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/pid_namespace.h>
48 #include <linux/smp.h>
49 #include <linux/threads.h>
50 #include <linux/timer.h>
51 #include <linux/rcupdate.h>
52 #include <linux/cpu.h>
53 #include <linux/cpuset.h>
54 #include <linux/percpu.h>
55 #include <linux/kthread.h>
56 #include <linux/seq_file.h>
57 #include <linux/sysctl.h>
58 #include <linux/syscalls.h>
59 #include <linux/times.h>
60 #include <linux/tsacct_kern.h>
61 #include <linux/kprobes.h>
62 #include <linux/delayacct.h>
63 #include <linux/reciprocal_div.h>
64 #include <linux/unistd.h>
65 #include <linux/pagemap.h>
68 #include <asm/irq_regs.h>
71 * Scheduler clock - returns current time in nanosec units.
72 * This is default implementation.
73 * Architectures and sub-architectures can override this.
75 unsigned long long __attribute__((weak)) sched_clock(void)
77 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Some helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / HZ))
104 #define NICE_0_LOAD SCHED_LOAD_SCALE
105 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
108 * These are the 'tuning knobs' of the scheduler:
110 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
111 * Timeslices get refilled after they expire.
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 static inline int rt_policy(int policy)
138 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
143 static inline int task_has_rt_policy(struct task_struct *p)
145 return rt_policy(p->policy);
149 * This is the priority-queue data structure of the RT scheduling class:
151 struct rt_prio_array {
152 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
153 struct list_head queue[MAX_RT_PRIO];
156 #ifdef CONFIG_FAIR_GROUP_SCHED
158 #include <linux/cgroup.h>
162 /* task group related information */
164 #ifdef CONFIG_FAIR_CGROUP_SCHED
165 struct cgroup_subsys_state css;
167 /* schedulable entities of this group on each cpu */
168 struct sched_entity **se;
169 /* runqueue "owned" by this group on each cpu */
170 struct cfs_rq **cfs_rq;
171 unsigned long shares;
172 /* spinlock to serialize modification to shares */
177 /* Default task group's sched entity on each cpu */
178 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
179 /* Default task group's cfs_rq on each cpu */
180 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
182 static struct sched_entity *init_sched_entity_p[NR_CPUS];
183 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
185 /* Default task group.
186 * Every task in system belong to this group at bootup.
188 struct task_group init_task_group = {
189 .se = init_sched_entity_p,
190 .cfs_rq = init_cfs_rq_p,
193 #ifdef CONFIG_FAIR_USER_SCHED
194 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
196 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
199 static int init_task_group_load = INIT_TASK_GRP_LOAD;
201 /* return group to which a task belongs */
202 static inline struct task_group *task_group(struct task_struct *p)
204 struct task_group *tg;
206 #ifdef CONFIG_FAIR_USER_SCHED
208 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
209 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
210 struct task_group, css);
212 tg = &init_task_group;
218 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
219 static inline void set_task_cfs_rq(struct task_struct *p)
221 p->se.cfs_rq = task_group(p)->cfs_rq[task_cpu(p)];
222 p->se.parent = task_group(p)->se[task_cpu(p)];
227 static inline void set_task_cfs_rq(struct task_struct *p) { }
229 #endif /* CONFIG_FAIR_GROUP_SCHED */
231 /* CFS-related fields in a runqueue */
233 struct load_weight load;
234 unsigned long nr_running;
239 struct rb_root tasks_timeline;
240 struct rb_node *rb_leftmost;
241 struct rb_node *rb_load_balance_curr;
242 /* 'curr' points to currently running entity on this cfs_rq.
243 * It is set to NULL otherwise (i.e when none are currently running).
245 struct sched_entity *curr;
247 unsigned long nr_spread_over;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
252 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
253 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
254 * (like users, containers etc.)
256 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
257 * list is used during load balance.
259 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
260 struct task_group *tg; /* group that "owns" this runqueue */
264 /* Real-Time classes' related field in a runqueue: */
266 struct rt_prio_array active;
267 int rt_load_balance_idx;
268 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
272 * This is the main, per-CPU runqueue data structure.
274 * Locking rule: those places that want to lock multiple runqueues
275 * (such as the load balancing or the thread migration code), lock
276 * acquire operations must be ordered by ascending &runqueue.
283 * nr_running and cpu_load should be in the same cacheline because
284 * remote CPUs use both these fields when doing load calculation.
286 unsigned long nr_running;
287 #define CPU_LOAD_IDX_MAX 5
288 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
289 unsigned char idle_at_tick;
291 unsigned char in_nohz_recently;
293 /* capture load from *all* tasks on this cpu: */
294 struct load_weight load;
295 unsigned long nr_load_updates;
299 #ifdef CONFIG_FAIR_GROUP_SCHED
300 /* list of leaf cfs_rq on this cpu: */
301 struct list_head leaf_cfs_rq_list;
306 * This is part of a global counter where only the total sum
307 * over all CPUs matters. A task can increase this counter on
308 * one CPU and if it got migrated afterwards it may decrease
309 * it on another CPU. Always updated under the runqueue lock:
311 unsigned long nr_uninterruptible;
313 struct task_struct *curr, *idle;
314 unsigned long next_balance;
315 struct mm_struct *prev_mm;
317 u64 clock, prev_clock_raw;
320 unsigned int clock_warps, clock_overflows;
322 unsigned int clock_deep_idle_events;
328 struct sched_domain *sd;
330 /* For active balancing */
333 /* cpu of this runqueue: */
336 struct task_struct *migration_thread;
337 struct list_head migration_queue;
340 #ifdef CONFIG_SCHEDSTATS
342 struct sched_info rq_sched_info;
344 /* sys_sched_yield() stats */
345 unsigned int yld_exp_empty;
346 unsigned int yld_act_empty;
347 unsigned int yld_both_empty;
348 unsigned int yld_count;
350 /* schedule() stats */
351 unsigned int sched_switch;
352 unsigned int sched_count;
353 unsigned int sched_goidle;
355 /* try_to_wake_up() stats */
356 unsigned int ttwu_count;
357 unsigned int ttwu_local;
360 unsigned int bkl_count;
362 struct lock_class_key rq_lock_key;
365 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
366 static DEFINE_MUTEX(sched_hotcpu_mutex);
368 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
370 rq->curr->sched_class->check_preempt_curr(rq, p);
373 static inline int cpu_of(struct rq *rq)
383 * Update the per-runqueue clock, as finegrained as the platform can give
384 * us, but without assuming monotonicity, etc.:
386 static void __update_rq_clock(struct rq *rq)
388 u64 prev_raw = rq->prev_clock_raw;
389 u64 now = sched_clock();
390 s64 delta = now - prev_raw;
391 u64 clock = rq->clock;
393 #ifdef CONFIG_SCHED_DEBUG
394 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
397 * Protect against sched_clock() occasionally going backwards:
399 if (unlikely(delta < 0)) {
404 * Catch too large forward jumps too:
406 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
407 if (clock < rq->tick_timestamp + TICK_NSEC)
408 clock = rq->tick_timestamp + TICK_NSEC;
411 rq->clock_overflows++;
413 if (unlikely(delta > rq->clock_max_delta))
414 rq->clock_max_delta = delta;
419 rq->prev_clock_raw = now;
423 static void update_rq_clock(struct rq *rq)
425 if (likely(smp_processor_id() == cpu_of(rq)))
426 __update_rq_clock(rq);
430 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
431 * See detach_destroy_domains: synchronize_sched for details.
433 * The domain tree of any CPU may only be accessed from within
434 * preempt-disabled sections.
436 #define for_each_domain(cpu, __sd) \
437 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
439 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
440 #define this_rq() (&__get_cpu_var(runqueues))
441 #define task_rq(p) cpu_rq(task_cpu(p))
442 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
445 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
447 #ifdef CONFIG_SCHED_DEBUG
448 # define const_debug __read_mostly
450 # define const_debug static const
454 * Debugging: various feature bits
457 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
458 SCHED_FEAT_START_DEBIT = 2,
459 SCHED_FEAT_TREE_AVG = 4,
460 SCHED_FEAT_APPROX_AVG = 8,
461 SCHED_FEAT_WAKEUP_PREEMPT = 16,
464 const_debug unsigned int sysctl_sched_features =
465 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
466 SCHED_FEAT_START_DEBIT * 1 |
467 SCHED_FEAT_TREE_AVG * 0 |
468 SCHED_FEAT_APPROX_AVG * 0 |
469 SCHED_FEAT_WAKEUP_PREEMPT * 1;
471 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
474 * Number of tasks to iterate in a single balance run.
475 * Limited because this is done with IRQs disabled.
477 const_debug unsigned int sysctl_sched_nr_migrate = 32;
480 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
481 * clock constructed from sched_clock():
483 unsigned long long cpu_clock(int cpu)
485 unsigned long long now;
489 local_irq_save(flags);
493 local_irq_restore(flags);
497 EXPORT_SYMBOL_GPL(cpu_clock);
499 #ifndef prepare_arch_switch
500 # define prepare_arch_switch(next) do { } while (0)
502 #ifndef finish_arch_switch
503 # define finish_arch_switch(prev) do { } while (0)
506 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
507 static inline int task_running(struct rq *rq, struct task_struct *p)
509 return rq->curr == p;
512 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
516 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
518 #ifdef CONFIG_DEBUG_SPINLOCK
519 /* this is a valid case when another task releases the spinlock */
520 rq->lock.owner = current;
523 * If we are tracking spinlock dependencies then we have to
524 * fix up the runqueue lock - which gets 'carried over' from
527 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
529 spin_unlock_irq(&rq->lock);
532 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
533 static inline int task_running(struct rq *rq, struct task_struct *p)
538 return rq->curr == p;
542 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
546 * We can optimise this out completely for !SMP, because the
547 * SMP rebalancing from interrupt is the only thing that cares
552 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
553 spin_unlock_irq(&rq->lock);
555 spin_unlock(&rq->lock);
559 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
563 * After ->oncpu is cleared, the task can be moved to a different CPU.
564 * We must ensure this doesn't happen until the switch is completely
570 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
574 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
577 * __task_rq_lock - lock the runqueue a given task resides on.
578 * Must be called interrupts disabled.
580 static inline struct rq *__task_rq_lock(struct task_struct *p)
584 struct rq *rq = task_rq(p);
585 spin_lock(&rq->lock);
586 if (likely(rq == task_rq(p)))
588 spin_unlock(&rq->lock);
593 * task_rq_lock - lock the runqueue a given task resides on and disable
594 * interrupts. Note the ordering: we can safely lookup the task_rq without
595 * explicitly disabling preemption.
597 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
603 local_irq_save(*flags);
605 spin_lock(&rq->lock);
606 if (likely(rq == task_rq(p)))
608 spin_unlock_irqrestore(&rq->lock, *flags);
612 static void __task_rq_unlock(struct rq *rq)
615 spin_unlock(&rq->lock);
618 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
621 spin_unlock_irqrestore(&rq->lock, *flags);
625 * this_rq_lock - lock this runqueue and disable interrupts.
627 static struct rq *this_rq_lock(void)
634 spin_lock(&rq->lock);
640 * We are going deep-idle (irqs are disabled):
642 void sched_clock_idle_sleep_event(void)
644 struct rq *rq = cpu_rq(smp_processor_id());
646 spin_lock(&rq->lock);
647 __update_rq_clock(rq);
648 spin_unlock(&rq->lock);
649 rq->clock_deep_idle_events++;
651 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
654 * We just idled delta nanoseconds (called with irqs disabled):
656 void sched_clock_idle_wakeup_event(u64 delta_ns)
658 struct rq *rq = cpu_rq(smp_processor_id());
659 u64 now = sched_clock();
661 rq->idle_clock += delta_ns;
663 * Override the previous timestamp and ignore all
664 * sched_clock() deltas that occured while we idled,
665 * and use the PM-provided delta_ns to advance the
668 spin_lock(&rq->lock);
669 rq->prev_clock_raw = now;
670 rq->clock += delta_ns;
671 spin_unlock(&rq->lock);
673 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
676 * resched_task - mark a task 'to be rescheduled now'.
678 * On UP this means the setting of the need_resched flag, on SMP it
679 * might also involve a cross-CPU call to trigger the scheduler on
684 #ifndef tsk_is_polling
685 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
688 static void resched_task(struct task_struct *p)
692 assert_spin_locked(&task_rq(p)->lock);
694 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
697 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
700 if (cpu == smp_processor_id())
703 /* NEED_RESCHED must be visible before we test polling */
705 if (!tsk_is_polling(p))
706 smp_send_reschedule(cpu);
709 static void resched_cpu(int cpu)
711 struct rq *rq = cpu_rq(cpu);
714 if (!spin_trylock_irqsave(&rq->lock, flags))
716 resched_task(cpu_curr(cpu));
717 spin_unlock_irqrestore(&rq->lock, flags);
720 static inline void resched_task(struct task_struct *p)
722 assert_spin_locked(&task_rq(p)->lock);
723 set_tsk_need_resched(p);
727 #if BITS_PER_LONG == 32
728 # define WMULT_CONST (~0UL)
730 # define WMULT_CONST (1UL << 32)
733 #define WMULT_SHIFT 32
736 * Shift right and round:
738 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
741 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
742 struct load_weight *lw)
746 if (unlikely(!lw->inv_weight))
747 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
749 tmp = (u64)delta_exec * weight;
751 * Check whether we'd overflow the 64-bit multiplication:
753 if (unlikely(tmp > WMULT_CONST))
754 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
757 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
759 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
762 static inline unsigned long
763 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
765 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
768 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
773 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
779 * To aid in avoiding the subversion of "niceness" due to uneven distribution
780 * of tasks with abnormal "nice" values across CPUs the contribution that
781 * each task makes to its run queue's load is weighted according to its
782 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
783 * scaled version of the new time slice allocation that they receive on time
787 #define WEIGHT_IDLEPRIO 2
788 #define WMULT_IDLEPRIO (1 << 31)
791 * Nice levels are multiplicative, with a gentle 10% change for every
792 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
793 * nice 1, it will get ~10% less CPU time than another CPU-bound task
794 * that remained on nice 0.
796 * The "10% effect" is relative and cumulative: from _any_ nice level,
797 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
798 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
799 * If a task goes up by ~10% and another task goes down by ~10% then
800 * the relative distance between them is ~25%.)
802 static const int prio_to_weight[40] = {
803 /* -20 */ 88761, 71755, 56483, 46273, 36291,
804 /* -15 */ 29154, 23254, 18705, 14949, 11916,
805 /* -10 */ 9548, 7620, 6100, 4904, 3906,
806 /* -5 */ 3121, 2501, 1991, 1586, 1277,
807 /* 0 */ 1024, 820, 655, 526, 423,
808 /* 5 */ 335, 272, 215, 172, 137,
809 /* 10 */ 110, 87, 70, 56, 45,
810 /* 15 */ 36, 29, 23, 18, 15,
814 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
816 * In cases where the weight does not change often, we can use the
817 * precalculated inverse to speed up arithmetics by turning divisions
818 * into multiplications:
820 static const u32 prio_to_wmult[40] = {
821 /* -20 */ 48388, 59856, 76040, 92818, 118348,
822 /* -15 */ 147320, 184698, 229616, 287308, 360437,
823 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
824 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
825 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
826 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
827 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
828 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
831 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
834 * runqueue iterator, to support SMP load-balancing between different
835 * scheduling classes, without having to expose their internal data
836 * structures to the load-balancing proper:
840 struct task_struct *(*start)(void *);
841 struct task_struct *(*next)(void *);
846 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
847 unsigned long max_load_move, struct sched_domain *sd,
848 enum cpu_idle_type idle, int *all_pinned,
849 int *this_best_prio, struct rq_iterator *iterator);
852 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
853 struct sched_domain *sd, enum cpu_idle_type idle,
854 struct rq_iterator *iterator);
857 #include "sched_stats.h"
858 #include "sched_idletask.c"
859 #include "sched_fair.c"
860 #include "sched_rt.c"
861 #ifdef CONFIG_SCHED_DEBUG
862 # include "sched_debug.c"
865 #define sched_class_highest (&rt_sched_class)
868 * Update delta_exec, delta_fair fields for rq.
870 * delta_fair clock advances at a rate inversely proportional to
871 * total load (rq->load.weight) on the runqueue, while
872 * delta_exec advances at the same rate as wall-clock (provided
875 * delta_exec / delta_fair is a measure of the (smoothened) load on this
876 * runqueue over any given interval. This (smoothened) load is used
877 * during load balance.
879 * This function is called /before/ updating rq->load
880 * and when switching tasks.
882 static inline void inc_load(struct rq *rq, const struct task_struct *p)
884 update_load_add(&rq->load, p->se.load.weight);
887 static inline void dec_load(struct rq *rq, const struct task_struct *p)
889 update_load_sub(&rq->load, p->se.load.weight);
892 static void inc_nr_running(struct task_struct *p, struct rq *rq)
898 static void dec_nr_running(struct task_struct *p, struct rq *rq)
904 static void set_load_weight(struct task_struct *p)
906 if (task_has_rt_policy(p)) {
907 p->se.load.weight = prio_to_weight[0] * 2;
908 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
913 * SCHED_IDLE tasks get minimal weight:
915 if (p->policy == SCHED_IDLE) {
916 p->se.load.weight = WEIGHT_IDLEPRIO;
917 p->se.load.inv_weight = WMULT_IDLEPRIO;
921 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
922 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
925 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
927 sched_info_queued(p);
928 p->sched_class->enqueue_task(rq, p, wakeup);
932 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
934 p->sched_class->dequeue_task(rq, p, sleep);
939 * __normal_prio - return the priority that is based on the static prio
941 static inline int __normal_prio(struct task_struct *p)
943 return p->static_prio;
947 * Calculate the expected normal priority: i.e. priority
948 * without taking RT-inheritance into account. Might be
949 * boosted by interactivity modifiers. Changes upon fork,
950 * setprio syscalls, and whenever the interactivity
951 * estimator recalculates.
953 static inline int normal_prio(struct task_struct *p)
957 if (task_has_rt_policy(p))
958 prio = MAX_RT_PRIO-1 - p->rt_priority;
960 prio = __normal_prio(p);
965 * Calculate the current priority, i.e. the priority
966 * taken into account by the scheduler. This value might
967 * be boosted by RT tasks, or might be boosted by
968 * interactivity modifiers. Will be RT if the task got
969 * RT-boosted. If not then it returns p->normal_prio.
971 static int effective_prio(struct task_struct *p)
973 p->normal_prio = normal_prio(p);
975 * If we are RT tasks or we were boosted to RT priority,
976 * keep the priority unchanged. Otherwise, update priority
977 * to the normal priority:
979 if (!rt_prio(p->prio))
980 return p->normal_prio;
985 * activate_task - move a task to the runqueue.
987 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
989 if (p->state == TASK_UNINTERRUPTIBLE)
990 rq->nr_uninterruptible--;
992 enqueue_task(rq, p, wakeup);
993 inc_nr_running(p, rq);
997 * deactivate_task - remove a task from the runqueue.
999 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1001 if (p->state == TASK_UNINTERRUPTIBLE)
1002 rq->nr_uninterruptible++;
1004 dequeue_task(rq, p, sleep);
1005 dec_nr_running(p, rq);
1009 * task_curr - is this task currently executing on a CPU?
1010 * @p: the task in question.
1012 inline int task_curr(const struct task_struct *p)
1014 return cpu_curr(task_cpu(p)) == p;
1017 /* Used instead of source_load when we know the type == 0 */
1018 unsigned long weighted_cpuload(const int cpu)
1020 return cpu_rq(cpu)->load.weight;
1023 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1026 task_thread_info(p)->cpu = cpu;
1034 * Is this task likely cache-hot:
1037 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1041 if (p->sched_class != &fair_sched_class)
1044 if (sysctl_sched_migration_cost == -1)
1046 if (sysctl_sched_migration_cost == 0)
1049 delta = now - p->se.exec_start;
1051 return delta < (s64)sysctl_sched_migration_cost;
1055 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1057 int old_cpu = task_cpu(p);
1058 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1059 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1060 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1063 clock_offset = old_rq->clock - new_rq->clock;
1065 #ifdef CONFIG_SCHEDSTATS
1066 if (p->se.wait_start)
1067 p->se.wait_start -= clock_offset;
1068 if (p->se.sleep_start)
1069 p->se.sleep_start -= clock_offset;
1070 if (p->se.block_start)
1071 p->se.block_start -= clock_offset;
1072 if (old_cpu != new_cpu) {
1073 schedstat_inc(p, se.nr_migrations);
1074 if (task_hot(p, old_rq->clock, NULL))
1075 schedstat_inc(p, se.nr_forced2_migrations);
1078 p->se.vruntime -= old_cfsrq->min_vruntime -
1079 new_cfsrq->min_vruntime;
1081 __set_task_cpu(p, new_cpu);
1084 struct migration_req {
1085 struct list_head list;
1087 struct task_struct *task;
1090 struct completion done;
1094 * The task's runqueue lock must be held.
1095 * Returns true if you have to wait for migration thread.
1098 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1100 struct rq *rq = task_rq(p);
1103 * If the task is not on a runqueue (and not running), then
1104 * it is sufficient to simply update the task's cpu field.
1106 if (!p->se.on_rq && !task_running(rq, p)) {
1107 set_task_cpu(p, dest_cpu);
1111 init_completion(&req->done);
1113 req->dest_cpu = dest_cpu;
1114 list_add(&req->list, &rq->migration_queue);
1120 * wait_task_inactive - wait for a thread to unschedule.
1122 * The caller must ensure that the task *will* unschedule sometime soon,
1123 * else this function might spin for a *long* time. This function can't
1124 * be called with interrupts off, or it may introduce deadlock with
1125 * smp_call_function() if an IPI is sent by the same process we are
1126 * waiting to become inactive.
1128 void wait_task_inactive(struct task_struct *p)
1130 unsigned long flags;
1136 * We do the initial early heuristics without holding
1137 * any task-queue locks at all. We'll only try to get
1138 * the runqueue lock when things look like they will
1144 * If the task is actively running on another CPU
1145 * still, just relax and busy-wait without holding
1148 * NOTE! Since we don't hold any locks, it's not
1149 * even sure that "rq" stays as the right runqueue!
1150 * But we don't care, since "task_running()" will
1151 * return false if the runqueue has changed and p
1152 * is actually now running somewhere else!
1154 while (task_running(rq, p))
1158 * Ok, time to look more closely! We need the rq
1159 * lock now, to be *sure*. If we're wrong, we'll
1160 * just go back and repeat.
1162 rq = task_rq_lock(p, &flags);
1163 running = task_running(rq, p);
1164 on_rq = p->se.on_rq;
1165 task_rq_unlock(rq, &flags);
1168 * Was it really running after all now that we
1169 * checked with the proper locks actually held?
1171 * Oops. Go back and try again..
1173 if (unlikely(running)) {
1179 * It's not enough that it's not actively running,
1180 * it must be off the runqueue _entirely_, and not
1183 * So if it wa still runnable (but just not actively
1184 * running right now), it's preempted, and we should
1185 * yield - it could be a while.
1187 if (unlikely(on_rq)) {
1188 schedule_timeout_uninterruptible(1);
1193 * Ahh, all good. It wasn't running, and it wasn't
1194 * runnable, which means that it will never become
1195 * running in the future either. We're all done!
1202 * kick_process - kick a running thread to enter/exit the kernel
1203 * @p: the to-be-kicked thread
1205 * Cause a process which is running on another CPU to enter
1206 * kernel-mode, without any delay. (to get signals handled.)
1208 * NOTE: this function doesnt have to take the runqueue lock,
1209 * because all it wants to ensure is that the remote task enters
1210 * the kernel. If the IPI races and the task has been migrated
1211 * to another CPU then no harm is done and the purpose has been
1214 void kick_process(struct task_struct *p)
1220 if ((cpu != smp_processor_id()) && task_curr(p))
1221 smp_send_reschedule(cpu);
1226 * Return a low guess at the load of a migration-source cpu weighted
1227 * according to the scheduling class and "nice" value.
1229 * We want to under-estimate the load of migration sources, to
1230 * balance conservatively.
1232 static unsigned long source_load(int cpu, int type)
1234 struct rq *rq = cpu_rq(cpu);
1235 unsigned long total = weighted_cpuload(cpu);
1240 return min(rq->cpu_load[type-1], total);
1244 * Return a high guess at the load of a migration-target cpu weighted
1245 * according to the scheduling class and "nice" value.
1247 static unsigned long target_load(int cpu, int type)
1249 struct rq *rq = cpu_rq(cpu);
1250 unsigned long total = weighted_cpuload(cpu);
1255 return max(rq->cpu_load[type-1], total);
1259 * Return the average load per task on the cpu's run queue
1261 static inline unsigned long cpu_avg_load_per_task(int cpu)
1263 struct rq *rq = cpu_rq(cpu);
1264 unsigned long total = weighted_cpuload(cpu);
1265 unsigned long n = rq->nr_running;
1267 return n ? total / n : SCHED_LOAD_SCALE;
1271 * find_idlest_group finds and returns the least busy CPU group within the
1274 static struct sched_group *
1275 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1277 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1278 unsigned long min_load = ULONG_MAX, this_load = 0;
1279 int load_idx = sd->forkexec_idx;
1280 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1283 unsigned long load, avg_load;
1287 /* Skip over this group if it has no CPUs allowed */
1288 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1291 local_group = cpu_isset(this_cpu, group->cpumask);
1293 /* Tally up the load of all CPUs in the group */
1296 for_each_cpu_mask(i, group->cpumask) {
1297 /* Bias balancing toward cpus of our domain */
1299 load = source_load(i, load_idx);
1301 load = target_load(i, load_idx);
1306 /* Adjust by relative CPU power of the group */
1307 avg_load = sg_div_cpu_power(group,
1308 avg_load * SCHED_LOAD_SCALE);
1311 this_load = avg_load;
1313 } else if (avg_load < min_load) {
1314 min_load = avg_load;
1317 } while (group = group->next, group != sd->groups);
1319 if (!idlest || 100*this_load < imbalance*min_load)
1325 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1328 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1331 unsigned long load, min_load = ULONG_MAX;
1335 /* Traverse only the allowed CPUs */
1336 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1338 for_each_cpu_mask(i, tmp) {
1339 load = weighted_cpuload(i);
1341 if (load < min_load || (load == min_load && i == this_cpu)) {
1351 * sched_balance_self: balance the current task (running on cpu) in domains
1352 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1355 * Balance, ie. select the least loaded group.
1357 * Returns the target CPU number, or the same CPU if no balancing is needed.
1359 * preempt must be disabled.
1361 static int sched_balance_self(int cpu, int flag)
1363 struct task_struct *t = current;
1364 struct sched_domain *tmp, *sd = NULL;
1366 for_each_domain(cpu, tmp) {
1368 * If power savings logic is enabled for a domain, stop there.
1370 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1372 if (tmp->flags & flag)
1378 struct sched_group *group;
1379 int new_cpu, weight;
1381 if (!(sd->flags & flag)) {
1387 group = find_idlest_group(sd, t, cpu);
1393 new_cpu = find_idlest_cpu(group, t, cpu);
1394 if (new_cpu == -1 || new_cpu == cpu) {
1395 /* Now try balancing at a lower domain level of cpu */
1400 /* Now try balancing at a lower domain level of new_cpu */
1403 weight = cpus_weight(span);
1404 for_each_domain(cpu, tmp) {
1405 if (weight <= cpus_weight(tmp->span))
1407 if (tmp->flags & flag)
1410 /* while loop will break here if sd == NULL */
1416 #endif /* CONFIG_SMP */
1419 * wake_idle() will wake a task on an idle cpu if task->cpu is
1420 * not idle and an idle cpu is available. The span of cpus to
1421 * search starts with cpus closest then further out as needed,
1422 * so we always favor a closer, idle cpu.
1424 * Returns the CPU we should wake onto.
1426 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1427 static int wake_idle(int cpu, struct task_struct *p)
1430 struct sched_domain *sd;
1434 * If it is idle, then it is the best cpu to run this task.
1436 * This cpu is also the best, if it has more than one task already.
1437 * Siblings must be also busy(in most cases) as they didn't already
1438 * pickup the extra load from this cpu and hence we need not check
1439 * sibling runqueue info. This will avoid the checks and cache miss
1440 * penalities associated with that.
1442 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1445 for_each_domain(cpu, sd) {
1446 if (sd->flags & SD_WAKE_IDLE) {
1447 cpus_and(tmp, sd->span, p->cpus_allowed);
1448 for_each_cpu_mask(i, tmp) {
1450 if (i != task_cpu(p)) {
1452 se.nr_wakeups_idle);
1464 static inline int wake_idle(int cpu, struct task_struct *p)
1471 * try_to_wake_up - wake up a thread
1472 * @p: the to-be-woken-up thread
1473 * @state: the mask of task states that can be woken
1474 * @sync: do a synchronous wakeup?
1476 * Put it on the run-queue if it's not already there. The "current"
1477 * thread is always on the run-queue (except when the actual
1478 * re-schedule is in progress), and as such you're allowed to do
1479 * the simpler "current->state = TASK_RUNNING" to mark yourself
1480 * runnable without the overhead of this.
1482 * returns failure only if the task is already active.
1484 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1486 int cpu, orig_cpu, this_cpu, success = 0;
1487 unsigned long flags;
1491 struct sched_domain *sd, *this_sd = NULL;
1492 unsigned long load, this_load;
1496 rq = task_rq_lock(p, &flags);
1497 old_state = p->state;
1498 if (!(old_state & state))
1506 this_cpu = smp_processor_id();
1509 if (unlikely(task_running(rq, p)))
1514 schedstat_inc(rq, ttwu_count);
1515 if (cpu == this_cpu) {
1516 schedstat_inc(rq, ttwu_local);
1520 for_each_domain(this_cpu, sd) {
1521 if (cpu_isset(cpu, sd->span)) {
1522 schedstat_inc(sd, ttwu_wake_remote);
1528 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1532 * Check for affine wakeup and passive balancing possibilities.
1535 int idx = this_sd->wake_idx;
1536 unsigned int imbalance;
1538 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1540 load = source_load(cpu, idx);
1541 this_load = target_load(this_cpu, idx);
1543 new_cpu = this_cpu; /* Wake to this CPU if we can */
1545 if (this_sd->flags & SD_WAKE_AFFINE) {
1546 unsigned long tl = this_load;
1547 unsigned long tl_per_task;
1550 * Attract cache-cold tasks on sync wakeups:
1552 if (sync && !task_hot(p, rq->clock, this_sd))
1555 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1556 tl_per_task = cpu_avg_load_per_task(this_cpu);
1559 * If sync wakeup then subtract the (maximum possible)
1560 * effect of the currently running task from the load
1561 * of the current CPU:
1564 tl -= current->se.load.weight;
1567 tl + target_load(cpu, idx) <= tl_per_task) ||
1568 100*(tl + p->se.load.weight) <= imbalance*load) {
1570 * This domain has SD_WAKE_AFFINE and
1571 * p is cache cold in this domain, and
1572 * there is no bad imbalance.
1574 schedstat_inc(this_sd, ttwu_move_affine);
1575 schedstat_inc(p, se.nr_wakeups_affine);
1581 * Start passive balancing when half the imbalance_pct
1584 if (this_sd->flags & SD_WAKE_BALANCE) {
1585 if (imbalance*this_load <= 100*load) {
1586 schedstat_inc(this_sd, ttwu_move_balance);
1587 schedstat_inc(p, se.nr_wakeups_passive);
1593 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1595 new_cpu = wake_idle(new_cpu, p);
1596 if (new_cpu != cpu) {
1597 set_task_cpu(p, new_cpu);
1598 task_rq_unlock(rq, &flags);
1599 /* might preempt at this point */
1600 rq = task_rq_lock(p, &flags);
1601 old_state = p->state;
1602 if (!(old_state & state))
1607 this_cpu = smp_processor_id();
1612 #endif /* CONFIG_SMP */
1613 schedstat_inc(p, se.nr_wakeups);
1615 schedstat_inc(p, se.nr_wakeups_sync);
1616 if (orig_cpu != cpu)
1617 schedstat_inc(p, se.nr_wakeups_migrate);
1618 if (cpu == this_cpu)
1619 schedstat_inc(p, se.nr_wakeups_local);
1621 schedstat_inc(p, se.nr_wakeups_remote);
1622 update_rq_clock(rq);
1623 activate_task(rq, p, 1);
1624 check_preempt_curr(rq, p);
1628 p->state = TASK_RUNNING;
1630 task_rq_unlock(rq, &flags);
1635 int fastcall wake_up_process(struct task_struct *p)
1637 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1638 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1640 EXPORT_SYMBOL(wake_up_process);
1642 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1644 return try_to_wake_up(p, state, 0);
1648 * Perform scheduler related setup for a newly forked process p.
1649 * p is forked by current.
1651 * __sched_fork() is basic setup used by init_idle() too:
1653 static void __sched_fork(struct task_struct *p)
1655 p->se.exec_start = 0;
1656 p->se.sum_exec_runtime = 0;
1657 p->se.prev_sum_exec_runtime = 0;
1659 #ifdef CONFIG_SCHEDSTATS
1660 p->se.wait_start = 0;
1661 p->se.sum_sleep_runtime = 0;
1662 p->se.sleep_start = 0;
1663 p->se.block_start = 0;
1664 p->se.sleep_max = 0;
1665 p->se.block_max = 0;
1667 p->se.slice_max = 0;
1671 INIT_LIST_HEAD(&p->run_list);
1674 #ifdef CONFIG_PREEMPT_NOTIFIERS
1675 INIT_HLIST_HEAD(&p->preempt_notifiers);
1679 * We mark the process as running here, but have not actually
1680 * inserted it onto the runqueue yet. This guarantees that
1681 * nobody will actually run it, and a signal or other external
1682 * event cannot wake it up and insert it on the runqueue either.
1684 p->state = TASK_RUNNING;
1688 * fork()/clone()-time setup:
1690 void sched_fork(struct task_struct *p, int clone_flags)
1692 int cpu = get_cpu();
1697 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1699 set_task_cpu(p, cpu);
1702 * Make sure we do not leak PI boosting priority to the child:
1704 p->prio = current->normal_prio;
1705 if (!rt_prio(p->prio))
1706 p->sched_class = &fair_sched_class;
1708 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1709 if (likely(sched_info_on()))
1710 memset(&p->sched_info, 0, sizeof(p->sched_info));
1712 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1715 #ifdef CONFIG_PREEMPT
1716 /* Want to start with kernel preemption disabled. */
1717 task_thread_info(p)->preempt_count = 1;
1723 * wake_up_new_task - wake up a newly created task for the first time.
1725 * This function will do some initial scheduler statistics housekeeping
1726 * that must be done for every newly created context, then puts the task
1727 * on the runqueue and wakes it.
1729 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1731 unsigned long flags;
1734 rq = task_rq_lock(p, &flags);
1735 BUG_ON(p->state != TASK_RUNNING);
1736 update_rq_clock(rq);
1738 p->prio = effective_prio(p);
1740 if (!p->sched_class->task_new || !current->se.on_rq) {
1741 activate_task(rq, p, 0);
1744 * Let the scheduling class do new task startup
1745 * management (if any):
1747 p->sched_class->task_new(rq, p);
1748 inc_nr_running(p, rq);
1750 check_preempt_curr(rq, p);
1751 task_rq_unlock(rq, &flags);
1754 #ifdef CONFIG_PREEMPT_NOTIFIERS
1757 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1758 * @notifier: notifier struct to register
1760 void preempt_notifier_register(struct preempt_notifier *notifier)
1762 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1764 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1767 * preempt_notifier_unregister - no longer interested in preemption notifications
1768 * @notifier: notifier struct to unregister
1770 * This is safe to call from within a preemption notifier.
1772 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1774 hlist_del(¬ifier->link);
1776 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1778 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1780 struct preempt_notifier *notifier;
1781 struct hlist_node *node;
1783 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1784 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1788 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1789 struct task_struct *next)
1791 struct preempt_notifier *notifier;
1792 struct hlist_node *node;
1794 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1795 notifier->ops->sched_out(notifier, next);
1800 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1805 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1806 struct task_struct *next)
1813 * prepare_task_switch - prepare to switch tasks
1814 * @rq: the runqueue preparing to switch
1815 * @prev: the current task that is being switched out
1816 * @next: the task we are going to switch to.
1818 * This is called with the rq lock held and interrupts off. It must
1819 * be paired with a subsequent finish_task_switch after the context
1822 * prepare_task_switch sets up locking and calls architecture specific
1826 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1827 struct task_struct *next)
1829 fire_sched_out_preempt_notifiers(prev, next);
1830 prepare_lock_switch(rq, next);
1831 prepare_arch_switch(next);
1835 * finish_task_switch - clean up after a task-switch
1836 * @rq: runqueue associated with task-switch
1837 * @prev: the thread we just switched away from.
1839 * finish_task_switch must be called after the context switch, paired
1840 * with a prepare_task_switch call before the context switch.
1841 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1842 * and do any other architecture-specific cleanup actions.
1844 * Note that we may have delayed dropping an mm in context_switch(). If
1845 * so, we finish that here outside of the runqueue lock. (Doing it
1846 * with the lock held can cause deadlocks; see schedule() for
1849 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1850 __releases(rq->lock)
1852 struct mm_struct *mm = rq->prev_mm;
1858 * A task struct has one reference for the use as "current".
1859 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1860 * schedule one last time. The schedule call will never return, and
1861 * the scheduled task must drop that reference.
1862 * The test for TASK_DEAD must occur while the runqueue locks are
1863 * still held, otherwise prev could be scheduled on another cpu, die
1864 * there before we look at prev->state, and then the reference would
1866 * Manfred Spraul <manfred@colorfullife.com>
1868 prev_state = prev->state;
1869 finish_arch_switch(prev);
1870 finish_lock_switch(rq, prev);
1871 fire_sched_in_preempt_notifiers(current);
1874 if (unlikely(prev_state == TASK_DEAD)) {
1876 * Remove function-return probe instances associated with this
1877 * task and put them back on the free list.
1879 kprobe_flush_task(prev);
1880 put_task_struct(prev);
1885 * schedule_tail - first thing a freshly forked thread must call.
1886 * @prev: the thread we just switched away from.
1888 asmlinkage void schedule_tail(struct task_struct *prev)
1889 __releases(rq->lock)
1891 struct rq *rq = this_rq();
1893 finish_task_switch(rq, prev);
1894 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1895 /* In this case, finish_task_switch does not reenable preemption */
1898 if (current->set_child_tid)
1899 put_user(task_pid_vnr(current), current->set_child_tid);
1903 * context_switch - switch to the new MM and the new
1904 * thread's register state.
1907 context_switch(struct rq *rq, struct task_struct *prev,
1908 struct task_struct *next)
1910 struct mm_struct *mm, *oldmm;
1912 prepare_task_switch(rq, prev, next);
1914 oldmm = prev->active_mm;
1916 * For paravirt, this is coupled with an exit in switch_to to
1917 * combine the page table reload and the switch backend into
1920 arch_enter_lazy_cpu_mode();
1922 if (unlikely(!mm)) {
1923 next->active_mm = oldmm;
1924 atomic_inc(&oldmm->mm_count);
1925 enter_lazy_tlb(oldmm, next);
1927 switch_mm(oldmm, mm, next);
1929 if (unlikely(!prev->mm)) {
1930 prev->active_mm = NULL;
1931 rq->prev_mm = oldmm;
1934 * Since the runqueue lock will be released by the next
1935 * task (which is an invalid locking op but in the case
1936 * of the scheduler it's an obvious special-case), so we
1937 * do an early lockdep release here:
1939 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1940 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1943 /* Here we just switch the register state and the stack. */
1944 switch_to(prev, next, prev);
1948 * this_rq must be evaluated again because prev may have moved
1949 * CPUs since it called schedule(), thus the 'rq' on its stack
1950 * frame will be invalid.
1952 finish_task_switch(this_rq(), prev);
1956 * nr_running, nr_uninterruptible and nr_context_switches:
1958 * externally visible scheduler statistics: current number of runnable
1959 * threads, current number of uninterruptible-sleeping threads, total
1960 * number of context switches performed since bootup.
1962 unsigned long nr_running(void)
1964 unsigned long i, sum = 0;
1966 for_each_online_cpu(i)
1967 sum += cpu_rq(i)->nr_running;
1972 unsigned long nr_uninterruptible(void)
1974 unsigned long i, sum = 0;
1976 for_each_possible_cpu(i)
1977 sum += cpu_rq(i)->nr_uninterruptible;
1980 * Since we read the counters lockless, it might be slightly
1981 * inaccurate. Do not allow it to go below zero though:
1983 if (unlikely((long)sum < 0))
1989 unsigned long long nr_context_switches(void)
1992 unsigned long long sum = 0;
1994 for_each_possible_cpu(i)
1995 sum += cpu_rq(i)->nr_switches;
2000 unsigned long nr_iowait(void)
2002 unsigned long i, sum = 0;
2004 for_each_possible_cpu(i)
2005 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2010 unsigned long nr_active(void)
2012 unsigned long i, running = 0, uninterruptible = 0;
2014 for_each_online_cpu(i) {
2015 running += cpu_rq(i)->nr_running;
2016 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2019 if (unlikely((long)uninterruptible < 0))
2020 uninterruptible = 0;
2022 return running + uninterruptible;
2026 * Update rq->cpu_load[] statistics. This function is usually called every
2027 * scheduler tick (TICK_NSEC).
2029 static void update_cpu_load(struct rq *this_rq)
2031 unsigned long this_load = this_rq->load.weight;
2034 this_rq->nr_load_updates++;
2036 /* Update our load: */
2037 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2038 unsigned long old_load, new_load;
2040 /* scale is effectively 1 << i now, and >> i divides by scale */
2042 old_load = this_rq->cpu_load[i];
2043 new_load = this_load;
2045 * Round up the averaging division if load is increasing. This
2046 * prevents us from getting stuck on 9 if the load is 10, for
2049 if (new_load > old_load)
2050 new_load += scale-1;
2051 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2058 * double_rq_lock - safely lock two runqueues
2060 * Note this does not disable interrupts like task_rq_lock,
2061 * you need to do so manually before calling.
2063 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2064 __acquires(rq1->lock)
2065 __acquires(rq2->lock)
2067 BUG_ON(!irqs_disabled());
2069 spin_lock(&rq1->lock);
2070 __acquire(rq2->lock); /* Fake it out ;) */
2073 spin_lock(&rq1->lock);
2074 spin_lock(&rq2->lock);
2076 spin_lock(&rq2->lock);
2077 spin_lock(&rq1->lock);
2080 update_rq_clock(rq1);
2081 update_rq_clock(rq2);
2085 * double_rq_unlock - safely unlock two runqueues
2087 * Note this does not restore interrupts like task_rq_unlock,
2088 * you need to do so manually after calling.
2090 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2091 __releases(rq1->lock)
2092 __releases(rq2->lock)
2094 spin_unlock(&rq1->lock);
2096 spin_unlock(&rq2->lock);
2098 __release(rq2->lock);
2102 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2104 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2105 __releases(this_rq->lock)
2106 __acquires(busiest->lock)
2107 __acquires(this_rq->lock)
2109 if (unlikely(!irqs_disabled())) {
2110 /* printk() doesn't work good under rq->lock */
2111 spin_unlock(&this_rq->lock);
2114 if (unlikely(!spin_trylock(&busiest->lock))) {
2115 if (busiest < this_rq) {
2116 spin_unlock(&this_rq->lock);
2117 spin_lock(&busiest->lock);
2118 spin_lock(&this_rq->lock);
2120 spin_lock(&busiest->lock);
2125 * If dest_cpu is allowed for this process, migrate the task to it.
2126 * This is accomplished by forcing the cpu_allowed mask to only
2127 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2128 * the cpu_allowed mask is restored.
2130 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2132 struct migration_req req;
2133 unsigned long flags;
2136 rq = task_rq_lock(p, &flags);
2137 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2138 || unlikely(cpu_is_offline(dest_cpu)))
2141 /* force the process onto the specified CPU */
2142 if (migrate_task(p, dest_cpu, &req)) {
2143 /* Need to wait for migration thread (might exit: take ref). */
2144 struct task_struct *mt = rq->migration_thread;
2146 get_task_struct(mt);
2147 task_rq_unlock(rq, &flags);
2148 wake_up_process(mt);
2149 put_task_struct(mt);
2150 wait_for_completion(&req.done);
2155 task_rq_unlock(rq, &flags);
2159 * sched_exec - execve() is a valuable balancing opportunity, because at
2160 * this point the task has the smallest effective memory and cache footprint.
2162 void sched_exec(void)
2164 int new_cpu, this_cpu = get_cpu();
2165 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2167 if (new_cpu != this_cpu)
2168 sched_migrate_task(current, new_cpu);
2172 * pull_task - move a task from a remote runqueue to the local runqueue.
2173 * Both runqueues must be locked.
2175 static void pull_task(struct rq *src_rq, struct task_struct *p,
2176 struct rq *this_rq, int this_cpu)
2178 deactivate_task(src_rq, p, 0);
2179 set_task_cpu(p, this_cpu);
2180 activate_task(this_rq, p, 0);
2182 * Note that idle threads have a prio of MAX_PRIO, for this test
2183 * to be always true for them.
2185 check_preempt_curr(this_rq, p);
2189 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2192 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2193 struct sched_domain *sd, enum cpu_idle_type idle,
2197 * We do not migrate tasks that are:
2198 * 1) running (obviously), or
2199 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2200 * 3) are cache-hot on their current CPU.
2202 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2203 schedstat_inc(p, se.nr_failed_migrations_affine);
2208 if (task_running(rq, p)) {
2209 schedstat_inc(p, se.nr_failed_migrations_running);
2214 * Aggressive migration if:
2215 * 1) task is cache cold, or
2216 * 2) too many balance attempts have failed.
2219 if (!task_hot(p, rq->clock, sd) ||
2220 sd->nr_balance_failed > sd->cache_nice_tries) {
2221 #ifdef CONFIG_SCHEDSTATS
2222 if (task_hot(p, rq->clock, sd)) {
2223 schedstat_inc(sd, lb_hot_gained[idle]);
2224 schedstat_inc(p, se.nr_forced_migrations);
2230 if (task_hot(p, rq->clock, sd)) {
2231 schedstat_inc(p, se.nr_failed_migrations_hot);
2237 static unsigned long
2238 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2239 unsigned long max_load_move, struct sched_domain *sd,
2240 enum cpu_idle_type idle, int *all_pinned,
2241 int *this_best_prio, struct rq_iterator *iterator)
2243 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2244 struct task_struct *p;
2245 long rem_load_move = max_load_move;
2247 if (max_load_move == 0)
2253 * Start the load-balancing iterator:
2255 p = iterator->start(iterator->arg);
2257 if (!p || loops++ > sysctl_sched_nr_migrate)
2260 * To help distribute high priority tasks across CPUs we don't
2261 * skip a task if it will be the highest priority task (i.e. smallest
2262 * prio value) on its new queue regardless of its load weight
2264 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2265 SCHED_LOAD_SCALE_FUZZ;
2266 if ((skip_for_load && p->prio >= *this_best_prio) ||
2267 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2268 p = iterator->next(iterator->arg);
2272 pull_task(busiest, p, this_rq, this_cpu);
2274 rem_load_move -= p->se.load.weight;
2277 * We only want to steal up to the prescribed amount of weighted load.
2279 if (rem_load_move > 0) {
2280 if (p->prio < *this_best_prio)
2281 *this_best_prio = p->prio;
2282 p = iterator->next(iterator->arg);
2287 * Right now, this is one of only two places pull_task() is called,
2288 * so we can safely collect pull_task() stats here rather than
2289 * inside pull_task().
2291 schedstat_add(sd, lb_gained[idle], pulled);
2294 *all_pinned = pinned;
2296 return max_load_move - rem_load_move;
2300 * move_tasks tries to move up to max_load_move weighted load from busiest to
2301 * this_rq, as part of a balancing operation within domain "sd".
2302 * Returns 1 if successful and 0 otherwise.
2304 * Called with both runqueues locked.
2306 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2307 unsigned long max_load_move,
2308 struct sched_domain *sd, enum cpu_idle_type idle,
2311 const struct sched_class *class = sched_class_highest;
2312 unsigned long total_load_moved = 0;
2313 int this_best_prio = this_rq->curr->prio;
2317 class->load_balance(this_rq, this_cpu, busiest,
2318 max_load_move - total_load_moved,
2319 sd, idle, all_pinned, &this_best_prio);
2320 class = class->next;
2321 } while (class && max_load_move > total_load_moved);
2323 return total_load_moved > 0;
2327 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2328 struct sched_domain *sd, enum cpu_idle_type idle,
2329 struct rq_iterator *iterator)
2331 struct task_struct *p = iterator->start(iterator->arg);
2335 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2336 pull_task(busiest, p, this_rq, this_cpu);
2338 * Right now, this is only the second place pull_task()
2339 * is called, so we can safely collect pull_task()
2340 * stats here rather than inside pull_task().
2342 schedstat_inc(sd, lb_gained[idle]);
2346 p = iterator->next(iterator->arg);
2353 * move_one_task tries to move exactly one task from busiest to this_rq, as
2354 * part of active balancing operations within "domain".
2355 * Returns 1 if successful and 0 otherwise.
2357 * Called with both runqueues locked.
2359 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2360 struct sched_domain *sd, enum cpu_idle_type idle)
2362 const struct sched_class *class;
2364 for (class = sched_class_highest; class; class = class->next)
2365 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2372 * find_busiest_group finds and returns the busiest CPU group within the
2373 * domain. It calculates and returns the amount of weighted load which
2374 * should be moved to restore balance via the imbalance parameter.
2376 static struct sched_group *
2377 find_busiest_group(struct sched_domain *sd, int this_cpu,
2378 unsigned long *imbalance, enum cpu_idle_type idle,
2379 int *sd_idle, cpumask_t *cpus, int *balance)
2381 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2382 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2383 unsigned long max_pull;
2384 unsigned long busiest_load_per_task, busiest_nr_running;
2385 unsigned long this_load_per_task, this_nr_running;
2386 int load_idx, group_imb = 0;
2387 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2388 int power_savings_balance = 1;
2389 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2390 unsigned long min_nr_running = ULONG_MAX;
2391 struct sched_group *group_min = NULL, *group_leader = NULL;
2394 max_load = this_load = total_load = total_pwr = 0;
2395 busiest_load_per_task = busiest_nr_running = 0;
2396 this_load_per_task = this_nr_running = 0;
2397 if (idle == CPU_NOT_IDLE)
2398 load_idx = sd->busy_idx;
2399 else if (idle == CPU_NEWLY_IDLE)
2400 load_idx = sd->newidle_idx;
2402 load_idx = sd->idle_idx;
2405 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2408 int __group_imb = 0;
2409 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2410 unsigned long sum_nr_running, sum_weighted_load;
2412 local_group = cpu_isset(this_cpu, group->cpumask);
2415 balance_cpu = first_cpu(group->cpumask);
2417 /* Tally up the load of all CPUs in the group */
2418 sum_weighted_load = sum_nr_running = avg_load = 0;
2420 min_cpu_load = ~0UL;
2422 for_each_cpu_mask(i, group->cpumask) {
2425 if (!cpu_isset(i, *cpus))
2430 if (*sd_idle && rq->nr_running)
2433 /* Bias balancing toward cpus of our domain */
2435 if (idle_cpu(i) && !first_idle_cpu) {
2440 load = target_load(i, load_idx);
2442 load = source_load(i, load_idx);
2443 if (load > max_cpu_load)
2444 max_cpu_load = load;
2445 if (min_cpu_load > load)
2446 min_cpu_load = load;
2450 sum_nr_running += rq->nr_running;
2451 sum_weighted_load += weighted_cpuload(i);
2455 * First idle cpu or the first cpu(busiest) in this sched group
2456 * is eligible for doing load balancing at this and above
2457 * domains. In the newly idle case, we will allow all the cpu's
2458 * to do the newly idle load balance.
2460 if (idle != CPU_NEWLY_IDLE && local_group &&
2461 balance_cpu != this_cpu && balance) {
2466 total_load += avg_load;
2467 total_pwr += group->__cpu_power;
2469 /* Adjust by relative CPU power of the group */
2470 avg_load = sg_div_cpu_power(group,
2471 avg_load * SCHED_LOAD_SCALE);
2473 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2476 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2479 this_load = avg_load;
2481 this_nr_running = sum_nr_running;
2482 this_load_per_task = sum_weighted_load;
2483 } else if (avg_load > max_load &&
2484 (sum_nr_running > group_capacity || __group_imb)) {
2485 max_load = avg_load;
2487 busiest_nr_running = sum_nr_running;
2488 busiest_load_per_task = sum_weighted_load;
2489 group_imb = __group_imb;
2492 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2494 * Busy processors will not participate in power savings
2497 if (idle == CPU_NOT_IDLE ||
2498 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2502 * If the local group is idle or completely loaded
2503 * no need to do power savings balance at this domain
2505 if (local_group && (this_nr_running >= group_capacity ||
2507 power_savings_balance = 0;
2510 * If a group is already running at full capacity or idle,
2511 * don't include that group in power savings calculations
2513 if (!power_savings_balance || sum_nr_running >= group_capacity
2518 * Calculate the group which has the least non-idle load.
2519 * This is the group from where we need to pick up the load
2522 if ((sum_nr_running < min_nr_running) ||
2523 (sum_nr_running == min_nr_running &&
2524 first_cpu(group->cpumask) <
2525 first_cpu(group_min->cpumask))) {
2527 min_nr_running = sum_nr_running;
2528 min_load_per_task = sum_weighted_load /
2533 * Calculate the group which is almost near its
2534 * capacity but still has some space to pick up some load
2535 * from other group and save more power
2537 if (sum_nr_running <= group_capacity - 1) {
2538 if (sum_nr_running > leader_nr_running ||
2539 (sum_nr_running == leader_nr_running &&
2540 first_cpu(group->cpumask) >
2541 first_cpu(group_leader->cpumask))) {
2542 group_leader = group;
2543 leader_nr_running = sum_nr_running;
2548 group = group->next;
2549 } while (group != sd->groups);
2551 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2554 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2556 if (this_load >= avg_load ||
2557 100*max_load <= sd->imbalance_pct*this_load)
2560 busiest_load_per_task /= busiest_nr_running;
2562 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2565 * We're trying to get all the cpus to the average_load, so we don't
2566 * want to push ourselves above the average load, nor do we wish to
2567 * reduce the max loaded cpu below the average load, as either of these
2568 * actions would just result in more rebalancing later, and ping-pong
2569 * tasks around. Thus we look for the minimum possible imbalance.
2570 * Negative imbalances (*we* are more loaded than anyone else) will
2571 * be counted as no imbalance for these purposes -- we can't fix that
2572 * by pulling tasks to us. Be careful of negative numbers as they'll
2573 * appear as very large values with unsigned longs.
2575 if (max_load <= busiest_load_per_task)
2579 * In the presence of smp nice balancing, certain scenarios can have
2580 * max load less than avg load(as we skip the groups at or below
2581 * its cpu_power, while calculating max_load..)
2583 if (max_load < avg_load) {
2585 goto small_imbalance;
2588 /* Don't want to pull so many tasks that a group would go idle */
2589 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2591 /* How much load to actually move to equalise the imbalance */
2592 *imbalance = min(max_pull * busiest->__cpu_power,
2593 (avg_load - this_load) * this->__cpu_power)
2597 * if *imbalance is less than the average load per runnable task
2598 * there is no gaurantee that any tasks will be moved so we'll have
2599 * a think about bumping its value to force at least one task to be
2602 if (*imbalance < busiest_load_per_task) {
2603 unsigned long tmp, pwr_now, pwr_move;
2607 pwr_move = pwr_now = 0;
2609 if (this_nr_running) {
2610 this_load_per_task /= this_nr_running;
2611 if (busiest_load_per_task > this_load_per_task)
2614 this_load_per_task = SCHED_LOAD_SCALE;
2616 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2617 busiest_load_per_task * imbn) {
2618 *imbalance = busiest_load_per_task;
2623 * OK, we don't have enough imbalance to justify moving tasks,
2624 * however we may be able to increase total CPU power used by
2628 pwr_now += busiest->__cpu_power *
2629 min(busiest_load_per_task, max_load);
2630 pwr_now += this->__cpu_power *
2631 min(this_load_per_task, this_load);
2632 pwr_now /= SCHED_LOAD_SCALE;
2634 /* Amount of load we'd subtract */
2635 tmp = sg_div_cpu_power(busiest,
2636 busiest_load_per_task * SCHED_LOAD_SCALE);
2638 pwr_move += busiest->__cpu_power *
2639 min(busiest_load_per_task, max_load - tmp);
2641 /* Amount of load we'd add */
2642 if (max_load * busiest->__cpu_power <
2643 busiest_load_per_task * SCHED_LOAD_SCALE)
2644 tmp = sg_div_cpu_power(this,
2645 max_load * busiest->__cpu_power);
2647 tmp = sg_div_cpu_power(this,
2648 busiest_load_per_task * SCHED_LOAD_SCALE);
2649 pwr_move += this->__cpu_power *
2650 min(this_load_per_task, this_load + tmp);
2651 pwr_move /= SCHED_LOAD_SCALE;
2653 /* Move if we gain throughput */
2654 if (pwr_move > pwr_now)
2655 *imbalance = busiest_load_per_task;
2661 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2662 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2665 if (this == group_leader && group_leader != group_min) {
2666 *imbalance = min_load_per_task;
2676 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2679 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2680 unsigned long imbalance, cpumask_t *cpus)
2682 struct rq *busiest = NULL, *rq;
2683 unsigned long max_load = 0;
2686 for_each_cpu_mask(i, group->cpumask) {
2689 if (!cpu_isset(i, *cpus))
2693 wl = weighted_cpuload(i);
2695 if (rq->nr_running == 1 && wl > imbalance)
2698 if (wl > max_load) {
2708 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2709 * so long as it is large enough.
2711 #define MAX_PINNED_INTERVAL 512
2714 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2715 * tasks if there is an imbalance.
2717 static int load_balance(int this_cpu, struct rq *this_rq,
2718 struct sched_domain *sd, enum cpu_idle_type idle,
2721 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2722 struct sched_group *group;
2723 unsigned long imbalance;
2725 cpumask_t cpus = CPU_MASK_ALL;
2726 unsigned long flags;
2729 * When power savings policy is enabled for the parent domain, idle
2730 * sibling can pick up load irrespective of busy siblings. In this case,
2731 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2732 * portraying it as CPU_NOT_IDLE.
2734 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2735 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2738 schedstat_inc(sd, lb_count[idle]);
2741 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2748 schedstat_inc(sd, lb_nobusyg[idle]);
2752 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2754 schedstat_inc(sd, lb_nobusyq[idle]);
2758 BUG_ON(busiest == this_rq);
2760 schedstat_add(sd, lb_imbalance[idle], imbalance);
2763 if (busiest->nr_running > 1) {
2765 * Attempt to move tasks. If find_busiest_group has found
2766 * an imbalance but busiest->nr_running <= 1, the group is
2767 * still unbalanced. ld_moved simply stays zero, so it is
2768 * correctly treated as an imbalance.
2770 local_irq_save(flags);
2771 double_rq_lock(this_rq, busiest);
2772 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2773 imbalance, sd, idle, &all_pinned);
2774 double_rq_unlock(this_rq, busiest);
2775 local_irq_restore(flags);
2778 * some other cpu did the load balance for us.
2780 if (ld_moved && this_cpu != smp_processor_id())
2781 resched_cpu(this_cpu);
2783 /* All tasks on this runqueue were pinned by CPU affinity */
2784 if (unlikely(all_pinned)) {
2785 cpu_clear(cpu_of(busiest), cpus);
2786 if (!cpus_empty(cpus))
2793 schedstat_inc(sd, lb_failed[idle]);
2794 sd->nr_balance_failed++;
2796 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2798 spin_lock_irqsave(&busiest->lock, flags);
2800 /* don't kick the migration_thread, if the curr
2801 * task on busiest cpu can't be moved to this_cpu
2803 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2804 spin_unlock_irqrestore(&busiest->lock, flags);
2806 goto out_one_pinned;
2809 if (!busiest->active_balance) {
2810 busiest->active_balance = 1;
2811 busiest->push_cpu = this_cpu;
2814 spin_unlock_irqrestore(&busiest->lock, flags);
2816 wake_up_process(busiest->migration_thread);
2819 * We've kicked active balancing, reset the failure
2822 sd->nr_balance_failed = sd->cache_nice_tries+1;
2825 sd->nr_balance_failed = 0;
2827 if (likely(!active_balance)) {
2828 /* We were unbalanced, so reset the balancing interval */
2829 sd->balance_interval = sd->min_interval;
2832 * If we've begun active balancing, start to back off. This
2833 * case may not be covered by the all_pinned logic if there
2834 * is only 1 task on the busy runqueue (because we don't call
2837 if (sd->balance_interval < sd->max_interval)
2838 sd->balance_interval *= 2;
2841 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2842 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2847 schedstat_inc(sd, lb_balanced[idle]);
2849 sd->nr_balance_failed = 0;
2852 /* tune up the balancing interval */
2853 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2854 (sd->balance_interval < sd->max_interval))
2855 sd->balance_interval *= 2;
2857 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2858 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2864 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2865 * tasks if there is an imbalance.
2867 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2868 * this_rq is locked.
2871 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2873 struct sched_group *group;
2874 struct rq *busiest = NULL;
2875 unsigned long imbalance;
2879 cpumask_t cpus = CPU_MASK_ALL;
2882 * When power savings policy is enabled for the parent domain, idle
2883 * sibling can pick up load irrespective of busy siblings. In this case,
2884 * let the state of idle sibling percolate up as IDLE, instead of
2885 * portraying it as CPU_NOT_IDLE.
2887 if (sd->flags & SD_SHARE_CPUPOWER &&
2888 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2891 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2893 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2894 &sd_idle, &cpus, NULL);
2896 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2900 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2903 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2907 BUG_ON(busiest == this_rq);
2909 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2912 if (busiest->nr_running > 1) {
2913 /* Attempt to move tasks */
2914 double_lock_balance(this_rq, busiest);
2915 /* this_rq->clock is already updated */
2916 update_rq_clock(busiest);
2917 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2918 imbalance, sd, CPU_NEWLY_IDLE,
2920 spin_unlock(&busiest->lock);
2922 if (unlikely(all_pinned)) {
2923 cpu_clear(cpu_of(busiest), cpus);
2924 if (!cpus_empty(cpus))
2930 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2931 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2932 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2935 sd->nr_balance_failed = 0;
2940 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2941 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2942 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2944 sd->nr_balance_failed = 0;
2950 * idle_balance is called by schedule() if this_cpu is about to become
2951 * idle. Attempts to pull tasks from other CPUs.
2953 static void idle_balance(int this_cpu, struct rq *this_rq)
2955 struct sched_domain *sd;
2956 int pulled_task = -1;
2957 unsigned long next_balance = jiffies + HZ;
2959 for_each_domain(this_cpu, sd) {
2960 unsigned long interval;
2962 if (!(sd->flags & SD_LOAD_BALANCE))
2965 if (sd->flags & SD_BALANCE_NEWIDLE)
2966 /* If we've pulled tasks over stop searching: */
2967 pulled_task = load_balance_newidle(this_cpu,
2970 interval = msecs_to_jiffies(sd->balance_interval);
2971 if (time_after(next_balance, sd->last_balance + interval))
2972 next_balance = sd->last_balance + interval;
2976 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2978 * We are going idle. next_balance may be set based on
2979 * a busy processor. So reset next_balance.
2981 this_rq->next_balance = next_balance;
2986 * active_load_balance is run by migration threads. It pushes running tasks
2987 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2988 * running on each physical CPU where possible, and avoids physical /
2989 * logical imbalances.
2991 * Called with busiest_rq locked.
2993 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2995 int target_cpu = busiest_rq->push_cpu;
2996 struct sched_domain *sd;
2997 struct rq *target_rq;
2999 /* Is there any task to move? */
3000 if (busiest_rq->nr_running <= 1)
3003 target_rq = cpu_rq(target_cpu);
3006 * This condition is "impossible", if it occurs
3007 * we need to fix it. Originally reported by
3008 * Bjorn Helgaas on a 128-cpu setup.
3010 BUG_ON(busiest_rq == target_rq);
3012 /* move a task from busiest_rq to target_rq */
3013 double_lock_balance(busiest_rq, target_rq);
3014 update_rq_clock(busiest_rq);
3015 update_rq_clock(target_rq);
3017 /* Search for an sd spanning us and the target CPU. */
3018 for_each_domain(target_cpu, sd) {
3019 if ((sd->flags & SD_LOAD_BALANCE) &&
3020 cpu_isset(busiest_cpu, sd->span))
3025 schedstat_inc(sd, alb_count);
3027 if (move_one_task(target_rq, target_cpu, busiest_rq,
3029 schedstat_inc(sd, alb_pushed);
3031 schedstat_inc(sd, alb_failed);
3033 spin_unlock(&target_rq->lock);
3038 atomic_t load_balancer;
3040 } nohz ____cacheline_aligned = {
3041 .load_balancer = ATOMIC_INIT(-1),
3042 .cpu_mask = CPU_MASK_NONE,
3046 * This routine will try to nominate the ilb (idle load balancing)
3047 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3048 * load balancing on behalf of all those cpus. If all the cpus in the system
3049 * go into this tickless mode, then there will be no ilb owner (as there is
3050 * no need for one) and all the cpus will sleep till the next wakeup event
3053 * For the ilb owner, tick is not stopped. And this tick will be used
3054 * for idle load balancing. ilb owner will still be part of
3057 * While stopping the tick, this cpu will become the ilb owner if there
3058 * is no other owner. And will be the owner till that cpu becomes busy
3059 * or if all cpus in the system stop their ticks at which point
3060 * there is no need for ilb owner.
3062 * When the ilb owner becomes busy, it nominates another owner, during the
3063 * next busy scheduler_tick()
3065 int select_nohz_load_balancer(int stop_tick)
3067 int cpu = smp_processor_id();
3070 cpu_set(cpu, nohz.cpu_mask);
3071 cpu_rq(cpu)->in_nohz_recently = 1;
3074 * If we are going offline and still the leader, give up!
3076 if (cpu_is_offline(cpu) &&
3077 atomic_read(&nohz.load_balancer) == cpu) {
3078 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3083 /* time for ilb owner also to sleep */
3084 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3085 if (atomic_read(&nohz.load_balancer) == cpu)
3086 atomic_set(&nohz.load_balancer, -1);
3090 if (atomic_read(&nohz.load_balancer) == -1) {
3091 /* make me the ilb owner */
3092 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3094 } else if (atomic_read(&nohz.load_balancer) == cpu)
3097 if (!cpu_isset(cpu, nohz.cpu_mask))
3100 cpu_clear(cpu, nohz.cpu_mask);
3102 if (atomic_read(&nohz.load_balancer) == cpu)
3103 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3110 static DEFINE_SPINLOCK(balancing);
3113 * It checks each scheduling domain to see if it is due to be balanced,
3114 * and initiates a balancing operation if so.
3116 * Balancing parameters are set up in arch_init_sched_domains.
3118 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3121 struct rq *rq = cpu_rq(cpu);
3122 unsigned long interval;
3123 struct sched_domain *sd;
3124 /* Earliest time when we have to do rebalance again */
3125 unsigned long next_balance = jiffies + 60*HZ;
3126 int update_next_balance = 0;
3128 for_each_domain(cpu, sd) {
3129 if (!(sd->flags & SD_LOAD_BALANCE))
3132 interval = sd->balance_interval;
3133 if (idle != CPU_IDLE)
3134 interval *= sd->busy_factor;
3136 /* scale ms to jiffies */
3137 interval = msecs_to_jiffies(interval);
3138 if (unlikely(!interval))
3140 if (interval > HZ*NR_CPUS/10)
3141 interval = HZ*NR_CPUS/10;
3144 if (sd->flags & SD_SERIALIZE) {
3145 if (!spin_trylock(&balancing))
3149 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3150 if (load_balance(cpu, rq, sd, idle, &balance)) {
3152 * We've pulled tasks over so either we're no
3153 * longer idle, or one of our SMT siblings is
3156 idle = CPU_NOT_IDLE;
3158 sd->last_balance = jiffies;
3160 if (sd->flags & SD_SERIALIZE)
3161 spin_unlock(&balancing);
3163 if (time_after(next_balance, sd->last_balance + interval)) {
3164 next_balance = sd->last_balance + interval;
3165 update_next_balance = 1;
3169 * Stop the load balance at this level. There is another
3170 * CPU in our sched group which is doing load balancing more
3178 * next_balance will be updated only when there is a need.
3179 * When the cpu is attached to null domain for ex, it will not be
3182 if (likely(update_next_balance))
3183 rq->next_balance = next_balance;
3187 * run_rebalance_domains is triggered when needed from the scheduler tick.
3188 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3189 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3191 static void run_rebalance_domains(struct softirq_action *h)
3193 int this_cpu = smp_processor_id();
3194 struct rq *this_rq = cpu_rq(this_cpu);
3195 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3196 CPU_IDLE : CPU_NOT_IDLE;
3198 rebalance_domains(this_cpu, idle);
3202 * If this cpu is the owner for idle load balancing, then do the
3203 * balancing on behalf of the other idle cpus whose ticks are
3206 if (this_rq->idle_at_tick &&
3207 atomic_read(&nohz.load_balancer) == this_cpu) {
3208 cpumask_t cpus = nohz.cpu_mask;
3212 cpu_clear(this_cpu, cpus);
3213 for_each_cpu_mask(balance_cpu, cpus) {
3215 * If this cpu gets work to do, stop the load balancing
3216 * work being done for other cpus. Next load
3217 * balancing owner will pick it up.
3222 rebalance_domains(balance_cpu, CPU_IDLE);
3224 rq = cpu_rq(balance_cpu);
3225 if (time_after(this_rq->next_balance, rq->next_balance))
3226 this_rq->next_balance = rq->next_balance;
3233 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3235 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3236 * idle load balancing owner or decide to stop the periodic load balancing,
3237 * if the whole system is idle.
3239 static inline void trigger_load_balance(struct rq *rq, int cpu)
3243 * If we were in the nohz mode recently and busy at the current
3244 * scheduler tick, then check if we need to nominate new idle
3247 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3248 rq->in_nohz_recently = 0;
3250 if (atomic_read(&nohz.load_balancer) == cpu) {
3251 cpu_clear(cpu, nohz.cpu_mask);
3252 atomic_set(&nohz.load_balancer, -1);
3255 if (atomic_read(&nohz.load_balancer) == -1) {
3257 * simple selection for now: Nominate the
3258 * first cpu in the nohz list to be the next
3261 * TBD: Traverse the sched domains and nominate
3262 * the nearest cpu in the nohz.cpu_mask.
3264 int ilb = first_cpu(nohz.cpu_mask);
3272 * If this cpu is idle and doing idle load balancing for all the
3273 * cpus with ticks stopped, is it time for that to stop?
3275 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3276 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3282 * If this cpu is idle and the idle load balancing is done by
3283 * someone else, then no need raise the SCHED_SOFTIRQ
3285 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3286 cpu_isset(cpu, nohz.cpu_mask))
3289 if (time_after_eq(jiffies, rq->next_balance))
3290 raise_softirq(SCHED_SOFTIRQ);
3293 #else /* CONFIG_SMP */
3296 * on UP we do not need to balance between CPUs:
3298 static inline void idle_balance(int cpu, struct rq *rq)
3304 DEFINE_PER_CPU(struct kernel_stat, kstat);
3306 EXPORT_PER_CPU_SYMBOL(kstat);
3309 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3310 * that have not yet been banked in case the task is currently running.
3312 unsigned long long task_sched_runtime(struct task_struct *p)
3314 unsigned long flags;
3318 rq = task_rq_lock(p, &flags);
3319 ns = p->se.sum_exec_runtime;
3320 if (rq->curr == p) {
3321 update_rq_clock(rq);
3322 delta_exec = rq->clock - p->se.exec_start;
3323 if ((s64)delta_exec > 0)
3326 task_rq_unlock(rq, &flags);
3332 * Account user cpu time to a process.
3333 * @p: the process that the cpu time gets accounted to
3334 * @cputime: the cpu time spent in user space since the last update
3336 void account_user_time(struct task_struct *p, cputime_t cputime)
3338 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3341 p->utime = cputime_add(p->utime, cputime);
3343 /* Add user time to cpustat. */
3344 tmp = cputime_to_cputime64(cputime);
3345 if (TASK_NICE(p) > 0)
3346 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3348 cpustat->user = cputime64_add(cpustat->user, tmp);
3352 * Account guest cpu time to a process.
3353 * @p: the process that the cpu time gets accounted to
3354 * @cputime: the cpu time spent in virtual machine since the last update
3356 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3359 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3361 tmp = cputime_to_cputime64(cputime);
3363 p->utime = cputime_add(p->utime, cputime);
3364 p->gtime = cputime_add(p->gtime, cputime);
3366 cpustat->user = cputime64_add(cpustat->user, tmp);
3367 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3371 * Account scaled user cpu time to a process.
3372 * @p: the process that the cpu time gets accounted to
3373 * @cputime: the cpu time spent in user space since the last update
3375 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3377 p->utimescaled = cputime_add(p->utimescaled, cputime);
3381 * Account system cpu time to a process.
3382 * @p: the process that the cpu time gets accounted to
3383 * @hardirq_offset: the offset to subtract from hardirq_count()
3384 * @cputime: the cpu time spent in kernel space since the last update
3386 void account_system_time(struct task_struct *p, int hardirq_offset,
3389 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3390 struct rq *rq = this_rq();
3393 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3394 return account_guest_time(p, cputime);
3396 p->stime = cputime_add(p->stime, cputime);
3398 /* Add system time to cpustat. */
3399 tmp = cputime_to_cputime64(cputime);
3400 if (hardirq_count() - hardirq_offset)
3401 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3402 else if (softirq_count())
3403 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3404 else if (p != rq->idle)
3405 cpustat->system = cputime64_add(cpustat->system, tmp);
3406 else if (atomic_read(&rq->nr_iowait) > 0)
3407 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3409 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3410 /* Account for system time used */
3411 acct_update_integrals(p);
3415 * Account scaled system cpu time to a process.
3416 * @p: the process that the cpu time gets accounted to
3417 * @hardirq_offset: the offset to subtract from hardirq_count()
3418 * @cputime: the cpu time spent in kernel space since the last update
3420 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3422 p->stimescaled = cputime_add(p->stimescaled, cputime);
3426 * Account for involuntary wait time.
3427 * @p: the process from which the cpu time has been stolen
3428 * @steal: the cpu time spent in involuntary wait
3430 void account_steal_time(struct task_struct *p, cputime_t steal)
3432 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3433 cputime64_t tmp = cputime_to_cputime64(steal);
3434 struct rq *rq = this_rq();
3436 if (p == rq->idle) {
3437 p->stime = cputime_add(p->stime, steal);
3438 if (atomic_read(&rq->nr_iowait) > 0)
3439 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3441 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3443 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3447 * This function gets called by the timer code, with HZ frequency.
3448 * We call it with interrupts disabled.
3450 * It also gets called by the fork code, when changing the parent's
3453 void scheduler_tick(void)
3455 int cpu = smp_processor_id();
3456 struct rq *rq = cpu_rq(cpu);
3457 struct task_struct *curr = rq->curr;
3458 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3460 spin_lock(&rq->lock);
3461 __update_rq_clock(rq);
3463 * Let rq->clock advance by at least TICK_NSEC:
3465 if (unlikely(rq->clock < next_tick))
3466 rq->clock = next_tick;
3467 rq->tick_timestamp = rq->clock;
3468 update_cpu_load(rq);
3469 if (curr != rq->idle) /* FIXME: needed? */
3470 curr->sched_class->task_tick(rq, curr);
3471 spin_unlock(&rq->lock);
3474 rq->idle_at_tick = idle_cpu(cpu);
3475 trigger_load_balance(rq, cpu);
3479 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3481 void fastcall add_preempt_count(int val)
3486 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3488 preempt_count() += val;
3490 * Spinlock count overflowing soon?
3492 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3495 EXPORT_SYMBOL(add_preempt_count);
3497 void fastcall sub_preempt_count(int val)
3502 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3505 * Is the spinlock portion underflowing?
3507 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3508 !(preempt_count() & PREEMPT_MASK)))
3511 preempt_count() -= val;
3513 EXPORT_SYMBOL(sub_preempt_count);
3518 * Print scheduling while atomic bug:
3520 static noinline void __schedule_bug(struct task_struct *prev)
3522 struct pt_regs *regs = get_irq_regs();
3524 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3525 prev->comm, prev->pid, preempt_count());
3527 debug_show_held_locks(prev);
3528 if (irqs_disabled())
3529 print_irqtrace_events(prev);
3538 * Various schedule()-time debugging checks and statistics:
3540 static inline void schedule_debug(struct task_struct *prev)
3543 * Test if we are atomic. Since do_exit() needs to call into
3544 * schedule() atomically, we ignore that path for now.
3545 * Otherwise, whine if we are scheduling when we should not be.
3547 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3548 __schedule_bug(prev);
3550 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3552 schedstat_inc(this_rq(), sched_count);
3553 #ifdef CONFIG_SCHEDSTATS
3554 if (unlikely(prev->lock_depth >= 0)) {
3555 schedstat_inc(this_rq(), bkl_count);
3556 schedstat_inc(prev, sched_info.bkl_count);
3562 * Pick up the highest-prio task:
3564 static inline struct task_struct *
3565 pick_next_task(struct rq *rq, struct task_struct *prev)
3567 const struct sched_class *class;
3568 struct task_struct *p;
3571 * Optimization: we know that if all tasks are in
3572 * the fair class we can call that function directly:
3574 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3575 p = fair_sched_class.pick_next_task(rq);
3580 class = sched_class_highest;
3582 p = class->pick_next_task(rq);
3586 * Will never be NULL as the idle class always
3587 * returns a non-NULL p:
3589 class = class->next;
3594 * schedule() is the main scheduler function.
3596 asmlinkage void __sched schedule(void)
3598 struct task_struct *prev, *next;
3605 cpu = smp_processor_id();
3609 switch_count = &prev->nivcsw;
3611 release_kernel_lock(prev);
3612 need_resched_nonpreemptible:
3614 schedule_debug(prev);
3617 * Do the rq-clock update outside the rq lock:
3619 local_irq_disable();
3620 __update_rq_clock(rq);
3621 spin_lock(&rq->lock);
3622 clear_tsk_need_resched(prev);
3624 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3625 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3626 unlikely(signal_pending(prev)))) {
3627 prev->state = TASK_RUNNING;
3629 deactivate_task(rq, prev, 1);
3631 switch_count = &prev->nvcsw;
3634 if (unlikely(!rq->nr_running))
3635 idle_balance(cpu, rq);
3637 prev->sched_class->put_prev_task(rq, prev);
3638 next = pick_next_task(rq, prev);
3640 sched_info_switch(prev, next);
3642 if (likely(prev != next)) {
3647 context_switch(rq, prev, next); /* unlocks the rq */
3649 spin_unlock_irq(&rq->lock);
3651 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3652 cpu = smp_processor_id();
3654 goto need_resched_nonpreemptible;
3656 preempt_enable_no_resched();
3657 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3660 EXPORT_SYMBOL(schedule);
3662 #ifdef CONFIG_PREEMPT
3664 * this is the entry point to schedule() from in-kernel preemption
3665 * off of preempt_enable. Kernel preemptions off return from interrupt
3666 * occur there and call schedule directly.
3668 asmlinkage void __sched preempt_schedule(void)
3670 struct thread_info *ti = current_thread_info();
3671 #ifdef CONFIG_PREEMPT_BKL
3672 struct task_struct *task = current;
3673 int saved_lock_depth;
3676 * If there is a non-zero preempt_count or interrupts are disabled,
3677 * we do not want to preempt the current task. Just return..
3679 if (likely(ti->preempt_count || irqs_disabled()))
3683 add_preempt_count(PREEMPT_ACTIVE);
3686 * We keep the big kernel semaphore locked, but we
3687 * clear ->lock_depth so that schedule() doesnt
3688 * auto-release the semaphore:
3690 #ifdef CONFIG_PREEMPT_BKL
3691 saved_lock_depth = task->lock_depth;
3692 task->lock_depth = -1;
3695 #ifdef CONFIG_PREEMPT_BKL
3696 task->lock_depth = saved_lock_depth;
3698 sub_preempt_count(PREEMPT_ACTIVE);
3701 * Check again in case we missed a preemption opportunity
3702 * between schedule and now.
3705 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3707 EXPORT_SYMBOL(preempt_schedule);
3710 * this is the entry point to schedule() from kernel preemption
3711 * off of irq context.
3712 * Note, that this is called and return with irqs disabled. This will
3713 * protect us against recursive calling from irq.
3715 asmlinkage void __sched preempt_schedule_irq(void)
3717 struct thread_info *ti = current_thread_info();
3718 #ifdef CONFIG_PREEMPT_BKL
3719 struct task_struct *task = current;
3720 int saved_lock_depth;
3722 /* Catch callers which need to be fixed */
3723 BUG_ON(ti->preempt_count || !irqs_disabled());
3726 add_preempt_count(PREEMPT_ACTIVE);
3729 * We keep the big kernel semaphore locked, but we
3730 * clear ->lock_depth so that schedule() doesnt
3731 * auto-release the semaphore:
3733 #ifdef CONFIG_PREEMPT_BKL
3734 saved_lock_depth = task->lock_depth;
3735 task->lock_depth = -1;
3739 local_irq_disable();
3740 #ifdef CONFIG_PREEMPT_BKL
3741 task->lock_depth = saved_lock_depth;
3743 sub_preempt_count(PREEMPT_ACTIVE);
3746 * Check again in case we missed a preemption opportunity
3747 * between schedule and now.
3750 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3753 #endif /* CONFIG_PREEMPT */
3755 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3758 return try_to_wake_up(curr->private, mode, sync);
3760 EXPORT_SYMBOL(default_wake_function);
3763 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3764 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3765 * number) then we wake all the non-exclusive tasks and one exclusive task.
3767 * There are circumstances in which we can try to wake a task which has already
3768 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3769 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3771 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3772 int nr_exclusive, int sync, void *key)
3774 wait_queue_t *curr, *next;
3776 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3777 unsigned flags = curr->flags;
3779 if (curr->func(curr, mode, sync, key) &&
3780 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3786 * __wake_up - wake up threads blocked on a waitqueue.
3788 * @mode: which threads
3789 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3790 * @key: is directly passed to the wakeup function
3792 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3793 int nr_exclusive, void *key)
3795 unsigned long flags;
3797 spin_lock_irqsave(&q->lock, flags);
3798 __wake_up_common(q, mode, nr_exclusive, 0, key);
3799 spin_unlock_irqrestore(&q->lock, flags);
3801 EXPORT_SYMBOL(__wake_up);
3804 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3806 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3808 __wake_up_common(q, mode, 1, 0, NULL);
3812 * __wake_up_sync - wake up threads blocked on a waitqueue.
3814 * @mode: which threads
3815 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3817 * The sync wakeup differs that the waker knows that it will schedule
3818 * away soon, so while the target thread will be woken up, it will not
3819 * be migrated to another CPU - ie. the two threads are 'synchronized'
3820 * with each other. This can prevent needless bouncing between CPUs.
3822 * On UP it can prevent extra preemption.
3825 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3827 unsigned long flags;
3833 if (unlikely(!nr_exclusive))
3836 spin_lock_irqsave(&q->lock, flags);
3837 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3838 spin_unlock_irqrestore(&q->lock, flags);
3840 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3842 void complete(struct completion *x)
3844 unsigned long flags;
3846 spin_lock_irqsave(&x->wait.lock, flags);
3848 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3850 spin_unlock_irqrestore(&x->wait.lock, flags);
3852 EXPORT_SYMBOL(complete);
3854 void complete_all(struct completion *x)
3856 unsigned long flags;
3858 spin_lock_irqsave(&x->wait.lock, flags);
3859 x->done += UINT_MAX/2;
3860 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3862 spin_unlock_irqrestore(&x->wait.lock, flags);
3864 EXPORT_SYMBOL(complete_all);
3866 static inline long __sched
3867 do_wait_for_common(struct completion *x, long timeout, int state)
3870 DECLARE_WAITQUEUE(wait, current);
3872 wait.flags |= WQ_FLAG_EXCLUSIVE;
3873 __add_wait_queue_tail(&x->wait, &wait);
3875 if (state == TASK_INTERRUPTIBLE &&
3876 signal_pending(current)) {
3877 __remove_wait_queue(&x->wait, &wait);
3878 return -ERESTARTSYS;
3880 __set_current_state(state);
3881 spin_unlock_irq(&x->wait.lock);
3882 timeout = schedule_timeout(timeout);
3883 spin_lock_irq(&x->wait.lock);
3885 __remove_wait_queue(&x->wait, &wait);
3889 __remove_wait_queue(&x->wait, &wait);
3896 wait_for_common(struct completion *x, long timeout, int state)
3900 spin_lock_irq(&x->wait.lock);
3901 timeout = do_wait_for_common(x, timeout, state);
3902 spin_unlock_irq(&x->wait.lock);
3906 void __sched wait_for_completion(struct completion *x)
3908 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3910 EXPORT_SYMBOL(wait_for_completion);
3912 unsigned long __sched
3913 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3915 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3917 EXPORT_SYMBOL(wait_for_completion_timeout);
3919 int __sched wait_for_completion_interruptible(struct completion *x)
3921 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3922 if (t == -ERESTARTSYS)
3926 EXPORT_SYMBOL(wait_for_completion_interruptible);
3928 unsigned long __sched
3929 wait_for_completion_interruptible_timeout(struct completion *x,
3930 unsigned long timeout)
3932 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3934 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3937 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3939 unsigned long flags;
3942 init_waitqueue_entry(&wait, current);
3944 __set_current_state(state);
3946 spin_lock_irqsave(&q->lock, flags);
3947 __add_wait_queue(q, &wait);
3948 spin_unlock(&q->lock);
3949 timeout = schedule_timeout(timeout);
3950 spin_lock_irq(&q->lock);
3951 __remove_wait_queue(q, &wait);
3952 spin_unlock_irqrestore(&q->lock, flags);
3957 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3959 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3961 EXPORT_SYMBOL(interruptible_sleep_on);
3964 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3966 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3968 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3970 void __sched sleep_on(wait_queue_head_t *q)
3972 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3974 EXPORT_SYMBOL(sleep_on);
3976 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3978 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3980 EXPORT_SYMBOL(sleep_on_timeout);
3982 #ifdef CONFIG_RT_MUTEXES
3985 * rt_mutex_setprio - set the current priority of a task
3987 * @prio: prio value (kernel-internal form)
3989 * This function changes the 'effective' priority of a task. It does
3990 * not touch ->normal_prio like __setscheduler().
3992 * Used by the rt_mutex code to implement priority inheritance logic.
3994 void rt_mutex_setprio(struct task_struct *p, int prio)
3996 unsigned long flags;
3997 int oldprio, on_rq, running;
4000 BUG_ON(prio < 0 || prio > MAX_PRIO);
4002 rq = task_rq_lock(p, &flags);
4003 update_rq_clock(rq);
4006 on_rq = p->se.on_rq;
4007 running = task_running(rq, p);
4009 dequeue_task(rq, p, 0);
4011 p->sched_class->put_prev_task(rq, p);
4015 p->sched_class = &rt_sched_class;
4017 p->sched_class = &fair_sched_class;
4023 p->sched_class->set_curr_task(rq);
4024 enqueue_task(rq, p, 0);
4026 * Reschedule if we are currently running on this runqueue and
4027 * our priority decreased, or if we are not currently running on
4028 * this runqueue and our priority is higher than the current's
4031 if (p->prio > oldprio)
4032 resched_task(rq->curr);
4034 check_preempt_curr(rq, p);
4037 task_rq_unlock(rq, &flags);
4042 void set_user_nice(struct task_struct *p, long nice)
4044 int old_prio, delta, on_rq;
4045 unsigned long flags;
4048 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4051 * We have to be careful, if called from sys_setpriority(),
4052 * the task might be in the middle of scheduling on another CPU.
4054 rq = task_rq_lock(p, &flags);
4055 update_rq_clock(rq);
4057 * The RT priorities are set via sched_setscheduler(), but we still
4058 * allow the 'normal' nice value to be set - but as expected
4059 * it wont have any effect on scheduling until the task is
4060 * SCHED_FIFO/SCHED_RR:
4062 if (task_has_rt_policy(p)) {
4063 p->static_prio = NICE_TO_PRIO(nice);
4066 on_rq = p->se.on_rq;
4068 dequeue_task(rq, p, 0);
4072 p->static_prio = NICE_TO_PRIO(nice);
4075 p->prio = effective_prio(p);
4076 delta = p->prio - old_prio;
4079 enqueue_task(rq, p, 0);
4082 * If the task increased its priority or is running and
4083 * lowered its priority, then reschedule its CPU:
4085 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4086 resched_task(rq->curr);
4089 task_rq_unlock(rq, &flags);
4091 EXPORT_SYMBOL(set_user_nice);
4094 * can_nice - check if a task can reduce its nice value
4098 int can_nice(const struct task_struct *p, const int nice)
4100 /* convert nice value [19,-20] to rlimit style value [1,40] */
4101 int nice_rlim = 20 - nice;
4103 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4104 capable(CAP_SYS_NICE));
4107 #ifdef __ARCH_WANT_SYS_NICE
4110 * sys_nice - change the priority of the current process.
4111 * @increment: priority increment
4113 * sys_setpriority is a more generic, but much slower function that
4114 * does similar things.
4116 asmlinkage long sys_nice(int increment)
4121 * Setpriority might change our priority at the same moment.
4122 * We don't have to worry. Conceptually one call occurs first
4123 * and we have a single winner.
4125 if (increment < -40)
4130 nice = PRIO_TO_NICE(current->static_prio) + increment;
4136 if (increment < 0 && !can_nice(current, nice))
4139 retval = security_task_setnice(current, nice);
4143 set_user_nice(current, nice);
4150 * task_prio - return the priority value of a given task.
4151 * @p: the task in question.
4153 * This is the priority value as seen by users in /proc.
4154 * RT tasks are offset by -200. Normal tasks are centered
4155 * around 0, value goes from -16 to +15.
4157 int task_prio(const struct task_struct *p)
4159 return p->prio - MAX_RT_PRIO;
4163 * task_nice - return the nice value of a given task.
4164 * @p: the task in question.
4166 int task_nice(const struct task_struct *p)
4168 return TASK_NICE(p);
4170 EXPORT_SYMBOL_GPL(task_nice);
4173 * idle_cpu - is a given cpu idle currently?
4174 * @cpu: the processor in question.
4176 int idle_cpu(int cpu)
4178 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4182 * idle_task - return the idle task for a given cpu.
4183 * @cpu: the processor in question.
4185 struct task_struct *idle_task(int cpu)
4187 return cpu_rq(cpu)->idle;
4191 * find_process_by_pid - find a process with a matching PID value.
4192 * @pid: the pid in question.
4194 static struct task_struct *find_process_by_pid(pid_t pid)
4196 return pid ? find_task_by_vpid(pid) : current;
4199 /* Actually do priority change: must hold rq lock. */
4201 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4203 BUG_ON(p->se.on_rq);
4206 switch (p->policy) {
4210 p->sched_class = &fair_sched_class;
4214 p->sched_class = &rt_sched_class;
4218 p->rt_priority = prio;
4219 p->normal_prio = normal_prio(p);
4220 /* we are holding p->pi_lock already */
4221 p->prio = rt_mutex_getprio(p);
4226 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4227 * @p: the task in question.
4228 * @policy: new policy.
4229 * @param: structure containing the new RT priority.
4231 * NOTE that the task may be already dead.
4233 int sched_setscheduler(struct task_struct *p, int policy,
4234 struct sched_param *param)
4236 int retval, oldprio, oldpolicy = -1, on_rq, running;
4237 unsigned long flags;
4240 /* may grab non-irq protected spin_locks */
4241 BUG_ON(in_interrupt());
4243 /* double check policy once rq lock held */
4245 policy = oldpolicy = p->policy;
4246 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4247 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4248 policy != SCHED_IDLE)
4251 * Valid priorities for SCHED_FIFO and SCHED_RR are
4252 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4253 * SCHED_BATCH and SCHED_IDLE is 0.
4255 if (param->sched_priority < 0 ||
4256 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4257 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4259 if (rt_policy(policy) != (param->sched_priority != 0))
4263 * Allow unprivileged RT tasks to decrease priority:
4265 if (!capable(CAP_SYS_NICE)) {
4266 if (rt_policy(policy)) {
4267 unsigned long rlim_rtprio;
4269 if (!lock_task_sighand(p, &flags))
4271 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4272 unlock_task_sighand(p, &flags);
4274 /* can't set/change the rt policy */
4275 if (policy != p->policy && !rlim_rtprio)
4278 /* can't increase priority */
4279 if (param->sched_priority > p->rt_priority &&
4280 param->sched_priority > rlim_rtprio)
4284 * Like positive nice levels, dont allow tasks to
4285 * move out of SCHED_IDLE either:
4287 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4290 /* can't change other user's priorities */
4291 if ((current->euid != p->euid) &&
4292 (current->euid != p->uid))
4296 retval = security_task_setscheduler(p, policy, param);
4300 * make sure no PI-waiters arrive (or leave) while we are
4301 * changing the priority of the task:
4303 spin_lock_irqsave(&p->pi_lock, flags);
4305 * To be able to change p->policy safely, the apropriate
4306 * runqueue lock must be held.
4308 rq = __task_rq_lock(p);
4309 /* recheck policy now with rq lock held */
4310 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4311 policy = oldpolicy = -1;
4312 __task_rq_unlock(rq);
4313 spin_unlock_irqrestore(&p->pi_lock, flags);
4316 update_rq_clock(rq);
4317 on_rq = p->se.on_rq;
4318 running = task_running(rq, p);
4320 deactivate_task(rq, p, 0);
4322 p->sched_class->put_prev_task(rq, p);
4326 __setscheduler(rq, p, policy, param->sched_priority);
4330 p->sched_class->set_curr_task(rq);
4331 activate_task(rq, p, 0);
4333 * Reschedule if we are currently running on this runqueue and
4334 * our priority decreased, or if we are not currently running on
4335 * this runqueue and our priority is higher than the current's
4338 if (p->prio > oldprio)
4339 resched_task(rq->curr);
4341 check_preempt_curr(rq, p);
4344 __task_rq_unlock(rq);
4345 spin_unlock_irqrestore(&p->pi_lock, flags);
4347 rt_mutex_adjust_pi(p);
4351 EXPORT_SYMBOL_GPL(sched_setscheduler);
4354 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4356 struct sched_param lparam;
4357 struct task_struct *p;
4360 if (!param || pid < 0)
4362 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4367 p = find_process_by_pid(pid);
4369 retval = sched_setscheduler(p, policy, &lparam);
4376 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4377 * @pid: the pid in question.
4378 * @policy: new policy.
4379 * @param: structure containing the new RT priority.
4381 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4382 struct sched_param __user *param)
4384 /* negative values for policy are not valid */
4388 return do_sched_setscheduler(pid, policy, param);
4392 * sys_sched_setparam - set/change the RT priority of a thread
4393 * @pid: the pid in question.
4394 * @param: structure containing the new RT priority.
4396 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4398 return do_sched_setscheduler(pid, -1, param);
4402 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4403 * @pid: the pid in question.
4405 asmlinkage long sys_sched_getscheduler(pid_t pid)
4407 struct task_struct *p;
4414 read_lock(&tasklist_lock);
4415 p = find_process_by_pid(pid);
4417 retval = security_task_getscheduler(p);
4421 read_unlock(&tasklist_lock);
4426 * sys_sched_getscheduler - get the RT priority of a thread
4427 * @pid: the pid in question.
4428 * @param: structure containing the RT priority.
4430 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4432 struct sched_param lp;
4433 struct task_struct *p;
4436 if (!param || pid < 0)
4439 read_lock(&tasklist_lock);
4440 p = find_process_by_pid(pid);
4445 retval = security_task_getscheduler(p);
4449 lp.sched_priority = p->rt_priority;
4450 read_unlock(&tasklist_lock);
4453 * This one might sleep, we cannot do it with a spinlock held ...
4455 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4460 read_unlock(&tasklist_lock);
4464 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4466 cpumask_t cpus_allowed;
4467 struct task_struct *p;
4470 mutex_lock(&sched_hotcpu_mutex);
4471 read_lock(&tasklist_lock);
4473 p = find_process_by_pid(pid);
4475 read_unlock(&tasklist_lock);
4476 mutex_unlock(&sched_hotcpu_mutex);
4481 * It is not safe to call set_cpus_allowed with the
4482 * tasklist_lock held. We will bump the task_struct's
4483 * usage count and then drop tasklist_lock.
4486 read_unlock(&tasklist_lock);
4489 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4490 !capable(CAP_SYS_NICE))
4493 retval = security_task_setscheduler(p, 0, NULL);
4497 cpus_allowed = cpuset_cpus_allowed(p);
4498 cpus_and(new_mask, new_mask, cpus_allowed);
4500 retval = set_cpus_allowed(p, new_mask);
4503 cpus_allowed = cpuset_cpus_allowed(p);
4504 if (!cpus_subset(new_mask, cpus_allowed)) {
4506 * We must have raced with a concurrent cpuset
4507 * update. Just reset the cpus_allowed to the
4508 * cpuset's cpus_allowed
4510 new_mask = cpus_allowed;
4516 mutex_unlock(&sched_hotcpu_mutex);
4520 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4521 cpumask_t *new_mask)
4523 if (len < sizeof(cpumask_t)) {
4524 memset(new_mask, 0, sizeof(cpumask_t));
4525 } else if (len > sizeof(cpumask_t)) {
4526 len = sizeof(cpumask_t);
4528 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4532 * sys_sched_setaffinity - set the cpu affinity of a process
4533 * @pid: pid of the process
4534 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4535 * @user_mask_ptr: user-space pointer to the new cpu mask
4537 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4538 unsigned long __user *user_mask_ptr)
4543 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4547 return sched_setaffinity(pid, new_mask);
4551 * Represents all cpu's present in the system
4552 * In systems capable of hotplug, this map could dynamically grow
4553 * as new cpu's are detected in the system via any platform specific
4554 * method, such as ACPI for e.g.
4557 cpumask_t cpu_present_map __read_mostly;
4558 EXPORT_SYMBOL(cpu_present_map);
4561 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4562 EXPORT_SYMBOL(cpu_online_map);
4564 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4565 EXPORT_SYMBOL(cpu_possible_map);
4568 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4570 struct task_struct *p;
4573 mutex_lock(&sched_hotcpu_mutex);
4574 read_lock(&tasklist_lock);
4577 p = find_process_by_pid(pid);
4581 retval = security_task_getscheduler(p);
4585 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4588 read_unlock(&tasklist_lock);
4589 mutex_unlock(&sched_hotcpu_mutex);
4595 * sys_sched_getaffinity - get the cpu affinity of a process
4596 * @pid: pid of the process
4597 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4598 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4600 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4601 unsigned long __user *user_mask_ptr)
4606 if (len < sizeof(cpumask_t))
4609 ret = sched_getaffinity(pid, &mask);
4613 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4616 return sizeof(cpumask_t);
4620 * sys_sched_yield - yield the current processor to other threads.
4622 * This function yields the current CPU to other tasks. If there are no
4623 * other threads running on this CPU then this function will return.
4625 asmlinkage long sys_sched_yield(void)
4627 struct rq *rq = this_rq_lock();
4629 schedstat_inc(rq, yld_count);
4630 current->sched_class->yield_task(rq);
4633 * Since we are going to call schedule() anyway, there's
4634 * no need to preempt or enable interrupts:
4636 __release(rq->lock);
4637 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4638 _raw_spin_unlock(&rq->lock);
4639 preempt_enable_no_resched();
4646 static void __cond_resched(void)
4648 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4649 __might_sleep(__FILE__, __LINE__);
4652 * The BKS might be reacquired before we have dropped
4653 * PREEMPT_ACTIVE, which could trigger a second
4654 * cond_resched() call.
4657 add_preempt_count(PREEMPT_ACTIVE);
4659 sub_preempt_count(PREEMPT_ACTIVE);
4660 } while (need_resched());
4663 int __sched cond_resched(void)
4665 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4666 system_state == SYSTEM_RUNNING) {
4672 EXPORT_SYMBOL(cond_resched);
4675 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4676 * call schedule, and on return reacquire the lock.
4678 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4679 * operations here to prevent schedule() from being called twice (once via
4680 * spin_unlock(), once by hand).
4682 int cond_resched_lock(spinlock_t *lock)
4686 if (need_lockbreak(lock)) {
4692 if (need_resched() && system_state == SYSTEM_RUNNING) {
4693 spin_release(&lock->dep_map, 1, _THIS_IP_);
4694 _raw_spin_unlock(lock);
4695 preempt_enable_no_resched();
4702 EXPORT_SYMBOL(cond_resched_lock);
4704 int __sched cond_resched_softirq(void)
4706 BUG_ON(!in_softirq());
4708 if (need_resched() && system_state == SYSTEM_RUNNING) {
4716 EXPORT_SYMBOL(cond_resched_softirq);
4719 * yield - yield the current processor to other threads.
4721 * This is a shortcut for kernel-space yielding - it marks the
4722 * thread runnable and calls sys_sched_yield().
4724 void __sched yield(void)
4726 set_current_state(TASK_RUNNING);
4729 EXPORT_SYMBOL(yield);
4732 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4733 * that process accounting knows that this is a task in IO wait state.
4735 * But don't do that if it is a deliberate, throttling IO wait (this task
4736 * has set its backing_dev_info: the queue against which it should throttle)
4738 void __sched io_schedule(void)
4740 struct rq *rq = &__raw_get_cpu_var(runqueues);
4742 delayacct_blkio_start();
4743 atomic_inc(&rq->nr_iowait);
4745 atomic_dec(&rq->nr_iowait);
4746 delayacct_blkio_end();
4748 EXPORT_SYMBOL(io_schedule);
4750 long __sched io_schedule_timeout(long timeout)
4752 struct rq *rq = &__raw_get_cpu_var(runqueues);
4755 delayacct_blkio_start();
4756 atomic_inc(&rq->nr_iowait);
4757 ret = schedule_timeout(timeout);
4758 atomic_dec(&rq->nr_iowait);
4759 delayacct_blkio_end();
4764 * sys_sched_get_priority_max - return maximum RT priority.
4765 * @policy: scheduling class.
4767 * this syscall returns the maximum rt_priority that can be used
4768 * by a given scheduling class.
4770 asmlinkage long sys_sched_get_priority_max(int policy)
4777 ret = MAX_USER_RT_PRIO-1;
4789 * sys_sched_get_priority_min - return minimum RT priority.
4790 * @policy: scheduling class.
4792 * this syscall returns the minimum rt_priority that can be used
4793 * by a given scheduling class.
4795 asmlinkage long sys_sched_get_priority_min(int policy)
4813 * sys_sched_rr_get_interval - return the default timeslice of a process.
4814 * @pid: pid of the process.
4815 * @interval: userspace pointer to the timeslice value.
4817 * this syscall writes the default timeslice value of a given process
4818 * into the user-space timespec buffer. A value of '0' means infinity.
4821 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4823 struct task_struct *p;
4824 unsigned int time_slice;
4832 read_lock(&tasklist_lock);
4833 p = find_process_by_pid(pid);
4837 retval = security_task_getscheduler(p);
4841 if (p->policy == SCHED_FIFO)
4843 else if (p->policy == SCHED_RR)
4844 time_slice = DEF_TIMESLICE;
4846 struct sched_entity *se = &p->se;
4847 unsigned long flags;
4850 rq = task_rq_lock(p, &flags);
4851 time_slice = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
4852 task_rq_unlock(rq, &flags);
4854 read_unlock(&tasklist_lock);
4855 jiffies_to_timespec(time_slice, &t);
4856 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4860 read_unlock(&tasklist_lock);
4864 static const char stat_nam[] = "RSDTtZX";
4866 static void show_task(struct task_struct *p)
4868 unsigned long free = 0;
4871 state = p->state ? __ffs(p->state) + 1 : 0;
4872 printk(KERN_INFO "%-13.13s %c", p->comm,
4873 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4874 #if BITS_PER_LONG == 32
4875 if (state == TASK_RUNNING)
4876 printk(KERN_CONT " running ");
4878 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4880 if (state == TASK_RUNNING)
4881 printk(KERN_CONT " running task ");
4883 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4885 #ifdef CONFIG_DEBUG_STACK_USAGE
4887 unsigned long *n = end_of_stack(p);
4890 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4893 printk(KERN_CONT "%5lu %5d %6d\n", free,
4894 task_pid_nr(p), task_pid_nr(p->parent));
4896 if (state != TASK_RUNNING)
4897 show_stack(p, NULL);
4900 void show_state_filter(unsigned long state_filter)
4902 struct task_struct *g, *p;
4904 #if BITS_PER_LONG == 32
4906 " task PC stack pid father\n");
4909 " task PC stack pid father\n");
4911 read_lock(&tasklist_lock);
4912 do_each_thread(g, p) {
4914 * reset the NMI-timeout, listing all files on a slow
4915 * console might take alot of time:
4917 touch_nmi_watchdog();
4918 if (!state_filter || (p->state & state_filter))
4920 } while_each_thread(g, p);
4922 touch_all_softlockup_watchdogs();
4924 #ifdef CONFIG_SCHED_DEBUG
4925 sysrq_sched_debug_show();
4927 read_unlock(&tasklist_lock);
4929 * Only show locks if all tasks are dumped:
4931 if (state_filter == -1)
4932 debug_show_all_locks();
4935 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4937 idle->sched_class = &idle_sched_class;
4941 * init_idle - set up an idle thread for a given CPU
4942 * @idle: task in question
4943 * @cpu: cpu the idle task belongs to
4945 * NOTE: this function does not set the idle thread's NEED_RESCHED
4946 * flag, to make booting more robust.
4948 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4950 struct rq *rq = cpu_rq(cpu);
4951 unsigned long flags;
4954 idle->se.exec_start = sched_clock();
4956 idle->prio = idle->normal_prio = MAX_PRIO;
4957 idle->cpus_allowed = cpumask_of_cpu(cpu);
4958 __set_task_cpu(idle, cpu);
4960 spin_lock_irqsave(&rq->lock, flags);
4961 rq->curr = rq->idle = idle;
4962 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4965 spin_unlock_irqrestore(&rq->lock, flags);
4967 /* Set the preempt count _outside_ the spinlocks! */
4968 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4969 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4971 task_thread_info(idle)->preempt_count = 0;
4974 * The idle tasks have their own, simple scheduling class:
4976 idle->sched_class = &idle_sched_class;
4980 * In a system that switches off the HZ timer nohz_cpu_mask
4981 * indicates which cpus entered this state. This is used
4982 * in the rcu update to wait only for active cpus. For system
4983 * which do not switch off the HZ timer nohz_cpu_mask should
4984 * always be CPU_MASK_NONE.
4986 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4989 * Increase the granularity value when there are more CPUs,
4990 * because with more CPUs the 'effective latency' as visible
4991 * to users decreases. But the relationship is not linear,
4992 * so pick a second-best guess by going with the log2 of the
4995 * This idea comes from the SD scheduler of Con Kolivas:
4997 static inline void sched_init_granularity(void)
4999 unsigned int factor = 1 + ilog2(num_online_cpus());
5000 const unsigned long limit = 200000000;
5002 sysctl_sched_min_granularity *= factor;
5003 if (sysctl_sched_min_granularity > limit)
5004 sysctl_sched_min_granularity = limit;
5006 sysctl_sched_latency *= factor;
5007 if (sysctl_sched_latency > limit)
5008 sysctl_sched_latency = limit;
5010 sysctl_sched_wakeup_granularity *= factor;
5011 sysctl_sched_batch_wakeup_granularity *= factor;
5016 * This is how migration works:
5018 * 1) we queue a struct migration_req structure in the source CPU's
5019 * runqueue and wake up that CPU's migration thread.
5020 * 2) we down() the locked semaphore => thread blocks.
5021 * 3) migration thread wakes up (implicitly it forces the migrated
5022 * thread off the CPU)
5023 * 4) it gets the migration request and checks whether the migrated
5024 * task is still in the wrong runqueue.
5025 * 5) if it's in the wrong runqueue then the migration thread removes
5026 * it and puts it into the right queue.
5027 * 6) migration thread up()s the semaphore.
5028 * 7) we wake up and the migration is done.
5032 * Change a given task's CPU affinity. Migrate the thread to a
5033 * proper CPU and schedule it away if the CPU it's executing on
5034 * is removed from the allowed bitmask.
5036 * NOTE: the caller must have a valid reference to the task, the
5037 * task must not exit() & deallocate itself prematurely. The
5038 * call is not atomic; no spinlocks may be held.
5040 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5042 struct migration_req req;
5043 unsigned long flags;
5047 rq = task_rq_lock(p, &flags);
5048 if (!cpus_intersects(new_mask, cpu_online_map)) {
5053 p->cpus_allowed = new_mask;
5054 /* Can the task run on the task's current CPU? If so, we're done */
5055 if (cpu_isset(task_cpu(p), new_mask))
5058 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5059 /* Need help from migration thread: drop lock and wait. */
5060 task_rq_unlock(rq, &flags);
5061 wake_up_process(rq->migration_thread);
5062 wait_for_completion(&req.done);
5063 tlb_migrate_finish(p->mm);
5067 task_rq_unlock(rq, &flags);
5071 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5074 * Move (not current) task off this cpu, onto dest cpu. We're doing
5075 * this because either it can't run here any more (set_cpus_allowed()
5076 * away from this CPU, or CPU going down), or because we're
5077 * attempting to rebalance this task on exec (sched_exec).
5079 * So we race with normal scheduler movements, but that's OK, as long
5080 * as the task is no longer on this CPU.
5082 * Returns non-zero if task was successfully migrated.
5084 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5086 struct rq *rq_dest, *rq_src;
5089 if (unlikely(cpu_is_offline(dest_cpu)))
5092 rq_src = cpu_rq(src_cpu);
5093 rq_dest = cpu_rq(dest_cpu);
5095 double_rq_lock(rq_src, rq_dest);
5096 /* Already moved. */
5097 if (task_cpu(p) != src_cpu)
5099 /* Affinity changed (again). */
5100 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5103 on_rq = p->se.on_rq;
5105 deactivate_task(rq_src, p, 0);
5107 set_task_cpu(p, dest_cpu);
5109 activate_task(rq_dest, p, 0);
5110 check_preempt_curr(rq_dest, p);
5114 double_rq_unlock(rq_src, rq_dest);
5119 * migration_thread - this is a highprio system thread that performs
5120 * thread migration by bumping thread off CPU then 'pushing' onto
5123 static int migration_thread(void *data)
5125 int cpu = (long)data;
5129 BUG_ON(rq->migration_thread != current);
5131 set_current_state(TASK_INTERRUPTIBLE);
5132 while (!kthread_should_stop()) {
5133 struct migration_req *req;
5134 struct list_head *head;
5136 spin_lock_irq(&rq->lock);
5138 if (cpu_is_offline(cpu)) {
5139 spin_unlock_irq(&rq->lock);
5143 if (rq->active_balance) {
5144 active_load_balance(rq, cpu);
5145 rq->active_balance = 0;
5148 head = &rq->migration_queue;
5150 if (list_empty(head)) {
5151 spin_unlock_irq(&rq->lock);
5153 set_current_state(TASK_INTERRUPTIBLE);
5156 req = list_entry(head->next, struct migration_req, list);
5157 list_del_init(head->next);
5159 spin_unlock(&rq->lock);
5160 __migrate_task(req->task, cpu, req->dest_cpu);
5163 complete(&req->done);
5165 __set_current_state(TASK_RUNNING);
5169 /* Wait for kthread_stop */
5170 set_current_state(TASK_INTERRUPTIBLE);
5171 while (!kthread_should_stop()) {
5173 set_current_state(TASK_INTERRUPTIBLE);
5175 __set_current_state(TASK_RUNNING);
5179 #ifdef CONFIG_HOTPLUG_CPU
5181 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5185 local_irq_disable();
5186 ret = __migrate_task(p, src_cpu, dest_cpu);
5192 * Figure out where task on dead CPU should go, use force if necessary.
5193 * NOTE: interrupts should be disabled by the caller
5195 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5197 unsigned long flags;
5204 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5205 cpus_and(mask, mask, p->cpus_allowed);
5206 dest_cpu = any_online_cpu(mask);
5208 /* On any allowed CPU? */
5209 if (dest_cpu == NR_CPUS)
5210 dest_cpu = any_online_cpu(p->cpus_allowed);
5212 /* No more Mr. Nice Guy. */
5213 if (dest_cpu == NR_CPUS) {
5214 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5216 * Try to stay on the same cpuset, where the
5217 * current cpuset may be a subset of all cpus.
5218 * The cpuset_cpus_allowed_locked() variant of
5219 * cpuset_cpus_allowed() will not block. It must be
5220 * called within calls to cpuset_lock/cpuset_unlock.
5222 rq = task_rq_lock(p, &flags);
5223 p->cpus_allowed = cpus_allowed;
5224 dest_cpu = any_online_cpu(p->cpus_allowed);
5225 task_rq_unlock(rq, &flags);
5228 * Don't tell them about moving exiting tasks or
5229 * kernel threads (both mm NULL), since they never
5232 if (p->mm && printk_ratelimit())
5233 printk(KERN_INFO "process %d (%s) no "
5234 "longer affine to cpu%d\n",
5235 task_pid_nr(p), p->comm, dead_cpu);
5237 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5241 * While a dead CPU has no uninterruptible tasks queued at this point,
5242 * it might still have a nonzero ->nr_uninterruptible counter, because
5243 * for performance reasons the counter is not stricly tracking tasks to
5244 * their home CPUs. So we just add the counter to another CPU's counter,
5245 * to keep the global sum constant after CPU-down:
5247 static void migrate_nr_uninterruptible(struct rq *rq_src)
5249 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5250 unsigned long flags;
5252 local_irq_save(flags);
5253 double_rq_lock(rq_src, rq_dest);
5254 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5255 rq_src->nr_uninterruptible = 0;
5256 double_rq_unlock(rq_src, rq_dest);
5257 local_irq_restore(flags);
5260 /* Run through task list and migrate tasks from the dead cpu. */
5261 static void migrate_live_tasks(int src_cpu)
5263 struct task_struct *p, *t;
5265 read_lock(&tasklist_lock);
5267 do_each_thread(t, p) {
5271 if (task_cpu(p) == src_cpu)
5272 move_task_off_dead_cpu(src_cpu, p);
5273 } while_each_thread(t, p);
5275 read_unlock(&tasklist_lock);
5279 * activate_idle_task - move idle task to the _front_ of runqueue.
5281 static void activate_idle_task(struct task_struct *p, struct rq *rq)
5283 update_rq_clock(rq);
5285 if (p->state == TASK_UNINTERRUPTIBLE)
5286 rq->nr_uninterruptible--;
5288 enqueue_task(rq, p, 0);
5289 inc_nr_running(p, rq);
5293 * Schedules idle task to be the next runnable task on current CPU.
5294 * It does so by boosting its priority to highest possible and adding it to
5295 * the _front_ of the runqueue. Used by CPU offline code.
5297 void sched_idle_next(void)
5299 int this_cpu = smp_processor_id();
5300 struct rq *rq = cpu_rq(this_cpu);
5301 struct task_struct *p = rq->idle;
5302 unsigned long flags;
5304 /* cpu has to be offline */
5305 BUG_ON(cpu_online(this_cpu));
5308 * Strictly not necessary since rest of the CPUs are stopped by now
5309 * and interrupts disabled on the current cpu.
5311 spin_lock_irqsave(&rq->lock, flags);
5313 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5315 /* Add idle task to the _front_ of its priority queue: */
5316 activate_idle_task(p, rq);
5318 spin_unlock_irqrestore(&rq->lock, flags);
5322 * Ensures that the idle task is using init_mm right before its cpu goes
5325 void idle_task_exit(void)
5327 struct mm_struct *mm = current->active_mm;
5329 BUG_ON(cpu_online(smp_processor_id()));
5332 switch_mm(mm, &init_mm, current);
5336 /* called under rq->lock with disabled interrupts */
5337 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5339 struct rq *rq = cpu_rq(dead_cpu);
5341 /* Must be exiting, otherwise would be on tasklist. */
5342 BUG_ON(!p->exit_state);
5344 /* Cannot have done final schedule yet: would have vanished. */
5345 BUG_ON(p->state == TASK_DEAD);
5350 * Drop lock around migration; if someone else moves it,
5351 * that's OK. No task can be added to this CPU, so iteration is
5354 spin_unlock_irq(&rq->lock);
5355 move_task_off_dead_cpu(dead_cpu, p);
5356 spin_lock_irq(&rq->lock);
5361 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5362 static void migrate_dead_tasks(unsigned int dead_cpu)
5364 struct rq *rq = cpu_rq(dead_cpu);
5365 struct task_struct *next;
5368 if (!rq->nr_running)
5370 update_rq_clock(rq);
5371 next = pick_next_task(rq, rq->curr);
5374 migrate_dead(dead_cpu, next);
5378 #endif /* CONFIG_HOTPLUG_CPU */
5380 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5382 static struct ctl_table sd_ctl_dir[] = {
5384 .procname = "sched_domain",
5390 static struct ctl_table sd_ctl_root[] = {
5392 .ctl_name = CTL_KERN,
5393 .procname = "kernel",
5395 .child = sd_ctl_dir,
5400 static struct ctl_table *sd_alloc_ctl_entry(int n)
5402 struct ctl_table *entry =
5403 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5408 static void sd_free_ctl_entry(struct ctl_table **tablep)
5410 struct ctl_table *entry;
5413 * In the intermediate directories, both the child directory and
5414 * procname are dynamically allocated and could fail but the mode
5415 * will always be set. In the lowest directory the names are
5416 * static strings and all have proc handlers.
5418 for (entry = *tablep; entry->mode; entry++) {
5420 sd_free_ctl_entry(&entry->child);
5421 if (entry->proc_handler == NULL)
5422 kfree(entry->procname);
5430 set_table_entry(struct ctl_table *entry,
5431 const char *procname, void *data, int maxlen,
5432 mode_t mode, proc_handler *proc_handler)
5434 entry->procname = procname;
5436 entry->maxlen = maxlen;
5438 entry->proc_handler = proc_handler;
5441 static struct ctl_table *
5442 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5444 struct ctl_table *table = sd_alloc_ctl_entry(12);
5449 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5450 sizeof(long), 0644, proc_doulongvec_minmax);
5451 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5452 sizeof(long), 0644, proc_doulongvec_minmax);
5453 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5454 sizeof(int), 0644, proc_dointvec_minmax);
5455 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5456 sizeof(int), 0644, proc_dointvec_minmax);
5457 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5458 sizeof(int), 0644, proc_dointvec_minmax);
5459 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5460 sizeof(int), 0644, proc_dointvec_minmax);
5461 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5462 sizeof(int), 0644, proc_dointvec_minmax);
5463 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5464 sizeof(int), 0644, proc_dointvec_minmax);
5465 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5466 sizeof(int), 0644, proc_dointvec_minmax);
5467 set_table_entry(&table[9], "cache_nice_tries",
5468 &sd->cache_nice_tries,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 set_table_entry(&table[10], "flags", &sd->flags,
5471 sizeof(int), 0644, proc_dointvec_minmax);
5472 /* &table[11] is terminator */
5477 static ctl_table * sd_alloc_ctl_cpu_table(int cpu)
5479 struct ctl_table *entry, *table;
5480 struct sched_domain *sd;
5481 int domain_num = 0, i;
5484 for_each_domain(cpu, sd)
5486 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5491 for_each_domain(cpu, sd) {
5492 snprintf(buf, 32, "domain%d", i);
5493 entry->procname = kstrdup(buf, GFP_KERNEL);
5495 entry->child = sd_alloc_ctl_domain_table(sd);
5502 static struct ctl_table_header *sd_sysctl_header;
5503 static void register_sched_domain_sysctl(void)
5505 int i, cpu_num = num_online_cpus();
5506 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5509 WARN_ON(sd_ctl_dir[0].child);
5510 sd_ctl_dir[0].child = entry;
5515 for_each_online_cpu(i) {
5516 snprintf(buf, 32, "cpu%d", i);
5517 entry->procname = kstrdup(buf, GFP_KERNEL);
5519 entry->child = sd_alloc_ctl_cpu_table(i);
5523 WARN_ON(sd_sysctl_header);
5524 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5527 /* may be called multiple times per register */
5528 static void unregister_sched_domain_sysctl(void)
5530 if (sd_sysctl_header)
5531 unregister_sysctl_table(sd_sysctl_header);
5532 sd_sysctl_header = NULL;
5533 if (sd_ctl_dir[0].child)
5534 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5537 static void register_sched_domain_sysctl(void)
5540 static void unregister_sched_domain_sysctl(void)
5546 * migration_call - callback that gets triggered when a CPU is added.
5547 * Here we can start up the necessary migration thread for the new CPU.
5549 static int __cpuinit
5550 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5552 struct task_struct *p;
5553 int cpu = (long)hcpu;
5554 unsigned long flags;
5558 case CPU_LOCK_ACQUIRE:
5559 mutex_lock(&sched_hotcpu_mutex);
5562 case CPU_UP_PREPARE:
5563 case CPU_UP_PREPARE_FROZEN:
5564 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5567 kthread_bind(p, cpu);
5568 /* Must be high prio: stop_machine expects to yield to it. */
5569 rq = task_rq_lock(p, &flags);
5570 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5571 task_rq_unlock(rq, &flags);
5572 cpu_rq(cpu)->migration_thread = p;
5576 case CPU_ONLINE_FROZEN:
5577 /* Strictly unnecessary, as first user will wake it. */
5578 wake_up_process(cpu_rq(cpu)->migration_thread);
5581 #ifdef CONFIG_HOTPLUG_CPU
5582 case CPU_UP_CANCELED:
5583 case CPU_UP_CANCELED_FROZEN:
5584 if (!cpu_rq(cpu)->migration_thread)
5586 /* Unbind it from offline cpu so it can run. Fall thru. */
5587 kthread_bind(cpu_rq(cpu)->migration_thread,
5588 any_online_cpu(cpu_online_map));
5589 kthread_stop(cpu_rq(cpu)->migration_thread);
5590 cpu_rq(cpu)->migration_thread = NULL;
5594 case CPU_DEAD_FROZEN:
5595 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5596 migrate_live_tasks(cpu);
5598 kthread_stop(rq->migration_thread);
5599 rq->migration_thread = NULL;
5600 /* Idle task back to normal (off runqueue, low prio) */
5601 spin_lock_irq(&rq->lock);
5602 update_rq_clock(rq);
5603 deactivate_task(rq, rq->idle, 0);
5604 rq->idle->static_prio = MAX_PRIO;
5605 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5606 rq->idle->sched_class = &idle_sched_class;
5607 migrate_dead_tasks(cpu);
5608 spin_unlock_irq(&rq->lock);
5610 migrate_nr_uninterruptible(rq);
5611 BUG_ON(rq->nr_running != 0);
5613 /* No need to migrate the tasks: it was best-effort if
5614 * they didn't take sched_hotcpu_mutex. Just wake up
5615 * the requestors. */
5616 spin_lock_irq(&rq->lock);
5617 while (!list_empty(&rq->migration_queue)) {
5618 struct migration_req *req;
5620 req = list_entry(rq->migration_queue.next,
5621 struct migration_req, list);
5622 list_del_init(&req->list);
5623 complete(&req->done);
5625 spin_unlock_irq(&rq->lock);
5628 case CPU_LOCK_RELEASE:
5629 mutex_unlock(&sched_hotcpu_mutex);
5635 /* Register at highest priority so that task migration (migrate_all_tasks)
5636 * happens before everything else.
5638 static struct notifier_block __cpuinitdata migration_notifier = {
5639 .notifier_call = migration_call,
5643 void __init migration_init(void)
5645 void *cpu = (void *)(long)smp_processor_id();
5648 /* Start one for the boot CPU: */
5649 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5650 BUG_ON(err == NOTIFY_BAD);
5651 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5652 register_cpu_notifier(&migration_notifier);
5658 /* Number of possible processor ids */
5659 int nr_cpu_ids __read_mostly = NR_CPUS;
5660 EXPORT_SYMBOL(nr_cpu_ids);
5662 #ifdef CONFIG_SCHED_DEBUG
5664 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5666 struct sched_group *group = sd->groups;
5667 cpumask_t groupmask;
5670 cpumask_scnprintf(str, NR_CPUS, sd->span);
5671 cpus_clear(groupmask);
5673 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5675 if (!(sd->flags & SD_LOAD_BALANCE)) {
5676 printk("does not load-balance\n");
5678 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5683 printk(KERN_CONT "span %s\n", str);
5685 if (!cpu_isset(cpu, sd->span)) {
5686 printk(KERN_ERR "ERROR: domain->span does not contain "
5689 if (!cpu_isset(cpu, group->cpumask)) {
5690 printk(KERN_ERR "ERROR: domain->groups does not contain"
5694 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5698 printk(KERN_ERR "ERROR: group is NULL\n");
5702 if (!group->__cpu_power) {
5703 printk(KERN_CONT "\n");
5704 printk(KERN_ERR "ERROR: domain->cpu_power not "
5709 if (!cpus_weight(group->cpumask)) {
5710 printk(KERN_CONT "\n");
5711 printk(KERN_ERR "ERROR: empty group\n");
5715 if (cpus_intersects(groupmask, group->cpumask)) {
5716 printk(KERN_CONT "\n");
5717 printk(KERN_ERR "ERROR: repeated CPUs\n");
5721 cpus_or(groupmask, groupmask, group->cpumask);
5723 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5724 printk(KERN_CONT " %s", str);
5726 group = group->next;
5727 } while (group != sd->groups);
5728 printk(KERN_CONT "\n");
5730 if (!cpus_equal(sd->span, groupmask))
5731 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5733 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5734 printk(KERN_ERR "ERROR: parent span is not a superset "
5735 "of domain->span\n");
5739 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5744 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5748 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5751 if (sched_domain_debug_one(sd, cpu, level))
5760 # define sched_domain_debug(sd, cpu) do { } while (0)
5763 static int sd_degenerate(struct sched_domain *sd)
5765 if (cpus_weight(sd->span) == 1)
5768 /* Following flags need at least 2 groups */
5769 if (sd->flags & (SD_LOAD_BALANCE |
5770 SD_BALANCE_NEWIDLE |
5774 SD_SHARE_PKG_RESOURCES)) {
5775 if (sd->groups != sd->groups->next)
5779 /* Following flags don't use groups */
5780 if (sd->flags & (SD_WAKE_IDLE |
5789 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5791 unsigned long cflags = sd->flags, pflags = parent->flags;
5793 if (sd_degenerate(parent))
5796 if (!cpus_equal(sd->span, parent->span))
5799 /* Does parent contain flags not in child? */
5800 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5801 if (cflags & SD_WAKE_AFFINE)
5802 pflags &= ~SD_WAKE_BALANCE;
5803 /* Flags needing groups don't count if only 1 group in parent */
5804 if (parent->groups == parent->groups->next) {
5805 pflags &= ~(SD_LOAD_BALANCE |
5806 SD_BALANCE_NEWIDLE |
5810 SD_SHARE_PKG_RESOURCES);
5812 if (~cflags & pflags)
5819 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5820 * hold the hotplug lock.
5822 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5824 struct rq *rq = cpu_rq(cpu);
5825 struct sched_domain *tmp;
5827 /* Remove the sched domains which do not contribute to scheduling. */
5828 for (tmp = sd; tmp; tmp = tmp->parent) {
5829 struct sched_domain *parent = tmp->parent;
5832 if (sd_parent_degenerate(tmp, parent)) {
5833 tmp->parent = parent->parent;
5835 parent->parent->child = tmp;
5839 if (sd && sd_degenerate(sd)) {
5845 sched_domain_debug(sd, cpu);
5847 rcu_assign_pointer(rq->sd, sd);
5850 /* cpus with isolated domains */
5851 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5853 /* Setup the mask of cpus configured for isolated domains */
5854 static int __init isolated_cpu_setup(char *str)
5856 int ints[NR_CPUS], i;
5858 str = get_options(str, ARRAY_SIZE(ints), ints);
5859 cpus_clear(cpu_isolated_map);
5860 for (i = 1; i <= ints[0]; i++)
5861 if (ints[i] < NR_CPUS)
5862 cpu_set(ints[i], cpu_isolated_map);
5866 __setup("isolcpus=", isolated_cpu_setup);
5869 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5870 * to a function which identifies what group(along with sched group) a CPU
5871 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5872 * (due to the fact that we keep track of groups covered with a cpumask_t).
5874 * init_sched_build_groups will build a circular linked list of the groups
5875 * covered by the given span, and will set each group's ->cpumask correctly,
5876 * and ->cpu_power to 0.
5879 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5880 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5881 struct sched_group **sg))
5883 struct sched_group *first = NULL, *last = NULL;
5884 cpumask_t covered = CPU_MASK_NONE;
5887 for_each_cpu_mask(i, span) {
5888 struct sched_group *sg;
5889 int group = group_fn(i, cpu_map, &sg);
5892 if (cpu_isset(i, covered))
5895 sg->cpumask = CPU_MASK_NONE;
5896 sg->__cpu_power = 0;
5898 for_each_cpu_mask(j, span) {
5899 if (group_fn(j, cpu_map, NULL) != group)
5902 cpu_set(j, covered);
5903 cpu_set(j, sg->cpumask);
5914 #define SD_NODES_PER_DOMAIN 16
5919 * find_next_best_node - find the next node to include in a sched_domain
5920 * @node: node whose sched_domain we're building
5921 * @used_nodes: nodes already in the sched_domain
5923 * Find the next node to include in a given scheduling domain. Simply
5924 * finds the closest node not already in the @used_nodes map.
5926 * Should use nodemask_t.
5928 static int find_next_best_node(int node, unsigned long *used_nodes)
5930 int i, n, val, min_val, best_node = 0;
5934 for (i = 0; i < MAX_NUMNODES; i++) {
5935 /* Start at @node */
5936 n = (node + i) % MAX_NUMNODES;
5938 if (!nr_cpus_node(n))
5941 /* Skip already used nodes */
5942 if (test_bit(n, used_nodes))
5945 /* Simple min distance search */
5946 val = node_distance(node, n);
5948 if (val < min_val) {
5954 set_bit(best_node, used_nodes);
5959 * sched_domain_node_span - get a cpumask for a node's sched_domain
5960 * @node: node whose cpumask we're constructing
5961 * @size: number of nodes to include in this span
5963 * Given a node, construct a good cpumask for its sched_domain to span. It
5964 * should be one that prevents unnecessary balancing, but also spreads tasks
5967 static cpumask_t sched_domain_node_span(int node)
5969 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5970 cpumask_t span, nodemask;
5974 bitmap_zero(used_nodes, MAX_NUMNODES);
5976 nodemask = node_to_cpumask(node);
5977 cpus_or(span, span, nodemask);
5978 set_bit(node, used_nodes);
5980 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5981 int next_node = find_next_best_node(node, used_nodes);
5983 nodemask = node_to_cpumask(next_node);
5984 cpus_or(span, span, nodemask);
5991 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5994 * SMT sched-domains:
5996 #ifdef CONFIG_SCHED_SMT
5997 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5998 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6000 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
6001 struct sched_group **sg)
6004 *sg = &per_cpu(sched_group_cpus, cpu);
6010 * multi-core sched-domains:
6012 #ifdef CONFIG_SCHED_MC
6013 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6014 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6017 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6018 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6019 struct sched_group **sg)
6022 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6023 cpus_and(mask, mask, *cpu_map);
6024 group = first_cpu(mask);
6026 *sg = &per_cpu(sched_group_core, group);
6029 #elif defined(CONFIG_SCHED_MC)
6030 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6031 struct sched_group **sg)
6034 *sg = &per_cpu(sched_group_core, cpu);
6039 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6040 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6042 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6043 struct sched_group **sg)
6046 #ifdef CONFIG_SCHED_MC
6047 cpumask_t mask = cpu_coregroup_map(cpu);
6048 cpus_and(mask, mask, *cpu_map);
6049 group = first_cpu(mask);
6050 #elif defined(CONFIG_SCHED_SMT)
6051 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6052 cpus_and(mask, mask, *cpu_map);
6053 group = first_cpu(mask);
6058 *sg = &per_cpu(sched_group_phys, group);
6064 * The init_sched_build_groups can't handle what we want to do with node
6065 * groups, so roll our own. Now each node has its own list of groups which
6066 * gets dynamically allocated.
6068 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6069 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6071 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6072 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6074 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6075 struct sched_group **sg)
6077 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6080 cpus_and(nodemask, nodemask, *cpu_map);
6081 group = first_cpu(nodemask);
6084 *sg = &per_cpu(sched_group_allnodes, group);
6088 static void init_numa_sched_groups_power(struct sched_group *group_head)
6090 struct sched_group *sg = group_head;
6096 for_each_cpu_mask(j, sg->cpumask) {
6097 struct sched_domain *sd;
6099 sd = &per_cpu(phys_domains, j);
6100 if (j != first_cpu(sd->groups->cpumask)) {
6102 * Only add "power" once for each
6108 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6111 } while (sg != group_head);
6116 /* Free memory allocated for various sched_group structures */
6117 static void free_sched_groups(const cpumask_t *cpu_map)
6121 for_each_cpu_mask(cpu, *cpu_map) {
6122 struct sched_group **sched_group_nodes
6123 = sched_group_nodes_bycpu[cpu];
6125 if (!sched_group_nodes)
6128 for (i = 0; i < MAX_NUMNODES; i++) {
6129 cpumask_t nodemask = node_to_cpumask(i);
6130 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6132 cpus_and(nodemask, nodemask, *cpu_map);
6133 if (cpus_empty(nodemask))
6143 if (oldsg != sched_group_nodes[i])
6146 kfree(sched_group_nodes);
6147 sched_group_nodes_bycpu[cpu] = NULL;
6151 static void free_sched_groups(const cpumask_t *cpu_map)
6157 * Initialize sched groups cpu_power.
6159 * cpu_power indicates the capacity of sched group, which is used while
6160 * distributing the load between different sched groups in a sched domain.
6161 * Typically cpu_power for all the groups in a sched domain will be same unless
6162 * there are asymmetries in the topology. If there are asymmetries, group
6163 * having more cpu_power will pickup more load compared to the group having
6166 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6167 * the maximum number of tasks a group can handle in the presence of other idle
6168 * or lightly loaded groups in the same sched domain.
6170 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6172 struct sched_domain *child;
6173 struct sched_group *group;
6175 WARN_ON(!sd || !sd->groups);
6177 if (cpu != first_cpu(sd->groups->cpumask))
6182 sd->groups->__cpu_power = 0;
6185 * For perf policy, if the groups in child domain share resources
6186 * (for example cores sharing some portions of the cache hierarchy
6187 * or SMT), then set this domain groups cpu_power such that each group
6188 * can handle only one task, when there are other idle groups in the
6189 * same sched domain.
6191 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6193 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6194 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6199 * add cpu_power of each child group to this groups cpu_power
6201 group = child->groups;
6203 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6204 group = group->next;
6205 } while (group != child->groups);
6209 * Build sched domains for a given set of cpus and attach the sched domains
6210 * to the individual cpus
6212 static int build_sched_domains(const cpumask_t *cpu_map)
6216 struct sched_group **sched_group_nodes = NULL;
6217 int sd_allnodes = 0;
6220 * Allocate the per-node list of sched groups
6222 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6224 if (!sched_group_nodes) {
6225 printk(KERN_WARNING "Can not alloc sched group node list\n");
6228 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6232 * Set up domains for cpus specified by the cpu_map.
6234 for_each_cpu_mask(i, *cpu_map) {
6235 struct sched_domain *sd = NULL, *p;
6236 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6238 cpus_and(nodemask, nodemask, *cpu_map);
6241 if (cpus_weight(*cpu_map) >
6242 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6243 sd = &per_cpu(allnodes_domains, i);
6244 *sd = SD_ALLNODES_INIT;
6245 sd->span = *cpu_map;
6246 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6252 sd = &per_cpu(node_domains, i);
6254 sd->span = sched_domain_node_span(cpu_to_node(i));
6258 cpus_and(sd->span, sd->span, *cpu_map);
6262 sd = &per_cpu(phys_domains, i);
6264 sd->span = nodemask;
6268 cpu_to_phys_group(i, cpu_map, &sd->groups);
6270 #ifdef CONFIG_SCHED_MC
6272 sd = &per_cpu(core_domains, i);
6274 sd->span = cpu_coregroup_map(i);
6275 cpus_and(sd->span, sd->span, *cpu_map);
6278 cpu_to_core_group(i, cpu_map, &sd->groups);
6281 #ifdef CONFIG_SCHED_SMT
6283 sd = &per_cpu(cpu_domains, i);
6284 *sd = SD_SIBLING_INIT;
6285 sd->span = per_cpu(cpu_sibling_map, i);
6286 cpus_and(sd->span, sd->span, *cpu_map);
6289 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6293 #ifdef CONFIG_SCHED_SMT
6294 /* Set up CPU (sibling) groups */
6295 for_each_cpu_mask(i, *cpu_map) {
6296 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6297 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6298 if (i != first_cpu(this_sibling_map))
6301 init_sched_build_groups(this_sibling_map, cpu_map,
6306 #ifdef CONFIG_SCHED_MC
6307 /* Set up multi-core groups */
6308 for_each_cpu_mask(i, *cpu_map) {
6309 cpumask_t this_core_map = cpu_coregroup_map(i);
6310 cpus_and(this_core_map, this_core_map, *cpu_map);
6311 if (i != first_cpu(this_core_map))
6313 init_sched_build_groups(this_core_map, cpu_map,
6314 &cpu_to_core_group);
6318 /* Set up physical groups */
6319 for (i = 0; i < MAX_NUMNODES; i++) {
6320 cpumask_t nodemask = node_to_cpumask(i);
6322 cpus_and(nodemask, nodemask, *cpu_map);
6323 if (cpus_empty(nodemask))
6326 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6330 /* Set up node groups */
6332 init_sched_build_groups(*cpu_map, cpu_map,
6333 &cpu_to_allnodes_group);
6335 for (i = 0; i < MAX_NUMNODES; i++) {
6336 /* Set up node groups */
6337 struct sched_group *sg, *prev;
6338 cpumask_t nodemask = node_to_cpumask(i);
6339 cpumask_t domainspan;
6340 cpumask_t covered = CPU_MASK_NONE;
6343 cpus_and(nodemask, nodemask, *cpu_map);
6344 if (cpus_empty(nodemask)) {
6345 sched_group_nodes[i] = NULL;
6349 domainspan = sched_domain_node_span(i);
6350 cpus_and(domainspan, domainspan, *cpu_map);
6352 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6354 printk(KERN_WARNING "Can not alloc domain group for "
6358 sched_group_nodes[i] = sg;
6359 for_each_cpu_mask(j, nodemask) {
6360 struct sched_domain *sd;
6362 sd = &per_cpu(node_domains, j);
6365 sg->__cpu_power = 0;
6366 sg->cpumask = nodemask;
6368 cpus_or(covered, covered, nodemask);
6371 for (j = 0; j < MAX_NUMNODES; j++) {
6372 cpumask_t tmp, notcovered;
6373 int n = (i + j) % MAX_NUMNODES;
6375 cpus_complement(notcovered, covered);
6376 cpus_and(tmp, notcovered, *cpu_map);
6377 cpus_and(tmp, tmp, domainspan);
6378 if (cpus_empty(tmp))
6381 nodemask = node_to_cpumask(n);
6382 cpus_and(tmp, tmp, nodemask);
6383 if (cpus_empty(tmp))
6386 sg = kmalloc_node(sizeof(struct sched_group),
6390 "Can not alloc domain group for node %d\n", j);
6393 sg->__cpu_power = 0;
6395 sg->next = prev->next;
6396 cpus_or(covered, covered, tmp);
6403 /* Calculate CPU power for physical packages and nodes */
6404 #ifdef CONFIG_SCHED_SMT
6405 for_each_cpu_mask(i, *cpu_map) {
6406 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6408 init_sched_groups_power(i, sd);
6411 #ifdef CONFIG_SCHED_MC
6412 for_each_cpu_mask(i, *cpu_map) {
6413 struct sched_domain *sd = &per_cpu(core_domains, i);
6415 init_sched_groups_power(i, sd);
6419 for_each_cpu_mask(i, *cpu_map) {
6420 struct sched_domain *sd = &per_cpu(phys_domains, i);
6422 init_sched_groups_power(i, sd);
6426 for (i = 0; i < MAX_NUMNODES; i++)
6427 init_numa_sched_groups_power(sched_group_nodes[i]);
6430 struct sched_group *sg;
6432 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6433 init_numa_sched_groups_power(sg);
6437 /* Attach the domains */
6438 for_each_cpu_mask(i, *cpu_map) {
6439 struct sched_domain *sd;
6440 #ifdef CONFIG_SCHED_SMT
6441 sd = &per_cpu(cpu_domains, i);
6442 #elif defined(CONFIG_SCHED_MC)
6443 sd = &per_cpu(core_domains, i);
6445 sd = &per_cpu(phys_domains, i);
6447 cpu_attach_domain(sd, i);
6454 free_sched_groups(cpu_map);
6459 static cpumask_t *doms_cur; /* current sched domains */
6460 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6463 * Special case: If a kmalloc of a doms_cur partition (array of
6464 * cpumask_t) fails, then fallback to a single sched domain,
6465 * as determined by the single cpumask_t fallback_doms.
6467 static cpumask_t fallback_doms;
6470 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6471 * For now this just excludes isolated cpus, but could be used to
6472 * exclude other special cases in the future.
6474 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6479 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6481 doms_cur = &fallback_doms;
6482 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6483 err = build_sched_domains(doms_cur);
6484 register_sched_domain_sysctl();
6489 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6491 free_sched_groups(cpu_map);
6495 * Detach sched domains from a group of cpus specified in cpu_map
6496 * These cpus will now be attached to the NULL domain
6498 static void detach_destroy_domains(const cpumask_t *cpu_map)
6502 unregister_sched_domain_sysctl();
6504 for_each_cpu_mask(i, *cpu_map)
6505 cpu_attach_domain(NULL, i);
6506 synchronize_sched();
6507 arch_destroy_sched_domains(cpu_map);
6511 * Partition sched domains as specified by the 'ndoms_new'
6512 * cpumasks in the array doms_new[] of cpumasks. This compares
6513 * doms_new[] to the current sched domain partitioning, doms_cur[].
6514 * It destroys each deleted domain and builds each new domain.
6516 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6517 * The masks don't intersect (don't overlap.) We should setup one
6518 * sched domain for each mask. CPUs not in any of the cpumasks will
6519 * not be load balanced. If the same cpumask appears both in the
6520 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6523 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6524 * ownership of it and will kfree it when done with it. If the caller
6525 * failed the kmalloc call, then it can pass in doms_new == NULL,
6526 * and partition_sched_domains() will fallback to the single partition
6529 * Call with hotplug lock held
6531 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6535 /* always unregister in case we don't destroy any domains */
6536 unregister_sched_domain_sysctl();
6538 if (doms_new == NULL) {
6540 doms_new = &fallback_doms;
6541 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6544 /* Destroy deleted domains */
6545 for (i = 0; i < ndoms_cur; i++) {
6546 for (j = 0; j < ndoms_new; j++) {
6547 if (cpus_equal(doms_cur[i], doms_new[j]))
6550 /* no match - a current sched domain not in new doms_new[] */
6551 detach_destroy_domains(doms_cur + i);
6556 /* Build new domains */
6557 for (i = 0; i < ndoms_new; i++) {
6558 for (j = 0; j < ndoms_cur; j++) {
6559 if (cpus_equal(doms_new[i], doms_cur[j]))
6562 /* no match - add a new doms_new */
6563 build_sched_domains(doms_new + i);
6568 /* Remember the new sched domains */
6569 if (doms_cur != &fallback_doms)
6571 doms_cur = doms_new;
6572 ndoms_cur = ndoms_new;
6574 register_sched_domain_sysctl();
6577 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6578 static int arch_reinit_sched_domains(void)
6582 mutex_lock(&sched_hotcpu_mutex);
6583 detach_destroy_domains(&cpu_online_map);
6584 err = arch_init_sched_domains(&cpu_online_map);
6585 mutex_unlock(&sched_hotcpu_mutex);
6590 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6594 if (buf[0] != '0' && buf[0] != '1')
6598 sched_smt_power_savings = (buf[0] == '1');
6600 sched_mc_power_savings = (buf[0] == '1');
6602 ret = arch_reinit_sched_domains();
6604 return ret ? ret : count;
6607 #ifdef CONFIG_SCHED_MC
6608 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6610 return sprintf(page, "%u\n", sched_mc_power_savings);
6612 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6613 const char *buf, size_t count)
6615 return sched_power_savings_store(buf, count, 0);
6617 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6618 sched_mc_power_savings_store);
6621 #ifdef CONFIG_SCHED_SMT
6622 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6624 return sprintf(page, "%u\n", sched_smt_power_savings);
6626 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6627 const char *buf, size_t count)
6629 return sched_power_savings_store(buf, count, 1);
6631 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6632 sched_smt_power_savings_store);
6635 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6639 #ifdef CONFIG_SCHED_SMT
6641 err = sysfs_create_file(&cls->kset.kobj,
6642 &attr_sched_smt_power_savings.attr);
6644 #ifdef CONFIG_SCHED_MC
6645 if (!err && mc_capable())
6646 err = sysfs_create_file(&cls->kset.kobj,
6647 &attr_sched_mc_power_savings.attr);
6654 * Force a reinitialization of the sched domains hierarchy. The domains
6655 * and groups cannot be updated in place without racing with the balancing
6656 * code, so we temporarily attach all running cpus to the NULL domain
6657 * which will prevent rebalancing while the sched domains are recalculated.
6659 static int update_sched_domains(struct notifier_block *nfb,
6660 unsigned long action, void *hcpu)
6663 case CPU_UP_PREPARE:
6664 case CPU_UP_PREPARE_FROZEN:
6665 case CPU_DOWN_PREPARE:
6666 case CPU_DOWN_PREPARE_FROZEN:
6667 detach_destroy_domains(&cpu_online_map);
6670 case CPU_UP_CANCELED:
6671 case CPU_UP_CANCELED_FROZEN:
6672 case CPU_DOWN_FAILED:
6673 case CPU_DOWN_FAILED_FROZEN:
6675 case CPU_ONLINE_FROZEN:
6677 case CPU_DEAD_FROZEN:
6679 * Fall through and re-initialise the domains.
6686 /* The hotplug lock is already held by cpu_up/cpu_down */
6687 arch_init_sched_domains(&cpu_online_map);
6692 void __init sched_init_smp(void)
6694 cpumask_t non_isolated_cpus;
6696 mutex_lock(&sched_hotcpu_mutex);
6697 arch_init_sched_domains(&cpu_online_map);
6698 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6699 if (cpus_empty(non_isolated_cpus))
6700 cpu_set(smp_processor_id(), non_isolated_cpus);
6701 mutex_unlock(&sched_hotcpu_mutex);
6702 /* XXX: Theoretical race here - CPU may be hotplugged now */
6703 hotcpu_notifier(update_sched_domains, 0);
6705 /* Move init over to a non-isolated CPU */
6706 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6708 sched_init_granularity();
6711 void __init sched_init_smp(void)
6713 sched_init_granularity();
6715 #endif /* CONFIG_SMP */
6717 int in_sched_functions(unsigned long addr)
6719 /* Linker adds these: start and end of __sched functions */
6720 extern char __sched_text_start[], __sched_text_end[];
6722 return in_lock_functions(addr) ||
6723 (addr >= (unsigned long)__sched_text_start
6724 && addr < (unsigned long)__sched_text_end);
6727 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6729 cfs_rq->tasks_timeline = RB_ROOT;
6730 #ifdef CONFIG_FAIR_GROUP_SCHED
6733 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6736 void __init sched_init(void)
6738 int highest_cpu = 0;
6741 for_each_possible_cpu(i) {
6742 struct rt_prio_array *array;
6746 spin_lock_init(&rq->lock);
6747 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6750 init_cfs_rq(&rq->cfs, rq);
6751 #ifdef CONFIG_FAIR_GROUP_SCHED
6752 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6754 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6755 struct sched_entity *se =
6756 &per_cpu(init_sched_entity, i);
6758 init_cfs_rq_p[i] = cfs_rq;
6759 init_cfs_rq(cfs_rq, rq);
6760 cfs_rq->tg = &init_task_group;
6761 list_add(&cfs_rq->leaf_cfs_rq_list,
6762 &rq->leaf_cfs_rq_list);
6764 init_sched_entity_p[i] = se;
6765 se->cfs_rq = &rq->cfs;
6767 se->load.weight = init_task_group_load;
6768 se->load.inv_weight =
6769 div64_64(1ULL<<32, init_task_group_load);
6772 init_task_group.shares = init_task_group_load;
6773 spin_lock_init(&init_task_group.lock);
6776 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6777 rq->cpu_load[j] = 0;
6780 rq->active_balance = 0;
6781 rq->next_balance = jiffies;
6784 rq->migration_thread = NULL;
6785 INIT_LIST_HEAD(&rq->migration_queue);
6787 atomic_set(&rq->nr_iowait, 0);
6789 array = &rq->rt.active;
6790 for (j = 0; j < MAX_RT_PRIO; j++) {
6791 INIT_LIST_HEAD(array->queue + j);
6792 __clear_bit(j, array->bitmap);
6795 /* delimiter for bitsearch: */
6796 __set_bit(MAX_RT_PRIO, array->bitmap);
6799 set_load_weight(&init_task);
6801 #ifdef CONFIG_PREEMPT_NOTIFIERS
6802 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6806 nr_cpu_ids = highest_cpu + 1;
6807 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6810 #ifdef CONFIG_RT_MUTEXES
6811 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6815 * The boot idle thread does lazy MMU switching as well:
6817 atomic_inc(&init_mm.mm_count);
6818 enter_lazy_tlb(&init_mm, current);
6821 * Make us the idle thread. Technically, schedule() should not be
6822 * called from this thread, however somewhere below it might be,
6823 * but because we are the idle thread, we just pick up running again
6824 * when this runqueue becomes "idle".
6826 init_idle(current, smp_processor_id());
6828 * During early bootup we pretend to be a normal task:
6830 current->sched_class = &fair_sched_class;
6833 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6834 void __might_sleep(char *file, int line)
6837 static unsigned long prev_jiffy; /* ratelimiting */
6839 if ((in_atomic() || irqs_disabled()) &&
6840 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6841 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6843 prev_jiffy = jiffies;
6844 printk(KERN_ERR "BUG: sleeping function called from invalid"
6845 " context at %s:%d\n", file, line);
6846 printk("in_atomic():%d, irqs_disabled():%d\n",
6847 in_atomic(), irqs_disabled());
6848 debug_show_held_locks(current);
6849 if (irqs_disabled())
6850 print_irqtrace_events(current);
6855 EXPORT_SYMBOL(__might_sleep);
6858 #ifdef CONFIG_MAGIC_SYSRQ
6859 static void normalize_task(struct rq *rq, struct task_struct *p)
6862 update_rq_clock(rq);
6863 on_rq = p->se.on_rq;
6865 deactivate_task(rq, p, 0);
6866 __setscheduler(rq, p, SCHED_NORMAL, 0);
6868 activate_task(rq, p, 0);
6869 resched_task(rq->curr);
6873 void normalize_rt_tasks(void)
6875 struct task_struct *g, *p;
6876 unsigned long flags;
6879 read_lock_irq(&tasklist_lock);
6880 do_each_thread(g, p) {
6882 * Only normalize user tasks:
6887 p->se.exec_start = 0;
6888 #ifdef CONFIG_SCHEDSTATS
6889 p->se.wait_start = 0;
6890 p->se.sleep_start = 0;
6891 p->se.block_start = 0;
6893 task_rq(p)->clock = 0;
6897 * Renice negative nice level userspace
6900 if (TASK_NICE(p) < 0 && p->mm)
6901 set_user_nice(p, 0);
6905 spin_lock_irqsave(&p->pi_lock, flags);
6906 rq = __task_rq_lock(p);
6908 normalize_task(rq, p);
6910 __task_rq_unlock(rq);
6911 spin_unlock_irqrestore(&p->pi_lock, flags);
6912 } while_each_thread(g, p);
6914 read_unlock_irq(&tasklist_lock);
6917 #endif /* CONFIG_MAGIC_SYSRQ */
6921 * These functions are only useful for the IA64 MCA handling.
6923 * They can only be called when the whole system has been
6924 * stopped - every CPU needs to be quiescent, and no scheduling
6925 * activity can take place. Using them for anything else would
6926 * be a serious bug, and as a result, they aren't even visible
6927 * under any other configuration.
6931 * curr_task - return the current task for a given cpu.
6932 * @cpu: the processor in question.
6934 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6936 struct task_struct *curr_task(int cpu)
6938 return cpu_curr(cpu);
6942 * set_curr_task - set the current task for a given cpu.
6943 * @cpu: the processor in question.
6944 * @p: the task pointer to set.
6946 * Description: This function must only be used when non-maskable interrupts
6947 * are serviced on a separate stack. It allows the architecture to switch the
6948 * notion of the current task on a cpu in a non-blocking manner. This function
6949 * must be called with all CPU's synchronized, and interrupts disabled, the
6950 * and caller must save the original value of the current task (see
6951 * curr_task() above) and restore that value before reenabling interrupts and
6952 * re-starting the system.
6954 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6956 void set_curr_task(int cpu, struct task_struct *p)
6963 #ifdef CONFIG_FAIR_GROUP_SCHED
6965 /* allocate runqueue etc for a new task group */
6966 struct task_group *sched_create_group(void)
6968 struct task_group *tg;
6969 struct cfs_rq *cfs_rq;
6970 struct sched_entity *se;
6974 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6976 return ERR_PTR(-ENOMEM);
6978 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6981 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6985 for_each_possible_cpu(i) {
6988 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
6993 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
6998 memset(cfs_rq, 0, sizeof(struct cfs_rq));
6999 memset(se, 0, sizeof(struct sched_entity));
7001 tg->cfs_rq[i] = cfs_rq;
7002 init_cfs_rq(cfs_rq, rq);
7006 se->cfs_rq = &rq->cfs;
7008 se->load.weight = NICE_0_LOAD;
7009 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7013 for_each_possible_cpu(i) {
7015 cfs_rq = tg->cfs_rq[i];
7016 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7019 tg->shares = NICE_0_LOAD;
7020 spin_lock_init(&tg->lock);
7025 for_each_possible_cpu(i) {
7027 kfree(tg->cfs_rq[i]);
7035 return ERR_PTR(-ENOMEM);
7038 /* rcu callback to free various structures associated with a task group */
7039 static void free_sched_group(struct rcu_head *rhp)
7041 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7042 struct cfs_rq *cfs_rq;
7043 struct sched_entity *se;
7046 /* now it should be safe to free those cfs_rqs */
7047 for_each_possible_cpu(i) {
7048 cfs_rq = tg->cfs_rq[i];
7060 /* Destroy runqueue etc associated with a task group */
7061 void sched_destroy_group(struct task_group *tg)
7063 struct cfs_rq *cfs_rq = NULL;
7066 for_each_possible_cpu(i) {
7067 cfs_rq = tg->cfs_rq[i];
7068 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7073 /* wait for possible concurrent references to cfs_rqs complete */
7074 call_rcu(&tg->rcu, free_sched_group);
7077 /* change task's runqueue when it moves between groups.
7078 * The caller of this function should have put the task in its new group
7079 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7080 * reflect its new group.
7082 void sched_move_task(struct task_struct *tsk)
7085 unsigned long flags;
7088 rq = task_rq_lock(tsk, &flags);
7090 if (tsk->sched_class != &fair_sched_class)
7093 update_rq_clock(rq);
7095 running = task_running(rq, tsk);
7096 on_rq = tsk->se.on_rq;
7099 dequeue_task(rq, tsk, 0);
7100 if (unlikely(running))
7101 tsk->sched_class->put_prev_task(rq, tsk);
7104 set_task_cfs_rq(tsk);
7107 if (unlikely(running))
7108 tsk->sched_class->set_curr_task(rq);
7109 enqueue_task(rq, tsk, 0);
7113 task_rq_unlock(rq, &flags);
7116 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7118 struct cfs_rq *cfs_rq = se->cfs_rq;
7119 struct rq *rq = cfs_rq->rq;
7122 spin_lock_irq(&rq->lock);
7126 dequeue_entity(cfs_rq, se, 0);
7128 se->load.weight = shares;
7129 se->load.inv_weight = div64_64((1ULL<<32), shares);
7132 enqueue_entity(cfs_rq, se, 0);
7134 spin_unlock_irq(&rq->lock);
7137 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7141 spin_lock(&tg->lock);
7142 if (tg->shares == shares)
7145 tg->shares = shares;
7146 for_each_possible_cpu(i)
7147 set_se_shares(tg->se[i], shares);
7150 spin_unlock(&tg->lock);
7154 unsigned long sched_group_shares(struct task_group *tg)
7159 #endif /* CONFIG_FAIR_GROUP_SCHED */
7161 #ifdef CONFIG_FAIR_CGROUP_SCHED
7163 /* return corresponding task_group object of a cgroup */
7164 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7166 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7167 struct task_group, css);
7170 static struct cgroup_subsys_state *
7171 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7173 struct task_group *tg;
7175 if (!cgrp->parent) {
7176 /* This is early initialization for the top cgroup */
7177 init_task_group.css.cgroup = cgrp;
7178 return &init_task_group.css;
7181 /* we support only 1-level deep hierarchical scheduler atm */
7182 if (cgrp->parent->parent)
7183 return ERR_PTR(-EINVAL);
7185 tg = sched_create_group();
7187 return ERR_PTR(-ENOMEM);
7189 /* Bind the cgroup to task_group object we just created */
7190 tg->css.cgroup = cgrp;
7195 static void cpu_cgroup_destroy(struct cgroup_subsys *ss,
7196 struct cgroup *cgrp)
7198 struct task_group *tg = cgroup_tg(cgrp);
7200 sched_destroy_group(tg);
7203 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss,
7204 struct cgroup *cgrp, struct task_struct *tsk)
7206 /* We don't support RT-tasks being in separate groups */
7207 if (tsk->sched_class != &fair_sched_class)
7214 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7215 struct cgroup *old_cont, struct task_struct *tsk)
7217 sched_move_task(tsk);
7220 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7223 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7226 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7228 struct task_group *tg = cgroup_tg(cgrp);
7230 return (u64) tg->shares;
7233 static u64 cpu_usage_read(struct cgroup *cgrp, struct cftype *cft)
7235 struct task_group *tg = cgroup_tg(cgrp);
7236 unsigned long flags;
7240 for_each_possible_cpu(i) {
7242 * Lock to prevent races with updating 64-bit counters
7245 spin_lock_irqsave(&cpu_rq(i)->lock, flags);
7246 res += tg->se[i]->sum_exec_runtime;
7247 spin_unlock_irqrestore(&cpu_rq(i)->lock, flags);
7249 /* Convert from ns to ms */
7250 do_div(res, NSEC_PER_MSEC);
7255 static struct cftype cpu_files[] = {
7258 .read_uint = cpu_shares_read_uint,
7259 .write_uint = cpu_shares_write_uint,
7263 .read_uint = cpu_usage_read,
7267 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7269 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7272 struct cgroup_subsys cpu_cgroup_subsys = {
7274 .create = cpu_cgroup_create,
7275 .destroy = cpu_cgroup_destroy,
7276 .can_attach = cpu_cgroup_can_attach,
7277 .attach = cpu_cgroup_attach,
7278 .populate = cpu_cgroup_populate,
7279 .subsys_id = cpu_cgroup_subsys_id,
7283 #endif /* CONFIG_FAIR_CGROUP_SCHED */