = nodemask;
6051 cpu_to_phys_group(i, cpu_map, &sd->groups);
6053 #ifdef CONFIG_SCHED_MC
6055 sd = &per_cpu(core_domains, i);
6057 sd->span = cpu_coregroup_map(i);
6058 cpus_and(sd->span, sd->span, *cpu_map);
6061 cpu_to_core_group(i, cpu_map, &sd->groups);
6064 #ifdef CONFIG_SCHED_SMT
6066 sd = &per_cpu(cpu_domains, i);
6067 *sd = SD_SIBLING_INIT;
6068 sd->span = cpu_sibling_map[i];
6069 cpus_and(sd->span, sd->span, *cpu_map);
6072 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6076 #ifdef CONFIG_SCHED_SMT
6077 /* Set up CPU (sibling) groups */
6078 for_each_cpu_mask(i, *cpu_map) {
6079 cpumask_t this_sibling_map = cpu_sibling_map[i];
6080 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6081 if (i != first_cpu(this_sibling_map))
6084 init_sched_build_groups(this_sibling_map, cpu_map,
6089 #ifdef CONFIG_SCHED_MC
6090 /* Set up multi-core groups */
6091 for_each_cpu_mask(i, *cpu_map) {
6092 cpumask_t this_core_map = cpu_coregroup_map(i);
6093 cpus_and(this_core_map, this_core_map, *cpu_map);
6094 if (i != first_cpu(this_core_map))
6096 init_sched_build_groups(this_core_map, cpu_map,
6097 &cpu_to_core_group);
6101 /* Set up physical groups */
6102 for (i = 0; i < MAX_NUMNODES; i++) {
6103 cpumask_t nodemask = node_to_cpumask(i);
6105 cpus_and(nodemask, nodemask, *cpu_map);
6106 if (cpus_empty(nodemask))
6109 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6113 /* Set up node groups */
6115 init_sched_build_groups(*cpu_map, cpu_map,
6116 &cpu_to_allnodes_group);
6118 for (i = 0; i < MAX_NUMNODES; i++) {
6119 /* Set up node groups */
6120 struct sched_group *sg, *prev;
6121 cpumask_t nodemask = node_to_cpumask(i);
6122 cpumask_t domainspan;
6123 cpumask_t covered = CPU_MASK_NONE;
6126 cpus_and(nodemask, nodemask, *cpu_map);
6127 if (cpus_empty(nodemask)) {
6128 sched_group_nodes[i] = NULL;
6132 domainspan = sched_domain_node_span(i);
6133 cpus_and(domainspan, domainspan, *cpu_map);
6135 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6137 printk(KERN_WARNING "Can not alloc domain group for "
6141 sched_group_nodes[i] = sg;
6142 for_each_cpu_mask(j, nodemask) {
6143 struct sched_domain *sd;
6145 sd = &per_cpu(node_domains, j);
6148 sg->__cpu_power = 0;
6149 sg->cpumask = nodemask;
6151 cpus_or(covered, covered, nodemask);
6154 for (j = 0; j < MAX_NUMNODES; j++) {
6155 cpumask_t tmp, notcovered;
6156 int n = (i + j) % MAX_NUMNODES;
6158 cpus_complement(notcovered, covered);
6159 cpus_and(tmp, notcovered, *cpu_map);
6160 cpus_and(tmp, tmp, domainspan);
6161 if (cpus_empty(tmp))
6164 nodemask = node_to_cpumask(n);
6165 cpus_and(tmp, tmp, nodemask);
6166 if (cpus_empty(tmp))
6169 sg = kmalloc_node(sizeof(struct sched_group),
6173 "Can not alloc domain group for node %d\n", j);
6176 sg->__cpu_power = 0;
6178 sg->next = prev->next;
6179 cpus_or(covered, covered, tmp);
6186 /* Calculate CPU power for physical packages and nodes */
6187 #ifdef CONFIG_SCHED_SMT
6188 for_each_cpu_mask(i, *cpu_map) {
6189 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6191 init_sched_groups_power(i, sd);
6194 #ifdef CONFIG_SCHED_MC
6195 for_each_cpu_mask(i, *cpu_map) {
6196 struct sched_domain *sd = &per_cpu(core_domains, i);
6198 init_sched_groups_power(i, sd);
6202 for_each_cpu_mask(i, *cpu_map) {
6203 struct sched_domain *sd = &per_cpu(phys_domains, i);
6205 init_sched_groups_power(i, sd);
6209 for (i = 0; i < MAX_NUMNODES; i++)
6210 init_numa_sched_groups_power(sched_group_nodes[i]);
6213 struct sched_group *sg;
6215 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6216 init_numa_sched_groups_power(sg);
6220 /* Attach the domains */
6221 for_each_cpu_mask(i, *cpu_map) {
6222 struct sched_domain *sd;
6223 #ifdef CONFIG_SCHED_SMT
6224 sd = &per_cpu(cpu_domains, i);
6225 #elif defined(CONFIG_SCHED_MC)
6226 sd = &per_cpu(core_domains, i);
6228 sd = &per_cpu(phys_domains, i);
6230 cpu_attach_domain(sd, i);
6237 free_sched_groups(cpu_map);
6242 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6244 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6246 cpumask_t cpu_default_map;
6250 * Setup mask for cpus without special case scheduling requirements.
6251 * For now this just excludes isolated cpus, but could be used to
6252 * exclude other special cases in the future.
6254 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6256 err = build_sched_domains(&cpu_default_map);
6261 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6263 free_sched_groups(cpu_map);
6267 * Detach sched domains from a group of cpus specified in cpu_map
6268 * These cpus will now be attached to the NULL domain
6270 static void detach_destroy_domains(const cpumask_t *cpu_map)
6274 for_each_cpu_mask(i, *cpu_map)
6275 cpu_attach_domain(NULL, i);
6276 synchronize_sched();
6277 arch_destroy_sched_domains(cpu_map);
6281 * Partition sched domains as specified by the cpumasks below.
6282 * This attaches all cpus from the cpumasks to the NULL domain,
6283 * waits for a RCU quiescent period, recalculates sched
6284 * domain information and then attaches them back to the
6285 * correct sched domains
6286 * Call with hotplug lock held
6288 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6290 cpumask_t change_map;
6293 cpus_and(*partition1, *partition1, cpu_online_map);
6294 cpus_and(*partition2, *partition2, cpu_online_map);
6295 cpus_or(change_map, *partition1, *partition2);
6297 /* Detach sched domains from all of the affected cpus */
6298 detach_destroy_domains(&change_map);
6299 if (!cpus_empty(*partition1))
6300 err = build_sched_domains(partition1);
6301 if (!err && !cpus_empty(*partition2))
6302 err = build_sched_domains(partition2);
6307 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6308 static int arch_reinit_sched_domains(void)
6312 mutex_lock(&sched_hotcpu_mutex);
6313 detach_destroy_domains(&cpu_online_map);
6314 err = arch_init_sched_domains(&cpu_online_map);
6315 mutex_unlock(&sched_hotcpu_mutex);
6320 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6324 if (buf[0] != '0' && buf[0] != '1')
6328 sched_smt_power_savings = (buf[0] == '1');
6330 sched_mc_power_savings = (buf[0] == '1');
6332 ret = arch_reinit_sched_domains();
6334 return ret ? ret : count;
6337 #ifdef CONFIG_SCHED_MC
6338 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6340 return sprintf(page, "%u\n", sched_mc_power_savings);
6342 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6343 const char *buf, size_t count)
6345 return sched_power_savings_store(buf, count, 0);
6347 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6348 sched_mc_power_savings_store);
6351 #ifdef CONFIG_SCHED_SMT
6352 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6354 return sprintf(page, "%u\n", sched_smt_power_savings);
6356 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6357 const char *buf, size_t count)
6359 return sched_power_savings_store(buf, count, 1);
6361 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6362 sched_smt_power_savings_store);
6365 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6369 #ifdef CONFIG_SCHED_SMT
6371 err = sysfs_create_file(&cls->kset.kobj,
6372 &attr_sched_smt_power_savings.attr);
6374 #ifdef CONFIG_SCHED_MC
6375 if (!err && mc_capable())
6376 err = sysfs_create_file(&cls->kset.kobj,
6377 &attr_sched_mc_power_savings.attr);
6384 * Force a reinitialization of the sched domains hierarchy. The domains
6385 * and groups cannot be updated in place without racing with the balancing
6386 * code, so we temporarily attach all running cpus to the NULL domain
6387 * which will prevent rebalancing while the sched domains are recalculated.
6389 static int update_sched_domains(struct notifier_block *nfb,
6390 unsigned long action, void *hcpu)
6393 case CPU_UP_PREPARE:
6394 case CPU_UP_PREPARE_FROZEN:
6395 case CPU_DOWN_PREPARE:
6396 case CPU_DOWN_PREPARE_FROZEN:
6397 detach_destroy_domains(&cpu_online_map);
6400 case CPU_UP_CANCELED:
6401 case CPU_UP_CANCELED_FROZEN:
6402 case CPU_DOWN_FAILED:
6403 case CPU_DOWN_FAILED_FROZEN:
6405 case CPU_ONLINE_FROZEN:
6407 case CPU_DEAD_FROZEN:
6409 * Fall through and re-initialise the domains.
6416 /* The hotplug lock is already held by cpu_up/cpu_down */
6417 arch_init_sched_domains(&cpu_online_map);
6422 void __init sched_init_smp(void)
6424 cpumask_t non_isolated_cpus;
6426 mutex_lock(&sched_hotcpu_mutex);
6427 arch_init_sched_domains(&cpu_online_map);
6428 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6429 if (cpus_empty(non_isolated_cpus))
6430 cpu_set(smp_processor_id(), non_isolated_cpus);
6431 mutex_unlock(&sched_hotcpu_mutex);
6432 /* XXX: Theoretical race here - CPU may be hotplugged now */
6433 hotcpu_notifier(update_sched_domains, 0);
6435 init_sched_domain_sysctl();
6437 /* Move init over to a non-isolated CPU */
6438 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6442 void __init sched_init_smp(void)
6445 #endif /* CONFIG_SMP */
6447 int in_sched_functions(unsigned long addr)
6449 /* Linker adds these: start and end of __sched functions */
6450 extern char __sched_text_start[], __sched_text_end[];
6452 return in_lock_functions(addr) ||
6453 (addr >= (unsigned long)__sched_text_start
6454 && addr < (unsigned long)__sched_text_end);
6457 static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6459 cfs_rq->tasks_timeline = RB_ROOT;
6460 cfs_rq->fair_clock = 1;
6461 #ifdef CONFIG_FAIR_GROUP_SCHED
6466 void __init sched_init(void)
6468 int highest_cpu = 0;
6472 * Link up the scheduling class hierarchy:
6474 rt_sched_class.next = &fair_sched_class;
6475 fair_sched_class.next = &idle_sched_class;
6476 idle_sched_class.next = NULL;
6478 for_each_possible_cpu(i) {
6479 struct rt_prio_array *array;
6483 spin_lock_init(&rq->lock);
6484 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6487 init_cfs_rq(&rq->cfs, rq);
6488 #ifdef CONFIG_FAIR_GROUP_SCHED
6489 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6490 list_add(&rq->cfs.leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6493 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6494 rq->cpu_load[j] = 0;
6497 rq->active_balance = 0;
6498 rq->next_balance = jiffies;
6501 rq->migration_thread = NULL;
6502 INIT_LIST_HEAD(&rq->migration_queue);
6504 atomic_set(&rq->nr_iowait, 0);
6506 array = &rq->rt.active;
6507 for (j = 0; j < MAX_RT_PRIO; j++) {
6508 INIT_LIST_HEAD(array->queue + j);
6509 __clear_bit(j, array->bitmap);
6512 /* delimiter for bitsearch: */
6513 __set_bit(MAX_RT_PRIO, array->bitmap);
6516 set_load_weight(&init_task);
6518 #ifdef CONFIG_PREEMPT_NOTIFIERS
6519 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6523 nr_cpu_ids = highest_cpu + 1;
6524 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6527 #ifdef CONFIG_RT_MUTEXES
6528 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6532 * The boot idle thread does lazy MMU switching as well:
6534 atomic_inc(&init_mm.mm_count);
6535 enter_lazy_tlb(&init_mm, current);
6538 * Make us the idle thread. Technically, schedule() should not be
6539 * called from this thread, however somewhere below it might be,
6540 * but because we are the idle thread, we just pick up running again
6541 * when this runqueue becomes "idle".
6543 init_idle(current, smp_processor_id());
6545 * During early bootup we pretend to be a normal task:
6547 current->sched_class = &fair_sched_class;
6550 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6551 void __might_sleep(char *file, int line)
6554 static unsigned long prev_jiffy; /* ratelimiting */
6556 if ((in_atomic() || irqs_disabled()) &&
6557 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6558 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6560 prev_jiffy = jiffies;
6561 printk(KERN_ERR "BUG: sleeping function called from invalid"
6562 " context at %s:%d\n", file, line);
6563 printk("in_atomic():%d, irqs_disabled():%d\n",
6564 in_atomic(), irqs_disabled());
6565 debug_show_held_locks(current);
6566 if (irqs_disabled())
6567 print_irqtrace_events(current);
6572 EXPORT_SYMBOL(__might_sleep);
6575 #ifdef CONFIG_MAGIC_SYSRQ
6576 void normalize_rt_tasks(void)
6578 struct task_struct *g, *p;
6579 unsigned long flags;
6583 read_lock_irq(&tasklist_lock);
6584 do_each_thread(g, p) {
6586 p->se.wait_runtime = 0;
6587 p->se.exec_start = 0;
6588 p->se.wait_start_fair = 0;
6589 p->se.sleep_start_fair = 0;
6590 #ifdef CONFIG_SCHEDSTATS
6591 p->se.wait_start = 0;
6592 p->se.sleep_start = 0;
6593 p->se.block_start = 0;
6595 task_rq(p)->cfs.fair_clock = 0;
6596 task_rq(p)->clock = 0;
6600 * Renice negative nice level userspace
6603 if (TASK_NICE(p) < 0 && p->mm)
6604 set_user_nice(p, 0);
6608 spin_lock_irqsave(&p->pi_lock, flags);
6609 rq = __task_rq_lock(p);
6612 * Do not touch the migration thread:
6614 if (p == rq->migration_thread)
6618 update_rq_clock(rq);
6619 on_rq = p->se.on_rq;
6621 deactivate_task(rq, p, 0);
6622 __setscheduler(rq, p, SCHED_NORMAL, 0);
6624 activate_task(rq, p, 0);
6625 resched_task(rq->curr);
6630 __task_rq_unlock(rq);
6631 spin_unlock_irqrestore(&p->pi_lock, flags);
6632 } while_each_thread(g, p);
6634 read_unlock_irq(&tasklist_lock);
6637 #endif /* CONFIG_MAGIC_SYSRQ */
6641 * These functions are only useful for the IA64 MCA handling.
6643 * They can only be called when the whole system has been
6644 * stopped - every CPU needs to be quiescent, and no scheduling
6645 * activity can take place. Using them for anything else would
6646 * be a serious bug, and as a result, they aren't even visible
6647 * under any other configuration.
6651 * curr_task - return the current task for a given cpu.
6652 * @cpu: the processor in question.
6654 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6656 struct task_struct *curr_task(int cpu)
6658 return cpu_curr(cpu);
6662 * set_curr_task - set the current task for a given cpu.
6663 * @cpu: the processor in question.
6664 * @p: the task pointer to set.
6666 * Description: This function must only be used when non-maskable interrupts
6667 * are serviced on a separate stack. It allows the architecture to switch the
6668 * notion of the current task on a cpu in a non-blocking manner. This function
6669 * must be called with all CPU's synchronized, and interrupts disabled, the
6670 * and caller must save the original value of the current task (see
6671 * curr_task() above) and restore that value before reenabling interrupts and
6672 * re-starting the system.
6674 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6676 void set_curr_task(int cpu, struct task_struct *p)