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/cpu_acct.h>
56 #include <linux/kthread.h>
57 #include <linux/seq_file.h>
58 #include <linux/sysctl.h>
59 #include <linux/syscalls.h>
60 #include <linux/times.h>
61 #include <linux/tsacct_kern.h>
62 #include <linux/kprobes.h>
63 #include <linux/delayacct.h>
64 #include <linux/reciprocal_div.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
69 #include <asm/irq_regs.h>
72 * Scheduler clock - returns current time in nanosec units.
73 * This is default implementation.
74 * Architectures and sub-architectures can override this.
76 unsigned long long __attribute__((weak)) sched_clock(void)
78 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
82 * Convert user-nice values [ -20 ... 0 ... 19 ]
83 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
86 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
87 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
88 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
91 * 'User priority' is the nice value converted to something we
92 * can work with better when scaling various scheduler parameters,
93 * it's a [ 0 ... 39 ] range.
95 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
96 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
97 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
100 * Some helpers for converting nanosecond timing to jiffy resolution
102 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
103 #define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
119 * Since cpu_power is a 'constant', we can use a reciprocal divide.
121 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
123 return reciprocal_divide(load, sg->reciprocal_cpu_power);
127 * Each time a sched group cpu_power is changed,
128 * we must compute its reciprocal value
130 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
132 sg->__cpu_power += val;
133 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
137 static inline int rt_policy(int policy)
139 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
144 static inline int task_has_rt_policy(struct task_struct *p)
146 return rt_policy(p->policy);
150 * This is the priority-queue data structure of the RT scheduling class:
152 struct rt_prio_array {
153 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
154 struct list_head queue[MAX_RT_PRIO];
157 #ifdef CONFIG_FAIR_GROUP_SCHED
159 #include <linux/cgroup.h>
163 /* task group related information */
165 #ifdef CONFIG_FAIR_CGROUP_SCHED
166 struct cgroup_subsys_state css;
168 /* schedulable entities of this group on each cpu */
169 struct sched_entity **se;
170 /* runqueue "owned" by this group on each cpu */
171 struct cfs_rq **cfs_rq;
172 unsigned long shares;
173 /* spinlock to serialize modification to shares */
178 /* Default task group's sched entity on each cpu */
179 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
180 /* Default task group's cfs_rq on each cpu */
181 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
183 static struct sched_entity *init_sched_entity_p[NR_CPUS];
184 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
186 /* Default task group.
187 * Every task in system belong to this group at bootup.
189 struct task_group init_task_group = {
190 .se = init_sched_entity_p,
191 .cfs_rq = init_cfs_rq_p,
194 #ifdef CONFIG_FAIR_USER_SCHED
195 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
197 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
200 static int init_task_group_load = INIT_TASK_GRP_LOAD;
202 /* return group to which a task belongs */
203 static inline struct task_group *task_group(struct task_struct *p)
205 struct task_group *tg;
207 #ifdef CONFIG_FAIR_USER_SCHED
209 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
210 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
211 struct task_group, css);
213 tg = &init_task_group;
219 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
220 static inline void set_task_cfs_rq(struct task_struct *p)
222 p->se.cfs_rq = task_group(p)->cfs_rq[task_cpu(p)];
223 p->se.parent = task_group(p)->se[task_cpu(p)];
228 static inline void set_task_cfs_rq(struct task_struct *p) { }
230 #endif /* CONFIG_FAIR_GROUP_SCHED */
232 /* CFS-related fields in a runqueue */
234 struct load_weight load;
235 unsigned long nr_running;
240 struct rb_root tasks_timeline;
241 struct rb_node *rb_leftmost;
242 struct rb_node *rb_load_balance_curr;
243 /* 'curr' points to currently running entity on this cfs_rq.
244 * It is set to NULL otherwise (i.e when none are currently running).
246 struct sched_entity *curr;
248 unsigned long nr_spread_over;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
253 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
254 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
255 * (like users, containers etc.)
257 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
258 * list is used during load balance.
260 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
261 struct task_group *tg; /* group that "owns" this runqueue */
265 /* Real-Time classes' related field in a runqueue: */
267 struct rt_prio_array active;
268 int rt_load_balance_idx;
269 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
273 * This is the main, per-CPU runqueue data structure.
275 * Locking rule: those places that want to lock multiple runqueues
276 * (such as the load balancing or the thread migration code), lock
277 * acquire operations must be ordered by ascending &runqueue.
284 * nr_running and cpu_load should be in the same cacheline because
285 * remote CPUs use both these fields when doing load calculation.
287 unsigned long nr_running;
288 #define CPU_LOAD_IDX_MAX 5
289 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
290 unsigned char idle_at_tick;
292 unsigned char in_nohz_recently;
294 /* capture load from *all* tasks on this cpu: */
295 struct load_weight load;
296 unsigned long nr_load_updates;
300 #ifdef CONFIG_FAIR_GROUP_SCHED
301 /* list of leaf cfs_rq on this cpu: */
302 struct list_head leaf_cfs_rq_list;
307 * This is part of a global counter where only the total sum
308 * over all CPUs matters. A task can increase this counter on
309 * one CPU and if it got migrated afterwards it may decrease
310 * it on another CPU. Always updated under the runqueue lock:
312 unsigned long nr_uninterruptible;
314 struct task_struct *curr, *idle;
315 unsigned long next_balance;
316 struct mm_struct *prev_mm;
318 u64 clock, prev_clock_raw;
321 unsigned int clock_warps, clock_overflows;
323 unsigned int clock_deep_idle_events;
329 struct sched_domain *sd;
331 /* For active balancing */
334 /* cpu of this runqueue: */
337 struct task_struct *migration_thread;
338 struct list_head migration_queue;
341 #ifdef CONFIG_SCHEDSTATS
343 struct sched_info rq_sched_info;
345 /* sys_sched_yield() stats */
346 unsigned int yld_exp_empty;
347 unsigned int yld_act_empty;
348 unsigned int yld_both_empty;
349 unsigned int yld_count;
351 /* schedule() stats */
352 unsigned int sched_switch;
353 unsigned int sched_count;
354 unsigned int sched_goidle;
356 /* try_to_wake_up() stats */
357 unsigned int ttwu_count;
358 unsigned int ttwu_local;
361 unsigned int bkl_count;
363 struct lock_class_key rq_lock_key;
366 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
367 static DEFINE_MUTEX(sched_hotcpu_mutex);
369 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
371 rq->curr->sched_class->check_preempt_curr(rq, p);
374 static inline int cpu_of(struct rq *rq)
384 * Update the per-runqueue clock, as finegrained as the platform can give
385 * us, but without assuming monotonicity, etc.:
387 static void __update_rq_clock(struct rq *rq)
389 u64 prev_raw = rq->prev_clock_raw;
390 u64 now = sched_clock();
391 s64 delta = now - prev_raw;
392 u64 clock = rq->clock;
394 #ifdef CONFIG_SCHED_DEBUG
395 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
398 * Protect against sched_clock() occasionally going backwards:
400 if (unlikely(delta < 0)) {
405 * Catch too large forward jumps too:
407 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
408 if (clock < rq->tick_timestamp + TICK_NSEC)
409 clock = rq->tick_timestamp + TICK_NSEC;
412 rq->clock_overflows++;
414 if (unlikely(delta > rq->clock_max_delta))
415 rq->clock_max_delta = delta;
420 rq->prev_clock_raw = now;
424 static void update_rq_clock(struct rq *rq)
426 if (likely(smp_processor_id() == cpu_of(rq)))
427 __update_rq_clock(rq);
431 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
432 * See detach_destroy_domains: synchronize_sched for details.
434 * The domain tree of any CPU may only be accessed from within
435 * preempt-disabled sections.
437 #define for_each_domain(cpu, __sd) \
438 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
440 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
441 #define this_rq() (&__get_cpu_var(runqueues))
442 #define task_rq(p) cpu_rq(task_cpu(p))
443 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
446 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
448 #ifdef CONFIG_SCHED_DEBUG
449 # define const_debug __read_mostly
451 # define const_debug static const
455 * Debugging: various feature bits
458 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
459 SCHED_FEAT_START_DEBIT = 2,
460 SCHED_FEAT_TREE_AVG = 4,
461 SCHED_FEAT_APPROX_AVG = 8,
462 SCHED_FEAT_WAKEUP_PREEMPT = 16,
465 const_debug unsigned int sysctl_sched_features =
466 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
467 SCHED_FEAT_START_DEBIT * 1 |
468 SCHED_FEAT_TREE_AVG * 0 |
469 SCHED_FEAT_APPROX_AVG * 0 |
470 SCHED_FEAT_WAKEUP_PREEMPT * 1;
472 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
475 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
476 * clock constructed from sched_clock():
478 unsigned long long cpu_clock(int cpu)
480 unsigned long long now;
484 local_irq_save(flags);
488 local_irq_restore(flags);
492 EXPORT_SYMBOL_GPL(cpu_clock);
494 #ifndef prepare_arch_switch
495 # define prepare_arch_switch(next) do { } while (0)
497 #ifndef finish_arch_switch
498 # define finish_arch_switch(prev) do { } while (0)
501 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
502 static inline int task_running(struct rq *rq, struct task_struct *p)
504 return rq->curr == p;
507 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
511 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
513 #ifdef CONFIG_DEBUG_SPINLOCK
514 /* this is a valid case when another task releases the spinlock */
515 rq->lock.owner = current;
518 * If we are tracking spinlock dependencies then we have to
519 * fix up the runqueue lock - which gets 'carried over' from
522 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
524 spin_unlock_irq(&rq->lock);
527 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
528 static inline int task_running(struct rq *rq, struct task_struct *p)
533 return rq->curr == p;
537 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
541 * We can optimise this out completely for !SMP, because the
542 * SMP rebalancing from interrupt is the only thing that cares
547 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
548 spin_unlock_irq(&rq->lock);
550 spin_unlock(&rq->lock);
554 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
558 * After ->oncpu is cleared, the task can be moved to a different CPU.
559 * We must ensure this doesn't happen until the switch is completely
565 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
569 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
572 * __task_rq_lock - lock the runqueue a given task resides on.
573 * Must be called interrupts disabled.
575 static inline struct rq *__task_rq_lock(struct task_struct *p)
579 struct rq *rq = task_rq(p);
580 spin_lock(&rq->lock);
581 if (likely(rq == task_rq(p)))
583 spin_unlock(&rq->lock);
588 * task_rq_lock - lock the runqueue a given task resides on and disable
589 * interrupts. Note the ordering: we can safely lookup the task_rq without
590 * explicitly disabling preemption.
592 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
598 local_irq_save(*flags);
600 spin_lock(&rq->lock);
601 if (likely(rq == task_rq(p)))
603 spin_unlock_irqrestore(&rq->lock, *flags);
607 static void __task_rq_unlock(struct rq *rq)
610 spin_unlock(&rq->lock);
613 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
616 spin_unlock_irqrestore(&rq->lock, *flags);
620 * this_rq_lock - lock this runqueue and disable interrupts.
622 static struct rq *this_rq_lock(void)
629 spin_lock(&rq->lock);
635 * We are going deep-idle (irqs are disabled):
637 void sched_clock_idle_sleep_event(void)
639 struct rq *rq = cpu_rq(smp_processor_id());
641 spin_lock(&rq->lock);
642 __update_rq_clock(rq);
643 spin_unlock(&rq->lock);
644 rq->clock_deep_idle_events++;
646 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
649 * We just idled delta nanoseconds (called with irqs disabled):
651 void sched_clock_idle_wakeup_event(u64 delta_ns)
653 struct rq *rq = cpu_rq(smp_processor_id());
654 u64 now = sched_clock();
656 rq->idle_clock += delta_ns;
658 * Override the previous timestamp and ignore all
659 * sched_clock() deltas that occured while we idled,
660 * and use the PM-provided delta_ns to advance the
663 spin_lock(&rq->lock);
664 rq->prev_clock_raw = now;
665 rq->clock += delta_ns;
666 spin_unlock(&rq->lock);
668 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
671 * resched_task - mark a task 'to be rescheduled now'.
673 * On UP this means the setting of the need_resched flag, on SMP it
674 * might also involve a cross-CPU call to trigger the scheduler on
679 #ifndef tsk_is_polling
680 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
683 static void resched_task(struct task_struct *p)
687 assert_spin_locked(&task_rq(p)->lock);
689 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
692 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
695 if (cpu == smp_processor_id())
698 /* NEED_RESCHED must be visible before we test polling */
700 if (!tsk_is_polling(p))
701 smp_send_reschedule(cpu);
704 static void resched_cpu(int cpu)
706 struct rq *rq = cpu_rq(cpu);
709 if (!spin_trylock_irqsave(&rq->lock, flags))
711 resched_task(cpu_curr(cpu));
712 spin_unlock_irqrestore(&rq->lock, flags);
715 static inline void resched_task(struct task_struct *p)
717 assert_spin_locked(&task_rq(p)->lock);
718 set_tsk_need_resched(p);
722 #if BITS_PER_LONG == 32
723 # define WMULT_CONST (~0UL)
725 # define WMULT_CONST (1UL << 32)
728 #define WMULT_SHIFT 32
731 * Shift right and round:
733 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
736 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
737 struct load_weight *lw)
741 if (unlikely(!lw->inv_weight))
742 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
744 tmp = (u64)delta_exec * weight;
746 * Check whether we'd overflow the 64-bit multiplication:
748 if (unlikely(tmp > WMULT_CONST))
749 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
752 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
754 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
757 static inline unsigned long
758 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
760 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
763 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
768 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
774 * To aid in avoiding the subversion of "niceness" due to uneven distribution
775 * of tasks with abnormal "nice" values across CPUs the contribution that
776 * each task makes to its run queue's load is weighted according to its
777 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
778 * scaled version of the new time slice allocation that they receive on time
782 #define WEIGHT_IDLEPRIO 2
783 #define WMULT_IDLEPRIO (1 << 31)
786 * Nice levels are multiplicative, with a gentle 10% change for every
787 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
788 * nice 1, it will get ~10% less CPU time than another CPU-bound task
789 * that remained on nice 0.
791 * The "10% effect" is relative and cumulative: from _any_ nice level,
792 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
793 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
794 * If a task goes up by ~10% and another task goes down by ~10% then
795 * the relative distance between them is ~25%.)
797 static const int prio_to_weight[40] = {
798 /* -20 */ 88761, 71755, 56483, 46273, 36291,
799 /* -15 */ 29154, 23254, 18705, 14949, 11916,
800 /* -10 */ 9548, 7620, 6100, 4904, 3906,
801 /* -5 */ 3121, 2501, 1991, 1586, 1277,
802 /* 0 */ 1024, 820, 655, 526, 423,
803 /* 5 */ 335, 272, 215, 172, 137,
804 /* 10 */ 110, 87, 70, 56, 45,
805 /* 15 */ 36, 29, 23, 18, 15,
809 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
811 * In cases where the weight does not change often, we can use the
812 * precalculated inverse to speed up arithmetics by turning divisions
813 * into multiplications:
815 static const u32 prio_to_wmult[40] = {
816 /* -20 */ 48388, 59856, 76040, 92818, 118348,
817 /* -15 */ 147320, 184698, 229616, 287308, 360437,
818 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
819 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
820 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
821 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
822 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
823 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
826 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
829 * runqueue iterator, to support SMP load-balancing between different
830 * scheduling classes, without having to expose their internal data
831 * structures to the load-balancing proper:
835 struct task_struct *(*start)(void *);
836 struct task_struct *(*next)(void *);
841 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
842 unsigned long max_load_move, struct sched_domain *sd,
843 enum cpu_idle_type idle, int *all_pinned,
844 int *this_best_prio, struct rq_iterator *iterator);
847 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
848 struct sched_domain *sd, enum cpu_idle_type idle,
849 struct rq_iterator *iterator);
852 #include "sched_stats.h"
853 #include "sched_idletask.c"
854 #include "sched_fair.c"
855 #include "sched_rt.c"
856 #ifdef CONFIG_SCHED_DEBUG
857 # include "sched_debug.c"
860 #define sched_class_highest (&rt_sched_class)
863 * Update delta_exec, delta_fair fields for rq.
865 * delta_fair clock advances at a rate inversely proportional to
866 * total load (rq->load.weight) on the runqueue, while
867 * delta_exec advances at the same rate as wall-clock (provided
870 * delta_exec / delta_fair is a measure of the (smoothened) load on this
871 * runqueue over any given interval. This (smoothened) load is used
872 * during load balance.
874 * This function is called /before/ updating rq->load
875 * and when switching tasks.
877 static inline void inc_load(struct rq *rq, const struct task_struct *p)
879 update_load_add(&rq->load, p->se.load.weight);
882 static inline void dec_load(struct rq *rq, const struct task_struct *p)
884 update_load_sub(&rq->load, p->se.load.weight);
887 static void inc_nr_running(struct task_struct *p, struct rq *rq)
893 static void dec_nr_running(struct task_struct *p, struct rq *rq)
899 static void set_load_weight(struct task_struct *p)
901 if (task_has_rt_policy(p)) {
902 p->se.load.weight = prio_to_weight[0] * 2;
903 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
908 * SCHED_IDLE tasks get minimal weight:
910 if (p->policy == SCHED_IDLE) {
911 p->se.load.weight = WEIGHT_IDLEPRIO;
912 p->se.load.inv_weight = WMULT_IDLEPRIO;
916 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
917 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
920 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
922 sched_info_queued(p);
923 p->sched_class->enqueue_task(rq, p, wakeup);
927 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
929 p->sched_class->dequeue_task(rq, p, sleep);
934 * __normal_prio - return the priority that is based on the static prio
936 static inline int __normal_prio(struct task_struct *p)
938 return p->static_prio;
942 * Calculate the expected normal priority: i.e. priority
943 * without taking RT-inheritance into account. Might be
944 * boosted by interactivity modifiers. Changes upon fork,
945 * setprio syscalls, and whenever the interactivity
946 * estimator recalculates.
948 static inline int normal_prio(struct task_struct *p)
952 if (task_has_rt_policy(p))
953 prio = MAX_RT_PRIO-1 - p->rt_priority;
955 prio = __normal_prio(p);
960 * Calculate the current priority, i.e. the priority
961 * taken into account by the scheduler. This value might
962 * be boosted by RT tasks, or might be boosted by
963 * interactivity modifiers. Will be RT if the task got
964 * RT-boosted. If not then it returns p->normal_prio.
966 static int effective_prio(struct task_struct *p)
968 p->normal_prio = normal_prio(p);
970 * If we are RT tasks or we were boosted to RT priority,
971 * keep the priority unchanged. Otherwise, update priority
972 * to the normal priority:
974 if (!rt_prio(p->prio))
975 return p->normal_prio;
980 * activate_task - move a task to the runqueue.
982 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
984 if (p->state == TASK_UNINTERRUPTIBLE)
985 rq->nr_uninterruptible--;
987 enqueue_task(rq, p, wakeup);
988 inc_nr_running(p, rq);
992 * deactivate_task - remove a task from the runqueue.
994 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
996 if (p->state == TASK_UNINTERRUPTIBLE)
997 rq->nr_uninterruptible++;
999 dequeue_task(rq, p, sleep);
1000 dec_nr_running(p, rq);
1004 * task_curr - is this task currently executing on a CPU?
1005 * @p: the task in question.
1007 inline int task_curr(const struct task_struct *p)
1009 return cpu_curr(task_cpu(p)) == p;
1012 /* Used instead of source_load when we know the type == 0 */
1013 unsigned long weighted_cpuload(const int cpu)
1015 return cpu_rq(cpu)->load.weight;
1018 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1021 task_thread_info(p)->cpu = cpu;
1029 * Is this task likely cache-hot:
1032 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1036 if (p->sched_class != &fair_sched_class)
1039 if (sysctl_sched_migration_cost == -1)
1041 if (sysctl_sched_migration_cost == 0)
1044 delta = now - p->se.exec_start;
1046 return delta < (s64)sysctl_sched_migration_cost;
1050 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1052 int old_cpu = task_cpu(p);
1053 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1054 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1055 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1058 clock_offset = old_rq->clock - new_rq->clock;
1060 #ifdef CONFIG_SCHEDSTATS
1061 if (p->se.wait_start)
1062 p->se.wait_start -= clock_offset;
1063 if (p->se.sleep_start)
1064 p->se.sleep_start -= clock_offset;
1065 if (p->se.block_start)
1066 p->se.block_start -= clock_offset;
1067 if (old_cpu != new_cpu) {
1068 schedstat_inc(p, se.nr_migrations);
1069 if (task_hot(p, old_rq->clock, NULL))
1070 schedstat_inc(p, se.nr_forced2_migrations);
1073 p->se.vruntime -= old_cfsrq->min_vruntime -
1074 new_cfsrq->min_vruntime;
1076 __set_task_cpu(p, new_cpu);
1079 struct migration_req {
1080 struct list_head list;
1082 struct task_struct *task;
1085 struct completion done;
1089 * The task's runqueue lock must be held.
1090 * Returns true if you have to wait for migration thread.
1093 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1095 struct rq *rq = task_rq(p);
1098 * If the task is not on a runqueue (and not running), then
1099 * it is sufficient to simply update the task's cpu field.
1101 if (!p->se.on_rq && !task_running(rq, p)) {
1102 set_task_cpu(p, dest_cpu);
1106 init_completion(&req->done);
1108 req->dest_cpu = dest_cpu;
1109 list_add(&req->list, &rq->migration_queue);
1115 * wait_task_inactive - wait for a thread to unschedule.
1117 * The caller must ensure that the task *will* unschedule sometime soon,
1118 * else this function might spin for a *long* time. This function can't
1119 * be called with interrupts off, or it may introduce deadlock with
1120 * smp_call_function() if an IPI is sent by the same process we are
1121 * waiting to become inactive.
1123 void wait_task_inactive(struct task_struct *p)
1125 unsigned long flags;
1131 * We do the initial early heuristics without holding
1132 * any task-queue locks at all. We'll only try to get
1133 * the runqueue lock when things look like they will
1139 * If the task is actively running on another CPU
1140 * still, just relax and busy-wait without holding
1143 * NOTE! Since we don't hold any locks, it's not
1144 * even sure that "rq" stays as the right runqueue!
1145 * But we don't care, since "task_running()" will
1146 * return false if the runqueue has changed and p
1147 * is actually now running somewhere else!
1149 while (task_running(rq, p))
1153 * Ok, time to look more closely! We need the rq
1154 * lock now, to be *sure*. If we're wrong, we'll
1155 * just go back and repeat.
1157 rq = task_rq_lock(p, &flags);
1158 running = task_running(rq, p);
1159 on_rq = p->se.on_rq;
1160 task_rq_unlock(rq, &flags);
1163 * Was it really running after all now that we
1164 * checked with the proper locks actually held?
1166 * Oops. Go back and try again..
1168 if (unlikely(running)) {
1174 * It's not enough that it's not actively running,
1175 * it must be off the runqueue _entirely_, and not
1178 * So if it wa still runnable (but just not actively
1179 * running right now), it's preempted, and we should
1180 * yield - it could be a while.
1182 if (unlikely(on_rq)) {
1183 schedule_timeout_uninterruptible(1);
1188 * Ahh, all good. It wasn't running, and it wasn't
1189 * runnable, which means that it will never become
1190 * running in the future either. We're all done!
1197 * kick_process - kick a running thread to enter/exit the kernel
1198 * @p: the to-be-kicked thread
1200 * Cause a process which is running on another CPU to enter
1201 * kernel-mode, without any delay. (to get signals handled.)
1203 * NOTE: this function doesnt have to take the runqueue lock,
1204 * because all it wants to ensure is that the remote task enters
1205 * the kernel. If the IPI races and the task has been migrated
1206 * to another CPU then no harm is done and the purpose has been
1209 void kick_process(struct task_struct *p)
1215 if ((cpu != smp_processor_id()) && task_curr(p))
1216 smp_send_reschedule(cpu);
1221 * Return a low guess at the load of a migration-source cpu weighted
1222 * according to the scheduling class and "nice" value.
1224 * We want to under-estimate the load of migration sources, to
1225 * balance conservatively.
1227 static unsigned long source_load(int cpu, int type)
1229 struct rq *rq = cpu_rq(cpu);
1230 unsigned long total = weighted_cpuload(cpu);
1235 return min(rq->cpu_load[type-1], total);
1239 * Return a high guess at the load of a migration-target cpu weighted
1240 * according to the scheduling class and "nice" value.
1242 static unsigned long target_load(int cpu, int type)
1244 struct rq *rq = cpu_rq(cpu);
1245 unsigned long total = weighted_cpuload(cpu);
1250 return max(rq->cpu_load[type-1], total);
1254 * Return the average load per task on the cpu's run queue
1256 static inline unsigned long cpu_avg_load_per_task(int cpu)
1258 struct rq *rq = cpu_rq(cpu);
1259 unsigned long total = weighted_cpuload(cpu);
1260 unsigned long n = rq->nr_running;
1262 return n ? total / n : SCHED_LOAD_SCALE;
1266 * find_idlest_group finds and returns the least busy CPU group within the
1269 static struct sched_group *
1270 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1272 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1273 unsigned long min_load = ULONG_MAX, this_load = 0;
1274 int load_idx = sd->forkexec_idx;
1275 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1278 unsigned long load, avg_load;
1282 /* Skip over this group if it has no CPUs allowed */
1283 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1286 local_group = cpu_isset(this_cpu, group->cpumask);
1288 /* Tally up the load of all CPUs in the group */
1291 for_each_cpu_mask(i, group->cpumask) {
1292 /* Bias balancing toward cpus of our domain */
1294 load = source_load(i, load_idx);
1296 load = target_load(i, load_idx);
1301 /* Adjust by relative CPU power of the group */
1302 avg_load = sg_div_cpu_power(group,
1303 avg_load * SCHED_LOAD_SCALE);
1306 this_load = avg_load;
1308 } else if (avg_load < min_load) {
1309 min_load = avg_load;
1312 } while (group = group->next, group != sd->groups);
1314 if (!idlest || 100*this_load < imbalance*min_load)
1320 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1323 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1326 unsigned long load, min_load = ULONG_MAX;
1330 /* Traverse only the allowed CPUs */
1331 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1333 for_each_cpu_mask(i, tmp) {
1334 load = weighted_cpuload(i);
1336 if (load < min_load || (load == min_load && i == this_cpu)) {
1346 * sched_balance_self: balance the current task (running on cpu) in domains
1347 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1350 * Balance, ie. select the least loaded group.
1352 * Returns the target CPU number, or the same CPU if no balancing is needed.
1354 * preempt must be disabled.
1356 static int sched_balance_self(int cpu, int flag)
1358 struct task_struct *t = current;
1359 struct sched_domain *tmp, *sd = NULL;
1361 for_each_domain(cpu, tmp) {
1363 * If power savings logic is enabled for a domain, stop there.
1365 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1367 if (tmp->flags & flag)
1373 struct sched_group *group;
1374 int new_cpu, weight;
1376 if (!(sd->flags & flag)) {
1382 group = find_idlest_group(sd, t, cpu);
1388 new_cpu = find_idlest_cpu(group, t, cpu);
1389 if (new_cpu == -1 || new_cpu == cpu) {
1390 /* Now try balancing at a lower domain level of cpu */
1395 /* Now try balancing at a lower domain level of new_cpu */
1398 weight = cpus_weight(span);
1399 for_each_domain(cpu, tmp) {
1400 if (weight <= cpus_weight(tmp->span))
1402 if (tmp->flags & flag)
1405 /* while loop will break here if sd == NULL */
1411 #endif /* CONFIG_SMP */
1414 * wake_idle() will wake a task on an idle cpu if task->cpu is
1415 * not idle and an idle cpu is available. The span of cpus to
1416 * search starts with cpus closest then further out as needed,
1417 * so we always favor a closer, idle cpu.
1419 * Returns the CPU we should wake onto.
1421 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1422 static int wake_idle(int cpu, struct task_struct *p)
1425 struct sched_domain *sd;
1429 * If it is idle, then it is the best cpu to run this task.
1431 * This cpu is also the best, if it has more than one task already.
1432 * Siblings must be also busy(in most cases) as they didn't already
1433 * pickup the extra load from this cpu and hence we need not check
1434 * sibling runqueue info. This will avoid the checks and cache miss
1435 * penalities associated with that.
1437 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1440 for_each_domain(cpu, sd) {
1441 if (sd->flags & SD_WAKE_IDLE) {
1442 cpus_and(tmp, sd->span, p->cpus_allowed);
1443 for_each_cpu_mask(i, tmp) {
1445 if (i != task_cpu(p)) {
1447 se.nr_wakeups_idle);
1459 static inline int wake_idle(int cpu, struct task_struct *p)
1466 * try_to_wake_up - wake up a thread
1467 * @p: the to-be-woken-up thread
1468 * @state: the mask of task states that can be woken
1469 * @sync: do a synchronous wakeup?
1471 * Put it on the run-queue if it's not already there. The "current"
1472 * thread is always on the run-queue (except when the actual
1473 * re-schedule is in progress), and as such you're allowed to do
1474 * the simpler "current->state = TASK_RUNNING" to mark yourself
1475 * runnable without the overhead of this.
1477 * returns failure only if the task is already active.
1479 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1481 int cpu, orig_cpu, this_cpu, success = 0;
1482 unsigned long flags;
1486 struct sched_domain *sd, *this_sd = NULL;
1487 unsigned long load, this_load;
1491 rq = task_rq_lock(p, &flags);
1492 old_state = p->state;
1493 if (!(old_state & state))
1501 this_cpu = smp_processor_id();
1504 if (unlikely(task_running(rq, p)))
1509 schedstat_inc(rq, ttwu_count);
1510 if (cpu == this_cpu) {
1511 schedstat_inc(rq, ttwu_local);
1515 for_each_domain(this_cpu, sd) {
1516 if (cpu_isset(cpu, sd->span)) {
1517 schedstat_inc(sd, ttwu_wake_remote);
1523 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1527 * Check for affine wakeup and passive balancing possibilities.
1530 int idx = this_sd->wake_idx;
1531 unsigned int imbalance;
1533 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1535 load = source_load(cpu, idx);
1536 this_load = target_load(this_cpu, idx);
1538 new_cpu = this_cpu; /* Wake to this CPU if we can */
1540 if (this_sd->flags & SD_WAKE_AFFINE) {
1541 unsigned long tl = this_load;
1542 unsigned long tl_per_task;
1545 * Attract cache-cold tasks on sync wakeups:
1547 if (sync && !task_hot(p, rq->clock, this_sd))
1550 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1551 tl_per_task = cpu_avg_load_per_task(this_cpu);
1554 * If sync wakeup then subtract the (maximum possible)
1555 * effect of the currently running task from the load
1556 * of the current CPU:
1559 tl -= current->se.load.weight;
1562 tl + target_load(cpu, idx) <= tl_per_task) ||
1563 100*(tl + p->se.load.weight) <= imbalance*load) {
1565 * This domain has SD_WAKE_AFFINE and
1566 * p is cache cold in this domain, and
1567 * there is no bad imbalance.
1569 schedstat_inc(this_sd, ttwu_move_affine);
1570 schedstat_inc(p, se.nr_wakeups_affine);
1576 * Start passive balancing when half the imbalance_pct
1579 if (this_sd->flags & SD_WAKE_BALANCE) {
1580 if (imbalance*this_load <= 100*load) {
1581 schedstat_inc(this_sd, ttwu_move_balance);
1582 schedstat_inc(p, se.nr_wakeups_passive);
1588 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1590 new_cpu = wake_idle(new_cpu, p);
1591 if (new_cpu != cpu) {
1592 set_task_cpu(p, new_cpu);
1593 task_rq_unlock(rq, &flags);
1594 /* might preempt at this point */
1595 rq = task_rq_lock(p, &flags);
1596 old_state = p->state;
1597 if (!(old_state & state))
1602 this_cpu = smp_processor_id();
1607 #endif /* CONFIG_SMP */
1608 schedstat_inc(p, se.nr_wakeups);
1610 schedstat_inc(p, se.nr_wakeups_sync);
1611 if (orig_cpu != cpu)
1612 schedstat_inc(p, se.nr_wakeups_migrate);
1613 if (cpu == this_cpu)
1614 schedstat_inc(p, se.nr_wakeups_local);
1616 schedstat_inc(p, se.nr_wakeups_remote);
1617 update_rq_clock(rq);
1618 activate_task(rq, p, 1);
1619 check_preempt_curr(rq, p);
1623 p->state = TASK_RUNNING;
1625 task_rq_unlock(rq, &flags);
1630 int fastcall wake_up_process(struct task_struct *p)
1632 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1633 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1635 EXPORT_SYMBOL(wake_up_process);
1637 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1639 return try_to_wake_up(p, state, 0);
1643 * Perform scheduler related setup for a newly forked process p.
1644 * p is forked by current.
1646 * __sched_fork() is basic setup used by init_idle() too:
1648 static void __sched_fork(struct task_struct *p)
1650 p->se.exec_start = 0;
1651 p->se.sum_exec_runtime = 0;
1652 p->se.prev_sum_exec_runtime = 0;
1654 #ifdef CONFIG_SCHEDSTATS
1655 p->se.wait_start = 0;
1656 p->se.sum_sleep_runtime = 0;
1657 p->se.sleep_start = 0;
1658 p->se.block_start = 0;
1659 p->se.sleep_max = 0;
1660 p->se.block_max = 0;
1662 p->se.slice_max = 0;
1666 INIT_LIST_HEAD(&p->run_list);
1669 #ifdef CONFIG_PREEMPT_NOTIFIERS
1670 INIT_HLIST_HEAD(&p->preempt_notifiers);
1674 * We mark the process as running here, but have not actually
1675 * inserted it onto the runqueue yet. This guarantees that
1676 * nobody will actually run it, and a signal or other external
1677 * event cannot wake it up and insert it on the runqueue either.
1679 p->state = TASK_RUNNING;
1683 * fork()/clone()-time setup:
1685 void sched_fork(struct task_struct *p, int clone_flags)
1687 int cpu = get_cpu();
1692 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1694 set_task_cpu(p, cpu);
1697 * Make sure we do not leak PI boosting priority to the child:
1699 p->prio = current->normal_prio;
1700 if (!rt_prio(p->prio))
1701 p->sched_class = &fair_sched_class;
1703 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1704 if (likely(sched_info_on()))
1705 memset(&p->sched_info, 0, sizeof(p->sched_info));
1707 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1710 #ifdef CONFIG_PREEMPT
1711 /* Want to start with kernel preemption disabled. */
1712 task_thread_info(p)->preempt_count = 1;
1718 * wake_up_new_task - wake up a newly created task for the first time.
1720 * This function will do some initial scheduler statistics housekeeping
1721 * that must be done for every newly created context, then puts the task
1722 * on the runqueue and wakes it.
1724 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1726 unsigned long flags;
1729 rq = task_rq_lock(p, &flags);
1730 BUG_ON(p->state != TASK_RUNNING);
1731 update_rq_clock(rq);
1733 p->prio = effective_prio(p);
1735 if (!p->sched_class->task_new || !current->se.on_rq) {
1736 activate_task(rq, p, 0);
1739 * Let the scheduling class do new task startup
1740 * management (if any):
1742 p->sched_class->task_new(rq, p);
1743 inc_nr_running(p, rq);
1745 check_preempt_curr(rq, p);
1746 task_rq_unlock(rq, &flags);
1749 #ifdef CONFIG_PREEMPT_NOTIFIERS
1752 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1753 * @notifier: notifier struct to register
1755 void preempt_notifier_register(struct preempt_notifier *notifier)
1757 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1759 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1762 * preempt_notifier_unregister - no longer interested in preemption notifications
1763 * @notifier: notifier struct to unregister
1765 * This is safe to call from within a preemption notifier.
1767 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1769 hlist_del(¬ifier->link);
1771 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1773 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1775 struct preempt_notifier *notifier;
1776 struct hlist_node *node;
1778 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1779 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1783 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1784 struct task_struct *next)
1786 struct preempt_notifier *notifier;
1787 struct hlist_node *node;
1789 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1790 notifier->ops->sched_out(notifier, next);
1795 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1800 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1801 struct task_struct *next)
1808 * prepare_task_switch - prepare to switch tasks
1809 * @rq: the runqueue preparing to switch
1810 * @prev: the current task that is being switched out
1811 * @next: the task we are going to switch to.
1813 * This is called with the rq lock held and interrupts off. It must
1814 * be paired with a subsequent finish_task_switch after the context
1817 * prepare_task_switch sets up locking and calls architecture specific
1821 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1822 struct task_struct *next)
1824 fire_sched_out_preempt_notifiers(prev, next);
1825 prepare_lock_switch(rq, next);
1826 prepare_arch_switch(next);
1830 * finish_task_switch - clean up after a task-switch
1831 * @rq: runqueue associated with task-switch
1832 * @prev: the thread we just switched away from.
1834 * finish_task_switch must be called after the context switch, paired
1835 * with a prepare_task_switch call before the context switch.
1836 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1837 * and do any other architecture-specific cleanup actions.
1839 * Note that we may have delayed dropping an mm in context_switch(). If
1840 * so, we finish that here outside of the runqueue lock. (Doing it
1841 * with the lock held can cause deadlocks; see schedule() for
1844 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1845 __releases(rq->lock)
1847 struct mm_struct *mm = rq->prev_mm;
1853 * A task struct has one reference for the use as "current".
1854 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1855 * schedule one last time. The schedule call will never return, and
1856 * the scheduled task must drop that reference.
1857 * The test for TASK_DEAD must occur while the runqueue locks are
1858 * still held, otherwise prev could be scheduled on another cpu, die
1859 * there before we look at prev->state, and then the reference would
1861 * Manfred Spraul <manfred@colorfullife.com>
1863 prev_state = prev->state;
1864 finish_arch_switch(prev);
1865 finish_lock_switch(rq, prev);
1866 fire_sched_in_preempt_notifiers(current);
1869 if (unlikely(prev_state == TASK_DEAD)) {
1871 * Remove function-return probe instances associated with this
1872 * task and put them back on the free list.
1874 kprobe_flush_task(prev);
1875 put_task_struct(prev);
1880 * schedule_tail - first thing a freshly forked thread must call.
1881 * @prev: the thread we just switched away from.
1883 asmlinkage void schedule_tail(struct task_struct *prev)
1884 __releases(rq->lock)
1886 struct rq *rq = this_rq();
1888 finish_task_switch(rq, prev);
1889 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1890 /* In this case, finish_task_switch does not reenable preemption */
1893 if (current->set_child_tid)
1894 put_user(task_pid_vnr(current), current->set_child_tid);
1898 * context_switch - switch to the new MM and the new
1899 * thread's register state.
1902 context_switch(struct rq *rq, struct task_struct *prev,
1903 struct task_struct *next)
1905 struct mm_struct *mm, *oldmm;
1907 prepare_task_switch(rq, prev, next);
1909 oldmm = prev->active_mm;
1911 * For paravirt, this is coupled with an exit in switch_to to
1912 * combine the page table reload and the switch backend into
1915 arch_enter_lazy_cpu_mode();
1917 if (unlikely(!mm)) {
1918 next->active_mm = oldmm;
1919 atomic_inc(&oldmm->mm_count);
1920 enter_lazy_tlb(oldmm, next);
1922 switch_mm(oldmm, mm, next);
1924 if (unlikely(!prev->mm)) {
1925 prev->active_mm = NULL;
1926 rq->prev_mm = oldmm;
1929 * Since the runqueue lock will be released by the next
1930 * task (which is an invalid locking op but in the case
1931 * of the scheduler it's an obvious special-case), so we
1932 * do an early lockdep release here:
1934 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1935 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1938 /* Here we just switch the register state and the stack. */
1939 switch_to(prev, next, prev);
1943 * this_rq must be evaluated again because prev may have moved
1944 * CPUs since it called schedule(), thus the 'rq' on its stack
1945 * frame will be invalid.
1947 finish_task_switch(this_rq(), prev);
1951 * nr_running, nr_uninterruptible and nr_context_switches:
1953 * externally visible scheduler statistics: current number of runnable
1954 * threads, current number of uninterruptible-sleeping threads, total
1955 * number of context switches performed since bootup.
1957 unsigned long nr_running(void)
1959 unsigned long i, sum = 0;
1961 for_each_online_cpu(i)
1962 sum += cpu_rq(i)->nr_running;
1967 unsigned long nr_uninterruptible(void)
1969 unsigned long i, sum = 0;
1971 for_each_possible_cpu(i)
1972 sum += cpu_rq(i)->nr_uninterruptible;
1975 * Since we read the counters lockless, it might be slightly
1976 * inaccurate. Do not allow it to go below zero though:
1978 if (unlikely((long)sum < 0))
1984 unsigned long long nr_context_switches(void)
1987 unsigned long long sum = 0;
1989 for_each_possible_cpu(i)
1990 sum += cpu_rq(i)->nr_switches;
1995 unsigned long nr_iowait(void)
1997 unsigned long i, sum = 0;
1999 for_each_possible_cpu(i)
2000 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2005 unsigned long nr_active(void)
2007 unsigned long i, running = 0, uninterruptible = 0;
2009 for_each_online_cpu(i) {
2010 running += cpu_rq(i)->nr_running;
2011 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2014 if (unlikely((long)uninterruptible < 0))
2015 uninterruptible = 0;
2017 return running + uninterruptible;
2021 * Update rq->cpu_load[] statistics. This function is usually called every
2022 * scheduler tick (TICK_NSEC).
2024 static void update_cpu_load(struct rq *this_rq)
2026 unsigned long this_load = this_rq->load.weight;
2029 this_rq->nr_load_updates++;
2031 /* Update our load: */
2032 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2033 unsigned long old_load, new_load;
2035 /* scale is effectively 1 << i now, and >> i divides by scale */
2037 old_load = this_rq->cpu_load[i];
2038 new_load = this_load;
2040 * Round up the averaging division if load is increasing. This
2041 * prevents us from getting stuck on 9 if the load is 10, for
2044 if (new_load > old_load)
2045 new_load += scale-1;
2046 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2053 * double_rq_lock - safely lock two runqueues
2055 * Note this does not disable interrupts like task_rq_lock,
2056 * you need to do so manually before calling.
2058 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2059 __acquires(rq1->lock)
2060 __acquires(rq2->lock)
2062 BUG_ON(!irqs_disabled());
2064 spin_lock(&rq1->lock);
2065 __acquire(rq2->lock); /* Fake it out ;) */
2068 spin_lock(&rq1->lock);
2069 spin_lock(&rq2->lock);
2071 spin_lock(&rq2->lock);
2072 spin_lock(&rq1->lock);
2075 update_rq_clock(rq1);
2076 update_rq_clock(rq2);
2080 * double_rq_unlock - safely unlock two runqueues
2082 * Note this does not restore interrupts like task_rq_unlock,
2083 * you need to do so manually after calling.
2085 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2086 __releases(rq1->lock)
2087 __releases(rq2->lock)
2089 spin_unlock(&rq1->lock);
2091 spin_unlock(&rq2->lock);
2093 __release(rq2->lock);
2097 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2099 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2100 __releases(this_rq->lock)
2101 __acquires(busiest->lock)
2102 __acquires(this_rq->lock)
2104 if (unlikely(!irqs_disabled())) {
2105 /* printk() doesn't work good under rq->lock */
2106 spin_unlock(&this_rq->lock);
2109 if (unlikely(!spin_trylock(&busiest->lock))) {
2110 if (busiest < this_rq) {
2111 spin_unlock(&this_rq->lock);
2112 spin_lock(&busiest->lock);
2113 spin_lock(&this_rq->lock);
2115 spin_lock(&busiest->lock);
2120 * If dest_cpu is allowed for this process, migrate the task to it.
2121 * This is accomplished by forcing the cpu_allowed mask to only
2122 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2123 * the cpu_allowed mask is restored.
2125 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2127 struct migration_req req;
2128 unsigned long flags;
2131 rq = task_rq_lock(p, &flags);
2132 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2133 || unlikely(cpu_is_offline(dest_cpu)))
2136 /* force the process onto the specified CPU */
2137 if (migrate_task(p, dest_cpu, &req)) {
2138 /* Need to wait for migration thread (might exit: take ref). */
2139 struct task_struct *mt = rq->migration_thread;
2141 get_task_struct(mt);
2142 task_rq_unlock(rq, &flags);
2143 wake_up_process(mt);
2144 put_task_struct(mt);
2145 wait_for_completion(&req.done);
2150 task_rq_unlock(rq, &flags);
2154 * sched_exec - execve() is a valuable balancing opportunity, because at
2155 * this point the task has the smallest effective memory and cache footprint.
2157 void sched_exec(void)
2159 int new_cpu, this_cpu = get_cpu();
2160 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2162 if (new_cpu != this_cpu)
2163 sched_migrate_task(current, new_cpu);
2167 * pull_task - move a task from a remote runqueue to the local runqueue.
2168 * Both runqueues must be locked.
2170 static void pull_task(struct rq *src_rq, struct task_struct *p,
2171 struct rq *this_rq, int this_cpu)
2173 deactivate_task(src_rq, p, 0);
2174 set_task_cpu(p, this_cpu);
2175 activate_task(this_rq, p, 0);
2177 * Note that idle threads have a prio of MAX_PRIO, for this test
2178 * to be always true for them.
2180 check_preempt_curr(this_rq, p);
2184 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2187 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2188 struct sched_domain *sd, enum cpu_idle_type idle,
2192 * We do not migrate tasks that are:
2193 * 1) running (obviously), or
2194 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2195 * 3) are cache-hot on their current CPU.
2197 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2198 schedstat_inc(p, se.nr_failed_migrations_affine);
2203 if (task_running(rq, p)) {
2204 schedstat_inc(p, se.nr_failed_migrations_running);
2209 * Aggressive migration if:
2210 * 1) task is cache cold, or
2211 * 2) too many balance attempts have failed.
2214 if (!task_hot(p, rq->clock, sd) ||
2215 sd->nr_balance_failed > sd->cache_nice_tries) {
2216 #ifdef CONFIG_SCHEDSTATS
2217 if (task_hot(p, rq->clock, sd)) {
2218 schedstat_inc(sd, lb_hot_gained[idle]);
2219 schedstat_inc(p, se.nr_forced_migrations);
2225 if (task_hot(p, rq->clock, sd)) {
2226 schedstat_inc(p, se.nr_failed_migrations_hot);
2232 static unsigned long
2233 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2234 unsigned long max_load_move, struct sched_domain *sd,
2235 enum cpu_idle_type idle, int *all_pinned,
2236 int *this_best_prio, struct rq_iterator *iterator)
2238 int pulled = 0, pinned = 0, skip_for_load;
2239 struct task_struct *p;
2240 long rem_load_move = max_load_move;
2242 if (max_load_move == 0)
2248 * Start the load-balancing iterator:
2250 p = iterator->start(iterator->arg);
2255 * To help distribute high priority tasks accross CPUs we don't
2256 * skip a task if it will be the highest priority task (i.e. smallest
2257 * prio value) on its new queue regardless of its load weight
2259 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2260 SCHED_LOAD_SCALE_FUZZ;
2261 if ((skip_for_load && p->prio >= *this_best_prio) ||
2262 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2263 p = iterator->next(iterator->arg);
2267 pull_task(busiest, p, this_rq, this_cpu);
2269 rem_load_move -= p->se.load.weight;
2272 * We only want to steal up to the prescribed number of tasks
2273 * and the prescribed amount of weighted load.
2275 if (rem_load_move > 0) {
2276 if (p->prio < *this_best_prio)
2277 *this_best_prio = p->prio;
2278 p = iterator->next(iterator->arg);
2283 * Right now, this is one of only two places pull_task() is called,
2284 * so we can safely collect pull_task() stats here rather than
2285 * inside pull_task().
2287 schedstat_add(sd, lb_gained[idle], pulled);
2290 *all_pinned = pinned;
2292 return max_load_move - rem_load_move;
2296 * move_tasks tries to move up to max_load_move weighted load from busiest to
2297 * this_rq, as part of a balancing operation within domain "sd".
2298 * Returns 1 if successful and 0 otherwise.
2300 * Called with both runqueues locked.
2302 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2303 unsigned long max_load_move,
2304 struct sched_domain *sd, enum cpu_idle_type idle,
2307 const struct sched_class *class = sched_class_highest;
2308 unsigned long total_load_moved = 0;
2309 int this_best_prio = this_rq->curr->prio;
2313 class->load_balance(this_rq, this_cpu, busiest,
2314 max_load_move - total_load_moved,
2315 sd, idle, all_pinned, &this_best_prio);
2316 class = class->next;
2317 } while (class && max_load_move > total_load_moved);
2319 return total_load_moved > 0;
2323 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2324 struct sched_domain *sd, enum cpu_idle_type idle,
2325 struct rq_iterator *iterator)
2327 struct task_struct *p = iterator->start(iterator->arg);
2331 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2332 pull_task(busiest, p, this_rq, this_cpu);
2334 * Right now, this is only the second place pull_task()
2335 * is called, so we can safely collect pull_task()
2336 * stats here rather than inside pull_task().
2338 schedstat_inc(sd, lb_gained[idle]);
2342 p = iterator->next(iterator->arg);
2349 * move_one_task tries to move exactly one task from busiest to this_rq, as
2350 * part of active balancing operations within "domain".
2351 * Returns 1 if successful and 0 otherwise.
2353 * Called with both runqueues locked.
2355 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2356 struct sched_domain *sd, enum cpu_idle_type idle)
2358 const struct sched_class *class;
2360 for (class = sched_class_highest; class; class = class->next)
2361 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2368 * find_busiest_group finds and returns the busiest CPU group within the
2369 * domain. It calculates and returns the amount of weighted load which
2370 * should be moved to restore balance via the imbalance parameter.
2372 static struct sched_group *
2373 find_busiest_group(struct sched_domain *sd, int this_cpu,
2374 unsigned long *imbalance, enum cpu_idle_type idle,
2375 int *sd_idle, cpumask_t *cpus, int *balance)
2377 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2378 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2379 unsigned long max_pull;
2380 unsigned long busiest_load_per_task, busiest_nr_running;
2381 unsigned long this_load_per_task, this_nr_running;
2382 int load_idx, group_imb = 0;
2383 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2384 int power_savings_balance = 1;
2385 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2386 unsigned long min_nr_running = ULONG_MAX;
2387 struct sched_group *group_min = NULL, *group_leader = NULL;
2390 max_load = this_load = total_load = total_pwr = 0;
2391 busiest_load_per_task = busiest_nr_running = 0;
2392 this_load_per_task = this_nr_running = 0;
2393 if (idle == CPU_NOT_IDLE)
2394 load_idx = sd->busy_idx;
2395 else if (idle == CPU_NEWLY_IDLE)
2396 load_idx = sd->newidle_idx;
2398 load_idx = sd->idle_idx;
2401 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2404 int __group_imb = 0;
2405 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2406 unsigned long sum_nr_running, sum_weighted_load;
2408 local_group = cpu_isset(this_cpu, group->cpumask);
2411 balance_cpu = first_cpu(group->cpumask);
2413 /* Tally up the load of all CPUs in the group */
2414 sum_weighted_load = sum_nr_running = avg_load = 0;
2416 min_cpu_load = ~0UL;
2418 for_each_cpu_mask(i, group->cpumask) {
2421 if (!cpu_isset(i, *cpus))
2426 if (*sd_idle && rq->nr_running)
2429 /* Bias balancing toward cpus of our domain */
2431 if (idle_cpu(i) && !first_idle_cpu) {
2436 load = target_load(i, load_idx);
2438 load = source_load(i, load_idx);
2439 if (load > max_cpu_load)
2440 max_cpu_load = load;
2441 if (min_cpu_load > load)
2442 min_cpu_load = load;
2446 sum_nr_running += rq->nr_running;
2447 sum_weighted_load += weighted_cpuload(i);
2451 * First idle cpu or the first cpu(busiest) in this sched group
2452 * is eligible for doing load balancing at this and above
2453 * domains. In the newly idle case, we will allow all the cpu's
2454 * to do the newly idle load balance.
2456 if (idle != CPU_NEWLY_IDLE && local_group &&
2457 balance_cpu != this_cpu && balance) {
2462 total_load += avg_load;
2463 total_pwr += group->__cpu_power;
2465 /* Adjust by relative CPU power of the group */
2466 avg_load = sg_div_cpu_power(group,
2467 avg_load * SCHED_LOAD_SCALE);
2469 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2472 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2475 this_load = avg_load;
2477 this_nr_running = sum_nr_running;
2478 this_load_per_task = sum_weighted_load;
2479 } else if (avg_load > max_load &&
2480 (sum_nr_running > group_capacity || __group_imb)) {
2481 max_load = avg_load;
2483 busiest_nr_running = sum_nr_running;
2484 busiest_load_per_task = sum_weighted_load;
2485 group_imb = __group_imb;
2488 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2490 * Busy processors will not participate in power savings
2493 if (idle == CPU_NOT_IDLE ||
2494 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2498 * If the local group is idle or completely loaded
2499 * no need to do power savings balance at this domain
2501 if (local_group && (this_nr_running >= group_capacity ||
2503 power_savings_balance = 0;
2506 * If a group is already running at full capacity or idle,
2507 * don't include that group in power savings calculations
2509 if (!power_savings_balance || sum_nr_running >= group_capacity
2514 * Calculate the group which has the least non-idle load.
2515 * This is the group from where we need to pick up the load
2518 if ((sum_nr_running < min_nr_running) ||
2519 (sum_nr_running == min_nr_running &&
2520 first_cpu(group->cpumask) <
2521 first_cpu(group_min->cpumask))) {
2523 min_nr_running = sum_nr_running;
2524 min_load_per_task = sum_weighted_load /
2529 * Calculate the group which is almost near its
2530 * capacity but still has some space to pick up some load
2531 * from other group and save more power
2533 if (sum_nr_running <= group_capacity - 1) {
2534 if (sum_nr_running > leader_nr_running ||
2535 (sum_nr_running == leader_nr_running &&
2536 first_cpu(group->cpumask) >
2537 first_cpu(group_leader->cpumask))) {
2538 group_leader = group;
2539 leader_nr_running = sum_nr_running;
2544 group = group->next;
2545 } while (group != sd->groups);
2547 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2550 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2552 if (this_load >= avg_load ||
2553 100*max_load <= sd->imbalance_pct*this_load)
2556 busiest_load_per_task /= busiest_nr_running;
2558 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2561 * We're trying to get all the cpus to the average_load, so we don't
2562 * want to push ourselves above the average load, nor do we wish to
2563 * reduce the max loaded cpu below the average load, as either of these
2564 * actions would just result in more rebalancing later, and ping-pong
2565 * tasks around. Thus we look for the minimum possible imbalance.
2566 * Negative imbalances (*we* are more loaded than anyone else) will
2567 * be counted as no imbalance for these purposes -- we can't fix that
2568 * by pulling tasks to us. Be careful of negative numbers as they'll
2569 * appear as very large values with unsigned longs.
2571 if (max_load <= busiest_load_per_task)
2575 * In the presence of smp nice balancing, certain scenarios can have
2576 * max load less than avg load(as we skip the groups at or below
2577 * its cpu_power, while calculating max_load..)
2579 if (max_load < avg_load) {
2581 goto small_imbalance;
2584 /* Don't want to pull so many tasks that a group would go idle */
2585 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2587 /* How much load to actually move to equalise the imbalance */
2588 *imbalance = min(max_pull * busiest->__cpu_power,
2589 (avg_load - this_load) * this->__cpu_power)
2593 * if *imbalance is less than the average load per runnable task
2594 * there is no gaurantee that any tasks will be moved so we'll have
2595 * a think about bumping its value to force at least one task to be
2598 if (*imbalance < busiest_load_per_task) {
2599 unsigned long tmp, pwr_now, pwr_move;
2603 pwr_move = pwr_now = 0;
2605 if (this_nr_running) {
2606 this_load_per_task /= this_nr_running;
2607 if (busiest_load_per_task > this_load_per_task)
2610 this_load_per_task = SCHED_LOAD_SCALE;
2612 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2613 busiest_load_per_task * imbn) {
2614 *imbalance = busiest_load_per_task;
2619 * OK, we don't have enough imbalance to justify moving tasks,
2620 * however we may be able to increase total CPU power used by
2624 pwr_now += busiest->__cpu_power *
2625 min(busiest_load_per_task, max_load);
2626 pwr_now += this->__cpu_power *
2627 min(this_load_per_task, this_load);
2628 pwr_now /= SCHED_LOAD_SCALE;
2630 /* Amount of load we'd subtract */
2631 tmp = sg_div_cpu_power(busiest,
2632 busiest_load_per_task * SCHED_LOAD_SCALE);
2634 pwr_move += busiest->__cpu_power *
2635 min(busiest_load_per_task, max_load - tmp);
2637 /* Amount of load we'd add */
2638 if (max_load * busiest->__cpu_power <
2639 busiest_load_per_task * SCHED_LOAD_SCALE)
2640 tmp = sg_div_cpu_power(this,
2641 max_load * busiest->__cpu_power);
2643 tmp = sg_div_cpu_power(this,
2644 busiest_load_per_task * SCHED_LOAD_SCALE);
2645 pwr_move += this->__cpu_power *
2646 min(this_load_per_task, this_load + tmp);
2647 pwr_move /= SCHED_LOAD_SCALE;
2649 /* Move if we gain throughput */
2650 if (pwr_move > pwr_now)
2651 *imbalance = busiest_load_per_task;
2657 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2658 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2661 if (this == group_leader && group_leader != group_min) {
2662 *imbalance = min_load_per_task;
2672 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2675 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2676 unsigned long imbalance, cpumask_t *cpus)
2678 struct rq *busiest = NULL, *rq;
2679 unsigned long max_load = 0;
2682 for_each_cpu_mask(i, group->cpumask) {
2685 if (!cpu_isset(i, *cpus))
2689 wl = weighted_cpuload(i);
2691 if (rq->nr_running == 1 && wl > imbalance)
2694 if (wl > max_load) {
2704 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2705 * so long as it is large enough.
2707 #define MAX_PINNED_INTERVAL 512
2710 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2711 * tasks if there is an imbalance.
2713 static int load_balance(int this_cpu, struct rq *this_rq,
2714 struct sched_domain *sd, enum cpu_idle_type idle,
2717 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2718 struct sched_group *group;
2719 unsigned long imbalance;
2721 cpumask_t cpus = CPU_MASK_ALL;
2722 unsigned long flags;
2725 * When power savings policy is enabled for the parent domain, idle
2726 * sibling can pick up load irrespective of busy siblings. In this case,
2727 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2728 * portraying it as CPU_NOT_IDLE.
2730 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2731 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2734 schedstat_inc(sd, lb_count[idle]);
2737 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2744 schedstat_inc(sd, lb_nobusyg[idle]);
2748 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2750 schedstat_inc(sd, lb_nobusyq[idle]);
2754 BUG_ON(busiest == this_rq);
2756 schedstat_add(sd, lb_imbalance[idle], imbalance);
2759 if (busiest->nr_running > 1) {
2761 * Attempt to move tasks. If find_busiest_group has found
2762 * an imbalance but busiest->nr_running <= 1, the group is
2763 * still unbalanced. ld_moved simply stays zero, so it is
2764 * correctly treated as an imbalance.
2766 local_irq_save(flags);
2767 double_rq_lock(this_rq, busiest);
2768 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2769 imbalance, sd, idle, &all_pinned);
2770 double_rq_unlock(this_rq, busiest);
2771 local_irq_restore(flags);
2774 * some other cpu did the load balance for us.
2776 if (ld_moved && this_cpu != smp_processor_id())
2777 resched_cpu(this_cpu);
2779 /* All tasks on this runqueue were pinned by CPU affinity */
2780 if (unlikely(all_pinned)) {
2781 cpu_clear(cpu_of(busiest), cpus);
2782 if (!cpus_empty(cpus))
2789 schedstat_inc(sd, lb_failed[idle]);
2790 sd->nr_balance_failed++;
2792 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2794 spin_lock_irqsave(&busiest->lock, flags);
2796 /* don't kick the migration_thread, if the curr
2797 * task on busiest cpu can't be moved to this_cpu
2799 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2800 spin_unlock_irqrestore(&busiest->lock, flags);
2802 goto out_one_pinned;
2805 if (!busiest->active_balance) {
2806 busiest->active_balance = 1;
2807 busiest->push_cpu = this_cpu;
2810 spin_unlock_irqrestore(&busiest->lock, flags);
2812 wake_up_process(busiest->migration_thread);
2815 * We've kicked active balancing, reset the failure
2818 sd->nr_balance_failed = sd->cache_nice_tries+1;
2821 sd->nr_balance_failed = 0;
2823 if (likely(!active_balance)) {
2824 /* We were unbalanced, so reset the balancing interval */
2825 sd->balance_interval = sd->min_interval;
2828 * If we've begun active balancing, start to back off. This
2829 * case may not be covered by the all_pinned logic if there
2830 * is only 1 task on the busy runqueue (because we don't call
2833 if (sd->balance_interval < sd->max_interval)
2834 sd->balance_interval *= 2;
2837 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2838 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2843 schedstat_inc(sd, lb_balanced[idle]);
2845 sd->nr_balance_failed = 0;
2848 /* tune up the balancing interval */
2849 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2850 (sd->balance_interval < sd->max_interval))
2851 sd->balance_interval *= 2;
2853 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2854 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2860 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2861 * tasks if there is an imbalance.
2863 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2864 * this_rq is locked.
2867 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2869 struct sched_group *group;
2870 struct rq *busiest = NULL;
2871 unsigned long imbalance;
2875 cpumask_t cpus = CPU_MASK_ALL;
2878 * When power savings policy is enabled for the parent domain, idle
2879 * sibling can pick up load irrespective of busy siblings. In this case,
2880 * let the state of idle sibling percolate up as IDLE, instead of
2881 * portraying it as CPU_NOT_IDLE.
2883 if (sd->flags & SD_SHARE_CPUPOWER &&
2884 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2887 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2889 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2890 &sd_idle, &cpus, NULL);
2892 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2896 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2899 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2903 BUG_ON(busiest == this_rq);
2905 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2908 if (busiest->nr_running > 1) {
2909 /* Attempt to move tasks */
2910 double_lock_balance(this_rq, busiest);
2911 /* this_rq->clock is already updated */
2912 update_rq_clock(busiest);
2913 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2914 imbalance, sd, CPU_NEWLY_IDLE,
2916 spin_unlock(&busiest->lock);
2918 if (unlikely(all_pinned)) {
2919 cpu_clear(cpu_of(busiest), cpus);
2920 if (!cpus_empty(cpus))
2926 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2927 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2928 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2931 sd->nr_balance_failed = 0;
2936 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2937 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2938 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2940 sd->nr_balance_failed = 0;
2946 * idle_balance is called by schedule() if this_cpu is about to become
2947 * idle. Attempts to pull tasks from other CPUs.
2949 static void idle_balance(int this_cpu, struct rq *this_rq)
2951 struct sched_domain *sd;
2952 int pulled_task = -1;
2953 unsigned long next_balance = jiffies + HZ;
2955 for_each_domain(this_cpu, sd) {
2956 unsigned long interval;
2958 if (!(sd->flags & SD_LOAD_BALANCE))
2961 if (sd->flags & SD_BALANCE_NEWIDLE)
2962 /* If we've pulled tasks over stop searching: */
2963 pulled_task = load_balance_newidle(this_cpu,
2966 interval = msecs_to_jiffies(sd->balance_interval);
2967 if (time_after(next_balance, sd->last_balance + interval))
2968 next_balance = sd->last_balance + interval;
2972 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2974 * We are going idle. next_balance may be set based on
2975 * a busy processor. So reset next_balance.
2977 this_rq->next_balance = next_balance;
2982 * active_load_balance is run by migration threads. It pushes running tasks
2983 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2984 * running on each physical CPU where possible, and avoids physical /
2985 * logical imbalances.
2987 * Called with busiest_rq locked.
2989 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2991 int target_cpu = busiest_rq->push_cpu;
2992 struct sched_domain *sd;
2993 struct rq *target_rq;
2995 /* Is there any task to move? */
2996 if (busiest_rq->nr_running <= 1)
2999 target_rq = cpu_rq(target_cpu);
3002 * This condition is "impossible", if it occurs
3003 * we need to fix it. Originally reported by
3004 * Bjorn Helgaas on a 128-cpu setup.
3006 BUG_ON(busiest_rq == target_rq);
3008 /* move a task from busiest_rq to target_rq */
3009 double_lock_balance(busiest_rq, target_rq);
3010 update_rq_clock(busiest_rq);
3011 update_rq_clock(target_rq);
3013 /* Search for an sd spanning us and the target CPU. */
3014 for_each_domain(target_cpu, sd) {
3015 if ((sd->flags & SD_LOAD_BALANCE) &&
3016 cpu_isset(busiest_cpu, sd->span))
3021 schedstat_inc(sd, alb_count);
3023 if (move_one_task(target_rq, target_cpu, busiest_rq,
3025 schedstat_inc(sd, alb_pushed);
3027 schedstat_inc(sd, alb_failed);
3029 spin_unlock(&target_rq->lock);
3034 atomic_t load_balancer;
3036 } nohz ____cacheline_aligned = {
3037 .load_balancer = ATOMIC_INIT(-1),
3038 .cpu_mask = CPU_MASK_NONE,
3042 * This routine will try to nominate the ilb (idle load balancing)
3043 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3044 * load balancing on behalf of all those cpus. If all the cpus in the system
3045 * go into this tickless mode, then there will be no ilb owner (as there is
3046 * no need for one) and all the cpus will sleep till the next wakeup event
3049 * For the ilb owner, tick is not stopped. And this tick will be used
3050 * for idle load balancing. ilb owner will still be part of
3053 * While stopping the tick, this cpu will become the ilb owner if there
3054 * is no other owner. And will be the owner till that cpu becomes busy
3055 * or if all cpus in the system stop their ticks at which point
3056 * there is no need for ilb owner.
3058 * When the ilb owner becomes busy, it nominates another owner, during the
3059 * next busy scheduler_tick()
3061 int select_nohz_load_balancer(int stop_tick)
3063 int cpu = smp_processor_id();
3066 cpu_set(cpu, nohz.cpu_mask);
3067 cpu_rq(cpu)->in_nohz_recently = 1;
3070 * If we are going offline and still the leader, give up!
3072 if (cpu_is_offline(cpu) &&
3073 atomic_read(&nohz.load_balancer) == cpu) {
3074 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3079 /* time for ilb owner also to sleep */
3080 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3081 if (atomic_read(&nohz.load_balancer) == cpu)
3082 atomic_set(&nohz.load_balancer, -1);
3086 if (atomic_read(&nohz.load_balancer) == -1) {
3087 /* make me the ilb owner */
3088 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3090 } else if (atomic_read(&nohz.load_balancer) == cpu)
3093 if (!cpu_isset(cpu, nohz.cpu_mask))
3096 cpu_clear(cpu, nohz.cpu_mask);
3098 if (atomic_read(&nohz.load_balancer) == cpu)
3099 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3106 static DEFINE_SPINLOCK(balancing);
3109 * It checks each scheduling domain to see if it is due to be balanced,
3110 * and initiates a balancing operation if so.
3112 * Balancing parameters are set up in arch_init_sched_domains.
3114 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3117 struct rq *rq = cpu_rq(cpu);
3118 unsigned long interval;
3119 struct sched_domain *sd;
3120 /* Earliest time when we have to do rebalance again */
3121 unsigned long next_balance = jiffies + 60*HZ;
3122 int update_next_balance = 0;
3124 for_each_domain(cpu, sd) {
3125 if (!(sd->flags & SD_LOAD_BALANCE))
3128 interval = sd->balance_interval;
3129 if (idle != CPU_IDLE)
3130 interval *= sd->busy_factor;
3132 /* scale ms to jiffies */
3133 interval = msecs_to_jiffies(interval);
3134 if (unlikely(!interval))
3136 if (interval > HZ*NR_CPUS/10)
3137 interval = HZ*NR_CPUS/10;
3140 if (sd->flags & SD_SERIALIZE) {
3141 if (!spin_trylock(&balancing))
3145 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3146 if (load_balance(cpu, rq, sd, idle, &balance)) {
3148 * We've pulled tasks over so either we're no
3149 * longer idle, or one of our SMT siblings is
3152 idle = CPU_NOT_IDLE;
3154 sd->last_balance = jiffies;
3156 if (sd->flags & SD_SERIALIZE)
3157 spin_unlock(&balancing);
3159 if (time_after(next_balance, sd->last_balance + interval)) {
3160 next_balance = sd->last_balance + interval;
3161 update_next_balance = 1;
3165 * Stop the load balance at this level. There is another
3166 * CPU in our sched group which is doing load balancing more
3174 * next_balance will be updated only when there is a need.
3175 * When the cpu is attached to null domain for ex, it will not be
3178 if (likely(update_next_balance))
3179 rq->next_balance = next_balance;
3183 * run_rebalance_domains is triggered when needed from the scheduler tick.
3184 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3185 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3187 static void run_rebalance_domains(struct softirq_action *h)
3189 int this_cpu = smp_processor_id();
3190 struct rq *this_rq = cpu_rq(this_cpu);
3191 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3192 CPU_IDLE : CPU_NOT_IDLE;
3194 rebalance_domains(this_cpu, idle);
3198 * If this cpu is the owner for idle load balancing, then do the
3199 * balancing on behalf of the other idle cpus whose ticks are
3202 if (this_rq->idle_at_tick &&
3203 atomic_read(&nohz.load_balancer) == this_cpu) {
3204 cpumask_t cpus = nohz.cpu_mask;
3208 cpu_clear(this_cpu, cpus);
3209 for_each_cpu_mask(balance_cpu, cpus) {
3211 * If this cpu gets work to do, stop the load balancing
3212 * work being done for other cpus. Next load
3213 * balancing owner will pick it up.
3218 rebalance_domains(balance_cpu, CPU_IDLE);
3220 rq = cpu_rq(balance_cpu);
3221 if (time_after(this_rq->next_balance, rq->next_balance))
3222 this_rq->next_balance = rq->next_balance;
3229 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3231 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3232 * idle load balancing owner or decide to stop the periodic load balancing,
3233 * if the whole system is idle.
3235 static inline void trigger_load_balance(struct rq *rq, int cpu)
3239 * If we were in the nohz mode recently and busy at the current
3240 * scheduler tick, then check if we need to nominate new idle
3243 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3244 rq->in_nohz_recently = 0;
3246 if (atomic_read(&nohz.load_balancer) == cpu) {
3247 cpu_clear(cpu, nohz.cpu_mask);
3248 atomic_set(&nohz.load_balancer, -1);
3251 if (atomic_read(&nohz.load_balancer) == -1) {
3253 * simple selection for now: Nominate the
3254 * first cpu in the nohz list to be the next
3257 * TBD: Traverse the sched domains and nominate
3258 * the nearest cpu in the nohz.cpu_mask.
3260 int ilb = first_cpu(nohz.cpu_mask);
3268 * If this cpu is idle and doing idle load balancing for all the
3269 * cpus with ticks stopped, is it time for that to stop?
3271 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3272 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3278 * If this cpu is idle and the idle load balancing is done by
3279 * someone else, then no need raise the SCHED_SOFTIRQ
3281 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3282 cpu_isset(cpu, nohz.cpu_mask))
3285 if (time_after_eq(jiffies, rq->next_balance))
3286 raise_softirq(SCHED_SOFTIRQ);
3289 #else /* CONFIG_SMP */
3292 * on UP we do not need to balance between CPUs:
3294 static inline void idle_balance(int cpu, struct rq *rq)
3300 DEFINE_PER_CPU(struct kernel_stat, kstat);
3302 EXPORT_PER_CPU_SYMBOL(kstat);
3305 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3306 * that have not yet been banked in case the task is currently running.
3308 unsigned long long task_sched_runtime(struct task_struct *p)
3310 unsigned long flags;
3314 rq = task_rq_lock(p, &flags);
3315 ns = p->se.sum_exec_runtime;
3316 if (rq->curr == p) {
3317 update_rq_clock(rq);
3318 delta_exec = rq->clock - p->se.exec_start;
3319 if ((s64)delta_exec > 0)
3322 task_rq_unlock(rq, &flags);
3328 * Account user cpu time to a process.
3329 * @p: the process that the cpu time gets accounted to
3330 * @cputime: the cpu time spent in user space since the last update
3332 void account_user_time(struct task_struct *p, cputime_t cputime)
3334 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3336 struct rq *rq = this_rq();
3338 p->utime = cputime_add(p->utime, cputime);
3341 cpuacct_charge(p, 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) {
3394 account_guest_time(p, cputime);
3398 p->stime = cputime_add(p->stime, cputime);
3400 /* Add system time to cpustat. */
3401 tmp = cputime_to_cputime64(cputime);
3402 if (hardirq_count() - hardirq_offset)
3403 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3404 else if (softirq_count())
3405 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3406 else if (p != rq->idle) {
3407 cpustat->system = cputime64_add(cpustat->system, tmp);
3408 cpuacct_charge(p, cputime);
3409 } else if (atomic_read(&rq->nr_iowait) > 0)
3410 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3412 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3413 /* Account for system time used */
3414 acct_update_integrals(p);
3418 * Account scaled system cpu time to a process.
3419 * @p: the process that the cpu time gets accounted to
3420 * @hardirq_offset: the offset to subtract from hardirq_count()
3421 * @cputime: the cpu time spent in kernel space since the last update
3423 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3425 p->stimescaled = cputime_add(p->stimescaled, cputime);
3429 * Account for involuntary wait time.
3430 * @p: the process from which the cpu time has been stolen
3431 * @steal: the cpu time spent in involuntary wait
3433 void account_steal_time(struct task_struct *p, cputime_t steal)
3435 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3436 cputime64_t tmp = cputime_to_cputime64(steal);
3437 struct rq *rq = this_rq();
3439 if (p == rq->idle) {
3440 p->stime = cputime_add(p->stime, steal);
3441 if (atomic_read(&rq->nr_iowait) > 0)
3442 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3444 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3446 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3447 cpuacct_charge(p, -tmp);
3452 * This function gets called by the timer code, with HZ frequency.
3453 * We call it with interrupts disabled.
3455 * It also gets called by the fork code, when changing the parent's
3458 void scheduler_tick(void)
3460 int cpu = smp_processor_id();
3461 struct rq *rq = cpu_rq(cpu);
3462 struct task_struct *curr = rq->curr;
3463 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3465 spin_lock(&rq->lock);
3466 __update_rq_clock(rq);
3468 * Let rq->clock advance by at least TICK_NSEC:
3470 if (unlikely(rq->clock < next_tick))
3471 rq->clock = next_tick;
3472 rq->tick_timestamp = rq->clock;
3473 update_cpu_load(rq);
3474 if (curr != rq->idle) /* FIXME: needed? */
3475 curr->sched_class->task_tick(rq, curr);
3476 spin_unlock(&rq->lock);
3479 rq->idle_at_tick = idle_cpu(cpu);
3480 trigger_load_balance(rq, cpu);
3484 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3486 void fastcall add_preempt_count(int val)
3491 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3493 preempt_count() += val;
3495 * Spinlock count overflowing soon?
3497 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3500 EXPORT_SYMBOL(add_preempt_count);
3502 void fastcall sub_preempt_count(int val)
3507 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3510 * Is the spinlock portion underflowing?
3512 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3513 !(preempt_count() & PREEMPT_MASK)))
3516 preempt_count() -= val;
3518 EXPORT_SYMBOL(sub_preempt_count);
3523 * Print scheduling while atomic bug:
3525 static noinline void __schedule_bug(struct task_struct *prev)
3527 struct pt_regs *regs = get_irq_regs();
3529 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3530 prev->comm, prev->pid, preempt_count());
3532 debug_show_held_locks(prev);
3533 if (irqs_disabled())
3534 print_irqtrace_events(prev);
3543 * Various schedule()-time debugging checks and statistics:
3545 static inline void schedule_debug(struct task_struct *prev)
3548 * Test if we are atomic. Since do_exit() needs to call into
3549 * schedule() atomically, we ignore that path for now.
3550 * Otherwise, whine if we are scheduling when we should not be.
3552 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3553 __schedule_bug(prev);
3555 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3557 schedstat_inc(this_rq(), sched_count);
3558 #ifdef CONFIG_SCHEDSTATS
3559 if (unlikely(prev->lock_depth >= 0)) {
3560 schedstat_inc(this_rq(), bkl_count);
3561 schedstat_inc(prev, sched_info.bkl_count);
3567 * Pick up the highest-prio task:
3569 static inline struct task_struct *
3570 pick_next_task(struct rq *rq, struct task_struct *prev)
3572 const struct sched_class *class;
3573 struct task_struct *p;
3576 * Optimization: we know that if all tasks are in
3577 * the fair class we can call that function directly:
3579 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3580 p = fair_sched_class.pick_next_task(rq);
3585 class = sched_class_highest;
3587 p = class->pick_next_task(rq);
3591 * Will never be NULL as the idle class always
3592 * returns a non-NULL p:
3594 class = class->next;
3599 * schedule() is the main scheduler function.
3601 asmlinkage void __sched schedule(void)
3603 struct task_struct *prev, *next;
3610 cpu = smp_processor_id();
3614 switch_count = &prev->nivcsw;
3616 release_kernel_lock(prev);
3617 need_resched_nonpreemptible:
3619 schedule_debug(prev);
3622 * Do the rq-clock update outside the rq lock:
3624 local_irq_disable();
3625 __update_rq_clock(rq);
3626 spin_lock(&rq->lock);
3627 clear_tsk_need_resched(prev);
3629 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3630 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3631 unlikely(signal_pending(prev)))) {
3632 prev->state = TASK_RUNNING;
3634 deactivate_task(rq, prev, 1);
3636 switch_count = &prev->nvcsw;
3639 if (unlikely(!rq->nr_running))
3640 idle_balance(cpu, rq);
3642 prev->sched_class->put_prev_task(rq, prev);
3643 next = pick_next_task(rq, prev);
3645 sched_info_switch(prev, next);
3647 if (likely(prev != next)) {
3652 context_switch(rq, prev, next); /* unlocks the rq */
3654 spin_unlock_irq(&rq->lock);
3656 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3657 cpu = smp_processor_id();
3659 goto need_resched_nonpreemptible;
3661 preempt_enable_no_resched();
3662 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3665 EXPORT_SYMBOL(schedule);
3667 #ifdef CONFIG_PREEMPT
3669 * this is the entry point to schedule() from in-kernel preemption
3670 * off of preempt_enable. Kernel preemptions off return from interrupt
3671 * occur there and call schedule directly.
3673 asmlinkage void __sched preempt_schedule(void)
3675 struct thread_info *ti = current_thread_info();
3676 #ifdef CONFIG_PREEMPT_BKL
3677 struct task_struct *task = current;
3678 int saved_lock_depth;
3681 * If there is a non-zero preempt_count or interrupts are disabled,
3682 * we do not want to preempt the current task. Just return..
3684 if (likely(ti->preempt_count || irqs_disabled()))
3688 add_preempt_count(PREEMPT_ACTIVE);
3691 * We keep the big kernel semaphore locked, but we
3692 * clear ->lock_depth so that schedule() doesnt
3693 * auto-release the semaphore:
3695 #ifdef CONFIG_PREEMPT_BKL
3696 saved_lock_depth = task->lock_depth;
3697 task->lock_depth = -1;
3700 #ifdef CONFIG_PREEMPT_BKL
3701 task->lock_depth = saved_lock_depth;
3703 sub_preempt_count(PREEMPT_ACTIVE);
3706 * Check again in case we missed a preemption opportunity
3707 * between schedule and now.
3710 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3712 EXPORT_SYMBOL(preempt_schedule);
3715 * this is the entry point to schedule() from kernel preemption
3716 * off of irq context.
3717 * Note, that this is called and return with irqs disabled. This will
3718 * protect us against recursive calling from irq.
3720 asmlinkage void __sched preempt_schedule_irq(void)
3722 struct thread_info *ti = current_thread_info();
3723 #ifdef CONFIG_PREEMPT_BKL
3724 struct task_struct *task = current;
3725 int saved_lock_depth;
3727 /* Catch callers which need to be fixed */
3728 BUG_ON(ti->preempt_count || !irqs_disabled());
3731 add_preempt_count(PREEMPT_ACTIVE);
3734 * We keep the big kernel semaphore locked, but we
3735 * clear ->lock_depth so that schedule() doesnt
3736 * auto-release the semaphore:
3738 #ifdef CONFIG_PREEMPT_BKL
3739 saved_lock_depth = task->lock_depth;
3740 task->lock_depth = -1;
3744 local_irq_disable();
3745 #ifdef CONFIG_PREEMPT_BKL
3746 task->lock_depth = saved_lock_depth;
3748 sub_preempt_count(PREEMPT_ACTIVE);
3751 * Check again in case we missed a preemption opportunity
3752 * between schedule and now.
3755 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3758 #endif /* CONFIG_PREEMPT */
3760 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3763 return try_to_wake_up(curr->private, mode, sync);
3765 EXPORT_SYMBOL(default_wake_function);
3768 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3769 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3770 * number) then we wake all the non-exclusive tasks and one exclusive task.
3772 * There are circumstances in which we can try to wake a task which has already
3773 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3774 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3776 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3777 int nr_exclusive, int sync, void *key)
3779 wait_queue_t *curr, *next;
3781 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3782 unsigned flags = curr->flags;
3784 if (curr->func(curr, mode, sync, key) &&
3785 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3791 * __wake_up - wake up threads blocked on a waitqueue.
3793 * @mode: which threads
3794 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3795 * @key: is directly passed to the wakeup function
3797 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3798 int nr_exclusive, void *key)
3800 unsigned long flags;
3802 spin_lock_irqsave(&q->lock, flags);
3803 __wake_up_common(q, mode, nr_exclusive, 0, key);
3804 spin_unlock_irqrestore(&q->lock, flags);
3806 EXPORT_SYMBOL(__wake_up);
3809 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3811 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3813 __wake_up_common(q, mode, 1, 0, NULL);
3817 * __wake_up_sync - wake up threads blocked on a waitqueue.
3819 * @mode: which threads
3820 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3822 * The sync wakeup differs that the waker knows that it will schedule
3823 * away soon, so while the target thread will be woken up, it will not
3824 * be migrated to another CPU - ie. the two threads are 'synchronized'
3825 * with each other. This can prevent needless bouncing between CPUs.
3827 * On UP it can prevent extra preemption.
3830 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3832 unsigned long flags;
3838 if (unlikely(!nr_exclusive))
3841 spin_lock_irqsave(&q->lock, flags);
3842 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3843 spin_unlock_irqrestore(&q->lock, flags);
3845 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3847 void complete(struct completion *x)
3849 unsigned long flags;
3851 spin_lock_irqsave(&x->wait.lock, flags);
3853 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3855 spin_unlock_irqrestore(&x->wait.lock, flags);
3857 EXPORT_SYMBOL(complete);
3859 void complete_all(struct completion *x)
3861 unsigned long flags;
3863 spin_lock_irqsave(&x->wait.lock, flags);
3864 x->done += UINT_MAX/2;
3865 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3867 spin_unlock_irqrestore(&x->wait.lock, flags);
3869 EXPORT_SYMBOL(complete_all);
3871 static inline long __sched
3872 do_wait_for_common(struct completion *x, long timeout, int state)
3875 DECLARE_WAITQUEUE(wait, current);
3877 wait.flags |= WQ_FLAG_EXCLUSIVE;
3878 __add_wait_queue_tail(&x->wait, &wait);
3880 if (state == TASK_INTERRUPTIBLE &&
3881 signal_pending(current)) {
3882 __remove_wait_queue(&x->wait, &wait);
3883 return -ERESTARTSYS;
3885 __set_current_state(state);
3886 spin_unlock_irq(&x->wait.lock);
3887 timeout = schedule_timeout(timeout);
3888 spin_lock_irq(&x->wait.lock);
3890 __remove_wait_queue(&x->wait, &wait);
3894 __remove_wait_queue(&x->wait, &wait);
3901 wait_for_common(struct completion *x, long timeout, int state)
3905 spin_lock_irq(&x->wait.lock);
3906 timeout = do_wait_for_common(x, timeout, state);
3907 spin_unlock_irq(&x->wait.lock);
3911 void __sched wait_for_completion(struct completion *x)
3913 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3915 EXPORT_SYMBOL(wait_for_completion);
3917 unsigned long __sched
3918 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3920 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3922 EXPORT_SYMBOL(wait_for_completion_timeout);
3924 int __sched wait_for_completion_interruptible(struct completion *x)
3926 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3927 if (t == -ERESTARTSYS)
3931 EXPORT_SYMBOL(wait_for_completion_interruptible);
3933 unsigned long __sched
3934 wait_for_completion_interruptible_timeout(struct completion *x,
3935 unsigned long timeout)
3937 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3939 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3942 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3944 unsigned long flags;
3947 init_waitqueue_entry(&wait, current);
3949 __set_current_state(state);
3951 spin_lock_irqsave(&q->lock, flags);
3952 __add_wait_queue(q, &wait);
3953 spin_unlock(&q->lock);
3954 timeout = schedule_timeout(timeout);
3955 spin_lock_irq(&q->lock);
3956 __remove_wait_queue(q, &wait);
3957 spin_unlock_irqrestore(&q->lock, flags);
3962 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3964 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3966 EXPORT_SYMBOL(interruptible_sleep_on);
3969 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3971 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3973 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3975 void __sched sleep_on(wait_queue_head_t *q)
3977 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3979 EXPORT_SYMBOL(sleep_on);
3981 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3983 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3985 EXPORT_SYMBOL(sleep_on_timeout);
3987 #ifdef CONFIG_RT_MUTEXES
3990 * rt_mutex_setprio - set the current priority of a task
3992 * @prio: prio value (kernel-internal form)
3994 * This function changes the 'effective' priority of a task. It does
3995 * not touch ->normal_prio like __setscheduler().
3997 * Used by the rt_mutex code to implement priority inheritance logic.
3999 void rt_mutex_setprio(struct task_struct *p, int prio)
4001 unsigned long flags;
4002 int oldprio, on_rq, running;
4005 BUG_ON(prio < 0 || prio > MAX_PRIO);
4007 rq = task_rq_lock(p, &flags);
4008 update_rq_clock(rq);
4011 on_rq = p->se.on_rq;
4012 running = task_running(rq, p);
4014 dequeue_task(rq, p, 0);
4016 p->sched_class->put_prev_task(rq, p);
4020 p->sched_class = &rt_sched_class;
4022 p->sched_class = &fair_sched_class;
4028 p->sched_class->set_curr_task(rq);
4029 enqueue_task(rq, p, 0);
4031 * Reschedule if we are currently running on this runqueue and
4032 * our priority decreased, or if we are not currently running on
4033 * this runqueue and our priority is higher than the current's
4036 if (p->prio > oldprio)
4037 resched_task(rq->curr);
4039 check_preempt_curr(rq, p);
4042 task_rq_unlock(rq, &flags);
4047 void set_user_nice(struct task_struct *p, long nice)
4049 int old_prio, delta, on_rq;
4050 unsigned long flags;
4053 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4056 * We have to be careful, if called from sys_setpriority(),
4057 * the task might be in the middle of scheduling on another CPU.
4059 rq = task_rq_lock(p, &flags);
4060 update_rq_clock(rq);
4062 * The RT priorities are set via sched_setscheduler(), but we still
4063 * allow the 'normal' nice value to be set - but as expected
4064 * it wont have any effect on scheduling until the task is
4065 * SCHED_FIFO/SCHED_RR:
4067 if (task_has_rt_policy(p)) {
4068 p->static_prio = NICE_TO_PRIO(nice);
4071 on_rq = p->se.on_rq;
4073 dequeue_task(rq, p, 0);
4077 p->static_prio = NICE_TO_PRIO(nice);
4080 p->prio = effective_prio(p);
4081 delta = p->prio - old_prio;
4084 enqueue_task(rq, p, 0);
4087 * If the task increased its priority or is running and
4088 * lowered its priority, then reschedule its CPU:
4090 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4091 resched_task(rq->curr);
4094 task_rq_unlock(rq, &flags);
4096 EXPORT_SYMBOL(set_user_nice);
4099 * can_nice - check if a task can reduce its nice value
4103 int can_nice(const struct task_struct *p, const int nice)
4105 /* convert nice value [19,-20] to rlimit style value [1,40] */
4106 int nice_rlim = 20 - nice;
4108 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4109 capable(CAP_SYS_NICE));
4112 #ifdef __ARCH_WANT_SYS_NICE
4115 * sys_nice - change the priority of the current process.
4116 * @increment: priority increment
4118 * sys_setpriority is a more generic, but much slower function that
4119 * does similar things.
4121 asmlinkage long sys_nice(int increment)
4126 * Setpriority might change our priority at the same moment.
4127 * We don't have to worry. Conceptually one call occurs first
4128 * and we have a single winner.
4130 if (increment < -40)
4135 nice = PRIO_TO_NICE(current->static_prio) + increment;
4141 if (increment < 0 && !can_nice(current, nice))
4144 retval = security_task_setnice(current, nice);
4148 set_user_nice(current, nice);
4155 * task_prio - return the priority value of a given task.
4156 * @p: the task in question.
4158 * This is the priority value as seen by users in /proc.
4159 * RT tasks are offset by -200. Normal tasks are centered
4160 * around 0, value goes from -16 to +15.
4162 int task_prio(const struct task_struct *p)
4164 return p->prio - MAX_RT_PRIO;
4168 * task_nice - return the nice value of a given task.
4169 * @p: the task in question.
4171 int task_nice(const struct task_struct *p)
4173 return TASK_NICE(p);
4175 EXPORT_SYMBOL_GPL(task_nice);
4178 * idle_cpu - is a given cpu idle currently?
4179 * @cpu: the processor in question.
4181 int idle_cpu(int cpu)
4183 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4187 * idle_task - return the idle task for a given cpu.
4188 * @cpu: the processor in question.
4190 struct task_struct *idle_task(int cpu)
4192 return cpu_rq(cpu)->idle;
4196 * find_process_by_pid - find a process with a matching PID value.
4197 * @pid: the pid in question.
4199 static struct task_struct *find_process_by_pid(pid_t pid)
4201 return pid ? find_task_by_vpid(pid) : current;
4204 /* Actually do priority change: must hold rq lock. */
4206 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4208 BUG_ON(p->se.on_rq);
4211 switch (p->policy) {
4215 p->sched_class = &fair_sched_class;
4219 p->sched_class = &rt_sched_class;
4223 p->rt_priority = prio;
4224 p->normal_prio = normal_prio(p);
4225 /* we are holding p->pi_lock already */
4226 p->prio = rt_mutex_getprio(p);
4231 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4232 * @p: the task in question.
4233 * @policy: new policy.
4234 * @param: structure containing the new RT priority.
4236 * NOTE that the task may be already dead.
4238 int sched_setscheduler(struct task_struct *p, int policy,
4239 struct sched_param *param)
4241 int retval, oldprio, oldpolicy = -1, on_rq, running;
4242 unsigned long flags;
4245 /* may grab non-irq protected spin_locks */
4246 BUG_ON(in_interrupt());
4248 /* double check policy once rq lock held */
4250 policy = oldpolicy = p->policy;
4251 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4252 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4253 policy != SCHED_IDLE)
4256 * Valid priorities for SCHED_FIFO and SCHED_RR are
4257 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4258 * SCHED_BATCH and SCHED_IDLE is 0.
4260 if (param->sched_priority < 0 ||
4261 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4262 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4264 if (rt_policy(policy) != (param->sched_priority != 0))
4268 * Allow unprivileged RT tasks to decrease priority:
4270 if (!capable(CAP_SYS_NICE)) {
4271 if (rt_policy(policy)) {
4272 unsigned long rlim_rtprio;
4274 if (!lock_task_sighand(p, &flags))
4276 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4277 unlock_task_sighand(p, &flags);
4279 /* can't set/change the rt policy */
4280 if (policy != p->policy && !rlim_rtprio)
4283 /* can't increase priority */
4284 if (param->sched_priority > p->rt_priority &&
4285 param->sched_priority > rlim_rtprio)
4289 * Like positive nice levels, dont allow tasks to
4290 * move out of SCHED_IDLE either:
4292 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4295 /* can't change other user's priorities */
4296 if ((current->euid != p->euid) &&
4297 (current->euid != p->uid))
4301 retval = security_task_setscheduler(p, policy, param);
4305 * make sure no PI-waiters arrive (or leave) while we are
4306 * changing the priority of the task:
4308 spin_lock_irqsave(&p->pi_lock, flags);
4310 * To be able to change p->policy safely, the apropriate
4311 * runqueue lock must be held.
4313 rq = __task_rq_lock(p);
4314 /* recheck policy now with rq lock held */
4315 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4316 policy = oldpolicy = -1;
4317 __task_rq_unlock(rq);
4318 spin_unlock_irqrestore(&p->pi_lock, flags);
4321 update_rq_clock(rq);
4322 on_rq = p->se.on_rq;
4323 running = task_running(rq, p);
4325 deactivate_task(rq, p, 0);
4327 p->sched_class->put_prev_task(rq, p);
4331 __setscheduler(rq, p, policy, param->sched_priority);
4335 p->sched_class->set_curr_task(rq);
4336 activate_task(rq, p, 0);
4338 * Reschedule if we are currently running on this runqueue and
4339 * our priority decreased, or if we are not currently running on
4340 * this runqueue and our priority is higher than the current's
4343 if (p->prio > oldprio)
4344 resched_task(rq->curr);
4346 check_preempt_curr(rq, p);
4349 __task_rq_unlock(rq);
4350 spin_unlock_irqrestore(&p->pi_lock, flags);
4352 rt_mutex_adjust_pi(p);
4356 EXPORT_SYMBOL_GPL(sched_setscheduler);
4359 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4361 struct sched_param lparam;
4362 struct task_struct *p;
4365 if (!param || pid < 0)
4367 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4372 p = find_process_by_pid(pid);
4374 retval = sched_setscheduler(p, policy, &lparam);
4381 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4382 * @pid: the pid in question.
4383 * @policy: new policy.
4384 * @param: structure containing the new RT priority.
4386 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4387 struct sched_param __user *param)
4389 /* negative values for policy are not valid */
4393 return do_sched_setscheduler(pid, policy, param);
4397 * sys_sched_setparam - set/change the RT priority of a thread
4398 * @pid: the pid in question.
4399 * @param: structure containing the new RT priority.
4401 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4403 return do_sched_setscheduler(pid, -1, param);
4407 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4408 * @pid: the pid in question.
4410 asmlinkage long sys_sched_getscheduler(pid_t pid)
4412 struct task_struct *p;
4419 read_lock(&tasklist_lock);
4420 p = find_process_by_pid(pid);
4422 retval = security_task_getscheduler(p);
4426 read_unlock(&tasklist_lock);
4431 * sys_sched_getscheduler - get the RT priority of a thread
4432 * @pid: the pid in question.
4433 * @param: structure containing the RT priority.
4435 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4437 struct sched_param lp;
4438 struct task_struct *p;
4441 if (!param || pid < 0)
4444 read_lock(&tasklist_lock);
4445 p = find_process_by_pid(pid);
4450 retval = security_task_getscheduler(p);
4454 lp.sched_priority = p->rt_priority;
4455 read_unlock(&tasklist_lock);
4458 * This one might sleep, we cannot do it with a spinlock held ...
4460 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4465 read_unlock(&tasklist_lock);
4469 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4471 cpumask_t cpus_allowed;
4472 struct task_struct *p;
4475 mutex_lock(&sched_hotcpu_mutex);
4476 read_lock(&tasklist_lock);
4478 p = find_process_by_pid(pid);
4480 read_unlock(&tasklist_lock);
4481 mutex_unlock(&sched_hotcpu_mutex);
4486 * It is not safe to call set_cpus_allowed with the
4487 * tasklist_lock held. We will bump the task_struct's
4488 * usage count and then drop tasklist_lock.
4491 read_unlock(&tasklist_lock);
4494 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4495 !capable(CAP_SYS_NICE))
4498 retval = security_task_setscheduler(p, 0, NULL);
4502 cpus_allowed = cpuset_cpus_allowed(p);
4503 cpus_and(new_mask, new_mask, cpus_allowed);
4505 retval = set_cpus_allowed(p, new_mask);
4508 cpus_allowed = cpuset_cpus_allowed(p);
4509 if (!cpus_subset(new_mask, cpus_allowed)) {
4511 * We must have raced with a concurrent cpuset
4512 * update. Just reset the cpus_allowed to the
4513 * cpuset's cpus_allowed
4515 new_mask = cpus_allowed;
4521 mutex_unlock(&sched_hotcpu_mutex);
4525 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4526 cpumask_t *new_mask)
4528 if (len < sizeof(cpumask_t)) {
4529 memset(new_mask, 0, sizeof(cpumask_t));
4530 } else if (len > sizeof(cpumask_t)) {
4531 len = sizeof(cpumask_t);
4533 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4537 * sys_sched_setaffinity - set the cpu affinity of a process
4538 * @pid: pid of the process
4539 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4540 * @user_mask_ptr: user-space pointer to the new cpu mask
4542 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4543 unsigned long __user *user_mask_ptr)
4548 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4552 return sched_setaffinity(pid, new_mask);
4556 * Represents all cpu's present in the system
4557 * In systems capable of hotplug, this map could dynamically grow
4558 * as new cpu's are detected in the system via any platform specific
4559 * method, such as ACPI for e.g.
4562 cpumask_t cpu_present_map __read_mostly;
4563 EXPORT_SYMBOL(cpu_present_map);
4566 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4567 EXPORT_SYMBOL(cpu_online_map);
4569 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4570 EXPORT_SYMBOL(cpu_possible_map);
4573 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4575 struct task_struct *p;
4578 mutex_lock(&sched_hotcpu_mutex);
4579 read_lock(&tasklist_lock);
4582 p = find_process_by_pid(pid);
4586 retval = security_task_getscheduler(p);
4590 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4593 read_unlock(&tasklist_lock);
4594 mutex_unlock(&sched_hotcpu_mutex);
4600 * sys_sched_getaffinity - get the cpu affinity of a process
4601 * @pid: pid of the process
4602 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4603 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4605 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4606 unsigned long __user *user_mask_ptr)
4611 if (len < sizeof(cpumask_t))
4614 ret = sched_getaffinity(pid, &mask);
4618 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4621 return sizeof(cpumask_t);
4625 * sys_sched_yield - yield the current processor to other threads.
4627 * This function yields the current CPU to other tasks. If there are no
4628 * other threads running on this CPU then this function will return.
4630 asmlinkage long sys_sched_yield(void)
4632 struct rq *rq = this_rq_lock();
4634 schedstat_inc(rq, yld_count);
4635 current->sched_class->yield_task(rq);
4638 * Since we are going to call schedule() anyway, there's
4639 * no need to preempt or enable interrupts:
4641 __release(rq->lock);
4642 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4643 _raw_spin_unlock(&rq->lock);
4644 preempt_enable_no_resched();
4651 static void __cond_resched(void)
4653 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4654 __might_sleep(__FILE__, __LINE__);
4657 * The BKS might be reacquired before we have dropped
4658 * PREEMPT_ACTIVE, which could trigger a second
4659 * cond_resched() call.
4662 add_preempt_count(PREEMPT_ACTIVE);
4664 sub_preempt_count(PREEMPT_ACTIVE);
4665 } while (need_resched());
4668 int __sched cond_resched(void)
4670 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4671 system_state == SYSTEM_RUNNING) {
4677 EXPORT_SYMBOL(cond_resched);
4680 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4681 * call schedule, and on return reacquire the lock.
4683 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4684 * operations here to prevent schedule() from being called twice (once via
4685 * spin_unlock(), once by hand).
4687 int cond_resched_lock(spinlock_t *lock)
4691 if (need_lockbreak(lock)) {
4697 if (need_resched() && system_state == SYSTEM_RUNNING) {
4698 spin_release(&lock->dep_map, 1, _THIS_IP_);
4699 _raw_spin_unlock(lock);
4700 preempt_enable_no_resched();
4707 EXPORT_SYMBOL(cond_resched_lock);
4709 int __sched cond_resched_softirq(void)
4711 BUG_ON(!in_softirq());
4713 if (need_resched() && system_state == SYSTEM_RUNNING) {
4721 EXPORT_SYMBOL(cond_resched_softirq);
4724 * yield - yield the current processor to other threads.
4726 * This is a shortcut for kernel-space yielding - it marks the
4727 * thread runnable and calls sys_sched_yield().
4729 void __sched yield(void)
4731 set_current_state(TASK_RUNNING);
4734 EXPORT_SYMBOL(yield);
4737 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4738 * that process accounting knows that this is a task in IO wait state.
4740 * But don't do that if it is a deliberate, throttling IO wait (this task
4741 * has set its backing_dev_info: the queue against which it should throttle)
4743 void __sched io_schedule(void)
4745 struct rq *rq = &__raw_get_cpu_var(runqueues);
4747 delayacct_blkio_start();
4748 atomic_inc(&rq->nr_iowait);
4750 atomic_dec(&rq->nr_iowait);
4751 delayacct_blkio_end();
4753 EXPORT_SYMBOL(io_schedule);
4755 long __sched io_schedule_timeout(long timeout)
4757 struct rq *rq = &__raw_get_cpu_var(runqueues);
4760 delayacct_blkio_start();
4761 atomic_inc(&rq->nr_iowait);
4762 ret = schedule_timeout(timeout);
4763 atomic_dec(&rq->nr_iowait);
4764 delayacct_blkio_end();
4769 * sys_sched_get_priority_max - return maximum RT priority.
4770 * @policy: scheduling class.
4772 * this syscall returns the maximum rt_priority that can be used
4773 * by a given scheduling class.
4775 asmlinkage long sys_sched_get_priority_max(int policy)
4782 ret = MAX_USER_RT_PRIO-1;
4794 * sys_sched_get_priority_min - return minimum RT priority.
4795 * @policy: scheduling class.
4797 * this syscall returns the minimum rt_priority that can be used
4798 * by a given scheduling class.
4800 asmlinkage long sys_sched_get_priority_min(int policy)
4818 * sys_sched_rr_get_interval - return the default timeslice of a process.
4819 * @pid: pid of the process.
4820 * @interval: userspace pointer to the timeslice value.
4822 * this syscall writes the default timeslice value of a given process
4823 * into the user-space timespec buffer. A value of '0' means infinity.
4826 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4828 struct task_struct *p;
4829 unsigned int time_slice;
4837 read_lock(&tasklist_lock);
4838 p = find_process_by_pid(pid);
4842 retval = security_task_getscheduler(p);
4846 if (p->policy == SCHED_FIFO)
4848 else if (p->policy == SCHED_RR)
4849 time_slice = DEF_TIMESLICE;
4851 struct sched_entity *se = &p->se;
4852 unsigned long flags;
4855 rq = task_rq_lock(p, &flags);
4856 time_slice = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
4857 task_rq_unlock(rq, &flags);
4859 read_unlock(&tasklist_lock);
4860 jiffies_to_timespec(time_slice, &t);
4861 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4865 read_unlock(&tasklist_lock);
4869 static const char stat_nam[] = "RSDTtZX";
4871 static void show_task(struct task_struct *p)
4873 unsigned long free = 0;
4876 state = p->state ? __ffs(p->state) + 1 : 0;
4877 printk(KERN_INFO "%-13.13s %c", p->comm,
4878 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4879 #if BITS_PER_LONG == 32
4880 if (state == TASK_RUNNING)
4881 printk(KERN_CONT " running ");
4883 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4885 if (state == TASK_RUNNING)
4886 printk(KERN_CONT " running task ");
4888 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4890 #ifdef CONFIG_DEBUG_STACK_USAGE
4892 unsigned long *n = end_of_stack(p);
4895 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4898 printk(KERN_CONT "%5lu %5d %6d\n", free,
4899 task_pid_nr(p), task_pid_nr(p->parent));
4901 if (state != TASK_RUNNING)
4902 show_stack(p, NULL);
4905 void show_state_filter(unsigned long state_filter)
4907 struct task_struct *g, *p;
4909 #if BITS_PER_LONG == 32
4911 " task PC stack pid father\n");
4914 " task PC stack pid father\n");
4916 read_lock(&tasklist_lock);
4917 do_each_thread(g, p) {
4919 * reset the NMI-timeout, listing all files on a slow
4920 * console might take alot of time:
4922 touch_nmi_watchdog();
4923 if (!state_filter || (p->state & state_filter))
4925 } while_each_thread(g, p);
4927 touch_all_softlockup_watchdogs();
4929 #ifdef CONFIG_SCHED_DEBUG
4930 sysrq_sched_debug_show();
4932 read_unlock(&tasklist_lock);
4934 * Only show locks if all tasks are dumped:
4936 if (state_filter == -1)
4937 debug_show_all_locks();
4940 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4942 idle->sched_class = &idle_sched_class;
4946 * init_idle - set up an idle thread for a given CPU
4947 * @idle: task in question
4948 * @cpu: cpu the idle task belongs to
4950 * NOTE: this function does not set the idle thread's NEED_RESCHED
4951 * flag, to make booting more robust.
4953 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4955 struct rq *rq = cpu_rq(cpu);
4956 unsigned long flags;
4959 idle->se.exec_start = sched_clock();
4961 idle->prio = idle->normal_prio = MAX_PRIO;
4962 idle->cpus_allowed = cpumask_of_cpu(cpu);
4963 __set_task_cpu(idle, cpu);
4965 spin_lock_irqsave(&rq->lock, flags);
4966 rq->curr = rq->idle = idle;
4967 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4970 spin_unlock_irqrestore(&rq->lock, flags);
4972 /* Set the preempt count _outside_ the spinlocks! */
4973 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4974 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4976 task_thread_info(idle)->preempt_count = 0;
4979 * The idle tasks have their own, simple scheduling class:
4981 idle->sched_class = &idle_sched_class;
4985 * In a system that switches off the HZ timer nohz_cpu_mask
4986 * indicates which cpus entered this state. This is used
4987 * in the rcu update to wait only for active cpus. For system
4988 * which do not switch off the HZ timer nohz_cpu_mask should
4989 * always be CPU_MASK_NONE.
4991 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4994 * Increase the granularity value when there are more CPUs,
4995 * because with more CPUs the 'effective latency' as visible
4996 * to users decreases. But the relationship is not linear,
4997 * so pick a second-best guess by going with the log2 of the
5000 * This idea comes from the SD scheduler of Con Kolivas:
5002 static inline void sched_init_granularity(void)
5004 unsigned int factor = 1 + ilog2(num_online_cpus());
5005 const unsigned long limit = 200000000;
5007 sysctl_sched_min_granularity *= factor;
5008 if (sysctl_sched_min_granularity > limit)
5009 sysctl_sched_min_granularity = limit;
5011 sysctl_sched_latency *= factor;
5012 if (sysctl_sched_latency > limit)
5013 sysctl_sched_latency = limit;
5015 sysctl_sched_wakeup_granularity *= factor;
5016 sysctl_sched_batch_wakeup_granularity *= factor;
5021 * This is how migration works:
5023 * 1) we queue a struct migration_req structure in the source CPU's
5024 * runqueue and wake up that CPU's migration thread.
5025 * 2) we down() the locked semaphore => thread blocks.
5026 * 3) migration thread wakes up (implicitly it forces the migrated
5027 * thread off the CPU)
5028 * 4) it gets the migration request and checks whether the migrated
5029 * task is still in the wrong runqueue.
5030 * 5) if it's in the wrong runqueue then the migration thread removes
5031 * it and puts it into the right queue.
5032 * 6) migration thread up()s the semaphore.
5033 * 7) we wake up and the migration is done.
5037 * Change a given task's CPU affinity. Migrate the thread to a
5038 * proper CPU and schedule it away if the CPU it's executing on
5039 * is removed from the allowed bitmask.
5041 * NOTE: the caller must have a valid reference to the task, the
5042 * task must not exit() & deallocate itself prematurely. The
5043 * call is not atomic; no spinlocks may be held.
5045 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5047 struct migration_req req;
5048 unsigned long flags;
5052 rq = task_rq_lock(p, &flags);
5053 if (!cpus_intersects(new_mask, cpu_online_map)) {
5058 p->cpus_allowed = new_mask;
5059 /* Can the task run on the task's current CPU? If so, we're done */
5060 if (cpu_isset(task_cpu(p), new_mask))
5063 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5064 /* Need help from migration thread: drop lock and wait. */
5065 task_rq_unlock(rq, &flags);
5066 wake_up_process(rq->migration_thread);
5067 wait_for_completion(&req.done);
5068 tlb_migrate_finish(p->mm);
5072 task_rq_unlock(rq, &flags);
5076 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5079 * Move (not current) task off this cpu, onto dest cpu. We're doing
5080 * this because either it can't run here any more (set_cpus_allowed()
5081 * away from this CPU, or CPU going down), or because we're
5082 * attempting to rebalance this task on exec (sched_exec).
5084 * So we race with normal scheduler movements, but that's OK, as long
5085 * as the task is no longer on this CPU.
5087 * Returns non-zero if task was successfully migrated.
5089 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5091 struct rq *rq_dest, *rq_src;
5094 if (unlikely(cpu_is_offline(dest_cpu)))
5097 rq_src = cpu_rq(src_cpu);
5098 rq_dest = cpu_rq(dest_cpu);
5100 double_rq_lock(rq_src, rq_dest);
5101 /* Already moved. */
5102 if (task_cpu(p) != src_cpu)
5104 /* Affinity changed (again). */
5105 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5108 on_rq = p->se.on_rq;
5110 deactivate_task(rq_src, p, 0);
5112 set_task_cpu(p, dest_cpu);
5114 activate_task(rq_dest, p, 0);
5115 check_preempt_curr(rq_dest, p);
5119 double_rq_unlock(rq_src, rq_dest);
5124 * migration_thread - this is a highprio system thread that performs
5125 * thread migration by bumping thread off CPU then 'pushing' onto
5128 static int migration_thread(void *data)
5130 int cpu = (long)data;
5134 BUG_ON(rq->migration_thread != current);
5136 set_current_state(TASK_INTERRUPTIBLE);
5137 while (!kthread_should_stop()) {
5138 struct migration_req *req;
5139 struct list_head *head;
5141 spin_lock_irq(&rq->lock);
5143 if (cpu_is_offline(cpu)) {
5144 spin_unlock_irq(&rq->lock);
5148 if (rq->active_balance) {
5149 active_load_balance(rq, cpu);
5150 rq->active_balance = 0;
5153 head = &rq->migration_queue;
5155 if (list_empty(head)) {
5156 spin_unlock_irq(&rq->lock);
5158 set_current_state(TASK_INTERRUPTIBLE);
5161 req = list_entry(head->next, struct migration_req, list);
5162 list_del_init(head->next);
5164 spin_unlock(&rq->lock);
5165 __migrate_task(req->task, cpu, req->dest_cpu);
5168 complete(&req->done);
5170 __set_current_state(TASK_RUNNING);
5174 /* Wait for kthread_stop */
5175 set_current_state(TASK_INTERRUPTIBLE);
5176 while (!kthread_should_stop()) {
5178 set_current_state(TASK_INTERRUPTIBLE);
5180 __set_current_state(TASK_RUNNING);
5184 #ifdef CONFIG_HOTPLUG_CPU
5186 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5190 local_irq_disable();
5191 ret = __migrate_task(p, src_cpu, dest_cpu);
5197 * Figure out where task on dead CPU should go, use force if necessary.
5198 * NOTE: interrupts should be disabled by the caller
5200 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5202 unsigned long flags;
5209 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5210 cpus_and(mask, mask, p->cpus_allowed);
5211 dest_cpu = any_online_cpu(mask);
5213 /* On any allowed CPU? */
5214 if (dest_cpu == NR_CPUS)
5215 dest_cpu = any_online_cpu(p->cpus_allowed);
5217 /* No more Mr. Nice Guy. */
5218 if (dest_cpu == NR_CPUS) {
5219 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5221 * Try to stay on the same cpuset, where the
5222 * current cpuset may be a subset of all cpus.
5223 * The cpuset_cpus_allowed_locked() variant of
5224 * cpuset_cpus_allowed() will not block. It must be
5225 * called within calls to cpuset_lock/cpuset_unlock.
5227 rq = task_rq_lock(p, &flags);
5228 p->cpus_allowed = cpus_allowed;
5229 dest_cpu = any_online_cpu(p->cpus_allowed);
5230 task_rq_unlock(rq, &flags);
5233 * Don't tell them about moving exiting tasks or
5234 * kernel threads (both mm NULL), since they never
5237 if (p->mm && printk_ratelimit())
5238 printk(KERN_INFO "process %d (%s) no "
5239 "longer affine to cpu%d\n",
5240 task_pid_nr(p), p->comm, dead_cpu);
5242 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5246 * While a dead CPU has no uninterruptible tasks queued at this point,
5247 * it might still have a nonzero ->nr_uninterruptible counter, because
5248 * for performance reasons the counter is not stricly tracking tasks to
5249 * their home CPUs. So we just add the counter to another CPU's counter,
5250 * to keep the global sum constant after CPU-down:
5252 static void migrate_nr_uninterruptible(struct rq *rq_src)
5254 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5255 unsigned long flags;
5257 local_irq_save(flags);
5258 double_rq_lock(rq_src, rq_dest);
5259 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5260 rq_src->nr_uninterruptible = 0;
5261 double_rq_unlock(rq_src, rq_dest);
5262 local_irq_restore(flags);
5265 /* Run through task list and migrate tasks from the dead cpu. */
5266 static void migrate_live_tasks(int src_cpu)
5268 struct task_struct *p, *t;
5270 read_lock(&tasklist_lock);
5272 do_each_thread(t, p) {
5276 if (task_cpu(p) == src_cpu)
5277 move_task_off_dead_cpu(src_cpu, p);
5278 } while_each_thread(t, p);
5280 read_unlock(&tasklist_lock);
5284 * activate_idle_task - move idle task to the _front_ of runqueue.
5286 static void activate_idle_task(struct task_struct *p, struct rq *rq)
5288 update_rq_clock(rq);
5290 if (p->state == TASK_UNINTERRUPTIBLE)
5291 rq->nr_uninterruptible--;
5293 enqueue_task(rq, p, 0);
5294 inc_nr_running(p, rq);
5298 * Schedules idle task to be the next runnable task on current CPU.
5299 * It does so by boosting its priority to highest possible and adding it to
5300 * the _front_ of the runqueue. Used by CPU offline code.
5302 void sched_idle_next(void)
5304 int this_cpu = smp_processor_id();
5305 struct rq *rq = cpu_rq(this_cpu);
5306 struct task_struct *p = rq->idle;
5307 unsigned long flags;
5309 /* cpu has to be offline */
5310 BUG_ON(cpu_online(this_cpu));
5313 * Strictly not necessary since rest of the CPUs are stopped by now
5314 * and interrupts disabled on the current cpu.
5316 spin_lock_irqsave(&rq->lock, flags);
5318 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5320 /* Add idle task to the _front_ of its priority queue: */
5321 activate_idle_task(p, rq);
5323 spin_unlock_irqrestore(&rq->lock, flags);
5327 * Ensures that the idle task is using init_mm right before its cpu goes
5330 void idle_task_exit(void)
5332 struct mm_struct *mm = current->active_mm;
5334 BUG_ON(cpu_online(smp_processor_id()));
5337 switch_mm(mm, &init_mm, current);
5341 /* called under rq->lock with disabled interrupts */
5342 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5344 struct rq *rq = cpu_rq(dead_cpu);
5346 /* Must be exiting, otherwise would be on tasklist. */
5347 BUG_ON(!p->exit_state);
5349 /* Cannot have done final schedule yet: would have vanished. */
5350 BUG_ON(p->state == TASK_DEAD);
5355 * Drop lock around migration; if someone else moves it,
5356 * that's OK. No task can be added to this CPU, so iteration is
5359 spin_unlock_irq(&rq->lock);
5360 move_task_off_dead_cpu(dead_cpu, p);
5361 spin_lock_irq(&rq->lock);
5366 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5367 static void migrate_dead_tasks(unsigned int dead_cpu)
5369 struct rq *rq = cpu_rq(dead_cpu);
5370 struct task_struct *next;
5373 if (!rq->nr_running)
5375 update_rq_clock(rq);
5376 next = pick_next_task(rq, rq->curr);
5379 migrate_dead(dead_cpu, next);
5383 #endif /* CONFIG_HOTPLUG_CPU */
5385 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5387 static struct ctl_table sd_ctl_dir[] = {
5389 .procname = "sched_domain",
5395 static struct ctl_table sd_ctl_root[] = {
5397 .ctl_name = CTL_KERN,
5398 .procname = "kernel",
5400 .child = sd_ctl_dir,
5405 static struct ctl_table *sd_alloc_ctl_entry(int n)
5407 struct ctl_table *entry =
5408 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5413 static void sd_free_ctl_entry(struct ctl_table **tablep)
5415 struct ctl_table *entry;
5418 * In the intermediate directories, both the child directory and
5419 * procname are dynamically allocated and could fail but the mode
5420 * will always be set. In the lowest directory the names are
5421 * static strings and all have proc handlers.
5423 for (entry = *tablep; entry->mode; entry++) {
5425 sd_free_ctl_entry(&entry->child);
5426 if (entry->proc_handler == NULL)
5427 kfree(entry->procname);
5435 set_table_entry(struct ctl_table *entry,
5436 const char *procname, void *data, int maxlen,
5437 mode_t mode, proc_handler *proc_handler)
5439 entry->procname = procname;
5441 entry->maxlen = maxlen;
5443 entry->proc_handler = proc_handler;
5446 static struct ctl_table *
5447 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5449 struct ctl_table *table = sd_alloc_ctl_entry(12);
5454 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5455 sizeof(long), 0644, proc_doulongvec_minmax);
5456 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5457 sizeof(long), 0644, proc_doulongvec_minmax);
5458 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5459 sizeof(int), 0644, proc_dointvec_minmax);
5460 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5461 sizeof(int), 0644, proc_dointvec_minmax);
5462 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5463 sizeof(int), 0644, proc_dointvec_minmax);
5464 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5465 sizeof(int), 0644, proc_dointvec_minmax);
5466 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5467 sizeof(int), 0644, proc_dointvec_minmax);
5468 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5471 sizeof(int), 0644, proc_dointvec_minmax);
5472 set_table_entry(&table[9], "cache_nice_tries",
5473 &sd->cache_nice_tries,
5474 sizeof(int), 0644, proc_dointvec_minmax);
5475 set_table_entry(&table[10], "flags", &sd->flags,
5476 sizeof(int), 0644, proc_dointvec_minmax);
5477 /* &table[11] is terminator */
5482 static ctl_table * sd_alloc_ctl_cpu_table(int cpu)
5484 struct ctl_table *entry, *table;
5485 struct sched_domain *sd;
5486 int domain_num = 0, i;
5489 for_each_domain(cpu, sd)
5491 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5496 for_each_domain(cpu, sd) {
5497 snprintf(buf, 32, "domain%d", i);
5498 entry->procname = kstrdup(buf, GFP_KERNEL);
5500 entry->child = sd_alloc_ctl_domain_table(sd);
5507 static struct ctl_table_header *sd_sysctl_header;
5508 static void register_sched_domain_sysctl(void)
5510 int i, cpu_num = num_online_cpus();
5511 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5514 WARN_ON(sd_ctl_dir[0].child);
5515 sd_ctl_dir[0].child = entry;
5520 for_each_online_cpu(i) {
5521 snprintf(buf, 32, "cpu%d", i);
5522 entry->procname = kstrdup(buf, GFP_KERNEL);
5524 entry->child = sd_alloc_ctl_cpu_table(i);
5528 WARN_ON(sd_sysctl_header);
5529 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5532 /* may be called multiple times per register */
5533 static void unregister_sched_domain_sysctl(void)
5535 if (sd_sysctl_header)
5536 unregister_sysctl_table(sd_sysctl_header);
5537 sd_sysctl_header = NULL;
5538 if (sd_ctl_dir[0].child)
5539 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5542 static void register_sched_domain_sysctl(void)
5545 static void unregister_sched_domain_sysctl(void)
5551 * migration_call - callback that gets triggered when a CPU is added.
5552 * Here we can start up the necessary migration thread for the new CPU.
5554 static int __cpuinit
5555 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5557 struct task_struct *p;
5558 int cpu = (long)hcpu;
5559 unsigned long flags;
5563 case CPU_LOCK_ACQUIRE:
5564 mutex_lock(&sched_hotcpu_mutex);
5567 case CPU_UP_PREPARE:
5568 case CPU_UP_PREPARE_FROZEN:
5569 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5572 kthread_bind(p, cpu);
5573 /* Must be high prio: stop_machine expects to yield to it. */
5574 rq = task_rq_lock(p, &flags);
5575 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5576 task_rq_unlock(rq, &flags);
5577 cpu_rq(cpu)->migration_thread = p;
5581 case CPU_ONLINE_FROZEN:
5582 /* Strictly unnecessary, as first user will wake it. */
5583 wake_up_process(cpu_rq(cpu)->migration_thread);
5586 #ifdef CONFIG_HOTPLUG_CPU
5587 case CPU_UP_CANCELED:
5588 case CPU_UP_CANCELED_FROZEN:
5589 if (!cpu_rq(cpu)->migration_thread)
5591 /* Unbind it from offline cpu so it can run. Fall thru. */
5592 kthread_bind(cpu_rq(cpu)->migration_thread,
5593 any_online_cpu(cpu_online_map));
5594 kthread_stop(cpu_rq(cpu)->migration_thread);
5595 cpu_rq(cpu)->migration_thread = NULL;
5599 case CPU_DEAD_FROZEN:
5600 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5601 migrate_live_tasks(cpu);
5603 kthread_stop(rq->migration_thread);
5604 rq->migration_thread = NULL;
5605 /* Idle task back to normal (off runqueue, low prio) */
5606 spin_lock_irq(&rq->lock);
5607 update_rq_clock(rq);
5608 deactivate_task(rq, rq->idle, 0);
5609 rq->idle->static_prio = MAX_PRIO;
5610 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5611 rq->idle->sched_class = &idle_sched_class;
5612 migrate_dead_tasks(cpu);
5613 spin_unlock_irq(&rq->lock);
5615 migrate_nr_uninterruptible(rq);
5616 BUG_ON(rq->nr_running != 0);
5618 /* No need to migrate the tasks: it was best-effort if
5619 * they didn't take sched_hotcpu_mutex. Just wake up
5620 * the requestors. */
5621 spin_lock_irq(&rq->lock);
5622 while (!list_empty(&rq->migration_queue)) {
5623 struct migration_req *req;
5625 req = list_entry(rq->migration_queue.next,
5626 struct migration_req, list);
5627 list_del_init(&req->list);
5628 complete(&req->done);
5630 spin_unlock_irq(&rq->lock);
5633 case CPU_LOCK_RELEASE:
5634 mutex_unlock(&sched_hotcpu_mutex);
5640 /* Register at highest priority so that task migration (migrate_all_tasks)
5641 * happens before everything else.
5643 static struct notifier_block __cpuinitdata migration_notifier = {
5644 .notifier_call = migration_call,
5648 int __init migration_init(void)
5650 void *cpu = (void *)(long)smp_processor_id();
5653 /* Start one for the boot CPU: */
5654 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5655 BUG_ON(err == NOTIFY_BAD);
5656 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5657 register_cpu_notifier(&migration_notifier);
5665 /* Number of possible processor ids */
5666 int nr_cpu_ids __read_mostly = NR_CPUS;
5667 EXPORT_SYMBOL(nr_cpu_ids);
5669 #ifdef CONFIG_SCHED_DEBUG
5671 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5673 struct sched_group *group = sd->groups;
5674 cpumask_t groupmask;
5677 cpumask_scnprintf(str, NR_CPUS, sd->span);
5678 cpus_clear(groupmask);
5680 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5682 if (!(sd->flags & SD_LOAD_BALANCE)) {
5683 printk("does not load-balance\n");
5685 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5690 printk(KERN_CONT "span %s\n", str);
5692 if (!cpu_isset(cpu, sd->span)) {
5693 printk(KERN_ERR "ERROR: domain->span does not contain "
5696 if (!cpu_isset(cpu, group->cpumask)) {
5697 printk(KERN_ERR "ERROR: domain->groups does not contain"
5701 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5705 printk(KERN_ERR "ERROR: group is NULL\n");
5709 if (!group->__cpu_power) {
5710 printk(KERN_CONT "\n");
5711 printk(KERN_ERR "ERROR: domain->cpu_power not "
5716 if (!cpus_weight(group->cpumask)) {
5717 printk(KERN_CONT "\n");
5718 printk(KERN_ERR "ERROR: empty group\n");
5722 if (cpus_intersects(groupmask, group->cpumask)) {
5723 printk(KERN_CONT "\n");
5724 printk(KERN_ERR "ERROR: repeated CPUs\n");
5728 cpus_or(groupmask, groupmask, group->cpumask);
5730 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5731 printk(KERN_CONT " %s", str);
5733 group = group->next;
5734 } while (group != sd->groups);
5735 printk(KERN_CONT "\n");
5737 if (!cpus_equal(sd->span, groupmask))
5738 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5740 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5741 printk(KERN_ERR "ERROR: parent span is not a superset "
5742 "of domain->span\n");
5746 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5751 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5755 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5758 if (sched_domain_debug_one(sd, cpu, level))
5767 # define sched_domain_debug(sd, cpu) do { } while (0)
5770 static int sd_degenerate(struct sched_domain *sd)
5772 if (cpus_weight(sd->span) == 1)
5775 /* Following flags need at least 2 groups */
5776 if (sd->flags & (SD_LOAD_BALANCE |
5777 SD_BALANCE_NEWIDLE |
5781 SD_SHARE_PKG_RESOURCES)) {
5782 if (sd->groups != sd->groups->next)
5786 /* Following flags don't use groups */
5787 if (sd->flags & (SD_WAKE_IDLE |
5796 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5798 unsigned long cflags = sd->flags, pflags = parent->flags;
5800 if (sd_degenerate(parent))
5803 if (!cpus_equal(sd->span, parent->span))
5806 /* Does parent contain flags not in child? */
5807 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5808 if (cflags & SD_WAKE_AFFINE)
5809 pflags &= ~SD_WAKE_BALANCE;
5810 /* Flags needing groups don't count if only 1 group in parent */
5811 if (parent->groups == parent->groups->next) {
5812 pflags &= ~(SD_LOAD_BALANCE |
5813 SD_BALANCE_NEWIDLE |
5817 SD_SHARE_PKG_RESOURCES);
5819 if (~cflags & pflags)
5826 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5827 * hold the hotplug lock.
5829 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5831 struct rq *rq = cpu_rq(cpu);
5832 struct sched_domain *tmp;
5834 /* Remove the sched domains which do not contribute to scheduling. */
5835 for (tmp = sd; tmp; tmp = tmp->parent) {
5836 struct sched_domain *parent = tmp->parent;
5839 if (sd_parent_degenerate(tmp, parent)) {
5840 tmp->parent = parent->parent;
5842 parent->parent->child = tmp;
5846 if (sd && sd_degenerate(sd)) {
5852 sched_domain_debug(sd, cpu);
5854 rcu_assign_pointer(rq->sd, sd);
5857 /* cpus with isolated domains */
5858 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5860 /* Setup the mask of cpus configured for isolated domains */
5861 static int __init isolated_cpu_setup(char *str)
5863 int ints[NR_CPUS], i;
5865 str = get_options(str, ARRAY_SIZE(ints), ints);
5866 cpus_clear(cpu_isolated_map);
5867 for (i = 1; i <= ints[0]; i++)
5868 if (ints[i] < NR_CPUS)
5869 cpu_set(ints[i], cpu_isolated_map);
5873 __setup("isolcpus=", isolated_cpu_setup);
5876 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5877 * to a function which identifies what group(along with sched group) a CPU
5878 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5879 * (due to the fact that we keep track of groups covered with a cpumask_t).
5881 * init_sched_build_groups will build a circular linked list of the groups
5882 * covered by the given span, and will set each group's ->cpumask correctly,
5883 * and ->cpu_power to 0.
5886 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5887 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5888 struct sched_group **sg))
5890 struct sched_group *first = NULL, *last = NULL;
5891 cpumask_t covered = CPU_MASK_NONE;
5894 for_each_cpu_mask(i, span) {
5895 struct sched_group *sg;
5896 int group = group_fn(i, cpu_map, &sg);
5899 if (cpu_isset(i, covered))
5902 sg->cpumask = CPU_MASK_NONE;
5903 sg->__cpu_power = 0;
5905 for_each_cpu_mask(j, span) {
5906 if (group_fn(j, cpu_map, NULL) != group)
5909 cpu_set(j, covered);
5910 cpu_set(j, sg->cpumask);
5921 #define SD_NODES_PER_DOMAIN 16
5926 * find_next_best_node - find the next node to include in a sched_domain
5927 * @node: node whose sched_domain we're building
5928 * @used_nodes: nodes already in the sched_domain
5930 * Find the next node to include in a given scheduling domain. Simply
5931 * finds the closest node not already in the @used_nodes map.
5933 * Should use nodemask_t.
5935 static int find_next_best_node(int node, unsigned long *used_nodes)
5937 int i, n, val, min_val, best_node = 0;
5941 for (i = 0; i < MAX_NUMNODES; i++) {
5942 /* Start at @node */
5943 n = (node + i) % MAX_NUMNODES;
5945 if (!nr_cpus_node(n))
5948 /* Skip already used nodes */
5949 if (test_bit(n, used_nodes))
5952 /* Simple min distance search */
5953 val = node_distance(node, n);
5955 if (val < min_val) {
5961 set_bit(best_node, used_nodes);
5966 * sched_domain_node_span - get a cpumask for a node's sched_domain
5967 * @node: node whose cpumask we're constructing
5968 * @size: number of nodes to include in this span
5970 * Given a node, construct a good cpumask for its sched_domain to span. It
5971 * should be one that prevents unnecessary balancing, but also spreads tasks
5974 static cpumask_t sched_domain_node_span(int node)
5976 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5977 cpumask_t span, nodemask;
5981 bitmap_zero(used_nodes, MAX_NUMNODES);
5983 nodemask = node_to_cpumask(node);
5984 cpus_or(span, span, nodemask);
5985 set_bit(node, used_nodes);
5987 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5988 int next_node = find_next_best_node(node, used_nodes);
5990 nodemask = node_to_cpumask(next_node);
5991 cpus_or(span, span, nodemask);
5998 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6001 * SMT sched-domains:
6003 #ifdef CONFIG_SCHED_SMT
6004 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6005 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6007 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
6008 struct sched_group **sg)
6011 *sg = &per_cpu(sched_group_cpus, cpu);
6017 * multi-core sched-domains:
6019 #ifdef CONFIG_SCHED_MC
6020 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6021 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6024 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6025 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6026 struct sched_group **sg)
6029 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6030 cpus_and(mask, mask, *cpu_map);
6031 group = first_cpu(mask);
6033 *sg = &per_cpu(sched_group_core, group);
6036 #elif defined(CONFIG_SCHED_MC)
6037 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
6038 struct sched_group **sg)
6041 *sg = &per_cpu(sched_group_core, cpu);
6046 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6047 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6049 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6050 struct sched_group **sg)
6053 #ifdef CONFIG_SCHED_MC
6054 cpumask_t mask = cpu_coregroup_map(cpu);
6055 cpus_and(mask, mask, *cpu_map);
6056 group = first_cpu(mask);
6057 #elif defined(CONFIG_SCHED_SMT)
6058 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6059 cpus_and(mask, mask, *cpu_map);
6060 group = first_cpu(mask);
6065 *sg = &per_cpu(sched_group_phys, group);
6071 * The init_sched_build_groups can't handle what we want to do with node
6072 * groups, so roll our own. Now each node has its own list of groups which
6073 * gets dynamically allocated.
6075 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6076 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6078 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6079 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6081 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6082 struct sched_group **sg)
6084 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6087 cpus_and(nodemask, nodemask, *cpu_map);
6088 group = first_cpu(nodemask);
6091 *sg = &per_cpu(sched_group_allnodes, group);
6095 static void init_numa_sched_groups_power(struct sched_group *group_head)
6097 struct sched_group *sg = group_head;
6103 for_each_cpu_mask(j, sg->cpumask) {
6104 struct sched_domain *sd;
6106 sd = &per_cpu(phys_domains, j);
6107 if (j != first_cpu(sd->groups->cpumask)) {
6109 * Only add "power" once for each
6115 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6118 } while (sg != group_head);
6123 /* Free memory allocated for various sched_group structures */
6124 static void free_sched_groups(const cpumask_t *cpu_map)
6128 for_each_cpu_mask(cpu, *cpu_map) {
6129 struct sched_group **sched_group_nodes
6130 = sched_group_nodes_bycpu[cpu];
6132 if (!sched_group_nodes)
6135 for (i = 0; i < MAX_NUMNODES; i++) {
6136 cpumask_t nodemask = node_to_cpumask(i);
6137 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6139 cpus_and(nodemask, nodemask, *cpu_map);
6140 if (cpus_empty(nodemask))
6150 if (oldsg != sched_group_nodes[i])
6153 kfree(sched_group_nodes);
6154 sched_group_nodes_bycpu[cpu] = NULL;
6158 static void free_sched_groups(const cpumask_t *cpu_map)
6164 * Initialize sched groups cpu_power.
6166 * cpu_power indicates the capacity of sched group, which is used while
6167 * distributing the load between different sched groups in a sched domain.
6168 * Typically cpu_power for all the groups in a sched domain will be same unless
6169 * there are asymmetries in the topology. If there are asymmetries, group
6170 * having more cpu_power will pickup more load compared to the group having
6173 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6174 * the maximum number of tasks a group can handle in the presence of other idle
6175 * or lightly loaded groups in the same sched domain.
6177 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6179 struct sched_domain *child;
6180 struct sched_group *group;
6182 WARN_ON(!sd || !sd->groups);
6184 if (cpu != first_cpu(sd->groups->cpumask))
6189 sd->groups->__cpu_power = 0;
6192 * For perf policy, if the groups in child domain share resources
6193 * (for example cores sharing some portions of the cache hierarchy
6194 * or SMT), then set this domain groups cpu_power such that each group
6195 * can handle only one task, when there are other idle groups in the
6196 * same sched domain.
6198 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6200 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6201 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6206 * add cpu_power of each child group to this groups cpu_power
6208 group = child->groups;
6210 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6211 group = group->next;
6212 } while (group != child->groups);
6216 * Build sched domains for a given set of cpus and attach the sched domains
6217 * to the individual cpus
6219 static int build_sched_domains(const cpumask_t *cpu_map)
6223 struct sched_group **sched_group_nodes = NULL;
6224 int sd_allnodes = 0;
6227 * Allocate the per-node list of sched groups
6229 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6231 if (!sched_group_nodes) {
6232 printk(KERN_WARNING "Can not alloc sched group node list\n");
6235 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6239 * Set up domains for cpus specified by the cpu_map.
6241 for_each_cpu_mask(i, *cpu_map) {
6242 struct sched_domain *sd = NULL, *p;
6243 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6245 cpus_and(nodemask, nodemask, *cpu_map);
6248 if (cpus_weight(*cpu_map) >
6249 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6250 sd = &per_cpu(allnodes_domains, i);
6251 *sd = SD_ALLNODES_INIT;
6252 sd->span = *cpu_map;
6253 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6259 sd = &per_cpu(node_domains, i);
6261 sd->span = sched_domain_node_span(cpu_to_node(i));
6265 cpus_and(sd->span, sd->span, *cpu_map);
6269 sd = &per_cpu(phys_domains, i);
6271 sd->span = nodemask;
6275 cpu_to_phys_group(i, cpu_map, &sd->groups);
6277 #ifdef CONFIG_SCHED_MC
6279 sd = &per_cpu(core_domains, i);
6281 sd->span = cpu_coregroup_map(i);
6282 cpus_and(sd->span, sd->span, *cpu_map);
6285 cpu_to_core_group(i, cpu_map, &sd->groups);
6288 #ifdef CONFIG_SCHED_SMT
6290 sd = &per_cpu(cpu_domains, i);
6291 *sd = SD_SIBLING_INIT;
6292 sd->span = per_cpu(cpu_sibling_map, i);
6293 cpus_and(sd->span, sd->span, *cpu_map);
6296 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6300 #ifdef CONFIG_SCHED_SMT
6301 /* Set up CPU (sibling) groups */
6302 for_each_cpu_mask(i, *cpu_map) {
6303 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6304 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6305 if (i != first_cpu(this_sibling_map))
6308 init_sched_build_groups(this_sibling_map, cpu_map,
6313 #ifdef CONFIG_SCHED_MC
6314 /* Set up multi-core groups */
6315 for_each_cpu_mask(i, *cpu_map) {
6316 cpumask_t this_core_map = cpu_coregroup_map(i);
6317 cpus_and(this_core_map, this_core_map, *cpu_map);
6318 if (i != first_cpu(this_core_map))
6320 init_sched_build_groups(this_core_map, cpu_map,
6321 &cpu_to_core_group);
6325 /* Set up physical groups */
6326 for (i = 0; i < MAX_NUMNODES; i++) {
6327 cpumask_t nodemask = node_to_cpumask(i);
6329 cpus_and(nodemask, nodemask, *cpu_map);
6330 if (cpus_empty(nodemask))
6333 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6337 /* Set up node groups */
6339 init_sched_build_groups(*cpu_map, cpu_map,
6340 &cpu_to_allnodes_group);
6342 for (i = 0; i < MAX_NUMNODES; i++) {
6343 /* Set up node groups */
6344 struct sched_group *sg, *prev;
6345 cpumask_t nodemask = node_to_cpumask(i);
6346 cpumask_t domainspan;
6347 cpumask_t covered = CPU_MASK_NONE;
6350 cpus_and(nodemask, nodemask, *cpu_map);
6351 if (cpus_empty(nodemask)) {
6352 sched_group_nodes[i] = NULL;
6356 domainspan = sched_domain_node_span(i);
6357 cpus_and(domainspan, domainspan, *cpu_map);
6359 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6361 printk(KERN_WARNING "Can not alloc domain group for "
6365 sched_group_nodes[i] = sg;
6366 for_each_cpu_mask(j, nodemask) {
6367 struct sched_domain *sd;
6369 sd = &per_cpu(node_domains, j);
6372 sg->__cpu_power = 0;
6373 sg->cpumask = nodemask;
6375 cpus_or(covered, covered, nodemask);
6378 for (j = 0; j < MAX_NUMNODES; j++) {
6379 cpumask_t tmp, notcovered;
6380 int n = (i + j) % MAX_NUMNODES;
6382 cpus_complement(notcovered, covered);
6383 cpus_and(tmp, notcovered, *cpu_map);
6384 cpus_and(tmp, tmp, domainspan);
6385 if (cpus_empty(tmp))
6388 nodemask = node_to_cpumask(n);
6389 cpus_and(tmp, tmp, nodemask);
6390 if (cpus_empty(tmp))
6393 sg = kmalloc_node(sizeof(struct sched_group),
6397 "Can not alloc domain group for node %d\n", j);
6400 sg->__cpu_power = 0;
6402 sg->next = prev->next;
6403 cpus_or(covered, covered, tmp);
6410 /* Calculate CPU power for physical packages and nodes */
6411 #ifdef CONFIG_SCHED_SMT
6412 for_each_cpu_mask(i, *cpu_map) {
6413 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6415 init_sched_groups_power(i, sd);
6418 #ifdef CONFIG_SCHED_MC
6419 for_each_cpu_mask(i, *cpu_map) {
6420 struct sched_domain *sd = &per_cpu(core_domains, i);
6422 init_sched_groups_power(i, sd);
6426 for_each_cpu_mask(i, *cpu_map) {
6427 struct sched_domain *sd = &per_cpu(phys_domains, i);
6429 init_sched_groups_power(i, sd);
6433 for (i = 0; i < MAX_NUMNODES; i++)
6434 init_numa_sched_groups_power(sched_group_nodes[i]);
6437 struct sched_group *sg;
6439 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6440 init_numa_sched_groups_power(sg);
6444 /* Attach the domains */
6445 for_each_cpu_mask(i, *cpu_map) {
6446 struct sched_domain *sd;
6447 #ifdef CONFIG_SCHED_SMT
6448 sd = &per_cpu(cpu_domains, i);
6449 #elif defined(CONFIG_SCHED_MC)
6450 sd = &per_cpu(core_domains, i);
6452 sd = &per_cpu(phys_domains, i);
6454 cpu_attach_domain(sd, i);
6461 free_sched_groups(cpu_map);
6466 static cpumask_t *doms_cur; /* current sched domains */
6467 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6470 * Special case: If a kmalloc of a doms_cur partition (array of
6471 * cpumask_t) fails, then fallback to a single sched domain,
6472 * as determined by the single cpumask_t fallback_doms.
6474 static cpumask_t fallback_doms;
6477 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6478 * For now this just excludes isolated cpus, but could be used to
6479 * exclude other special cases in the future.
6481 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6486 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6488 doms_cur = &fallback_doms;
6489 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6490 err = build_sched_domains(doms_cur);
6491 register_sched_domain_sysctl();
6496 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6498 free_sched_groups(cpu_map);
6502 * Detach sched domains from a group of cpus specified in cpu_map
6503 * These cpus will now be attached to the NULL domain
6505 static void detach_destroy_domains(const cpumask_t *cpu_map)
6509 unregister_sched_domain_sysctl();
6511 for_each_cpu_mask(i, *cpu_map)
6512 cpu_attach_domain(NULL, i);
6513 synchronize_sched();
6514 arch_destroy_sched_domains(cpu_map);
6518 * Partition sched domains as specified by the 'ndoms_new'
6519 * cpumasks in the array doms_new[] of cpumasks. This compares
6520 * doms_new[] to the current sched domain partitioning, doms_cur[].
6521 * It destroys each deleted domain and builds each new domain.
6523 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6524 * The masks don't intersect (don't overlap.) We should setup one
6525 * sched domain for each mask. CPUs not in any of the cpumasks will
6526 * not be load balanced. If the same cpumask appears both in the
6527 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6530 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6531 * ownership of it and will kfree it when done with it. If the caller
6532 * failed the kmalloc call, then it can pass in doms_new == NULL,
6533 * and partition_sched_domains() will fallback to the single partition
6536 * Call with hotplug lock held
6538 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6542 /* always unregister in case we don't destroy any domains */
6543 unregister_sched_domain_sysctl();
6545 if (doms_new == NULL) {
6547 doms_new = &fallback_doms;
6548 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6551 /* Destroy deleted domains */
6552 for (i = 0; i < ndoms_cur; i++) {
6553 for (j = 0; j < ndoms_new; j++) {
6554 if (cpus_equal(doms_cur[i], doms_new[j]))
6557 /* no match - a current sched domain not in new doms_new[] */
6558 detach_destroy_domains(doms_cur + i);
6563 /* Build new domains */
6564 for (i = 0; i < ndoms_new; i++) {
6565 for (j = 0; j < ndoms_cur; j++) {
6566 if (cpus_equal(doms_new[i], doms_cur[j]))
6569 /* no match - add a new doms_new */
6570 build_sched_domains(doms_new + i);
6575 /* Remember the new sched domains */
6576 if (doms_cur != &fallback_doms)
6578 doms_cur = doms_new;
6579 ndoms_cur = ndoms_new;
6581 register_sched_domain_sysctl();
6584 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6585 static int arch_reinit_sched_domains(void)
6589 mutex_lock(&sched_hotcpu_mutex);
6590 detach_destroy_domains(&cpu_online_map);
6591 err = arch_init_sched_domains(&cpu_online_map);
6592 mutex_unlock(&sched_hotcpu_mutex);
6597 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6601 if (buf[0] != '0' && buf[0] != '1')
6605 sched_smt_power_savings = (buf[0] == '1');
6607 sched_mc_power_savings = (buf[0] == '1');
6609 ret = arch_reinit_sched_domains();
6611 return ret ? ret : count;
6614 #ifdef CONFIG_SCHED_MC
6615 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6617 return sprintf(page, "%u\n", sched_mc_power_savings);
6619 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6620 const char *buf, size_t count)
6622 return sched_power_savings_store(buf, count, 0);
6624 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6625 sched_mc_power_savings_store);
6628 #ifdef CONFIG_SCHED_SMT
6629 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6631 return sprintf(page, "%u\n", sched_smt_power_savings);
6633 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6634 const char *buf, size_t count)
6636 return sched_power_savings_store(buf, count, 1);
6638 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6639 sched_smt_power_savings_store);
6642 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6646 #ifdef CONFIG_SCHED_SMT
6648 err = sysfs_create_file(&cls->kset.kobj,
6649 &attr_sched_smt_power_savings.attr);
6651 #ifdef CONFIG_SCHED_MC
6652 if (!err && mc_capable())
6653 err = sysfs_create_file(&cls->kset.kobj,
6654 &attr_sched_mc_power_savings.attr);
6661 * Force a reinitialization of the sched domains hierarchy. The domains
6662 * and groups cannot be updated in place without racing with the balancing
6663 * code, so we temporarily attach all running cpus to the NULL domain
6664 * which will prevent rebalancing while the sched domains are recalculated.
6666 static int update_sched_domains(struct notifier_block *nfb,
6667 unsigned long action, void *hcpu)
6670 case CPU_UP_PREPARE:
6671 case CPU_UP_PREPARE_FROZEN:
6672 case CPU_DOWN_PREPARE:
6673 case CPU_DOWN_PREPARE_FROZEN:
6674 detach_destroy_domains(&cpu_online_map);
6677 case CPU_UP_CANCELED:
6678 case CPU_UP_CANCELED_FROZEN:
6679 case CPU_DOWN_FAILED:
6680 case CPU_DOWN_FAILED_FROZEN:
6682 case CPU_ONLINE_FROZEN:
6684 case CPU_DEAD_FROZEN:
6686 * Fall through and re-initialise the domains.
6693 /* The hotplug lock is already held by cpu_up/cpu_down */
6694 arch_init_sched_domains(&cpu_online_map);
6699 void __init sched_init_smp(void)
6701 cpumask_t non_isolated_cpus;
6703 mutex_lock(&sched_hotcpu_mutex);
6704 arch_init_sched_domains(&cpu_online_map);
6705 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6706 if (cpus_empty(non_isolated_cpus))
6707 cpu_set(smp_processor_id(), non_isolated_cpus);
6708 mutex_unlock(&sched_hotcpu_mutex);
6709 /* XXX: Theoretical race here - CPU may be hotplugged now */
6710 hotcpu_notifier(update_sched_domains, 0);
6712 /* Move init over to a non-isolated CPU */
6713 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6715 sched_init_granularity();
6718 void __init sched_init_smp(void)
6720 sched_init_granularity();
6722 #endif /* CONFIG_SMP */
6724 int in_sched_functions(unsigned long addr)
6726 /* Linker adds these: start and end of __sched functions */
6727 extern char __sched_text_start[], __sched_text_end[];
6729 return in_lock_functions(addr) ||
6730 (addr >= (unsigned long)__sched_text_start
6731 && addr < (unsigned long)__sched_text_end);
6734 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6736 cfs_rq->tasks_timeline = RB_ROOT;
6737 #ifdef CONFIG_FAIR_GROUP_SCHED
6740 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6743 void __init sched_init(void)
6745 int highest_cpu = 0;
6748 for_each_possible_cpu(i) {
6749 struct rt_prio_array *array;
6753 spin_lock_init(&rq->lock);
6754 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6757 init_cfs_rq(&rq->cfs, rq);
6758 #ifdef CONFIG_FAIR_GROUP_SCHED
6759 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6761 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6762 struct sched_entity *se =
6763 &per_cpu(init_sched_entity, i);
6765 init_cfs_rq_p[i] = cfs_rq;
6766 init_cfs_rq(cfs_rq, rq);
6767 cfs_rq->tg = &init_task_group;
6768 list_add(&cfs_rq->leaf_cfs_rq_list,
6769 &rq->leaf_cfs_rq_list);
6771 init_sched_entity_p[i] = se;
6772 se->cfs_rq = &rq->cfs;
6774 se->load.weight = init_task_group_load;
6775 se->load.inv_weight =
6776 div64_64(1ULL<<32, init_task_group_load);
6779 init_task_group.shares = init_task_group_load;
6780 spin_lock_init(&init_task_group.lock);
6783 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6784 rq->cpu_load[j] = 0;
6787 rq->active_balance = 0;
6788 rq->next_balance = jiffies;
6791 rq->migration_thread = NULL;
6792 INIT_LIST_HEAD(&rq->migration_queue);
6794 atomic_set(&rq->nr_iowait, 0);
6796 array = &rq->rt.active;
6797 for (j = 0; j < MAX_RT_PRIO; j++) {
6798 INIT_LIST_HEAD(array->queue + j);
6799 __clear_bit(j, array->bitmap);
6802 /* delimiter for bitsearch: */
6803 __set_bit(MAX_RT_PRIO, array->bitmap);
6806 set_load_weight(&init_task);
6808 #ifdef CONFIG_PREEMPT_NOTIFIERS
6809 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6813 nr_cpu_ids = highest_cpu + 1;
6814 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6817 #ifdef CONFIG_RT_MUTEXES
6818 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6822 * The boot idle thread does lazy MMU switching as well:
6824 atomic_inc(&init_mm.mm_count);
6825 enter_lazy_tlb(&init_mm, current);
6828 * Make us the idle thread. Technically, schedule() should not be
6829 * called from this thread, however somewhere below it might be,
6830 * but because we are the idle thread, we just pick up running again
6831 * when this runqueue becomes "idle".
6833 init_idle(current, smp_processor_id());
6835 * During early bootup we pretend to be a normal task:
6837 current->sched_class = &fair_sched_class;
6840 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6841 void __might_sleep(char *file, int line)
6844 static unsigned long prev_jiffy; /* ratelimiting */
6846 if ((in_atomic() || irqs_disabled()) &&
6847 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6848 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6850 prev_jiffy = jiffies;
6851 printk(KERN_ERR "BUG: sleeping function called from invalid"
6852 " context at %s:%d\n", file, line);
6853 printk("in_atomic():%d, irqs_disabled():%d\n",
6854 in_atomic(), irqs_disabled());
6855 debug_show_held_locks(current);
6856 if (irqs_disabled())
6857 print_irqtrace_events(current);
6862 EXPORT_SYMBOL(__might_sleep);
6865 #ifdef CONFIG_MAGIC_SYSRQ
6866 static void normalize_task(struct rq *rq, struct task_struct *p)
6869 update_rq_clock(rq);
6870 on_rq = p->se.on_rq;
6872 deactivate_task(rq, p, 0);
6873 __setscheduler(rq, p, SCHED_NORMAL, 0);
6875 activate_task(rq, p, 0);
6876 resched_task(rq->curr);
6880 void normalize_rt_tasks(void)
6882 struct task_struct *g, *p;
6883 unsigned long flags;
6886 read_lock_irq(&tasklist_lock);
6887 do_each_thread(g, p) {
6889 * Only normalize user tasks:
6894 p->se.exec_start = 0;
6895 #ifdef CONFIG_SCHEDSTATS
6896 p->se.wait_start = 0;
6897 p->se.sleep_start = 0;
6898 p->se.block_start = 0;
6900 task_rq(p)->clock = 0;
6904 * Renice negative nice level userspace
6907 if (TASK_NICE(p) < 0 && p->mm)
6908 set_user_nice(p, 0);
6912 spin_lock_irqsave(&p->pi_lock, flags);
6913 rq = __task_rq_lock(p);
6915 normalize_task(rq, p);
6917 __task_rq_unlock(rq);
6918 spin_unlock_irqrestore(&p->pi_lock, flags);
6919 } while_each_thread(g, p);
6921 read_unlock_irq(&tasklist_lock);
6924 #endif /* CONFIG_MAGIC_SYSRQ */
6928 * These functions are only useful for the IA64 MCA handling.
6930 * They can only be called when the whole system has been
6931 * stopped - every CPU needs to be quiescent, and no scheduling
6932 * activity can take place. Using them for anything else would
6933 * be a serious bug, and as a result, they aren't even visible
6934 * under any other configuration.
6938 * curr_task - return the current task for a given cpu.
6939 * @cpu: the processor in question.
6941 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6943 struct task_struct *curr_task(int cpu)
6945 return cpu_curr(cpu);
6949 * set_curr_task - set the current task for a given cpu.
6950 * @cpu: the processor in question.
6951 * @p: the task pointer to set.
6953 * Description: This function must only be used when non-maskable interrupts
6954 * are serviced on a separate stack. It allows the architecture to switch the
6955 * notion of the current task on a cpu in a non-blocking manner. This function
6956 * must be called with all CPU's synchronized, and interrupts disabled, the
6957 * and caller must save the original value of the current task (see
6958 * curr_task() above) and restore that value before reenabling interrupts and
6959 * re-starting the system.
6961 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6963 void set_curr_task(int cpu, struct task_struct *p)
6970 #ifdef CONFIG_FAIR_GROUP_SCHED
6972 /* allocate runqueue etc for a new task group */
6973 struct task_group *sched_create_group(void)
6975 struct task_group *tg;
6976 struct cfs_rq *cfs_rq;
6977 struct sched_entity *se;
6981 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6983 return ERR_PTR(-ENOMEM);
6985 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6988 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6992 for_each_possible_cpu(i) {
6995 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
7000 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
7005 memset(cfs_rq, 0, sizeof(struct cfs_rq));
7006 memset(se, 0, sizeof(struct sched_entity));
7008 tg->cfs_rq[i] = cfs_rq;
7009 init_cfs_rq(cfs_rq, rq);
7013 se->cfs_rq = &rq->cfs;
7015 se->load.weight = NICE_0_LOAD;
7016 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7020 for_each_possible_cpu(i) {
7022 cfs_rq = tg->cfs_rq[i];
7023 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7026 tg->shares = NICE_0_LOAD;
7027 spin_lock_init(&tg->lock);
7032 for_each_possible_cpu(i) {
7034 kfree(tg->cfs_rq[i]);
7042 return ERR_PTR(-ENOMEM);
7045 /* rcu callback to free various structures associated with a task group */
7046 static void free_sched_group(struct rcu_head *rhp)
7048 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7049 struct cfs_rq *cfs_rq;
7050 struct sched_entity *se;
7053 /* now it should be safe to free those cfs_rqs */
7054 for_each_possible_cpu(i) {
7055 cfs_rq = tg->cfs_rq[i];
7067 /* Destroy runqueue etc associated with a task group */
7068 void sched_destroy_group(struct task_group *tg)
7070 struct cfs_rq *cfs_rq = NULL;
7073 for_each_possible_cpu(i) {
7074 cfs_rq = tg->cfs_rq[i];
7075 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7080 /* wait for possible concurrent references to cfs_rqs complete */
7081 call_rcu(&tg->rcu, free_sched_group);
7084 /* change task's runqueue when it moves between groups.
7085 * The caller of this function should have put the task in its new group
7086 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7087 * reflect its new group.
7089 void sched_move_task(struct task_struct *tsk)
7092 unsigned long flags;
7095 rq = task_rq_lock(tsk, &flags);
7097 if (tsk->sched_class != &fair_sched_class)
7100 update_rq_clock(rq);
7102 running = task_running(rq, tsk);
7103 on_rq = tsk->se.on_rq;
7106 dequeue_task(rq, tsk, 0);
7107 if (unlikely(running))
7108 tsk->sched_class->put_prev_task(rq, tsk);
7111 set_task_cfs_rq(tsk);
7114 if (unlikely(running))
7115 tsk->sched_class->set_curr_task(rq);
7116 enqueue_task(rq, tsk, 0);
7120 task_rq_unlock(rq, &flags);
7123 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7125 struct cfs_rq *cfs_rq = se->cfs_rq;
7126 struct rq *rq = cfs_rq->rq;
7129 spin_lock_irq(&rq->lock);
7133 dequeue_entity(cfs_rq, se, 0);
7135 se->load.weight = shares;
7136 se->load.inv_weight = div64_64((1ULL<<32), shares);
7139 enqueue_entity(cfs_rq, se, 0);
7141 spin_unlock_irq(&rq->lock);
7144 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7148 spin_lock(&tg->lock);
7149 if (tg->shares == shares)
7152 tg->shares = shares;
7153 for_each_possible_cpu(i)
7154 set_se_shares(tg->se[i], shares);
7157 spin_unlock(&tg->lock);
7161 unsigned long sched_group_shares(struct task_group *tg)
7166 #endif /* CONFIG_FAIR_GROUP_SCHED */
7168 #ifdef CONFIG_FAIR_CGROUP_SCHED
7170 /* return corresponding task_group object of a cgroup */
7171 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7173 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7174 struct task_group, css);
7177 static struct cgroup_subsys_state *
7178 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7180 struct task_group *tg;
7182 if (!cgrp->parent) {
7183 /* This is early initialization for the top cgroup */
7184 init_task_group.css.cgroup = cgrp;
7185 return &init_task_group.css;
7188 /* we support only 1-level deep hierarchical scheduler atm */
7189 if (cgrp->parent->parent)
7190 return ERR_PTR(-EINVAL);
7192 tg = sched_create_group();
7194 return ERR_PTR(-ENOMEM);
7196 /* Bind the cgroup to task_group object we just created */
7197 tg->css.cgroup = cgrp;
7202 static void cpu_cgroup_destroy(struct cgroup_subsys *ss,
7203 struct cgroup *cgrp)
7205 struct task_group *tg = cgroup_tg(cgrp);
7207 sched_destroy_group(tg);
7210 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss,
7211 struct cgroup *cgrp, struct task_struct *tsk)
7213 /* We don't support RT-tasks being in separate groups */
7214 if (tsk->sched_class != &fair_sched_class)
7221 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7222 struct cgroup *old_cont, struct task_struct *tsk)
7224 sched_move_task(tsk);
7227 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7230 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7233 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7235 struct task_group *tg = cgroup_tg(cgrp);
7237 return (u64) tg->shares;
7240 static u64 cpu_usage_read(struct cgroup *cgrp, struct cftype *cft)
7242 struct task_group *tg = cgroup_tg(cgrp);
7243 unsigned long flags;
7247 for_each_possible_cpu(i) {
7249 * Lock to prevent races with updating 64-bit counters
7252 spin_lock_irqsave(&cpu_rq(i)->lock, flags);
7253 res += tg->se[i]->sum_exec_runtime;
7254 spin_unlock_irqrestore(&cpu_rq(i)->lock, flags);
7256 /* Convert from ns to ms */
7257 do_div(res, NSEC_PER_MSEC);
7262 static struct cftype cpu_files[] = {
7265 .read_uint = cpu_shares_read_uint,
7266 .write_uint = cpu_shares_write_uint,
7270 .read_uint = cpu_usage_read,
7274 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7276 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7279 struct cgroup_subsys cpu_cgroup_subsys = {
7281 .create = cpu_cgroup_create,
7282 .destroy = cpu_cgroup_destroy,
7283 .can_attach = cpu_cgroup_can_attach,
7284 .attach = cpu_cgroup_attach,
7285 .populate = cpu_cgroup_populate,
7286 .subsys_id = cpu_cgroup_subsys_id,
7290 #endif /* CONFIG_FAIR_CGROUP_SCHED */