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 * (1000000000 / 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) / (1000000000 / HZ))
103 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / 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 */
177 /* Default task group's sched entity on each cpu */
178 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
179 /* Default task group's cfs_rq on each cpu */
180 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
182 static struct sched_entity *init_sched_entity_p[NR_CPUS];
183 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
185 /* Default task group.
186 * Every task in system belong to this group at bootup.
188 struct task_group init_task_group = {
189 .se = init_sched_entity_p,
190 .cfs_rq = init_cfs_rq_p,
193 #ifdef CONFIG_FAIR_USER_SCHED
194 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
196 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
199 static int init_task_group_load = INIT_TASK_GRP_LOAD;
201 /* return group to which a task belongs */
202 static inline struct task_group *task_group(struct task_struct *p)
204 struct task_group *tg;
206 #ifdef CONFIG_FAIR_USER_SCHED
208 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
209 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
210 struct task_group, css);
212 tg = &init_task_group;
218 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
219 static inline void set_task_cfs_rq(struct task_struct *p)
221 p->se.cfs_rq = task_group(p)->cfs_rq[task_cpu(p)];
222 p->se.parent = task_group(p)->se[task_cpu(p)];
227 static inline void set_task_cfs_rq(struct task_struct *p) { }
229 #endif /* CONFIG_FAIR_GROUP_SCHED */
231 /* CFS-related fields in a runqueue */
233 struct load_weight load;
234 unsigned long nr_running;
239 struct rb_root tasks_timeline;
240 struct rb_node *rb_leftmost;
241 struct rb_node *rb_load_balance_curr;
242 /* 'curr' points to currently running entity on this cfs_rq.
243 * It is set to NULL otherwise (i.e when none are currently running).
245 struct sched_entity *curr;
247 unsigned long nr_spread_over;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
252 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
253 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
254 * (like users, containers etc.)
256 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
257 * list is used during load balance.
259 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
260 struct task_group *tg; /* group that "owns" this runqueue */
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,
463 SCHED_FEAT_PREEMPT_RESTRICT = 32,
466 const_debug unsigned int sysctl_sched_features =
467 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
468 SCHED_FEAT_START_DEBIT * 1 |
469 SCHED_FEAT_TREE_AVG * 0 |
470 SCHED_FEAT_APPROX_AVG * 0 |
471 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
472 SCHED_FEAT_PREEMPT_RESTRICT * 1;
474 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
477 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
478 * clock constructed from sched_clock():
480 unsigned long long cpu_clock(int cpu)
482 unsigned long long now;
486 local_irq_save(flags);
490 local_irq_restore(flags);
494 EXPORT_SYMBOL_GPL(cpu_clock);
496 #ifndef prepare_arch_switch
497 # define prepare_arch_switch(next) do { } while (0)
499 #ifndef finish_arch_switch
500 # define finish_arch_switch(prev) do { } while (0)
503 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
504 static inline int task_running(struct rq *rq, struct task_struct *p)
506 return rq->curr == p;
509 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
513 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
515 #ifdef CONFIG_DEBUG_SPINLOCK
516 /* this is a valid case when another task releases the spinlock */
517 rq->lock.owner = current;
520 * If we are tracking spinlock dependencies then we have to
521 * fix up the runqueue lock - which gets 'carried over' from
524 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
526 spin_unlock_irq(&rq->lock);
529 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
530 static inline int task_running(struct rq *rq, struct task_struct *p)
535 return rq->curr == p;
539 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
543 * We can optimise this out completely for !SMP, because the
544 * SMP rebalancing from interrupt is the only thing that cares
549 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
550 spin_unlock_irq(&rq->lock);
552 spin_unlock(&rq->lock);
556 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
560 * After ->oncpu is cleared, the task can be moved to a different CPU.
561 * We must ensure this doesn't happen until the switch is completely
567 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
571 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
574 * __task_rq_lock - lock the runqueue a given task resides on.
575 * Must be called interrupts disabled.
577 static inline struct rq *__task_rq_lock(struct task_struct *p)
581 struct rq *rq = task_rq(p);
582 spin_lock(&rq->lock);
583 if (likely(rq == task_rq(p)))
585 spin_unlock(&rq->lock);
590 * task_rq_lock - lock the runqueue a given task resides on and disable
591 * interrupts. Note the ordering: we can safely lookup the task_rq without
592 * explicitly disabling preemption.
594 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
600 local_irq_save(*flags);
602 spin_lock(&rq->lock);
603 if (likely(rq == task_rq(p)))
605 spin_unlock_irqrestore(&rq->lock, *flags);
609 static void __task_rq_unlock(struct rq *rq)
612 spin_unlock(&rq->lock);
615 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
618 spin_unlock_irqrestore(&rq->lock, *flags);
622 * this_rq_lock - lock this runqueue and disable interrupts.
624 static struct rq *this_rq_lock(void)
631 spin_lock(&rq->lock);
637 * We are going deep-idle (irqs are disabled):
639 void sched_clock_idle_sleep_event(void)
641 struct rq *rq = cpu_rq(smp_processor_id());
643 spin_lock(&rq->lock);
644 __update_rq_clock(rq);
645 spin_unlock(&rq->lock);
646 rq->clock_deep_idle_events++;
648 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
651 * We just idled delta nanoseconds (called with irqs disabled):
653 void sched_clock_idle_wakeup_event(u64 delta_ns)
655 struct rq *rq = cpu_rq(smp_processor_id());
656 u64 now = sched_clock();
658 rq->idle_clock += delta_ns;
660 * Override the previous timestamp and ignore all
661 * sched_clock() deltas that occured while we idled,
662 * and use the PM-provided delta_ns to advance the
665 spin_lock(&rq->lock);
666 rq->prev_clock_raw = now;
667 rq->clock += delta_ns;
668 spin_unlock(&rq->lock);
670 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
673 * resched_task - mark a task 'to be rescheduled now'.
675 * On UP this means the setting of the need_resched flag, on SMP it
676 * might also involve a cross-CPU call to trigger the scheduler on
681 #ifndef tsk_is_polling
682 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
685 static void resched_task(struct task_struct *p)
689 assert_spin_locked(&task_rq(p)->lock);
691 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
694 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
697 if (cpu == smp_processor_id())
700 /* NEED_RESCHED must be visible before we test polling */
702 if (!tsk_is_polling(p))
703 smp_send_reschedule(cpu);
706 static void resched_cpu(int cpu)
708 struct rq *rq = cpu_rq(cpu);
711 if (!spin_trylock_irqsave(&rq->lock, flags))
713 resched_task(cpu_curr(cpu));
714 spin_unlock_irqrestore(&rq->lock, flags);
717 static inline void resched_task(struct task_struct *p)
719 assert_spin_locked(&task_rq(p)->lock);
720 set_tsk_need_resched(p);
724 #if BITS_PER_LONG == 32
725 # define WMULT_CONST (~0UL)
727 # define WMULT_CONST (1UL << 32)
730 #define WMULT_SHIFT 32
733 * Shift right and round:
735 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
738 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
739 struct load_weight *lw)
743 if (unlikely(!lw->inv_weight))
744 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
746 tmp = (u64)delta_exec * weight;
748 * Check whether we'd overflow the 64-bit multiplication:
750 if (unlikely(tmp > WMULT_CONST))
751 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
754 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
756 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
759 static inline unsigned long
760 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
762 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
765 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
770 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
776 * To aid in avoiding the subversion of "niceness" due to uneven distribution
777 * of tasks with abnormal "nice" values across CPUs the contribution that
778 * each task makes to its run queue's load is weighted according to its
779 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
780 * scaled version of the new time slice allocation that they receive on time
784 #define WEIGHT_IDLEPRIO 2
785 #define WMULT_IDLEPRIO (1 << 31)
788 * Nice levels are multiplicative, with a gentle 10% change for every
789 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
790 * nice 1, it will get ~10% less CPU time than another CPU-bound task
791 * that remained on nice 0.
793 * The "10% effect" is relative and cumulative: from _any_ nice level,
794 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
795 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
796 * If a task goes up by ~10% and another task goes down by ~10% then
797 * the relative distance between them is ~25%.)
799 static const int prio_to_weight[40] = {
800 /* -20 */ 88761, 71755, 56483, 46273, 36291,
801 /* -15 */ 29154, 23254, 18705, 14949, 11916,
802 /* -10 */ 9548, 7620, 6100, 4904, 3906,
803 /* -5 */ 3121, 2501, 1991, 1586, 1277,
804 /* 0 */ 1024, 820, 655, 526, 423,
805 /* 5 */ 335, 272, 215, 172, 137,
806 /* 10 */ 110, 87, 70, 56, 45,
807 /* 15 */ 36, 29, 23, 18, 15,
811 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
813 * In cases where the weight does not change often, we can use the
814 * precalculated inverse to speed up arithmetics by turning divisions
815 * into multiplications:
817 static const u32 prio_to_wmult[40] = {
818 /* -20 */ 48388, 59856, 76040, 92818, 118348,
819 /* -15 */ 147320, 184698, 229616, 287308, 360437,
820 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
821 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
822 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
823 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
824 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
825 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
828 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
831 * runqueue iterator, to support SMP load-balancing between different
832 * scheduling classes, without having to expose their internal data
833 * structures to the load-balancing proper:
837 struct task_struct *(*start)(void *);
838 struct task_struct *(*next)(void *);
841 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
842 unsigned long max_nr_move, unsigned long max_load_move,
843 struct sched_domain *sd, enum cpu_idle_type idle,
844 int *all_pinned, unsigned long *load_moved,
845 int *this_best_prio, struct rq_iterator *iterator);
847 #include "sched_stats.h"
848 #include "sched_idletask.c"
849 #include "sched_fair.c"
850 #include "sched_rt.c"
851 #ifdef CONFIG_SCHED_DEBUG
852 # include "sched_debug.c"
855 #define sched_class_highest (&rt_sched_class)
858 * Update delta_exec, delta_fair fields for rq.
860 * delta_fair clock advances at a rate inversely proportional to
861 * total load (rq->load.weight) on the runqueue, while
862 * delta_exec advances at the same rate as wall-clock (provided
865 * delta_exec / delta_fair is a measure of the (smoothened) load on this
866 * runqueue over any given interval. This (smoothened) load is used
867 * during load balance.
869 * This function is called /before/ updating rq->load
870 * and when switching tasks.
872 static inline void inc_load(struct rq *rq, const struct task_struct *p)
874 update_load_add(&rq->load, p->se.load.weight);
877 static inline void dec_load(struct rq *rq, const struct task_struct *p)
879 update_load_sub(&rq->load, p->se.load.weight);
882 static void inc_nr_running(struct task_struct *p, struct rq *rq)
888 static void dec_nr_running(struct task_struct *p, struct rq *rq)
894 static void set_load_weight(struct task_struct *p)
896 if (task_has_rt_policy(p)) {
897 p->se.load.weight = prio_to_weight[0] * 2;
898 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
903 * SCHED_IDLE tasks get minimal weight:
905 if (p->policy == SCHED_IDLE) {
906 p->se.load.weight = WEIGHT_IDLEPRIO;
907 p->se.load.inv_weight = WMULT_IDLEPRIO;
911 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
912 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
915 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
917 sched_info_queued(p);
918 p->sched_class->enqueue_task(rq, p, wakeup);
922 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
924 p->sched_class->dequeue_task(rq, p, sleep);
929 * __normal_prio - return the priority that is based on the static prio
931 static inline int __normal_prio(struct task_struct *p)
933 return p->static_prio;
937 * Calculate the expected normal priority: i.e. priority
938 * without taking RT-inheritance into account. Might be
939 * boosted by interactivity modifiers. Changes upon fork,
940 * setprio syscalls, and whenever the interactivity
941 * estimator recalculates.
943 static inline int normal_prio(struct task_struct *p)
947 if (task_has_rt_policy(p))
948 prio = MAX_RT_PRIO-1 - p->rt_priority;
950 prio = __normal_prio(p);
955 * Calculate the current priority, i.e. the priority
956 * taken into account by the scheduler. This value might
957 * be boosted by RT tasks, or might be boosted by
958 * interactivity modifiers. Will be RT if the task got
959 * RT-boosted. If not then it returns p->normal_prio.
961 static int effective_prio(struct task_struct *p)
963 p->normal_prio = normal_prio(p);
965 * If we are RT tasks or we were boosted to RT priority,
966 * keep the priority unchanged. Otherwise, update priority
967 * to the normal priority:
969 if (!rt_prio(p->prio))
970 return p->normal_prio;
975 * activate_task - move a task to the runqueue.
977 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
979 if (p->state == TASK_UNINTERRUPTIBLE)
980 rq->nr_uninterruptible--;
982 enqueue_task(rq, p, wakeup);
983 inc_nr_running(p, rq);
987 * deactivate_task - remove a task from the runqueue.
989 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
991 if (p->state == TASK_UNINTERRUPTIBLE)
992 rq->nr_uninterruptible++;
994 dequeue_task(rq, p, sleep);
995 dec_nr_running(p, rq);
999 * task_curr - is this task currently executing on a CPU?
1000 * @p: the task in question.
1002 inline int task_curr(const struct task_struct *p)
1004 return cpu_curr(task_cpu(p)) == p;
1007 /* Used instead of source_load when we know the type == 0 */
1008 unsigned long weighted_cpuload(const int cpu)
1010 return cpu_rq(cpu)->load.weight;
1013 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1016 task_thread_info(p)->cpu = cpu;
1024 * Is this task likely cache-hot:
1027 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1031 if (p->sched_class != &fair_sched_class)
1034 if (sysctl_sched_migration_cost == -1)
1036 if (sysctl_sched_migration_cost == 0)
1039 delta = now - p->se.exec_start;
1041 return delta < (s64)sysctl_sched_migration_cost;
1045 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1047 int old_cpu = task_cpu(p);
1048 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1049 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1050 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1053 clock_offset = old_rq->clock - new_rq->clock;
1055 #ifdef CONFIG_SCHEDSTATS
1056 if (p->se.wait_start)
1057 p->se.wait_start -= clock_offset;
1058 if (p->se.sleep_start)
1059 p->se.sleep_start -= clock_offset;
1060 if (p->se.block_start)
1061 p->se.block_start -= clock_offset;
1062 if (old_cpu != new_cpu) {
1063 schedstat_inc(p, se.nr_migrations);
1064 if (task_hot(p, old_rq->clock, NULL))
1065 schedstat_inc(p, se.nr_forced2_migrations);
1068 p->se.vruntime -= old_cfsrq->min_vruntime -
1069 new_cfsrq->min_vruntime;
1071 __set_task_cpu(p, new_cpu);
1074 struct migration_req {
1075 struct list_head list;
1077 struct task_struct *task;
1080 struct completion done;
1084 * The task's runqueue lock must be held.
1085 * Returns true if you have to wait for migration thread.
1088 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1090 struct rq *rq = task_rq(p);
1093 * If the task is not on a runqueue (and not running), then
1094 * it is sufficient to simply update the task's cpu field.
1096 if (!p->se.on_rq && !task_running(rq, p)) {
1097 set_task_cpu(p, dest_cpu);
1101 init_completion(&req->done);
1103 req->dest_cpu = dest_cpu;
1104 list_add(&req->list, &rq->migration_queue);
1110 * wait_task_inactive - wait for a thread to unschedule.
1112 * The caller must ensure that the task *will* unschedule sometime soon,
1113 * else this function might spin for a *long* time. This function can't
1114 * be called with interrupts off, or it may introduce deadlock with
1115 * smp_call_function() if an IPI is sent by the same process we are
1116 * waiting to become inactive.
1118 void wait_task_inactive(struct task_struct *p)
1120 unsigned long flags;
1126 * We do the initial early heuristics without holding
1127 * any task-queue locks at all. We'll only try to get
1128 * the runqueue lock when things look like they will
1134 * If the task is actively running on another CPU
1135 * still, just relax and busy-wait without holding
1138 * NOTE! Since we don't hold any locks, it's not
1139 * even sure that "rq" stays as the right runqueue!
1140 * But we don't care, since "task_running()" will
1141 * return false if the runqueue has changed and p
1142 * is actually now running somewhere else!
1144 while (task_running(rq, p))
1148 * Ok, time to look more closely! We need the rq
1149 * lock now, to be *sure*. If we're wrong, we'll
1150 * just go back and repeat.
1152 rq = task_rq_lock(p, &flags);
1153 running = task_running(rq, p);
1154 on_rq = p->se.on_rq;
1155 task_rq_unlock(rq, &flags);
1158 * Was it really running after all now that we
1159 * checked with the proper locks actually held?
1161 * Oops. Go back and try again..
1163 if (unlikely(running)) {
1169 * It's not enough that it's not actively running,
1170 * it must be off the runqueue _entirely_, and not
1173 * So if it wa still runnable (but just not actively
1174 * running right now), it's preempted, and we should
1175 * yield - it could be a while.
1177 if (unlikely(on_rq)) {
1178 schedule_timeout_uninterruptible(1);
1183 * Ahh, all good. It wasn't running, and it wasn't
1184 * runnable, which means that it will never become
1185 * running in the future either. We're all done!
1192 * kick_process - kick a running thread to enter/exit the kernel
1193 * @p: the to-be-kicked thread
1195 * Cause a process which is running on another CPU to enter
1196 * kernel-mode, without any delay. (to get signals handled.)
1198 * NOTE: this function doesnt have to take the runqueue lock,
1199 * because all it wants to ensure is that the remote task enters
1200 * the kernel. If the IPI races and the task has been migrated
1201 * to another CPU then no harm is done and the purpose has been
1204 void kick_process(struct task_struct *p)
1210 if ((cpu != smp_processor_id()) && task_curr(p))
1211 smp_send_reschedule(cpu);
1216 * Return a low guess at the load of a migration-source cpu weighted
1217 * according to the scheduling class and "nice" value.
1219 * We want to under-estimate the load of migration sources, to
1220 * balance conservatively.
1222 static unsigned long source_load(int cpu, int type)
1224 struct rq *rq = cpu_rq(cpu);
1225 unsigned long total = weighted_cpuload(cpu);
1230 return min(rq->cpu_load[type-1], total);
1234 * Return a high guess at the load of a migration-target cpu weighted
1235 * according to the scheduling class and "nice" value.
1237 static unsigned long target_load(int cpu, int type)
1239 struct rq *rq = cpu_rq(cpu);
1240 unsigned long total = weighted_cpuload(cpu);
1245 return max(rq->cpu_load[type-1], total);
1249 * Return the average load per task on the cpu's run queue
1251 static inline unsigned long cpu_avg_load_per_task(int cpu)
1253 struct rq *rq = cpu_rq(cpu);
1254 unsigned long total = weighted_cpuload(cpu);
1255 unsigned long n = rq->nr_running;
1257 return n ? total / n : SCHED_LOAD_SCALE;
1261 * find_idlest_group finds and returns the least busy CPU group within the
1264 static struct sched_group *
1265 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1267 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1268 unsigned long min_load = ULONG_MAX, this_load = 0;
1269 int load_idx = sd->forkexec_idx;
1270 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1273 unsigned long load, avg_load;
1277 /* Skip over this group if it has no CPUs allowed */
1278 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1281 local_group = cpu_isset(this_cpu, group->cpumask);
1283 /* Tally up the load of all CPUs in the group */
1286 for_each_cpu_mask(i, group->cpumask) {
1287 /* Bias balancing toward cpus of our domain */
1289 load = source_load(i, load_idx);
1291 load = target_load(i, load_idx);
1296 /* Adjust by relative CPU power of the group */
1297 avg_load = sg_div_cpu_power(group,
1298 avg_load * SCHED_LOAD_SCALE);
1301 this_load = avg_load;
1303 } else if (avg_load < min_load) {
1304 min_load = avg_load;
1307 } while (group = group->next, group != sd->groups);
1309 if (!idlest || 100*this_load < imbalance*min_load)
1315 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1318 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1321 unsigned long load, min_load = ULONG_MAX;
1325 /* Traverse only the allowed CPUs */
1326 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1328 for_each_cpu_mask(i, tmp) {
1329 load = weighted_cpuload(i);
1331 if (load < min_load || (load == min_load && i == this_cpu)) {
1341 * sched_balance_self: balance the current task (running on cpu) in domains
1342 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1345 * Balance, ie. select the least loaded group.
1347 * Returns the target CPU number, or the same CPU if no balancing is needed.
1349 * preempt must be disabled.
1351 static int sched_balance_self(int cpu, int flag)
1353 struct task_struct *t = current;
1354 struct sched_domain *tmp, *sd = NULL;
1356 for_each_domain(cpu, tmp) {
1358 * If power savings logic is enabled for a domain, stop there.
1360 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1362 if (tmp->flags & flag)
1368 struct sched_group *group;
1369 int new_cpu, weight;
1371 if (!(sd->flags & flag)) {
1377 group = find_idlest_group(sd, t, cpu);
1383 new_cpu = find_idlest_cpu(group, t, cpu);
1384 if (new_cpu == -1 || new_cpu == cpu) {
1385 /* Now try balancing at a lower domain level of cpu */
1390 /* Now try balancing at a lower domain level of new_cpu */
1393 weight = cpus_weight(span);
1394 for_each_domain(cpu, tmp) {
1395 if (weight <= cpus_weight(tmp->span))
1397 if (tmp->flags & flag)
1400 /* while loop will break here if sd == NULL */
1406 #endif /* CONFIG_SMP */
1409 * wake_idle() will wake a task on an idle cpu if task->cpu is
1410 * not idle and an idle cpu is available. The span of cpus to
1411 * search starts with cpus closest then further out as needed,
1412 * so we always favor a closer, idle cpu.
1414 * Returns the CPU we should wake onto.
1416 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1417 static int wake_idle(int cpu, struct task_struct *p)
1420 struct sched_domain *sd;
1424 * If it is idle, then it is the best cpu to run this task.
1426 * This cpu is also the best, if it has more than one task already.
1427 * Siblings must be also busy(in most cases) as they didn't already
1428 * pickup the extra load from this cpu and hence we need not check
1429 * sibling runqueue info. This will avoid the checks and cache miss
1430 * penalities associated with that.
1432 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1435 for_each_domain(cpu, sd) {
1436 if (sd->flags & SD_WAKE_IDLE) {
1437 cpus_and(tmp, sd->span, p->cpus_allowed);
1438 for_each_cpu_mask(i, tmp) {
1440 if (i != task_cpu(p)) {
1442 se.nr_wakeups_idle);
1454 static inline int wake_idle(int cpu, struct task_struct *p)
1461 * try_to_wake_up - wake up a thread
1462 * @p: the to-be-woken-up thread
1463 * @state: the mask of task states that can be woken
1464 * @sync: do a synchronous wakeup?
1466 * Put it on the run-queue if it's not already there. The "current"
1467 * thread is always on the run-queue (except when the actual
1468 * re-schedule is in progress), and as such you're allowed to do
1469 * the simpler "current->state = TASK_RUNNING" to mark yourself
1470 * runnable without the overhead of this.
1472 * returns failure only if the task is already active.
1474 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1476 int cpu, orig_cpu, this_cpu, success = 0;
1477 unsigned long flags;
1481 struct sched_domain *sd, *this_sd = NULL;
1482 unsigned long load, this_load;
1486 rq = task_rq_lock(p, &flags);
1487 old_state = p->state;
1488 if (!(old_state & state))
1496 this_cpu = smp_processor_id();
1499 if (unlikely(task_running(rq, p)))
1504 schedstat_inc(rq, ttwu_count);
1505 if (cpu == this_cpu) {
1506 schedstat_inc(rq, ttwu_local);
1510 for_each_domain(this_cpu, sd) {
1511 if (cpu_isset(cpu, sd->span)) {
1512 schedstat_inc(sd, ttwu_wake_remote);
1518 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1522 * Check for affine wakeup and passive balancing possibilities.
1525 int idx = this_sd->wake_idx;
1526 unsigned int imbalance;
1528 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1530 load = source_load(cpu, idx);
1531 this_load = target_load(this_cpu, idx);
1533 new_cpu = this_cpu; /* Wake to this CPU if we can */
1535 if (this_sd->flags & SD_WAKE_AFFINE) {
1536 unsigned long tl = this_load;
1537 unsigned long tl_per_task;
1540 * Attract cache-cold tasks on sync wakeups:
1542 if (sync && !task_hot(p, rq->clock, this_sd))
1545 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1546 tl_per_task = cpu_avg_load_per_task(this_cpu);
1549 * If sync wakeup then subtract the (maximum possible)
1550 * effect of the currently running task from the load
1551 * of the current CPU:
1554 tl -= current->se.load.weight;
1557 tl + target_load(cpu, idx) <= tl_per_task) ||
1558 100*(tl + p->se.load.weight) <= imbalance*load) {
1560 * This domain has SD_WAKE_AFFINE and
1561 * p is cache cold in this domain, and
1562 * there is no bad imbalance.
1564 schedstat_inc(this_sd, ttwu_move_affine);
1565 schedstat_inc(p, se.nr_wakeups_affine);
1571 * Start passive balancing when half the imbalance_pct
1574 if (this_sd->flags & SD_WAKE_BALANCE) {
1575 if (imbalance*this_load <= 100*load) {
1576 schedstat_inc(this_sd, ttwu_move_balance);
1577 schedstat_inc(p, se.nr_wakeups_passive);
1583 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1585 new_cpu = wake_idle(new_cpu, p);
1586 if (new_cpu != cpu) {
1587 set_task_cpu(p, new_cpu);
1588 task_rq_unlock(rq, &flags);
1589 /* might preempt at this point */
1590 rq = task_rq_lock(p, &flags);
1591 old_state = p->state;
1592 if (!(old_state & state))
1597 this_cpu = smp_processor_id();
1602 #endif /* CONFIG_SMP */
1603 schedstat_inc(p, se.nr_wakeups);
1605 schedstat_inc(p, se.nr_wakeups_sync);
1606 if (orig_cpu != cpu)
1607 schedstat_inc(p, se.nr_wakeups_migrate);
1608 if (cpu == this_cpu)
1609 schedstat_inc(p, se.nr_wakeups_local);
1611 schedstat_inc(p, se.nr_wakeups_remote);
1612 update_rq_clock(rq);
1613 activate_task(rq, p, 1);
1614 check_preempt_curr(rq, p);
1618 p->state = TASK_RUNNING;
1620 task_rq_unlock(rq, &flags);
1625 int fastcall wake_up_process(struct task_struct *p)
1627 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1628 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1630 EXPORT_SYMBOL(wake_up_process);
1632 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1634 return try_to_wake_up(p, state, 0);
1638 * Perform scheduler related setup for a newly forked process p.
1639 * p is forked by current.
1641 * __sched_fork() is basic setup used by init_idle() too:
1643 static void __sched_fork(struct task_struct *p)
1645 p->se.exec_start = 0;
1646 p->se.sum_exec_runtime = 0;
1647 p->se.prev_sum_exec_runtime = 0;
1649 #ifdef CONFIG_SCHEDSTATS
1650 p->se.wait_start = 0;
1651 p->se.sum_sleep_runtime = 0;
1652 p->se.sleep_start = 0;
1653 p->se.block_start = 0;
1654 p->se.sleep_max = 0;
1655 p->se.block_max = 0;
1657 p->se.slice_max = 0;
1661 INIT_LIST_HEAD(&p->run_list);
1664 #ifdef CONFIG_PREEMPT_NOTIFIERS
1665 INIT_HLIST_HEAD(&p->preempt_notifiers);
1669 * We mark the process as running here, but have not actually
1670 * inserted it onto the runqueue yet. This guarantees that
1671 * nobody will actually run it, and a signal or other external
1672 * event cannot wake it up and insert it on the runqueue either.
1674 p->state = TASK_RUNNING;
1678 * fork()/clone()-time setup:
1680 void sched_fork(struct task_struct *p, int clone_flags)
1682 int cpu = get_cpu();
1687 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1689 set_task_cpu(p, cpu);
1692 * Make sure we do not leak PI boosting priority to the child:
1694 p->prio = current->normal_prio;
1695 if (!rt_prio(p->prio))
1696 p->sched_class = &fair_sched_class;
1698 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1699 if (likely(sched_info_on()))
1700 memset(&p->sched_info, 0, sizeof(p->sched_info));
1702 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1705 #ifdef CONFIG_PREEMPT
1706 /* Want to start with kernel preemption disabled. */
1707 task_thread_info(p)->preempt_count = 1;
1713 * wake_up_new_task - wake up a newly created task for the first time.
1715 * This function will do some initial scheduler statistics housekeeping
1716 * that must be done for every newly created context, then puts the task
1717 * on the runqueue and wakes it.
1719 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1721 unsigned long flags;
1724 rq = task_rq_lock(p, &flags);
1725 BUG_ON(p->state != TASK_RUNNING);
1726 update_rq_clock(rq);
1728 p->prio = effective_prio(p);
1730 if (!p->sched_class->task_new || !current->se.on_rq) {
1731 activate_task(rq, p, 0);
1734 * Let the scheduling class do new task startup
1735 * management (if any):
1737 p->sched_class->task_new(rq, p);
1738 inc_nr_running(p, rq);
1740 check_preempt_curr(rq, p);
1741 task_rq_unlock(rq, &flags);
1744 #ifdef CONFIG_PREEMPT_NOTIFIERS
1747 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1748 * @notifier: notifier struct to register
1750 void preempt_notifier_register(struct preempt_notifier *notifier)
1752 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1754 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1757 * preempt_notifier_unregister - no longer interested in preemption notifications
1758 * @notifier: notifier struct to unregister
1760 * This is safe to call from within a preemption notifier.
1762 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1764 hlist_del(¬ifier->link);
1766 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1768 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1770 struct preempt_notifier *notifier;
1771 struct hlist_node *node;
1773 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1774 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1778 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1779 struct task_struct *next)
1781 struct preempt_notifier *notifier;
1782 struct hlist_node *node;
1784 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1785 notifier->ops->sched_out(notifier, next);
1790 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1795 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1796 struct task_struct *next)
1803 * prepare_task_switch - prepare to switch tasks
1804 * @rq: the runqueue preparing to switch
1805 * @prev: the current task that is being switched out
1806 * @next: the task we are going to switch to.
1808 * This is called with the rq lock held and interrupts off. It must
1809 * be paired with a subsequent finish_task_switch after the context
1812 * prepare_task_switch sets up locking and calls architecture specific
1816 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1817 struct task_struct *next)
1819 fire_sched_out_preempt_notifiers(prev, next);
1820 prepare_lock_switch(rq, next);
1821 prepare_arch_switch(next);
1825 * finish_task_switch - clean up after a task-switch
1826 * @rq: runqueue associated with task-switch
1827 * @prev: the thread we just switched away from.
1829 * finish_task_switch must be called after the context switch, paired
1830 * with a prepare_task_switch call before the context switch.
1831 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1832 * and do any other architecture-specific cleanup actions.
1834 * Note that we may have delayed dropping an mm in context_switch(). If
1835 * so, we finish that here outside of the runqueue lock. (Doing it
1836 * with the lock held can cause deadlocks; see schedule() for
1839 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1840 __releases(rq->lock)
1842 struct mm_struct *mm = rq->prev_mm;
1848 * A task struct has one reference for the use as "current".
1849 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1850 * schedule one last time. The schedule call will never return, and
1851 * the scheduled task must drop that reference.
1852 * The test for TASK_DEAD must occur while the runqueue locks are
1853 * still held, otherwise prev could be scheduled on another cpu, die
1854 * there before we look at prev->state, and then the reference would
1856 * Manfred Spraul <manfred@colorfullife.com>
1858 prev_state = prev->state;
1859 finish_arch_switch(prev);
1860 finish_lock_switch(rq, prev);
1861 fire_sched_in_preempt_notifiers(current);
1864 if (unlikely(prev_state == TASK_DEAD)) {
1866 * Remove function-return probe instances associated with this
1867 * task and put them back on the free list.
1869 kprobe_flush_task(prev);
1870 put_task_struct(prev);
1875 * schedule_tail - first thing a freshly forked thread must call.
1876 * @prev: the thread we just switched away from.
1878 asmlinkage void schedule_tail(struct task_struct *prev)
1879 __releases(rq->lock)
1881 struct rq *rq = this_rq();
1883 finish_task_switch(rq, prev);
1884 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1885 /* In this case, finish_task_switch does not reenable preemption */
1888 if (current->set_child_tid)
1889 put_user(task_pid_vnr(current), current->set_child_tid);
1893 * context_switch - switch to the new MM and the new
1894 * thread's register state.
1897 context_switch(struct rq *rq, struct task_struct *prev,
1898 struct task_struct *next)
1900 struct mm_struct *mm, *oldmm;
1902 prepare_task_switch(rq, prev, next);
1904 oldmm = prev->active_mm;
1906 * For paravirt, this is coupled with an exit in switch_to to
1907 * combine the page table reload and the switch backend into
1910 arch_enter_lazy_cpu_mode();
1912 if (unlikely(!mm)) {
1913 next->active_mm = oldmm;
1914 atomic_inc(&oldmm->mm_count);
1915 enter_lazy_tlb(oldmm, next);
1917 switch_mm(oldmm, mm, next);
1919 if (unlikely(!prev->mm)) {
1920 prev->active_mm = NULL;
1921 rq->prev_mm = oldmm;
1924 * Since the runqueue lock will be released by the next
1925 * task (which is an invalid locking op but in the case
1926 * of the scheduler it's an obvious special-case), so we
1927 * do an early lockdep release here:
1929 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1930 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1933 /* Here we just switch the register state and the stack. */
1934 switch_to(prev, next, prev);
1938 * this_rq must be evaluated again because prev may have moved
1939 * CPUs since it called schedule(), thus the 'rq' on its stack
1940 * frame will be invalid.
1942 finish_task_switch(this_rq(), prev);
1946 * nr_running, nr_uninterruptible and nr_context_switches:
1948 * externally visible scheduler statistics: current number of runnable
1949 * threads, current number of uninterruptible-sleeping threads, total
1950 * number of context switches performed since bootup.
1952 unsigned long nr_running(void)
1954 unsigned long i, sum = 0;
1956 for_each_online_cpu(i)
1957 sum += cpu_rq(i)->nr_running;
1962 unsigned long nr_uninterruptible(void)
1964 unsigned long i, sum = 0;
1966 for_each_possible_cpu(i)
1967 sum += cpu_rq(i)->nr_uninterruptible;
1970 * Since we read the counters lockless, it might be slightly
1971 * inaccurate. Do not allow it to go below zero though:
1973 if (unlikely((long)sum < 0))
1979 unsigned long long nr_context_switches(void)
1982 unsigned long long sum = 0;
1984 for_each_possible_cpu(i)
1985 sum += cpu_rq(i)->nr_switches;
1990 unsigned long nr_iowait(void)
1992 unsigned long i, sum = 0;
1994 for_each_possible_cpu(i)
1995 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2000 unsigned long nr_active(void)
2002 unsigned long i, running = 0, uninterruptible = 0;
2004 for_each_online_cpu(i) {
2005 running += cpu_rq(i)->nr_running;
2006 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2009 if (unlikely((long)uninterruptible < 0))
2010 uninterruptible = 0;
2012 return running + uninterruptible;
2016 * Update rq->cpu_load[] statistics. This function is usually called every
2017 * scheduler tick (TICK_NSEC).
2019 static void update_cpu_load(struct rq *this_rq)
2021 unsigned long this_load = this_rq->load.weight;
2024 this_rq->nr_load_updates++;
2026 /* Update our load: */
2027 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2028 unsigned long old_load, new_load;
2030 /* scale is effectively 1 << i now, and >> i divides by scale */
2032 old_load = this_rq->cpu_load[i];
2033 new_load = this_load;
2035 * Round up the averaging division if load is increasing. This
2036 * prevents us from getting stuck on 9 if the load is 10, for
2039 if (new_load > old_load)
2040 new_load += scale-1;
2041 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2048 * double_rq_lock - safely lock two runqueues
2050 * Note this does not disable interrupts like task_rq_lock,
2051 * you need to do so manually before calling.
2053 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2054 __acquires(rq1->lock)
2055 __acquires(rq2->lock)
2057 BUG_ON(!irqs_disabled());
2059 spin_lock(&rq1->lock);
2060 __acquire(rq2->lock); /* Fake it out ;) */
2063 spin_lock(&rq1->lock);
2064 spin_lock(&rq2->lock);
2066 spin_lock(&rq2->lock);
2067 spin_lock(&rq1->lock);
2070 update_rq_clock(rq1);
2071 update_rq_clock(rq2);
2075 * double_rq_unlock - safely unlock two runqueues
2077 * Note this does not restore interrupts like task_rq_unlock,
2078 * you need to do so manually after calling.
2080 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2081 __releases(rq1->lock)
2082 __releases(rq2->lock)
2084 spin_unlock(&rq1->lock);
2086 spin_unlock(&rq2->lock);
2088 __release(rq2->lock);
2092 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2094 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2095 __releases(this_rq->lock)
2096 __acquires(busiest->lock)
2097 __acquires(this_rq->lock)
2099 if (unlikely(!irqs_disabled())) {
2100 /* printk() doesn't work good under rq->lock */
2101 spin_unlock(&this_rq->lock);
2104 if (unlikely(!spin_trylock(&busiest->lock))) {
2105 if (busiest < this_rq) {
2106 spin_unlock(&this_rq->lock);
2107 spin_lock(&busiest->lock);
2108 spin_lock(&this_rq->lock);
2110 spin_lock(&busiest->lock);
2115 * If dest_cpu is allowed for this process, migrate the task to it.
2116 * This is accomplished by forcing the cpu_allowed mask to only
2117 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2118 * the cpu_allowed mask is restored.
2120 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2122 struct migration_req req;
2123 unsigned long flags;
2126 rq = task_rq_lock(p, &flags);
2127 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2128 || unlikely(cpu_is_offline(dest_cpu)))
2131 /* force the process onto the specified CPU */
2132 if (migrate_task(p, dest_cpu, &req)) {
2133 /* Need to wait for migration thread (might exit: take ref). */
2134 struct task_struct *mt = rq->migration_thread;
2136 get_task_struct(mt);
2137 task_rq_unlock(rq, &flags);
2138 wake_up_process(mt);
2139 put_task_struct(mt);
2140 wait_for_completion(&req.done);
2145 task_rq_unlock(rq, &flags);
2149 * sched_exec - execve() is a valuable balancing opportunity, because at
2150 * this point the task has the smallest effective memory and cache footprint.
2152 void sched_exec(void)
2154 int new_cpu, this_cpu = get_cpu();
2155 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2157 if (new_cpu != this_cpu)
2158 sched_migrate_task(current, new_cpu);
2162 * pull_task - move a task from a remote runqueue to the local runqueue.
2163 * Both runqueues must be locked.
2165 static void pull_task(struct rq *src_rq, struct task_struct *p,
2166 struct rq *this_rq, int this_cpu)
2168 deactivate_task(src_rq, p, 0);
2169 set_task_cpu(p, this_cpu);
2170 activate_task(this_rq, p, 0);
2172 * Note that idle threads have a prio of MAX_PRIO, for this test
2173 * to be always true for them.
2175 check_preempt_curr(this_rq, p);
2179 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2182 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2183 struct sched_domain *sd, enum cpu_idle_type idle,
2187 * We do not migrate tasks that are:
2188 * 1) running (obviously), or
2189 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2190 * 3) are cache-hot on their current CPU.
2192 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2193 schedstat_inc(p, se.nr_failed_migrations_affine);
2198 if (task_running(rq, p)) {
2199 schedstat_inc(p, se.nr_failed_migrations_running);
2204 * Aggressive migration if:
2205 * 1) task is cache cold, or
2206 * 2) too many balance attempts have failed.
2209 if (!task_hot(p, rq->clock, sd) ||
2210 sd->nr_balance_failed > sd->cache_nice_tries) {
2211 #ifdef CONFIG_SCHEDSTATS
2212 if (task_hot(p, rq->clock, sd)) {
2213 schedstat_inc(sd, lb_hot_gained[idle]);
2214 schedstat_inc(p, se.nr_forced_migrations);
2220 if (task_hot(p, rq->clock, sd)) {
2221 schedstat_inc(p, se.nr_failed_migrations_hot);
2227 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2228 unsigned long max_nr_move, unsigned long max_load_move,
2229 struct sched_domain *sd, enum cpu_idle_type idle,
2230 int *all_pinned, unsigned long *load_moved,
2231 int *this_best_prio, struct rq_iterator *iterator)
2233 int pulled = 0, pinned = 0, skip_for_load;
2234 struct task_struct *p;
2235 long rem_load_move = max_load_move;
2237 if (max_nr_move == 0 || max_load_move == 0)
2243 * Start the load-balancing iterator:
2245 p = iterator->start(iterator->arg);
2250 * To help distribute high priority tasks accross CPUs we don't
2251 * skip a task if it will be the highest priority task (i.e. smallest
2252 * prio value) on its new queue regardless of its load weight
2254 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2255 SCHED_LOAD_SCALE_FUZZ;
2256 if ((skip_for_load && p->prio >= *this_best_prio) ||
2257 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2258 p = iterator->next(iterator->arg);
2262 pull_task(busiest, p, this_rq, this_cpu);
2264 rem_load_move -= p->se.load.weight;
2267 * We only want to steal up to the prescribed number of tasks
2268 * and the prescribed amount of weighted load.
2270 if (pulled < max_nr_move && rem_load_move > 0) {
2271 if (p->prio < *this_best_prio)
2272 *this_best_prio = p->prio;
2273 p = iterator->next(iterator->arg);
2278 * Right now, this is the only place pull_task() is called,
2279 * so we can safely collect pull_task() stats here rather than
2280 * inside pull_task().
2282 schedstat_add(sd, lb_gained[idle], pulled);
2285 *all_pinned = pinned;
2286 *load_moved = max_load_move - rem_load_move;
2291 * move_tasks tries to move up to max_load_move weighted load from busiest to
2292 * this_rq, as part of a balancing operation within domain "sd".
2293 * Returns 1 if successful and 0 otherwise.
2295 * Called with both runqueues locked.
2297 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2298 unsigned long max_load_move,
2299 struct sched_domain *sd, enum cpu_idle_type idle,
2302 const struct sched_class *class = sched_class_highest;
2303 unsigned long total_load_moved = 0;
2304 int this_best_prio = this_rq->curr->prio;
2308 class->load_balance(this_rq, this_cpu, busiest,
2309 ULONG_MAX, max_load_move - total_load_moved,
2310 sd, idle, all_pinned, &this_best_prio);
2311 class = class->next;
2312 } while (class && max_load_move > total_load_moved);
2314 return total_load_moved > 0;
2318 * move_one_task tries to move exactly one task from busiest to this_rq, as
2319 * part of active balancing operations within "domain".
2320 * Returns 1 if successful and 0 otherwise.
2322 * Called with both runqueues locked.
2324 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2325 struct sched_domain *sd, enum cpu_idle_type idle)
2327 const struct sched_class *class;
2328 int this_best_prio = MAX_PRIO;
2330 for (class = sched_class_highest; class; class = class->next)
2331 if (class->load_balance(this_rq, this_cpu, busiest,
2332 1, ULONG_MAX, sd, idle, NULL,
2340 * find_busiest_group finds and returns the busiest CPU group within the
2341 * domain. It calculates and returns the amount of weighted load which
2342 * should be moved to restore balance via the imbalance parameter.
2344 static struct sched_group *
2345 find_busiest_group(struct sched_domain *sd, int this_cpu,
2346 unsigned long *imbalance, enum cpu_idle_type idle,
2347 int *sd_idle, cpumask_t *cpus, int *balance)
2349 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2350 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2351 unsigned long max_pull;
2352 unsigned long busiest_load_per_task, busiest_nr_running;
2353 unsigned long this_load_per_task, this_nr_running;
2354 int load_idx, group_imb = 0;
2355 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2356 int power_savings_balance = 1;
2357 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2358 unsigned long min_nr_running = ULONG_MAX;
2359 struct sched_group *group_min = NULL, *group_leader = NULL;
2362 max_load = this_load = total_load = total_pwr = 0;
2363 busiest_load_per_task = busiest_nr_running = 0;
2364 this_load_per_task = this_nr_running = 0;
2365 if (idle == CPU_NOT_IDLE)
2366 load_idx = sd->busy_idx;
2367 else if (idle == CPU_NEWLY_IDLE)
2368 load_idx = sd->newidle_idx;
2370 load_idx = sd->idle_idx;
2373 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2376 int __group_imb = 0;
2377 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2378 unsigned long sum_nr_running, sum_weighted_load;
2380 local_group = cpu_isset(this_cpu, group->cpumask);
2383 balance_cpu = first_cpu(group->cpumask);
2385 /* Tally up the load of all CPUs in the group */
2386 sum_weighted_load = sum_nr_running = avg_load = 0;
2388 min_cpu_load = ~0UL;
2390 for_each_cpu_mask(i, group->cpumask) {
2393 if (!cpu_isset(i, *cpus))
2398 if (*sd_idle && rq->nr_running)
2401 /* Bias balancing toward cpus of our domain */
2403 if (idle_cpu(i) && !first_idle_cpu) {
2408 load = target_load(i, load_idx);
2410 load = source_load(i, load_idx);
2411 if (load > max_cpu_load)
2412 max_cpu_load = load;
2413 if (min_cpu_load > load)
2414 min_cpu_load = load;
2418 sum_nr_running += rq->nr_running;
2419 sum_weighted_load += weighted_cpuload(i);
2423 * First idle cpu or the first cpu(busiest) in this sched group
2424 * is eligible for doing load balancing at this and above
2425 * domains. In the newly idle case, we will allow all the cpu's
2426 * to do the newly idle load balance.
2428 if (idle != CPU_NEWLY_IDLE && local_group &&
2429 balance_cpu != this_cpu && balance) {
2434 total_load += avg_load;
2435 total_pwr += group->__cpu_power;
2437 /* Adjust by relative CPU power of the group */
2438 avg_load = sg_div_cpu_power(group,
2439 avg_load * SCHED_LOAD_SCALE);
2441 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2444 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2447 this_load = avg_load;
2449 this_nr_running = sum_nr_running;
2450 this_load_per_task = sum_weighted_load;
2451 } else if (avg_load > max_load &&
2452 (sum_nr_running > group_capacity || __group_imb)) {
2453 max_load = avg_load;
2455 busiest_nr_running = sum_nr_running;
2456 busiest_load_per_task = sum_weighted_load;
2457 group_imb = __group_imb;
2460 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2462 * Busy processors will not participate in power savings
2465 if (idle == CPU_NOT_IDLE ||
2466 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2470 * If the local group is idle or completely loaded
2471 * no need to do power savings balance at this domain
2473 if (local_group && (this_nr_running >= group_capacity ||
2475 power_savings_balance = 0;
2478 * If a group is already running at full capacity or idle,
2479 * don't include that group in power savings calculations
2481 if (!power_savings_balance || sum_nr_running >= group_capacity
2486 * Calculate the group which has the least non-idle load.
2487 * This is the group from where we need to pick up the load
2490 if ((sum_nr_running < min_nr_running) ||
2491 (sum_nr_running == min_nr_running &&
2492 first_cpu(group->cpumask) <
2493 first_cpu(group_min->cpumask))) {
2495 min_nr_running = sum_nr_running;
2496 min_load_per_task = sum_weighted_load /
2501 * Calculate the group which is almost near its
2502 * capacity but still has some space to pick up some load
2503 * from other group and save more power
2505 if (sum_nr_running <= group_capacity - 1) {
2506 if (sum_nr_running > leader_nr_running ||
2507 (sum_nr_running == leader_nr_running &&
2508 first_cpu(group->cpumask) >
2509 first_cpu(group_leader->cpumask))) {
2510 group_leader = group;
2511 leader_nr_running = sum_nr_running;
2516 group = group->next;
2517 } while (group != sd->groups);
2519 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2522 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2524 if (this_load >= avg_load ||
2525 100*max_load <= sd->imbalance_pct*this_load)
2528 busiest_load_per_task /= busiest_nr_running;
2530 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2533 * We're trying to get all the cpus to the average_load, so we don't
2534 * want to push ourselves above the average load, nor do we wish to
2535 * reduce the max loaded cpu below the average load, as either of these
2536 * actions would just result in more rebalancing later, and ping-pong
2537 * tasks around. Thus we look for the minimum possible imbalance.
2538 * Negative imbalances (*we* are more loaded than anyone else) will
2539 * be counted as no imbalance for these purposes -- we can't fix that
2540 * by pulling tasks to us. Be careful of negative numbers as they'll
2541 * appear as very large values with unsigned longs.
2543 if (max_load <= busiest_load_per_task)
2547 * In the presence of smp nice balancing, certain scenarios can have
2548 * max load less than avg load(as we skip the groups at or below
2549 * its cpu_power, while calculating max_load..)
2551 if (max_load < avg_load) {
2553 goto small_imbalance;
2556 /* Don't want to pull so many tasks that a group would go idle */
2557 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2559 /* How much load to actually move to equalise the imbalance */
2560 *imbalance = min(max_pull * busiest->__cpu_power,
2561 (avg_load - this_load) * this->__cpu_power)
2565 * if *imbalance is less than the average load per runnable task
2566 * there is no gaurantee that any tasks will be moved so we'll have
2567 * a think about bumping its value to force at least one task to be
2570 if (*imbalance < busiest_load_per_task) {
2571 unsigned long tmp, pwr_now, pwr_move;
2575 pwr_move = pwr_now = 0;
2577 if (this_nr_running) {
2578 this_load_per_task /= this_nr_running;
2579 if (busiest_load_per_task > this_load_per_task)
2582 this_load_per_task = SCHED_LOAD_SCALE;
2584 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2585 busiest_load_per_task * imbn) {
2586 *imbalance = busiest_load_per_task;
2591 * OK, we don't have enough imbalance to justify moving tasks,
2592 * however we may be able to increase total CPU power used by
2596 pwr_now += busiest->__cpu_power *
2597 min(busiest_load_per_task, max_load);
2598 pwr_now += this->__cpu_power *
2599 min(this_load_per_task, this_load);
2600 pwr_now /= SCHED_LOAD_SCALE;
2602 /* Amount of load we'd subtract */
2603 tmp = sg_div_cpu_power(busiest,
2604 busiest_load_per_task * SCHED_LOAD_SCALE);
2606 pwr_move += busiest->__cpu_power *
2607 min(busiest_load_per_task, max_load - tmp);
2609 /* Amount of load we'd add */
2610 if (max_load * busiest->__cpu_power <
2611 busiest_load_per_task * SCHED_LOAD_SCALE)
2612 tmp = sg_div_cpu_power(this,
2613 max_load * busiest->__cpu_power);
2615 tmp = sg_div_cpu_power(this,
2616 busiest_load_per_task * SCHED_LOAD_SCALE);
2617 pwr_move += this->__cpu_power *
2618 min(this_load_per_task, this_load + tmp);
2619 pwr_move /= SCHED_LOAD_SCALE;
2621 /* Move if we gain throughput */
2622 if (pwr_move > pwr_now)
2623 *imbalance = busiest_load_per_task;
2629 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2630 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2633 if (this == group_leader && group_leader != group_min) {
2634 *imbalance = min_load_per_task;
2644 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2647 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2648 unsigned long imbalance, cpumask_t *cpus)
2650 struct rq *busiest = NULL, *rq;
2651 unsigned long max_load = 0;
2654 for_each_cpu_mask(i, group->cpumask) {
2657 if (!cpu_isset(i, *cpus))
2661 wl = weighted_cpuload(i);
2663 if (rq->nr_running == 1 && wl > imbalance)
2666 if (wl > max_load) {
2676 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2677 * so long as it is large enough.
2679 #define MAX_PINNED_INTERVAL 512
2682 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2683 * tasks if there is an imbalance.
2685 static int load_balance(int this_cpu, struct rq *this_rq,
2686 struct sched_domain *sd, enum cpu_idle_type idle,
2689 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2690 struct sched_group *group;
2691 unsigned long imbalance;
2693 cpumask_t cpus = CPU_MASK_ALL;
2694 unsigned long flags;
2697 * When power savings policy is enabled for the parent domain, idle
2698 * sibling can pick up load irrespective of busy siblings. In this case,
2699 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2700 * portraying it as CPU_NOT_IDLE.
2702 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2703 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2706 schedstat_inc(sd, lb_count[idle]);
2709 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2716 schedstat_inc(sd, lb_nobusyg[idle]);
2720 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2722 schedstat_inc(sd, lb_nobusyq[idle]);
2726 BUG_ON(busiest == this_rq);
2728 schedstat_add(sd, lb_imbalance[idle], imbalance);
2731 if (busiest->nr_running > 1) {
2733 * Attempt to move tasks. If find_busiest_group has found
2734 * an imbalance but busiest->nr_running <= 1, the group is
2735 * still unbalanced. ld_moved simply stays zero, so it is
2736 * correctly treated as an imbalance.
2738 local_irq_save(flags);
2739 double_rq_lock(this_rq, busiest);
2740 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2741 imbalance, sd, idle, &all_pinned);
2742 double_rq_unlock(this_rq, busiest);
2743 local_irq_restore(flags);
2746 * some other cpu did the load balance for us.
2748 if (ld_moved && this_cpu != smp_processor_id())
2749 resched_cpu(this_cpu);
2751 /* All tasks on this runqueue were pinned by CPU affinity */
2752 if (unlikely(all_pinned)) {
2753 cpu_clear(cpu_of(busiest), cpus);
2754 if (!cpus_empty(cpus))
2761 schedstat_inc(sd, lb_failed[idle]);
2762 sd->nr_balance_failed++;
2764 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2766 spin_lock_irqsave(&busiest->lock, flags);
2768 /* don't kick the migration_thread, if the curr
2769 * task on busiest cpu can't be moved to this_cpu
2771 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2772 spin_unlock_irqrestore(&busiest->lock, flags);
2774 goto out_one_pinned;
2777 if (!busiest->active_balance) {
2778 busiest->active_balance = 1;
2779 busiest->push_cpu = this_cpu;
2782 spin_unlock_irqrestore(&busiest->lock, flags);
2784 wake_up_process(busiest->migration_thread);
2787 * We've kicked active balancing, reset the failure
2790 sd->nr_balance_failed = sd->cache_nice_tries+1;
2793 sd->nr_balance_failed = 0;
2795 if (likely(!active_balance)) {
2796 /* We were unbalanced, so reset the balancing interval */
2797 sd->balance_interval = sd->min_interval;
2800 * If we've begun active balancing, start to back off. This
2801 * case may not be covered by the all_pinned logic if there
2802 * is only 1 task on the busy runqueue (because we don't call
2805 if (sd->balance_interval < sd->max_interval)
2806 sd->balance_interval *= 2;
2809 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2810 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2815 schedstat_inc(sd, lb_balanced[idle]);
2817 sd->nr_balance_failed = 0;
2820 /* tune up the balancing interval */
2821 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2822 (sd->balance_interval < sd->max_interval))
2823 sd->balance_interval *= 2;
2825 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2826 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2832 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2833 * tasks if there is an imbalance.
2835 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2836 * this_rq is locked.
2839 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2841 struct sched_group *group;
2842 struct rq *busiest = NULL;
2843 unsigned long imbalance;
2847 cpumask_t cpus = CPU_MASK_ALL;
2850 * When power savings policy is enabled for the parent domain, idle
2851 * sibling can pick up load irrespective of busy siblings. In this case,
2852 * let the state of idle sibling percolate up as IDLE, instead of
2853 * portraying it as CPU_NOT_IDLE.
2855 if (sd->flags & SD_SHARE_CPUPOWER &&
2856 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2859 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2861 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2862 &sd_idle, &cpus, NULL);
2864 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2868 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2871 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2875 BUG_ON(busiest == this_rq);
2877 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2880 if (busiest->nr_running > 1) {
2881 /* Attempt to move tasks */
2882 double_lock_balance(this_rq, busiest);
2883 /* this_rq->clock is already updated */
2884 update_rq_clock(busiest);
2885 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2886 imbalance, sd, CPU_NEWLY_IDLE,
2888 spin_unlock(&busiest->lock);
2890 if (unlikely(all_pinned)) {
2891 cpu_clear(cpu_of(busiest), cpus);
2892 if (!cpus_empty(cpus))
2898 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2899 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2900 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2903 sd->nr_balance_failed = 0;
2908 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2909 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2910 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2912 sd->nr_balance_failed = 0;
2918 * idle_balance is called by schedule() if this_cpu is about to become
2919 * idle. Attempts to pull tasks from other CPUs.
2921 static void idle_balance(int this_cpu, struct rq *this_rq)
2923 struct sched_domain *sd;
2924 int pulled_task = -1;
2925 unsigned long next_balance = jiffies + HZ;
2927 for_each_domain(this_cpu, sd) {
2928 unsigned long interval;
2930 if (!(sd->flags & SD_LOAD_BALANCE))
2933 if (sd->flags & SD_BALANCE_NEWIDLE)
2934 /* If we've pulled tasks over stop searching: */
2935 pulled_task = load_balance_newidle(this_cpu,
2938 interval = msecs_to_jiffies(sd->balance_interval);
2939 if (time_after(next_balance, sd->last_balance + interval))
2940 next_balance = sd->last_balance + interval;
2944 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2946 * We are going idle. next_balance may be set based on
2947 * a busy processor. So reset next_balance.
2949 this_rq->next_balance = next_balance;
2954 * active_load_balance is run by migration threads. It pushes running tasks
2955 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2956 * running on each physical CPU where possible, and avoids physical /
2957 * logical imbalances.
2959 * Called with busiest_rq locked.
2961 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2963 int target_cpu = busiest_rq->push_cpu;
2964 struct sched_domain *sd;
2965 struct rq *target_rq;
2967 /* Is there any task to move? */
2968 if (busiest_rq->nr_running <= 1)
2971 target_rq = cpu_rq(target_cpu);
2974 * This condition is "impossible", if it occurs
2975 * we need to fix it. Originally reported by
2976 * Bjorn Helgaas on a 128-cpu setup.
2978 BUG_ON(busiest_rq == target_rq);
2980 /* move a task from busiest_rq to target_rq */
2981 double_lock_balance(busiest_rq, target_rq);
2982 update_rq_clock(busiest_rq);
2983 update_rq_clock(target_rq);
2985 /* Search for an sd spanning us and the target CPU. */
2986 for_each_domain(target_cpu, sd) {
2987 if ((sd->flags & SD_LOAD_BALANCE) &&
2988 cpu_isset(busiest_cpu, sd->span))
2993 schedstat_inc(sd, alb_count);
2995 if (move_one_task(target_rq, target_cpu, busiest_rq,
2997 schedstat_inc(sd, alb_pushed);
2999 schedstat_inc(sd, alb_failed);
3001 spin_unlock(&target_rq->lock);
3006 atomic_t load_balancer;
3008 } nohz ____cacheline_aligned = {
3009 .load_balancer = ATOMIC_INIT(-1),
3010 .cpu_mask = CPU_MASK_NONE,
3014 * This routine will try to nominate the ilb (idle load balancing)
3015 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3016 * load balancing on behalf of all those cpus. If all the cpus in the system
3017 * go into this tickless mode, then there will be no ilb owner (as there is
3018 * no need for one) and all the cpus will sleep till the next wakeup event
3021 * For the ilb owner, tick is not stopped. And this tick will be used
3022 * for idle load balancing. ilb owner will still be part of
3025 * While stopping the tick, this cpu will become the ilb owner if there
3026 * is no other owner. And will be the owner till that cpu becomes busy
3027 * or if all cpus in the system stop their ticks at which point
3028 * there is no need for ilb owner.
3030 * When the ilb owner becomes busy, it nominates another owner, during the
3031 * next busy scheduler_tick()
3033 int select_nohz_load_balancer(int stop_tick)
3035 int cpu = smp_processor_id();
3038 cpu_set(cpu, nohz.cpu_mask);
3039 cpu_rq(cpu)->in_nohz_recently = 1;
3042 * If we are going offline and still the leader, give up!
3044 if (cpu_is_offline(cpu) &&
3045 atomic_read(&nohz.load_balancer) == cpu) {
3046 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3051 /* time for ilb owner also to sleep */
3052 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3053 if (atomic_read(&nohz.load_balancer) == cpu)
3054 atomic_set(&nohz.load_balancer, -1);
3058 if (atomic_read(&nohz.load_balancer) == -1) {
3059 /* make me the ilb owner */
3060 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3062 } else if (atomic_read(&nohz.load_balancer) == cpu)
3065 if (!cpu_isset(cpu, nohz.cpu_mask))
3068 cpu_clear(cpu, nohz.cpu_mask);
3070 if (atomic_read(&nohz.load_balancer) == cpu)
3071 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3078 static DEFINE_SPINLOCK(balancing);
3081 * It checks each scheduling domain to see if it is due to be balanced,
3082 * and initiates a balancing operation if so.
3084 * Balancing parameters are set up in arch_init_sched_domains.
3086 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3089 struct rq *rq = cpu_rq(cpu);
3090 unsigned long interval;
3091 struct sched_domain *sd;
3092 /* Earliest time when we have to do rebalance again */
3093 unsigned long next_balance = jiffies + 60*HZ;
3094 int update_next_balance = 0;
3096 for_each_domain(cpu, sd) {
3097 if (!(sd->flags & SD_LOAD_BALANCE))
3100 interval = sd->balance_interval;
3101 if (idle != CPU_IDLE)
3102 interval *= sd->busy_factor;
3104 /* scale ms to jiffies */
3105 interval = msecs_to_jiffies(interval);
3106 if (unlikely(!interval))
3108 if (interval > HZ*NR_CPUS/10)
3109 interval = HZ*NR_CPUS/10;
3112 if (sd->flags & SD_SERIALIZE) {
3113 if (!spin_trylock(&balancing))
3117 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3118 if (load_balance(cpu, rq, sd, idle, &balance)) {
3120 * We've pulled tasks over so either we're no
3121 * longer idle, or one of our SMT siblings is
3124 idle = CPU_NOT_IDLE;
3126 sd->last_balance = jiffies;
3128 if (sd->flags & SD_SERIALIZE)
3129 spin_unlock(&balancing);
3131 if (time_after(next_balance, sd->last_balance + interval)) {
3132 next_balance = sd->last_balance + interval;
3133 update_next_balance = 1;
3137 * Stop the load balance at this level. There is another
3138 * CPU in our sched group which is doing load balancing more
3146 * next_balance will be updated only when there is a need.
3147 * When the cpu is attached to null domain for ex, it will not be
3150 if (likely(update_next_balance))
3151 rq->next_balance = next_balance;
3155 * run_rebalance_domains is triggered when needed from the scheduler tick.
3156 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3157 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3159 static void run_rebalance_domains(struct softirq_action *h)
3161 int this_cpu = smp_processor_id();
3162 struct rq *this_rq = cpu_rq(this_cpu);
3163 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3164 CPU_IDLE : CPU_NOT_IDLE;
3166 rebalance_domains(this_cpu, idle);
3170 * If this cpu is the owner for idle load balancing, then do the
3171 * balancing on behalf of the other idle cpus whose ticks are
3174 if (this_rq->idle_at_tick &&
3175 atomic_read(&nohz.load_balancer) == this_cpu) {
3176 cpumask_t cpus = nohz.cpu_mask;
3180 cpu_clear(this_cpu, cpus);
3181 for_each_cpu_mask(balance_cpu, cpus) {
3183 * If this cpu gets work to do, stop the load balancing
3184 * work being done for other cpus. Next load
3185 * balancing owner will pick it up.
3190 rebalance_domains(balance_cpu, CPU_IDLE);
3192 rq = cpu_rq(balance_cpu);
3193 if (time_after(this_rq->next_balance, rq->next_balance))
3194 this_rq->next_balance = rq->next_balance;
3201 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3203 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3204 * idle load balancing owner or decide to stop the periodic load balancing,
3205 * if the whole system is idle.
3207 static inline void trigger_load_balance(struct rq *rq, int cpu)
3211 * If we were in the nohz mode recently and busy at the current
3212 * scheduler tick, then check if we need to nominate new idle
3215 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3216 rq->in_nohz_recently = 0;
3218 if (atomic_read(&nohz.load_balancer) == cpu) {
3219 cpu_clear(cpu, nohz.cpu_mask);
3220 atomic_set(&nohz.load_balancer, -1);
3223 if (atomic_read(&nohz.load_balancer) == -1) {
3225 * simple selection for now: Nominate the
3226 * first cpu in the nohz list to be the next
3229 * TBD: Traverse the sched domains and nominate
3230 * the nearest cpu in the nohz.cpu_mask.
3232 int ilb = first_cpu(nohz.cpu_mask);
3240 * If this cpu is idle and doing idle load balancing for all the
3241 * cpus with ticks stopped, is it time for that to stop?
3243 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3244 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3250 * If this cpu is idle and the idle load balancing is done by
3251 * someone else, then no need raise the SCHED_SOFTIRQ
3253 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3254 cpu_isset(cpu, nohz.cpu_mask))
3257 if (time_after_eq(jiffies, rq->next_balance))
3258 raise_softirq(SCHED_SOFTIRQ);
3261 #else /* CONFIG_SMP */
3264 * on UP we do not need to balance between CPUs:
3266 static inline void idle_balance(int cpu, struct rq *rq)
3270 /* Avoid "used but not defined" warning on UP */
3271 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3272 unsigned long max_nr_move, unsigned long max_load_move,
3273 struct sched_domain *sd, enum cpu_idle_type idle,
3274 int *all_pinned, unsigned long *load_moved,
3275 int *this_best_prio, struct rq_iterator *iterator)
3284 DEFINE_PER_CPU(struct kernel_stat, kstat);
3286 EXPORT_PER_CPU_SYMBOL(kstat);
3289 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3290 * that have not yet been banked in case the task is currently running.
3292 unsigned long long task_sched_runtime(struct task_struct *p)
3294 unsigned long flags;
3298 rq = task_rq_lock(p, &flags);
3299 ns = p->se.sum_exec_runtime;
3300 if (rq->curr == p) {
3301 update_rq_clock(rq);
3302 delta_exec = rq->clock - p->se.exec_start;
3303 if ((s64)delta_exec > 0)
3306 task_rq_unlock(rq, &flags);
3312 * Account user cpu time to a process.
3313 * @p: the process that the cpu time gets accounted to
3314 * @cputime: the cpu time spent in user space since the last update
3316 void account_user_time(struct task_struct *p, cputime_t cputime)
3318 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3320 struct rq *rq = this_rq();
3322 p->utime = cputime_add(p->utime, cputime);
3325 cpuacct_charge(p, cputime);
3327 /* Add user time to cpustat. */
3328 tmp = cputime_to_cputime64(cputime);
3329 if (TASK_NICE(p) > 0)
3330 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3332 cpustat->user = cputime64_add(cpustat->user, tmp);
3336 * Account guest cpu time to a process.
3337 * @p: the process that the cpu time gets accounted to
3338 * @cputime: the cpu time spent in virtual machine since the last update
3340 void account_guest_time(struct task_struct *p, cputime_t cputime)
3343 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3345 tmp = cputime_to_cputime64(cputime);
3347 p->utime = cputime_add(p->utime, cputime);
3348 p->gtime = cputime_add(p->gtime, cputime);
3350 cpustat->user = cputime64_add(cpustat->user, tmp);
3351 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3355 * Account scaled user cpu time to a process.
3356 * @p: the process that the cpu time gets accounted to
3357 * @cputime: the cpu time spent in user space since the last update
3359 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3361 p->utimescaled = cputime_add(p->utimescaled, cputime);
3365 * Account system cpu time to a process.
3366 * @p: the process that the cpu time gets accounted to
3367 * @hardirq_offset: the offset to subtract from hardirq_count()
3368 * @cputime: the cpu time spent in kernel space since the last update
3370 void account_system_time(struct task_struct *p, int hardirq_offset,
3373 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3374 struct rq *rq = this_rq();
3377 if (p->flags & PF_VCPU) {
3378 account_guest_time(p, cputime);
3382 p->stime = cputime_add(p->stime, cputime);
3384 /* Add system time to cpustat. */
3385 tmp = cputime_to_cputime64(cputime);
3386 if (hardirq_count() - hardirq_offset)
3387 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3388 else if (softirq_count())
3389 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3390 else if (p != rq->idle) {
3391 cpustat->system = cputime64_add(cpustat->system, tmp);
3392 cpuacct_charge(p, cputime);
3393 } else if (atomic_read(&rq->nr_iowait) > 0)
3394 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3396 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3397 /* Account for system time used */
3398 acct_update_integrals(p);
3402 * Account scaled system cpu time to a process.
3403 * @p: the process that the cpu time gets accounted to
3404 * @hardirq_offset: the offset to subtract from hardirq_count()
3405 * @cputime: the cpu time spent in kernel space since the last update
3407 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3409 p->stimescaled = cputime_add(p->stimescaled, cputime);
3413 * Account for involuntary wait time.
3414 * @p: the process from which the cpu time has been stolen
3415 * @steal: the cpu time spent in involuntary wait
3417 void account_steal_time(struct task_struct *p, cputime_t steal)
3419 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3420 cputime64_t tmp = cputime_to_cputime64(steal);
3421 struct rq *rq = this_rq();
3423 if (p == rq->idle) {
3424 p->stime = cputime_add(p->stime, steal);
3425 if (atomic_read(&rq->nr_iowait) > 0)
3426 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3428 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3430 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3431 cpuacct_charge(p, -tmp);
3436 * This function gets called by the timer code, with HZ frequency.
3437 * We call it with interrupts disabled.
3439 * It also gets called by the fork code, when changing the parent's
3442 void scheduler_tick(void)
3444 int cpu = smp_processor_id();
3445 struct rq *rq = cpu_rq(cpu);
3446 struct task_struct *curr = rq->curr;
3447 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3449 spin_lock(&rq->lock);
3450 __update_rq_clock(rq);
3452 * Let rq->clock advance by at least TICK_NSEC:
3454 if (unlikely(rq->clock < next_tick))
3455 rq->clock = next_tick;
3456 rq->tick_timestamp = rq->clock;
3457 update_cpu_load(rq);
3458 if (curr != rq->idle) /* FIXME: needed? */
3459 curr->sched_class->task_tick(rq, curr);
3460 spin_unlock(&rq->lock);
3463 rq->idle_at_tick = idle_cpu(cpu);
3464 trigger_load_balance(rq, cpu);
3468 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3470 void fastcall add_preempt_count(int val)
3475 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3477 preempt_count() += val;
3479 * Spinlock count overflowing soon?
3481 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3484 EXPORT_SYMBOL(add_preempt_count);
3486 void fastcall sub_preempt_count(int val)
3491 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3494 * Is the spinlock portion underflowing?
3496 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3497 !(preempt_count() & PREEMPT_MASK)))
3500 preempt_count() -= val;
3502 EXPORT_SYMBOL(sub_preempt_count);
3507 * Print scheduling while atomic bug:
3509 static noinline void __schedule_bug(struct task_struct *prev)
3511 struct pt_regs *regs = get_irq_regs();
3513 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3514 prev->comm, prev->pid, preempt_count());
3516 debug_show_held_locks(prev);
3517 if (irqs_disabled())
3518 print_irqtrace_events(prev);
3527 * Various schedule()-time debugging checks and statistics:
3529 static inline void schedule_debug(struct task_struct *prev)
3532 * Test if we are atomic. Since do_exit() needs to call into
3533 * schedule() atomically, we ignore that path for now.
3534 * Otherwise, whine if we are scheduling when we should not be.
3536 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3537 __schedule_bug(prev);
3539 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3541 schedstat_inc(this_rq(), sched_count);
3542 #ifdef CONFIG_SCHEDSTATS
3543 if (unlikely(prev->lock_depth >= 0)) {
3544 schedstat_inc(this_rq(), bkl_count);
3545 schedstat_inc(prev, sched_info.bkl_count);
3551 * Pick up the highest-prio task:
3553 static inline struct task_struct *
3554 pick_next_task(struct rq *rq, struct task_struct *prev)
3556 const struct sched_class *class;
3557 struct task_struct *p;
3560 * Optimization: we know that if all tasks are in
3561 * the fair class we can call that function directly:
3563 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3564 p = fair_sched_class.pick_next_task(rq);
3569 class = sched_class_highest;
3571 p = class->pick_next_task(rq);
3575 * Will never be NULL as the idle class always
3576 * returns a non-NULL p:
3578 class = class->next;
3583 * schedule() is the main scheduler function.
3585 asmlinkage void __sched schedule(void)
3587 struct task_struct *prev, *next;
3594 cpu = smp_processor_id();
3598 switch_count = &prev->nivcsw;
3600 release_kernel_lock(prev);
3601 need_resched_nonpreemptible:
3603 schedule_debug(prev);
3606 * Do the rq-clock update outside the rq lock:
3608 local_irq_disable();
3609 __update_rq_clock(rq);
3610 spin_lock(&rq->lock);
3611 clear_tsk_need_resched(prev);
3613 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3614 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3615 unlikely(signal_pending(prev)))) {
3616 prev->state = TASK_RUNNING;
3618 deactivate_task(rq, prev, 1);
3620 switch_count = &prev->nvcsw;
3623 if (unlikely(!rq->nr_running))
3624 idle_balance(cpu, rq);
3626 prev->sched_class->put_prev_task(rq, prev);
3627 next = pick_next_task(rq, prev);
3629 sched_info_switch(prev, next);
3631 if (likely(prev != next)) {
3636 context_switch(rq, prev, next); /* unlocks the rq */
3638 spin_unlock_irq(&rq->lock);
3640 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3641 cpu = smp_processor_id();
3643 goto need_resched_nonpreemptible;
3645 preempt_enable_no_resched();
3646 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3649 EXPORT_SYMBOL(schedule);
3651 #ifdef CONFIG_PREEMPT
3653 * this is the entry point to schedule() from in-kernel preemption
3654 * off of preempt_enable. Kernel preemptions off return from interrupt
3655 * occur there and call schedule directly.
3657 asmlinkage void __sched preempt_schedule(void)
3659 struct thread_info *ti = current_thread_info();
3660 #ifdef CONFIG_PREEMPT_BKL
3661 struct task_struct *task = current;
3662 int saved_lock_depth;
3665 * If there is a non-zero preempt_count or interrupts are disabled,
3666 * we do not want to preempt the current task. Just return..
3668 if (likely(ti->preempt_count || irqs_disabled()))
3672 add_preempt_count(PREEMPT_ACTIVE);
3675 * We keep the big kernel semaphore locked, but we
3676 * clear ->lock_depth so that schedule() doesnt
3677 * auto-release the semaphore:
3679 #ifdef CONFIG_PREEMPT_BKL
3680 saved_lock_depth = task->lock_depth;
3681 task->lock_depth = -1;
3684 #ifdef CONFIG_PREEMPT_BKL
3685 task->lock_depth = saved_lock_depth;
3687 sub_preempt_count(PREEMPT_ACTIVE);
3690 * Check again in case we missed a preemption opportunity
3691 * between schedule and now.
3694 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3696 EXPORT_SYMBOL(preempt_schedule);
3699 * this is the entry point to schedule() from kernel preemption
3700 * off of irq context.
3701 * Note, that this is called and return with irqs disabled. This will
3702 * protect us against recursive calling from irq.
3704 asmlinkage void __sched preempt_schedule_irq(void)
3706 struct thread_info *ti = current_thread_info();
3707 #ifdef CONFIG_PREEMPT_BKL
3708 struct task_struct *task = current;
3709 int saved_lock_depth;
3711 /* Catch callers which need to be fixed */
3712 BUG_ON(ti->preempt_count || !irqs_disabled());
3715 add_preempt_count(PREEMPT_ACTIVE);
3718 * We keep the big kernel semaphore locked, but we
3719 * clear ->lock_depth so that schedule() doesnt
3720 * auto-release the semaphore:
3722 #ifdef CONFIG_PREEMPT_BKL
3723 saved_lock_depth = task->lock_depth;
3724 task->lock_depth = -1;
3728 local_irq_disable();
3729 #ifdef CONFIG_PREEMPT_BKL
3730 task->lock_depth = saved_lock_depth;
3732 sub_preempt_count(PREEMPT_ACTIVE);
3735 * Check again in case we missed a preemption opportunity
3736 * between schedule and now.
3739 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3742 #endif /* CONFIG_PREEMPT */
3744 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3747 return try_to_wake_up(curr->private, mode, sync);
3749 EXPORT_SYMBOL(default_wake_function);
3752 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3753 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3754 * number) then we wake all the non-exclusive tasks and one exclusive task.
3756 * There are circumstances in which we can try to wake a task which has already
3757 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3758 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3760 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3761 int nr_exclusive, int sync, void *key)
3763 wait_queue_t *curr, *next;
3765 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3766 unsigned flags = curr->flags;
3768 if (curr->func(curr, mode, sync, key) &&
3769 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3775 * __wake_up - wake up threads blocked on a waitqueue.
3777 * @mode: which threads
3778 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3779 * @key: is directly passed to the wakeup function
3781 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3782 int nr_exclusive, void *key)
3784 unsigned long flags;
3786 spin_lock_irqsave(&q->lock, flags);
3787 __wake_up_common(q, mode, nr_exclusive, 0, key);
3788 spin_unlock_irqrestore(&q->lock, flags);
3790 EXPORT_SYMBOL(__wake_up);
3793 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3795 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3797 __wake_up_common(q, mode, 1, 0, NULL);
3801 * __wake_up_sync - wake up threads blocked on a waitqueue.
3803 * @mode: which threads
3804 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3806 * The sync wakeup differs that the waker knows that it will schedule
3807 * away soon, so while the target thread will be woken up, it will not
3808 * be migrated to another CPU - ie. the two threads are 'synchronized'
3809 * with each other. This can prevent needless bouncing between CPUs.
3811 * On UP it can prevent extra preemption.
3814 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3816 unsigned long flags;
3822 if (unlikely(!nr_exclusive))
3825 spin_lock_irqsave(&q->lock, flags);
3826 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3827 spin_unlock_irqrestore(&q->lock, flags);
3829 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3831 void complete(struct completion *x)
3833 unsigned long flags;
3835 spin_lock_irqsave(&x->wait.lock, flags);
3837 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3839 spin_unlock_irqrestore(&x->wait.lock, flags);
3841 EXPORT_SYMBOL(complete);
3843 void complete_all(struct completion *x)
3845 unsigned long flags;
3847 spin_lock_irqsave(&x->wait.lock, flags);
3848 x->done += UINT_MAX/2;
3849 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3851 spin_unlock_irqrestore(&x->wait.lock, flags);
3853 EXPORT_SYMBOL(complete_all);
3855 static inline long __sched
3856 do_wait_for_common(struct completion *x, long timeout, int state)
3859 DECLARE_WAITQUEUE(wait, current);
3861 wait.flags |= WQ_FLAG_EXCLUSIVE;
3862 __add_wait_queue_tail(&x->wait, &wait);
3864 if (state == TASK_INTERRUPTIBLE &&
3865 signal_pending(current)) {
3866 __remove_wait_queue(&x->wait, &wait);
3867 return -ERESTARTSYS;
3869 __set_current_state(state);
3870 spin_unlock_irq(&x->wait.lock);
3871 timeout = schedule_timeout(timeout);
3872 spin_lock_irq(&x->wait.lock);
3874 __remove_wait_queue(&x->wait, &wait);
3878 __remove_wait_queue(&x->wait, &wait);
3885 wait_for_common(struct completion *x, long timeout, int state)
3889 spin_lock_irq(&x->wait.lock);
3890 timeout = do_wait_for_common(x, timeout, state);
3891 spin_unlock_irq(&x->wait.lock);
3895 void __sched wait_for_completion(struct completion *x)
3897 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3899 EXPORT_SYMBOL(wait_for_completion);
3901 unsigned long __sched
3902 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3904 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3906 EXPORT_SYMBOL(wait_for_completion_timeout);
3908 int __sched wait_for_completion_interruptible(struct completion *x)
3910 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3911 if (t == -ERESTARTSYS)
3915 EXPORT_SYMBOL(wait_for_completion_interruptible);
3917 unsigned long __sched
3918 wait_for_completion_interruptible_timeout(struct completion *x,
3919 unsigned long timeout)
3921 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3923 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3926 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3928 unsigned long flags;
3931 init_waitqueue_entry(&wait, current);
3933 __set_current_state(state);
3935 spin_lock_irqsave(&q->lock, flags);
3936 __add_wait_queue(q, &wait);
3937 spin_unlock(&q->lock);
3938 timeout = schedule_timeout(timeout);
3939 spin_lock_irq(&q->lock);
3940 __remove_wait_queue(q, &wait);
3941 spin_unlock_irqrestore(&q->lock, flags);
3946 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3948 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3950 EXPORT_SYMBOL(interruptible_sleep_on);
3953 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3955 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3957 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3959 void __sched sleep_on(wait_queue_head_t *q)
3961 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3963 EXPORT_SYMBOL(sleep_on);
3965 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3967 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3969 EXPORT_SYMBOL(sleep_on_timeout);
3971 #ifdef CONFIG_RT_MUTEXES
3974 * rt_mutex_setprio - set the current priority of a task
3976 * @prio: prio value (kernel-internal form)
3978 * This function changes the 'effective' priority of a task. It does
3979 * not touch ->normal_prio like __setscheduler().
3981 * Used by the rt_mutex code to implement priority inheritance logic.
3983 void rt_mutex_setprio(struct task_struct *p, int prio)
3985 unsigned long flags;
3986 int oldprio, on_rq, running;
3989 BUG_ON(prio < 0 || prio > MAX_PRIO);
3991 rq = task_rq_lock(p, &flags);
3992 update_rq_clock(rq);
3995 on_rq = p->se.on_rq;
3996 running = task_running(rq, p);
3998 dequeue_task(rq, p, 0);
4000 p->sched_class->put_prev_task(rq, p);
4004 p->sched_class = &rt_sched_class;
4006 p->sched_class = &fair_sched_class;
4012 p->sched_class->set_curr_task(rq);
4013 enqueue_task(rq, p, 0);
4015 * Reschedule if we are currently running on this runqueue and
4016 * our priority decreased, or if we are not currently running on
4017 * this runqueue and our priority is higher than the current's
4020 if (p->prio > oldprio)
4021 resched_task(rq->curr);
4023 check_preempt_curr(rq, p);
4026 task_rq_unlock(rq, &flags);
4031 void set_user_nice(struct task_struct *p, long nice)
4033 int old_prio, delta, on_rq;
4034 unsigned long flags;
4037 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4040 * We have to be careful, if called from sys_setpriority(),
4041 * the task might be in the middle of scheduling on another CPU.
4043 rq = task_rq_lock(p, &flags);
4044 update_rq_clock(rq);
4046 * The RT priorities are set via sched_setscheduler(), but we still
4047 * allow the 'normal' nice value to be set - but as expected
4048 * it wont have any effect on scheduling until the task is
4049 * SCHED_FIFO/SCHED_RR:
4051 if (task_has_rt_policy(p)) {
4052 p->static_prio = NICE_TO_PRIO(nice);
4055 on_rq = p->se.on_rq;
4057 dequeue_task(rq, p, 0);
4061 p->static_prio = NICE_TO_PRIO(nice);
4064 p->prio = effective_prio(p);
4065 delta = p->prio - old_prio;
4068 enqueue_task(rq, p, 0);
4071 * If the task increased its priority or is running and
4072 * lowered its priority, then reschedule its CPU:
4074 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4075 resched_task(rq->curr);
4078 task_rq_unlock(rq, &flags);
4080 EXPORT_SYMBOL(set_user_nice);
4083 * can_nice - check if a task can reduce its nice value
4087 int can_nice(const struct task_struct *p, const int nice)
4089 /* convert nice value [19,-20] to rlimit style value [1,40] */
4090 int nice_rlim = 20 - nice;
4092 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4093 capable(CAP_SYS_NICE));
4096 #ifdef __ARCH_WANT_SYS_NICE
4099 * sys_nice - change the priority of the current process.
4100 * @increment: priority increment
4102 * sys_setpriority is a more generic, but much slower function that
4103 * does similar things.
4105 asmlinkage long sys_nice(int increment)
4110 * Setpriority might change our priority at the same moment.
4111 * We don't have to worry. Conceptually one call occurs first
4112 * and we have a single winner.
4114 if (increment < -40)
4119 nice = PRIO_TO_NICE(current->static_prio) + increment;
4125 if (increment < 0 && !can_nice(current, nice))
4128 retval = security_task_setnice(current, nice);
4132 set_user_nice(current, nice);
4139 * task_prio - return the priority value of a given task.
4140 * @p: the task in question.
4142 * This is the priority value as seen by users in /proc.
4143 * RT tasks are offset by -200. Normal tasks are centered
4144 * around 0, value goes from -16 to +15.
4146 int task_prio(const struct task_struct *p)
4148 return p->prio - MAX_RT_PRIO;
4152 * task_nice - return the nice value of a given task.
4153 * @p: the task in question.
4155 int task_nice(const struct task_struct *p)
4157 return TASK_NICE(p);
4159 EXPORT_SYMBOL_GPL(task_nice);
4162 * idle_cpu - is a given cpu idle currently?
4163 * @cpu: the processor in question.
4165 int idle_cpu(int cpu)
4167 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4171 * idle_task - return the idle task for a given cpu.
4172 * @cpu: the processor in question.
4174 struct task_struct *idle_task(int cpu)
4176 return cpu_rq(cpu)->idle;
4180 * find_process_by_pid - find a process with a matching PID value.
4181 * @pid: the pid in question.
4183 static struct task_struct *find_process_by_pid(pid_t pid)
4185 return pid ? find_task_by_vpid(pid) : current;
4188 /* Actually do priority change: must hold rq lock. */
4190 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4192 BUG_ON(p->se.on_rq);
4195 switch (p->policy) {
4199 p->sched_class = &fair_sched_class;
4203 p->sched_class = &rt_sched_class;
4207 p->rt_priority = prio;
4208 p->normal_prio = normal_prio(p);
4209 /* we are holding p->pi_lock already */
4210 p->prio = rt_mutex_getprio(p);
4215 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4216 * @p: the task in question.
4217 * @policy: new policy.
4218 * @param: structure containing the new RT priority.
4220 * NOTE that the task may be already dead.
4222 int sched_setscheduler(struct task_struct *p, int policy,
4223 struct sched_param *param)
4225 int retval, oldprio, oldpolicy = -1, on_rq, running;
4226 unsigned long flags;
4229 /* may grab non-irq protected spin_locks */
4230 BUG_ON(in_interrupt());
4232 /* double check policy once rq lock held */
4234 policy = oldpolicy = p->policy;
4235 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4236 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4237 policy != SCHED_IDLE)
4240 * Valid priorities for SCHED_FIFO and SCHED_RR are
4241 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4242 * SCHED_BATCH and SCHED_IDLE is 0.
4244 if (param->sched_priority < 0 ||
4245 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4246 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4248 if (rt_policy(policy) != (param->sched_priority != 0))
4252 * Allow unprivileged RT tasks to decrease priority:
4254 if (!capable(CAP_SYS_NICE)) {
4255 if (rt_policy(policy)) {
4256 unsigned long rlim_rtprio;
4258 if (!lock_task_sighand(p, &flags))
4260 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4261 unlock_task_sighand(p, &flags);
4263 /* can't set/change the rt policy */
4264 if (policy != p->policy && !rlim_rtprio)
4267 /* can't increase priority */
4268 if (param->sched_priority > p->rt_priority &&
4269 param->sched_priority > rlim_rtprio)
4273 * Like positive nice levels, dont allow tasks to
4274 * move out of SCHED_IDLE either:
4276 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4279 /* can't change other user's priorities */
4280 if ((current->euid != p->euid) &&
4281 (current->euid != p->uid))
4285 retval = security_task_setscheduler(p, policy, param);
4289 * make sure no PI-waiters arrive (or leave) while we are
4290 * changing the priority of the task:
4292 spin_lock_irqsave(&p->pi_lock, flags);
4294 * To be able to change p->policy safely, the apropriate
4295 * runqueue lock must be held.
4297 rq = __task_rq_lock(p);
4298 /* recheck policy now with rq lock held */
4299 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4300 policy = oldpolicy = -1;
4301 __task_rq_unlock(rq);
4302 spin_unlock_irqrestore(&p->pi_lock, flags);
4305 update_rq_clock(rq);
4306 on_rq = p->se.on_rq;
4307 running = task_running(rq, p);
4309 deactivate_task(rq, p, 0);
4311 p->sched_class->put_prev_task(rq, p);
4315 __setscheduler(rq, p, policy, param->sched_priority);
4319 p->sched_class->set_curr_task(rq);
4320 activate_task(rq, p, 0);
4322 * Reschedule if we are currently running on this runqueue and
4323 * our priority decreased, or if we are not currently running on
4324 * this runqueue and our priority is higher than the current's
4327 if (p->prio > oldprio)
4328 resched_task(rq->curr);
4330 check_preempt_curr(rq, p);
4333 __task_rq_unlock(rq);
4334 spin_unlock_irqrestore(&p->pi_lock, flags);
4336 rt_mutex_adjust_pi(p);
4340 EXPORT_SYMBOL_GPL(sched_setscheduler);
4343 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4345 struct sched_param lparam;
4346 struct task_struct *p;
4349 if (!param || pid < 0)
4351 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4356 p = find_process_by_pid(pid);
4358 retval = sched_setscheduler(p, policy, &lparam);
4365 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4366 * @pid: the pid in question.
4367 * @policy: new policy.
4368 * @param: structure containing the new RT priority.
4370 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4371 struct sched_param __user *param)
4373 /* negative values for policy are not valid */
4377 return do_sched_setscheduler(pid, policy, param);
4381 * sys_sched_setparam - set/change the RT priority of a thread
4382 * @pid: the pid in question.
4383 * @param: structure containing the new RT priority.
4385 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4387 return do_sched_setscheduler(pid, -1, param);
4391 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4392 * @pid: the pid in question.
4394 asmlinkage long sys_sched_getscheduler(pid_t pid)
4396 struct task_struct *p;
4403 read_lock(&tasklist_lock);
4404 p = find_process_by_pid(pid);
4406 retval = security_task_getscheduler(p);
4410 read_unlock(&tasklist_lock);
4415 * sys_sched_getscheduler - get the RT priority of a thread
4416 * @pid: the pid in question.
4417 * @param: structure containing the RT priority.
4419 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4421 struct sched_param lp;
4422 struct task_struct *p;
4425 if (!param || pid < 0)
4428 read_lock(&tasklist_lock);
4429 p = find_process_by_pid(pid);
4434 retval = security_task_getscheduler(p);
4438 lp.sched_priority = p->rt_priority;
4439 read_unlock(&tasklist_lock);
4442 * This one might sleep, we cannot do it with a spinlock held ...
4444 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4449 read_unlock(&tasklist_lock);
4453 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4455 cpumask_t cpus_allowed;
4456 struct task_struct *p;
4459 mutex_lock(&sched_hotcpu_mutex);
4460 read_lock(&tasklist_lock);
4462 p = find_process_by_pid(pid);
4464 read_unlock(&tasklist_lock);
4465 mutex_unlock(&sched_hotcpu_mutex);
4470 * It is not safe to call set_cpus_allowed with the
4471 * tasklist_lock held. We will bump the task_struct's
4472 * usage count and then drop tasklist_lock.
4475 read_unlock(&tasklist_lock);
4478 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4479 !capable(CAP_SYS_NICE))
4482 retval = security_task_setscheduler(p, 0, NULL);
4486 cpus_allowed = cpuset_cpus_allowed(p);
4487 cpus_and(new_mask, new_mask, cpus_allowed);
4489 retval = set_cpus_allowed(p, new_mask);
4492 cpus_allowed = cpuset_cpus_allowed(p);
4493 if (!cpus_subset(new_mask, cpus_allowed)) {
4495 * We must have raced with a concurrent cpuset
4496 * update. Just reset the cpus_allowed to the
4497 * cpuset's cpus_allowed
4499 new_mask = cpus_allowed;
4505 mutex_unlock(&sched_hotcpu_mutex);
4509 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4510 cpumask_t *new_mask)
4512 if (len < sizeof(cpumask_t)) {
4513 memset(new_mask, 0, sizeof(cpumask_t));
4514 } else if (len > sizeof(cpumask_t)) {
4515 len = sizeof(cpumask_t);
4517 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4521 * sys_sched_setaffinity - set the cpu affinity of a process
4522 * @pid: pid of the process
4523 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4524 * @user_mask_ptr: user-space pointer to the new cpu mask
4526 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4527 unsigned long __user *user_mask_ptr)
4532 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4536 return sched_setaffinity(pid, new_mask);
4540 * Represents all cpu's present in the system
4541 * In systems capable of hotplug, this map could dynamically grow
4542 * as new cpu's are detected in the system via any platform specific
4543 * method, such as ACPI for e.g.
4546 cpumask_t cpu_present_map __read_mostly;
4547 EXPORT_SYMBOL(cpu_present_map);
4550 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4551 EXPORT_SYMBOL(cpu_online_map);
4553 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4554 EXPORT_SYMBOL(cpu_possible_map);
4557 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4559 struct task_struct *p;
4562 mutex_lock(&sched_hotcpu_mutex);
4563 read_lock(&tasklist_lock);
4566 p = find_process_by_pid(pid);
4570 retval = security_task_getscheduler(p);
4574 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4577 read_unlock(&tasklist_lock);
4578 mutex_unlock(&sched_hotcpu_mutex);
4584 * sys_sched_getaffinity - get the cpu affinity of a process
4585 * @pid: pid of the process
4586 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4587 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4589 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4590 unsigned long __user *user_mask_ptr)
4595 if (len < sizeof(cpumask_t))
4598 ret = sched_getaffinity(pid, &mask);
4602 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4605 return sizeof(cpumask_t);
4609 * sys_sched_yield - yield the current processor to other threads.
4611 * This function yields the current CPU to other tasks. If there are no
4612 * other threads running on this CPU then this function will return.
4614 asmlinkage long sys_sched_yield(void)
4616 struct rq *rq = this_rq_lock();
4618 schedstat_inc(rq, yld_count);
4619 current->sched_class->yield_task(rq);
4622 * Since we are going to call schedule() anyway, there's
4623 * no need to preempt or enable interrupts:
4625 __release(rq->lock);
4626 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4627 _raw_spin_unlock(&rq->lock);
4628 preempt_enable_no_resched();
4635 static void __cond_resched(void)
4637 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4638 __might_sleep(__FILE__, __LINE__);
4641 * The BKS might be reacquired before we have dropped
4642 * PREEMPT_ACTIVE, which could trigger a second
4643 * cond_resched() call.
4646 add_preempt_count(PREEMPT_ACTIVE);
4648 sub_preempt_count(PREEMPT_ACTIVE);
4649 } while (need_resched());
4652 int __sched cond_resched(void)
4654 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4655 system_state == SYSTEM_RUNNING) {
4661 EXPORT_SYMBOL(cond_resched);
4664 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4665 * call schedule, and on return reacquire the lock.
4667 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4668 * operations here to prevent schedule() from being called twice (once via
4669 * spin_unlock(), once by hand).
4671 int cond_resched_lock(spinlock_t *lock)
4675 if (need_lockbreak(lock)) {
4681 if (need_resched() && system_state == SYSTEM_RUNNING) {
4682 spin_release(&lock->dep_map, 1, _THIS_IP_);
4683 _raw_spin_unlock(lock);
4684 preempt_enable_no_resched();
4691 EXPORT_SYMBOL(cond_resched_lock);
4693 int __sched cond_resched_softirq(void)
4695 BUG_ON(!in_softirq());
4697 if (need_resched() && system_state == SYSTEM_RUNNING) {
4705 EXPORT_SYMBOL(cond_resched_softirq);
4708 * yield - yield the current processor to other threads.
4710 * This is a shortcut for kernel-space yielding - it marks the
4711 * thread runnable and calls sys_sched_yield().
4713 void __sched yield(void)
4715 set_current_state(TASK_RUNNING);
4718 EXPORT_SYMBOL(yield);
4721 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4722 * that process accounting knows that this is a task in IO wait state.
4724 * But don't do that if it is a deliberate, throttling IO wait (this task
4725 * has set its backing_dev_info: the queue against which it should throttle)
4727 void __sched io_schedule(void)
4729 struct rq *rq = &__raw_get_cpu_var(runqueues);
4731 delayacct_blkio_start();
4732 atomic_inc(&rq->nr_iowait);
4734 atomic_dec(&rq->nr_iowait);
4735 delayacct_blkio_end();
4737 EXPORT_SYMBOL(io_schedule);
4739 long __sched io_schedule_timeout(long timeout)
4741 struct rq *rq = &__raw_get_cpu_var(runqueues);
4744 delayacct_blkio_start();
4745 atomic_inc(&rq->nr_iowait);
4746 ret = schedule_timeout(timeout);
4747 atomic_dec(&rq->nr_iowait);
4748 delayacct_blkio_end();
4753 * sys_sched_get_priority_max - return maximum RT priority.
4754 * @policy: scheduling class.
4756 * this syscall returns the maximum rt_priority that can be used
4757 * by a given scheduling class.
4759 asmlinkage long sys_sched_get_priority_max(int policy)
4766 ret = MAX_USER_RT_PRIO-1;
4778 * sys_sched_get_priority_min - return minimum RT priority.
4779 * @policy: scheduling class.
4781 * this syscall returns the minimum rt_priority that can be used
4782 * by a given scheduling class.
4784 asmlinkage long sys_sched_get_priority_min(int policy)
4802 * sys_sched_rr_get_interval - return the default timeslice of a process.
4803 * @pid: pid of the process.
4804 * @interval: userspace pointer to the timeslice value.
4806 * this syscall writes the default timeslice value of a given process
4807 * into the user-space timespec buffer. A value of '0' means infinity.
4810 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4812 struct task_struct *p;
4813 unsigned int time_slice;
4821 read_lock(&tasklist_lock);
4822 p = find_process_by_pid(pid);
4826 retval = security_task_getscheduler(p);
4830 if (p->policy == SCHED_FIFO)
4832 else if (p->policy == SCHED_RR)
4833 time_slice = DEF_TIMESLICE;
4835 struct sched_entity *se = &p->se;
4836 unsigned long flags;
4839 rq = task_rq_lock(p, &flags);
4840 time_slice = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
4841 task_rq_unlock(rq, &flags);
4843 read_unlock(&tasklist_lock);
4844 jiffies_to_timespec(time_slice, &t);
4845 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4849 read_unlock(&tasklist_lock);
4853 static const char stat_nam[] = "RSDTtZX";
4855 static void show_task(struct task_struct *p)
4857 unsigned long free = 0;
4860 state = p->state ? __ffs(p->state) + 1 : 0;
4861 printk(KERN_INFO "%-13.13s %c", p->comm,
4862 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4863 #if BITS_PER_LONG == 32
4864 if (state == TASK_RUNNING)
4865 printk(KERN_CONT " running ");
4867 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4869 if (state == TASK_RUNNING)
4870 printk(KERN_CONT " running task ");
4872 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4874 #ifdef CONFIG_DEBUG_STACK_USAGE
4876 unsigned long *n = end_of_stack(p);
4879 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4882 printk(KERN_CONT "%5lu %5d %6d\n", free,
4883 task_pid_nr(p), task_pid_nr(p->parent));
4885 if (state != TASK_RUNNING)
4886 show_stack(p, NULL);
4889 void show_state_filter(unsigned long state_filter)
4891 struct task_struct *g, *p;
4893 #if BITS_PER_LONG == 32
4895 " task PC stack pid father\n");
4898 " task PC stack pid father\n");
4900 read_lock(&tasklist_lock);
4901 do_each_thread(g, p) {
4903 * reset the NMI-timeout, listing all files on a slow
4904 * console might take alot of time:
4906 touch_nmi_watchdog();
4907 if (!state_filter || (p->state & state_filter))
4909 } while_each_thread(g, p);
4911 touch_all_softlockup_watchdogs();
4913 #ifdef CONFIG_SCHED_DEBUG
4914 sysrq_sched_debug_show();
4916 read_unlock(&tasklist_lock);
4918 * Only show locks if all tasks are dumped:
4920 if (state_filter == -1)
4921 debug_show_all_locks();
4924 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4926 idle->sched_class = &idle_sched_class;
4930 * init_idle - set up an idle thread for a given CPU
4931 * @idle: task in question
4932 * @cpu: cpu the idle task belongs to
4934 * NOTE: this function does not set the idle thread's NEED_RESCHED
4935 * flag, to make booting more robust.
4937 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4939 struct rq *rq = cpu_rq(cpu);
4940 unsigned long flags;
4943 idle->se.exec_start = sched_clock();
4945 idle->prio = idle->normal_prio = MAX_PRIO;
4946 idle->cpus_allowed = cpumask_of_cpu(cpu);
4947 __set_task_cpu(idle, cpu);
4949 spin_lock_irqsave(&rq->lock, flags);
4950 rq->curr = rq->idle = idle;
4951 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4954 spin_unlock_irqrestore(&rq->lock, flags);
4956 /* Set the preempt count _outside_ the spinlocks! */
4957 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4958 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4960 task_thread_info(idle)->preempt_count = 0;
4963 * The idle tasks have their own, simple scheduling class:
4965 idle->sched_class = &idle_sched_class;
4969 * In a system that switches off the HZ timer nohz_cpu_mask
4970 * indicates which cpus entered this state. This is used
4971 * in the rcu update to wait only for active cpus. For system
4972 * which do not switch off the HZ timer nohz_cpu_mask should
4973 * always be CPU_MASK_NONE.
4975 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4979 * This is how migration works:
4981 * 1) we queue a struct migration_req structure in the source CPU's
4982 * runqueue and wake up that CPU's migration thread.
4983 * 2) we down() the locked semaphore => thread blocks.
4984 * 3) migration thread wakes up (implicitly it forces the migrated
4985 * thread off the CPU)
4986 * 4) it gets the migration request and checks whether the migrated
4987 * task is still in the wrong runqueue.
4988 * 5) if it's in the wrong runqueue then the migration thread removes
4989 * it and puts it into the right queue.
4990 * 6) migration thread up()s the semaphore.
4991 * 7) we wake up and the migration is done.
4995 * Change a given task's CPU affinity. Migrate the thread to a
4996 * proper CPU and schedule it away if the CPU it's executing on
4997 * is removed from the allowed bitmask.
4999 * NOTE: the caller must have a valid reference to the task, the
5000 * task must not exit() & deallocate itself prematurely. The
5001 * call is not atomic; no spinlocks may be held.
5003 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5005 struct migration_req req;
5006 unsigned long flags;
5010 rq = task_rq_lock(p, &flags);
5011 if (!cpus_intersects(new_mask, cpu_online_map)) {
5016 p->cpus_allowed = new_mask;
5017 /* Can the task run on the task's current CPU? If so, we're done */
5018 if (cpu_isset(task_cpu(p), new_mask))
5021 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5022 /* Need help from migration thread: drop lock and wait. */
5023 task_rq_unlock(rq, &flags);
5024 wake_up_process(rq->migration_thread);
5025 wait_for_completion(&req.done);
5026 tlb_migrate_finish(p->mm);
5030 task_rq_unlock(rq, &flags);
5034 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5037 * Move (not current) task off this cpu, onto dest cpu. We're doing
5038 * this because either it can't run here any more (set_cpus_allowed()
5039 * away from this CPU, or CPU going down), or because we're
5040 * attempting to rebalance this task on exec (sched_exec).
5042 * So we race with normal scheduler movements, but that's OK, as long
5043 * as the task is no longer on this CPU.
5045 * Returns non-zero if task was successfully migrated.
5047 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5049 struct rq *rq_dest, *rq_src;
5052 if (unlikely(cpu_is_offline(dest_cpu)))
5055 rq_src = cpu_rq(src_cpu);
5056 rq_dest = cpu_rq(dest_cpu);
5058 double_rq_lock(rq_src, rq_dest);
5059 /* Already moved. */
5060 if (task_cpu(p) != src_cpu)
5062 /* Affinity changed (again). */
5063 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5066 on_rq = p->se.on_rq;
5068 deactivate_task(rq_src, p, 0);
5070 set_task_cpu(p, dest_cpu);
5072 activate_task(rq_dest, p, 0);
5073 check_preempt_curr(rq_dest, p);
5077 double_rq_unlock(rq_src, rq_dest);
5082 * migration_thread - this is a highprio system thread that performs
5083 * thread migration by bumping thread off CPU then 'pushing' onto
5086 static int migration_thread(void *data)
5088 int cpu = (long)data;
5092 BUG_ON(rq->migration_thread != current);
5094 set_current_state(TASK_INTERRUPTIBLE);
5095 while (!kthread_should_stop()) {
5096 struct migration_req *req;
5097 struct list_head *head;
5099 spin_lock_irq(&rq->lock);
5101 if (cpu_is_offline(cpu)) {
5102 spin_unlock_irq(&rq->lock);
5106 if (rq->active_balance) {
5107 active_load_balance(rq, cpu);
5108 rq->active_balance = 0;
5111 head = &rq->migration_queue;
5113 if (list_empty(head)) {
5114 spin_unlock_irq(&rq->lock);
5116 set_current_state(TASK_INTERRUPTIBLE);
5119 req = list_entry(head->next, struct migration_req, list);
5120 list_del_init(head->next);
5122 spin_unlock(&rq->lock);
5123 __migrate_task(req->task, cpu, req->dest_cpu);
5126 complete(&req->done);
5128 __set_current_state(TASK_RUNNING);
5132 /* Wait for kthread_stop */
5133 set_current_state(TASK_INTERRUPTIBLE);
5134 while (!kthread_should_stop()) {
5136 set_current_state(TASK_INTERRUPTIBLE);
5138 __set_current_state(TASK_RUNNING);
5142 #ifdef CONFIG_HOTPLUG_CPU
5144 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5148 local_irq_disable();
5149 ret = __migrate_task(p, src_cpu, dest_cpu);
5155 * Figure out where task on dead CPU should go, use force if necessary.
5156 * NOTE: interrupts should be disabled by the caller
5158 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5160 unsigned long flags;
5167 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5168 cpus_and(mask, mask, p->cpus_allowed);
5169 dest_cpu = any_online_cpu(mask);
5171 /* On any allowed CPU? */
5172 if (dest_cpu == NR_CPUS)
5173 dest_cpu = any_online_cpu(p->cpus_allowed);
5175 /* No more Mr. Nice Guy. */
5176 if (dest_cpu == NR_CPUS) {
5177 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5179 * Try to stay on the same cpuset, where the
5180 * current cpuset may be a subset of all cpus.
5181 * The cpuset_cpus_allowed_locked() variant of
5182 * cpuset_cpus_allowed() will not block. It must be
5183 * called within calls to cpuset_lock/cpuset_unlock.
5185 rq = task_rq_lock(p, &flags);
5186 p->cpus_allowed = cpus_allowed;
5187 dest_cpu = any_online_cpu(p->cpus_allowed);
5188 task_rq_unlock(rq, &flags);
5191 * Don't tell them about moving exiting tasks or
5192 * kernel threads (both mm NULL), since they never
5195 if (p->mm && printk_ratelimit())
5196 printk(KERN_INFO "process %d (%s) no "
5197 "longer affine to cpu%d\n",
5198 task_pid_nr(p), p->comm, dead_cpu);
5200 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5204 * While a dead CPU has no uninterruptible tasks queued at this point,
5205 * it might still have a nonzero ->nr_uninterruptible counter, because
5206 * for performance reasons the counter is not stricly tracking tasks to
5207 * their home CPUs. So we just add the counter to another CPU's counter,
5208 * to keep the global sum constant after CPU-down:
5210 static void migrate_nr_uninterruptible(struct rq *rq_src)
5212 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5213 unsigned long flags;
5215 local_irq_save(flags);
5216 double_rq_lock(rq_src, rq_dest);
5217 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5218 rq_src->nr_uninterruptible = 0;
5219 double_rq_unlock(rq_src, rq_dest);
5220 local_irq_restore(flags);
5223 /* Run through task list and migrate tasks from the dead cpu. */
5224 static void migrate_live_tasks(int src_cpu)
5226 struct task_struct *p, *t;
5228 read_lock(&tasklist_lock);
5230 do_each_thread(t, p) {
5234 if (task_cpu(p) == src_cpu)
5235 move_task_off_dead_cpu(src_cpu, p);
5236 } while_each_thread(t, p);
5238 read_unlock(&tasklist_lock);
5242 * activate_idle_task - move idle task to the _front_ of runqueue.
5244 static void activate_idle_task(struct task_struct *p, struct rq *rq)
5246 update_rq_clock(rq);
5248 if (p->state == TASK_UNINTERRUPTIBLE)
5249 rq->nr_uninterruptible--;
5251 enqueue_task(rq, p, 0);
5252 inc_nr_running(p, rq);
5256 * Schedules idle task to be the next runnable task on current CPU.
5257 * It does so by boosting its priority to highest possible and adding it to
5258 * the _front_ of the runqueue. Used by CPU offline code.
5260 void sched_idle_next(void)
5262 int this_cpu = smp_processor_id();
5263 struct rq *rq = cpu_rq(this_cpu);
5264 struct task_struct *p = rq->idle;
5265 unsigned long flags;
5267 /* cpu has to be offline */
5268 BUG_ON(cpu_online(this_cpu));
5271 * Strictly not necessary since rest of the CPUs are stopped by now
5272 * and interrupts disabled on the current cpu.
5274 spin_lock_irqsave(&rq->lock, flags);
5276 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5278 /* Add idle task to the _front_ of its priority queue: */
5279 activate_idle_task(p, rq);
5281 spin_unlock_irqrestore(&rq->lock, flags);
5285 * Ensures that the idle task is using init_mm right before its cpu goes
5288 void idle_task_exit(void)
5290 struct mm_struct *mm = current->active_mm;
5292 BUG_ON(cpu_online(smp_processor_id()));
5295 switch_mm(mm, &init_mm, current);
5299 /* called under rq->lock with disabled interrupts */
5300 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5302 struct rq *rq = cpu_rq(dead_cpu);
5304 /* Must be exiting, otherwise would be on tasklist. */
5305 BUG_ON(!p->exit_state);
5307 /* Cannot have done final schedule yet: would have vanished. */
5308 BUG_ON(p->state == TASK_DEAD);
5313 * Drop lock around migration; if someone else moves it,
5314 * that's OK. No task can be added to this CPU, so iteration is
5317 spin_unlock_irq(&rq->lock);
5318 move_task_off_dead_cpu(dead_cpu, p);
5319 spin_lock_irq(&rq->lock);
5324 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5325 static void migrate_dead_tasks(unsigned int dead_cpu)
5327 struct rq *rq = cpu_rq(dead_cpu);
5328 struct task_struct *next;
5331 if (!rq->nr_running)
5333 update_rq_clock(rq);
5334 next = pick_next_task(rq, rq->curr);
5337 migrate_dead(dead_cpu, next);
5341 #endif /* CONFIG_HOTPLUG_CPU */
5343 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5345 static struct ctl_table sd_ctl_dir[] = {
5347 .procname = "sched_domain",
5353 static struct ctl_table sd_ctl_root[] = {
5355 .ctl_name = CTL_KERN,
5356 .procname = "kernel",
5358 .child = sd_ctl_dir,
5363 static struct ctl_table *sd_alloc_ctl_entry(int n)
5365 struct ctl_table *entry =
5366 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5371 static void sd_free_ctl_entry(struct ctl_table **tablep)
5373 struct ctl_table *entry;
5376 * In the intermediate directories, both the child directory and
5377 * procname are dynamically allocated and could fail but the mode
5378 * will always be set. In the lowest directory the names are
5379 * static strings and all have proc handlers.
5381 for (entry = *tablep; entry->mode; entry++) {
5383 sd_free_ctl_entry(&entry->child);
5384 if (entry->proc_handler == NULL)
5385 kfree(entry->procname);
5393 set_table_entry(struct ctl_table *entry,
5394 const char *procname, void *data, int maxlen,
5395 mode_t mode, proc_handler *proc_handler)
5397 entry->procname = procname;
5399 entry->maxlen = maxlen;
5401 entry->proc_handler = proc_handler;
5404 static struct ctl_table *
5405 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5407 struct ctl_table *table = sd_alloc_ctl_entry(12);
5412 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5413 sizeof(long), 0644, proc_doulongvec_minmax);
5414 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5415 sizeof(long), 0644, proc_doulongvec_minmax);
5416 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5417 sizeof(int), 0644, proc_dointvec_minmax);
5418 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5419 sizeof(int), 0644, proc_dointvec_minmax);
5420 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5421 sizeof(int), 0644, proc_dointvec_minmax);
5422 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5423 sizeof(int), 0644, proc_dointvec_minmax);
5424 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5425 sizeof(int), 0644, proc_dointvec_minmax);
5426 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5427 sizeof(int), 0644, proc_dointvec_minmax);
5428 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5429 sizeof(int), 0644, proc_dointvec_minmax);
5430 set_table_entry(&table[9], "cache_nice_tries",
5431 &sd->cache_nice_tries,
5432 sizeof(int), 0644, proc_dointvec_minmax);
5433 set_table_entry(&table[10], "flags", &sd->flags,
5434 sizeof(int), 0644, proc_dointvec_minmax);
5435 /* &table[11] is terminator */
5440 static ctl_table * sd_alloc_ctl_cpu_table(int cpu)
5442 struct ctl_table *entry, *table;
5443 struct sched_domain *sd;
5444 int domain_num = 0, i;
5447 for_each_domain(cpu, sd)
5449 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5454 for_each_domain(cpu, sd) {
5455 snprintf(buf, 32, "domain%d", i);
5456 entry->procname = kstrdup(buf, GFP_KERNEL);
5458 entry->child = sd_alloc_ctl_domain_table(sd);
5465 static struct ctl_table_header *sd_sysctl_header;
5466 static void register_sched_domain_sysctl(void)
5468 int i, cpu_num = num_online_cpus();
5469 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5472 WARN_ON(sd_ctl_dir[0].child);
5473 sd_ctl_dir[0].child = entry;
5478 for_each_online_cpu(i) {
5479 snprintf(buf, 32, "cpu%d", i);
5480 entry->procname = kstrdup(buf, GFP_KERNEL);
5482 entry->child = sd_alloc_ctl_cpu_table(i);
5486 WARN_ON(sd_sysctl_header);
5487 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5490 /* may be called multiple times per register */
5491 static void unregister_sched_domain_sysctl(void)
5493 if (sd_sysctl_header)
5494 unregister_sysctl_table(sd_sysctl_header);
5495 sd_sysctl_header = NULL;
5496 if (sd_ctl_dir[0].child)
5497 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5500 static void register_sched_domain_sysctl(void)
5503 static void unregister_sched_domain_sysctl(void)
5509 * migration_call - callback that gets triggered when a CPU is added.
5510 * Here we can start up the necessary migration thread for the new CPU.
5512 static int __cpuinit
5513 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5515 struct task_struct *p;
5516 int cpu = (long)hcpu;
5517 unsigned long flags;
5521 case CPU_LOCK_ACQUIRE:
5522 mutex_lock(&sched_hotcpu_mutex);
5525 case CPU_UP_PREPARE:
5526 case CPU_UP_PREPARE_FROZEN:
5527 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5530 kthread_bind(p, cpu);
5531 /* Must be high prio: stop_machine expects to yield to it. */
5532 rq = task_rq_lock(p, &flags);
5533 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5534 task_rq_unlock(rq, &flags);
5535 cpu_rq(cpu)->migration_thread = p;
5539 case CPU_ONLINE_FROZEN:
5540 /* Strictly unnecessary, as first user will wake it. */
5541 wake_up_process(cpu_rq(cpu)->migration_thread);
5544 #ifdef CONFIG_HOTPLUG_CPU
5545 case CPU_UP_CANCELED:
5546 case CPU_UP_CANCELED_FROZEN:
5547 if (!cpu_rq(cpu)->migration_thread)
5549 /* Unbind it from offline cpu so it can run. Fall thru. */
5550 kthread_bind(cpu_rq(cpu)->migration_thread,
5551 any_online_cpu(cpu_online_map));
5552 kthread_stop(cpu_rq(cpu)->migration_thread);
5553 cpu_rq(cpu)->migration_thread = NULL;
5557 case CPU_DEAD_FROZEN:
5558 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5559 migrate_live_tasks(cpu);
5561 kthread_stop(rq->migration_thread);
5562 rq->migration_thread = NULL;
5563 /* Idle task back to normal (off runqueue, low prio) */
5564 spin_lock_irq(&rq->lock);
5565 update_rq_clock(rq);
5566 deactivate_task(rq, rq->idle, 0);
5567 rq->idle->static_prio = MAX_PRIO;
5568 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5569 rq->idle->sched_class = &idle_sched_class;
5570 migrate_dead_tasks(cpu);
5571 spin_unlock_irq(&rq->lock);
5573 migrate_nr_uninterruptible(rq);
5574 BUG_ON(rq->nr_running != 0);
5576 /* No need to migrate the tasks: it was best-effort if
5577 * they didn't take sched_hotcpu_mutex. Just wake up
5578 * the requestors. */
5579 spin_lock_irq(&rq->lock);
5580 while (!list_empty(&rq->migration_queue)) {
5581 struct migration_req *req;
5583 req = list_entry(rq->migration_queue.next,
5584 struct migration_req, list);
5585 list_del_init(&req->list);
5586 complete(&req->done);
5588 spin_unlock_irq(&rq->lock);
5591 case CPU_LOCK_RELEASE:
5592 mutex_unlock(&sched_hotcpu_mutex);
5598 /* Register at highest priority so that task migration (migrate_all_tasks)
5599 * happens before everything else.
5601 static struct notifier_block __cpuinitdata migration_notifier = {
5602 .notifier_call = migration_call,
5606 int __init migration_init(void)
5608 void *cpu = (void *)(long)smp_processor_id();
5611 /* Start one for the boot CPU: */
5612 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5613 BUG_ON(err == NOTIFY_BAD);
5614 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5615 register_cpu_notifier(&migration_notifier);
5623 /* Number of possible processor ids */
5624 int nr_cpu_ids __read_mostly = NR_CPUS;
5625 EXPORT_SYMBOL(nr_cpu_ids);
5627 #ifdef CONFIG_SCHED_DEBUG
5629 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5631 struct sched_group *group = sd->groups;
5632 cpumask_t groupmask;
5635 cpumask_scnprintf(str, NR_CPUS, sd->span);
5636 cpus_clear(groupmask);
5638 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5640 if (!(sd->flags & SD_LOAD_BALANCE)) {
5641 printk("does not load-balance\n");
5643 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5648 printk(KERN_CONT "span %s\n", str);
5650 if (!cpu_isset(cpu, sd->span)) {
5651 printk(KERN_ERR "ERROR: domain->span does not contain "
5654 if (!cpu_isset(cpu, group->cpumask)) {
5655 printk(KERN_ERR "ERROR: domain->groups does not contain"
5659 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5663 printk(KERN_ERR "ERROR: group is NULL\n");
5667 if (!group->__cpu_power) {
5668 printk(KERN_CONT "\n");
5669 printk(KERN_ERR "ERROR: domain->cpu_power not "
5674 if (!cpus_weight(group->cpumask)) {
5675 printk(KERN_CONT "\n");
5676 printk(KERN_ERR "ERROR: empty group\n");
5680 if (cpus_intersects(groupmask, group->cpumask)) {
5681 printk(KERN_CONT "\n");
5682 printk(KERN_ERR "ERROR: repeated CPUs\n");
5686 cpus_or(groupmask, groupmask, group->cpumask);
5688 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5689 printk(KERN_CONT " %s", str);
5691 group = group->next;
5692 } while (group != sd->groups);
5693 printk(KERN_CONT "\n");
5695 if (!cpus_equal(sd->span, groupmask))
5696 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5698 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5699 printk(KERN_ERR "ERROR: parent span is not a superset "
5700 "of domain->span\n");
5704 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5709 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5713 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5716 if (sched_domain_debug_one(sd, cpu, level))
5725 # define sched_domain_debug(sd, cpu) do { } while (0)
5728 static int sd_degenerate(struct sched_domain *sd)
5730 if (cpus_weight(sd->span) == 1)
5733 /* Following flags need at least 2 groups */
5734 if (sd->flags & (SD_LOAD_BALANCE |
5735 SD_BALANCE_NEWIDLE |
5739 SD_SHARE_PKG_RESOURCES)) {
5740 if (sd->groups != sd->groups->next)
5744 /* Following flags don't use groups */
5745 if (sd->flags & (SD_WAKE_IDLE |
5754 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5756 unsigned long cflags = sd->flags, pflags = parent->flags;
5758 if (sd_degenerate(parent))
5761 if (!cpus_equal(sd->span, parent->span))
5764 /* Does parent contain flags not in child? */
5765 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5766 if (cflags & SD_WAKE_AFFINE)
5767 pflags &= ~SD_WAKE_BALANCE;
5768 /* Flags needing groups don't count if only 1 group in parent */
5769 if (parent->groups == parent->groups->next) {
5770 pflags &= ~(SD_LOAD_BALANCE |
5771 SD_BALANCE_NEWIDLE |
5775 SD_SHARE_PKG_RESOURCES);
5777 if (~cflags & pflags)
5784 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5785 * hold the hotplug lock.
5787 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5789 struct rq *rq = cpu_rq(cpu);
5790 struct sched_domain *tmp;
5792 /* Remove the sched domains which do not contribute to scheduling. */
5793 for (tmp = sd; tmp; tmp = tmp->parent) {
5794 struct sched_domain *parent = tmp->parent;
5797 if (sd_parent_degenerate(tmp, parent)) {
5798 tmp->parent = parent->parent;
5800 parent->parent->child = tmp;
5804 if (sd && sd_degenerate(sd)) {
5810 sched_domain_debug(sd, cpu);
5812 rcu_assign_pointer(rq->sd, sd);
5815 /* cpus with isolated domains */
5816 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5818 /* Setup the mask of cpus configured for isolated domains */
5819 static int __init isolated_cpu_setup(char *str)
5821 int ints[NR_CPUS], i;
5823 str = get_options(str, ARRAY_SIZE(ints), ints);
5824 cpus_clear(cpu_isolated_map);
5825 for (i = 1; i <= ints[0]; i++)
5826 if (ints[i] < NR_CPUS)
5827 cpu_set(ints[i], cpu_isolated_map);
5831 __setup("isolcpus=", isolated_cpu_setup);
5834 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5835 * to a function which identifies what group(along with sched group) a CPU
5836 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5837 * (due to the fact that we keep track of groups covered with a cpumask_t).
5839 * init_sched_build_groups will build a circular linked list of the groups
5840 * covered by the given span, and will set each group's ->cpumask correctly,
5841 * and ->cpu_power to 0.
5844 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5845 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5846 struct sched_group **sg))
5848 struct sched_group *first = NULL, *last = NULL;
5849 cpumask_t covered = CPU_MASK_NONE;
5852 for_each_cpu_mask(i, span) {
5853 struct sched_group *sg;
5854 int group = group_fn(i, cpu_map, &sg);
5857 if (cpu_isset(i, covered))
5860 sg->cpumask = CPU_MASK_NONE;
5861 sg->__cpu_power = 0;
5863 for_each_cpu_mask(j, span) {
5864 if (group_fn(j, cpu_map, NULL) != group)
5867 cpu_set(j, covered);
5868 cpu_set(j, sg->cpumask);
5879 #define SD_NODES_PER_DOMAIN 16
5884 * find_next_best_node - find the next node to include in a sched_domain
5885 * @node: node whose sched_domain we're building
5886 * @used_nodes: nodes already in the sched_domain
5888 * Find the next node to include in a given scheduling domain. Simply
5889 * finds the closest node not already in the @used_nodes map.
5891 * Should use nodemask_t.
5893 static int find_next_best_node(int node, unsigned long *used_nodes)
5895 int i, n, val, min_val, best_node = 0;
5899 for (i = 0; i < MAX_NUMNODES; i++) {
5900 /* Start at @node */
5901 n = (node + i) % MAX_NUMNODES;
5903 if (!nr_cpus_node(n))
5906 /* Skip already used nodes */
5907 if (test_bit(n, used_nodes))
5910 /* Simple min distance search */
5911 val = node_distance(node, n);
5913 if (val < min_val) {
5919 set_bit(best_node, used_nodes);
5924 * sched_domain_node_span - get a cpumask for a node's sched_domain
5925 * @node: node whose cpumask we're constructing
5926 * @size: number of nodes to include in this span
5928 * Given a node, construct a good cpumask for its sched_domain to span. It
5929 * should be one that prevents unnecessary balancing, but also spreads tasks
5932 static cpumask_t sched_domain_node_span(int node)
5934 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5935 cpumask_t span, nodemask;
5939 bitmap_zero(used_nodes, MAX_NUMNODES);
5941 nodemask = node_to_cpumask(node);
5942 cpus_or(span, span, nodemask);
5943 set_bit(node, used_nodes);
5945 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5946 int next_node = find_next_best_node(node, used_nodes);
5948 nodemask = node_to_cpumask(next_node);
5949 cpus_or(span, span, nodemask);
5956 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5959 * SMT sched-domains:
5961 #ifdef CONFIG_SCHED_SMT
5962 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5963 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5965 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5966 struct sched_group **sg)
5969 *sg = &per_cpu(sched_group_cpus, cpu);
5975 * multi-core sched-domains:
5977 #ifdef CONFIG_SCHED_MC
5978 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5979 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5982 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5983 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5984 struct sched_group **sg)
5987 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
5988 cpus_and(mask, mask, *cpu_map);
5989 group = first_cpu(mask);
5991 *sg = &per_cpu(sched_group_core, group);
5994 #elif defined(CONFIG_SCHED_MC)
5995 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5996 struct sched_group **sg)
5999 *sg = &per_cpu(sched_group_core, cpu);
6004 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6005 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6007 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
6008 struct sched_group **sg)
6011 #ifdef CONFIG_SCHED_MC
6012 cpumask_t mask = cpu_coregroup_map(cpu);
6013 cpus_and(mask, mask, *cpu_map);
6014 group = first_cpu(mask);
6015 #elif defined(CONFIG_SCHED_SMT)
6016 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6017 cpus_and(mask, mask, *cpu_map);
6018 group = first_cpu(mask);
6023 *sg = &per_cpu(sched_group_phys, group);
6029 * The init_sched_build_groups can't handle what we want to do with node
6030 * groups, so roll our own. Now each node has its own list of groups which
6031 * gets dynamically allocated.
6033 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6034 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6036 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6037 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6039 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6040 struct sched_group **sg)
6042 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6045 cpus_and(nodemask, nodemask, *cpu_map);
6046 group = first_cpu(nodemask);
6049 *sg = &per_cpu(sched_group_allnodes, group);
6053 static void init_numa_sched_groups_power(struct sched_group *group_head)
6055 struct sched_group *sg = group_head;
6061 for_each_cpu_mask(j, sg->cpumask) {
6062 struct sched_domain *sd;
6064 sd = &per_cpu(phys_domains, j);
6065 if (j != first_cpu(sd->groups->cpumask)) {
6067 * Only add "power" once for each
6073 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6076 } while (sg != group_head);
6081 /* Free memory allocated for various sched_group structures */
6082 static void free_sched_groups(const cpumask_t *cpu_map)
6086 for_each_cpu_mask(cpu, *cpu_map) {
6087 struct sched_group **sched_group_nodes
6088 = sched_group_nodes_bycpu[cpu];
6090 if (!sched_group_nodes)
6093 for (i = 0; i < MAX_NUMNODES; i++) {
6094 cpumask_t nodemask = node_to_cpumask(i);
6095 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6097 cpus_and(nodemask, nodemask, *cpu_map);
6098 if (cpus_empty(nodemask))
6108 if (oldsg != sched_group_nodes[i])
6111 kfree(sched_group_nodes);
6112 sched_group_nodes_bycpu[cpu] = NULL;
6116 static void free_sched_groups(const cpumask_t *cpu_map)
6122 * Initialize sched groups cpu_power.
6124 * cpu_power indicates the capacity of sched group, which is used while
6125 * distributing the load between different sched groups in a sched domain.
6126 * Typically cpu_power for all the groups in a sched domain will be same unless
6127 * there are asymmetries in the topology. If there are asymmetries, group
6128 * having more cpu_power will pickup more load compared to the group having
6131 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6132 * the maximum number of tasks a group can handle in the presence of other idle
6133 * or lightly loaded groups in the same sched domain.
6135 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6137 struct sched_domain *child;
6138 struct sched_group *group;
6140 WARN_ON(!sd || !sd->groups);
6142 if (cpu != first_cpu(sd->groups->cpumask))
6147 sd->groups->__cpu_power = 0;
6150 * For perf policy, if the groups in child domain share resources
6151 * (for example cores sharing some portions of the cache hierarchy
6152 * or SMT), then set this domain groups cpu_power such that each group
6153 * can handle only one task, when there are other idle groups in the
6154 * same sched domain.
6156 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6158 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6159 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6164 * add cpu_power of each child group to this groups cpu_power
6166 group = child->groups;
6168 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6169 group = group->next;
6170 } while (group != child->groups);
6174 * Build sched domains for a given set of cpus and attach the sched domains
6175 * to the individual cpus
6177 static int build_sched_domains(const cpumask_t *cpu_map)
6181 struct sched_group **sched_group_nodes = NULL;
6182 int sd_allnodes = 0;
6185 * Allocate the per-node list of sched groups
6187 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6189 if (!sched_group_nodes) {
6190 printk(KERN_WARNING "Can not alloc sched group node list\n");
6193 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6197 * Set up domains for cpus specified by the cpu_map.
6199 for_each_cpu_mask(i, *cpu_map) {
6200 struct sched_domain *sd = NULL, *p;
6201 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6203 cpus_and(nodemask, nodemask, *cpu_map);
6206 if (cpus_weight(*cpu_map) >
6207 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6208 sd = &per_cpu(allnodes_domains, i);
6209 *sd = SD_ALLNODES_INIT;
6210 sd->span = *cpu_map;
6211 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6217 sd = &per_cpu(node_domains, i);
6219 sd->span = sched_domain_node_span(cpu_to_node(i));
6223 cpus_and(sd->span, sd->span, *cpu_map);
6227 sd = &per_cpu(phys_domains, i);
6229 sd->span = nodemask;
6233 cpu_to_phys_group(i, cpu_map, &sd->groups);
6235 #ifdef CONFIG_SCHED_MC
6237 sd = &per_cpu(core_domains, i);
6239 sd->span = cpu_coregroup_map(i);
6240 cpus_and(sd->span, sd->span, *cpu_map);
6243 cpu_to_core_group(i, cpu_map, &sd->groups);
6246 #ifdef CONFIG_SCHED_SMT
6248 sd = &per_cpu(cpu_domains, i);
6249 *sd = SD_SIBLING_INIT;
6250 sd->span = per_cpu(cpu_sibling_map, i);
6251 cpus_and(sd->span, sd->span, *cpu_map);
6254 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6258 #ifdef CONFIG_SCHED_SMT
6259 /* Set up CPU (sibling) groups */
6260 for_each_cpu_mask(i, *cpu_map) {
6261 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6262 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6263 if (i != first_cpu(this_sibling_map))
6266 init_sched_build_groups(this_sibling_map, cpu_map,
6271 #ifdef CONFIG_SCHED_MC
6272 /* Set up multi-core groups */
6273 for_each_cpu_mask(i, *cpu_map) {
6274 cpumask_t this_core_map = cpu_coregroup_map(i);
6275 cpus_and(this_core_map, this_core_map, *cpu_map);
6276 if (i != first_cpu(this_core_map))
6278 init_sched_build_groups(this_core_map, cpu_map,
6279 &cpu_to_core_group);
6283 /* Set up physical groups */
6284 for (i = 0; i < MAX_NUMNODES; i++) {
6285 cpumask_t nodemask = node_to_cpumask(i);
6287 cpus_and(nodemask, nodemask, *cpu_map);
6288 if (cpus_empty(nodemask))
6291 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6295 /* Set up node groups */
6297 init_sched_build_groups(*cpu_map, cpu_map,
6298 &cpu_to_allnodes_group);
6300 for (i = 0; i < MAX_NUMNODES; i++) {
6301 /* Set up node groups */
6302 struct sched_group *sg, *prev;
6303 cpumask_t nodemask = node_to_cpumask(i);
6304 cpumask_t domainspan;
6305 cpumask_t covered = CPU_MASK_NONE;
6308 cpus_and(nodemask, nodemask, *cpu_map);
6309 if (cpus_empty(nodemask)) {
6310 sched_group_nodes[i] = NULL;
6314 domainspan = sched_domain_node_span(i);
6315 cpus_and(domainspan, domainspan, *cpu_map);
6317 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6319 printk(KERN_WARNING "Can not alloc domain group for "
6323 sched_group_nodes[i] = sg;
6324 for_each_cpu_mask(j, nodemask) {
6325 struct sched_domain *sd;
6327 sd = &per_cpu(node_domains, j);
6330 sg->__cpu_power = 0;
6331 sg->cpumask = nodemask;
6333 cpus_or(covered, covered, nodemask);
6336 for (j = 0; j < MAX_NUMNODES; j++) {
6337 cpumask_t tmp, notcovered;
6338 int n = (i + j) % MAX_NUMNODES;
6340 cpus_complement(notcovered, covered);
6341 cpus_and(tmp, notcovered, *cpu_map);
6342 cpus_and(tmp, tmp, domainspan);
6343 if (cpus_empty(tmp))
6346 nodemask = node_to_cpumask(n);
6347 cpus_and(tmp, tmp, nodemask);
6348 if (cpus_empty(tmp))
6351 sg = kmalloc_node(sizeof(struct sched_group),
6355 "Can not alloc domain group for node %d\n", j);
6358 sg->__cpu_power = 0;
6360 sg->next = prev->next;
6361 cpus_or(covered, covered, tmp);
6368 /* Calculate CPU power for physical packages and nodes */
6369 #ifdef CONFIG_SCHED_SMT
6370 for_each_cpu_mask(i, *cpu_map) {
6371 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6373 init_sched_groups_power(i, sd);
6376 #ifdef CONFIG_SCHED_MC
6377 for_each_cpu_mask(i, *cpu_map) {
6378 struct sched_domain *sd = &per_cpu(core_domains, i);
6380 init_sched_groups_power(i, sd);
6384 for_each_cpu_mask(i, *cpu_map) {
6385 struct sched_domain *sd = &per_cpu(phys_domains, i);
6387 init_sched_groups_power(i, sd);
6391 for (i = 0; i < MAX_NUMNODES; i++)
6392 init_numa_sched_groups_power(sched_group_nodes[i]);
6395 struct sched_group *sg;
6397 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6398 init_numa_sched_groups_power(sg);
6402 /* Attach the domains */
6403 for_each_cpu_mask(i, *cpu_map) {
6404 struct sched_domain *sd;
6405 #ifdef CONFIG_SCHED_SMT
6406 sd = &per_cpu(cpu_domains, i);
6407 #elif defined(CONFIG_SCHED_MC)
6408 sd = &per_cpu(core_domains, i);
6410 sd = &per_cpu(phys_domains, i);
6412 cpu_attach_domain(sd, i);
6419 free_sched_groups(cpu_map);
6424 static cpumask_t *doms_cur; /* current sched domains */
6425 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6428 * Special case: If a kmalloc of a doms_cur partition (array of
6429 * cpumask_t) fails, then fallback to a single sched domain,
6430 * as determined by the single cpumask_t fallback_doms.
6432 static cpumask_t fallback_doms;
6435 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6436 * For now this just excludes isolated cpus, but could be used to
6437 * exclude other special cases in the future.
6439 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6444 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6446 doms_cur = &fallback_doms;
6447 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6448 err = build_sched_domains(doms_cur);
6449 register_sched_domain_sysctl();
6454 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6456 free_sched_groups(cpu_map);
6460 * Detach sched domains from a group of cpus specified in cpu_map
6461 * These cpus will now be attached to the NULL domain
6463 static void detach_destroy_domains(const cpumask_t *cpu_map)
6467 unregister_sched_domain_sysctl();
6469 for_each_cpu_mask(i, *cpu_map)
6470 cpu_attach_domain(NULL, i);
6471 synchronize_sched();
6472 arch_destroy_sched_domains(cpu_map);
6476 * Partition sched domains as specified by the 'ndoms_new'
6477 * cpumasks in the array doms_new[] of cpumasks. This compares
6478 * doms_new[] to the current sched domain partitioning, doms_cur[].
6479 * It destroys each deleted domain and builds each new domain.
6481 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6482 * The masks don't intersect (don't overlap.) We should setup one
6483 * sched domain for each mask. CPUs not in any of the cpumasks will
6484 * not be load balanced. If the same cpumask appears both in the
6485 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6488 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6489 * ownership of it and will kfree it when done with it. If the caller
6490 * failed the kmalloc call, then it can pass in doms_new == NULL,
6491 * and partition_sched_domains() will fallback to the single partition
6494 * Call with hotplug lock held
6496 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6500 /* always unregister in case we don't destroy any domains */
6501 unregister_sched_domain_sysctl();
6503 if (doms_new == NULL) {
6505 doms_new = &fallback_doms;
6506 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6509 /* Destroy deleted domains */
6510 for (i = 0; i < ndoms_cur; i++) {
6511 for (j = 0; j < ndoms_new; j++) {
6512 if (cpus_equal(doms_cur[i], doms_new[j]))
6515 /* no match - a current sched domain not in new doms_new[] */
6516 detach_destroy_domains(doms_cur + i);
6521 /* Build new domains */
6522 for (i = 0; i < ndoms_new; i++) {
6523 for (j = 0; j < ndoms_cur; j++) {
6524 if (cpus_equal(doms_new[i], doms_cur[j]))
6527 /* no match - add a new doms_new */
6528 build_sched_domains(doms_new + i);
6533 /* Remember the new sched domains */
6534 if (doms_cur != &fallback_doms)
6536 doms_cur = doms_new;
6537 ndoms_cur = ndoms_new;
6539 register_sched_domain_sysctl();
6542 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6543 static int arch_reinit_sched_domains(void)
6547 mutex_lock(&sched_hotcpu_mutex);
6548 detach_destroy_domains(&cpu_online_map);
6549 err = arch_init_sched_domains(&cpu_online_map);
6550 mutex_unlock(&sched_hotcpu_mutex);
6555 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6559 if (buf[0] != '0' && buf[0] != '1')
6563 sched_smt_power_savings = (buf[0] == '1');
6565 sched_mc_power_savings = (buf[0] == '1');
6567 ret = arch_reinit_sched_domains();
6569 return ret ? ret : count;
6572 #ifdef CONFIG_SCHED_MC
6573 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6575 return sprintf(page, "%u\n", sched_mc_power_savings);
6577 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6578 const char *buf, size_t count)
6580 return sched_power_savings_store(buf, count, 0);
6582 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6583 sched_mc_power_savings_store);
6586 #ifdef CONFIG_SCHED_SMT
6587 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6589 return sprintf(page, "%u\n", sched_smt_power_savings);
6591 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6592 const char *buf, size_t count)
6594 return sched_power_savings_store(buf, count, 1);
6596 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6597 sched_smt_power_savings_store);
6600 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6604 #ifdef CONFIG_SCHED_SMT
6606 err = sysfs_create_file(&cls->kset.kobj,
6607 &attr_sched_smt_power_savings.attr);
6609 #ifdef CONFIG_SCHED_MC
6610 if (!err && mc_capable())
6611 err = sysfs_create_file(&cls->kset.kobj,
6612 &attr_sched_mc_power_savings.attr);
6619 * Force a reinitialization of the sched domains hierarchy. The domains
6620 * and groups cannot be updated in place without racing with the balancing
6621 * code, so we temporarily attach all running cpus to the NULL domain
6622 * which will prevent rebalancing while the sched domains are recalculated.
6624 static int update_sched_domains(struct notifier_block *nfb,
6625 unsigned long action, void *hcpu)
6628 case CPU_UP_PREPARE:
6629 case CPU_UP_PREPARE_FROZEN:
6630 case CPU_DOWN_PREPARE:
6631 case CPU_DOWN_PREPARE_FROZEN:
6632 detach_destroy_domains(&cpu_online_map);
6635 case CPU_UP_CANCELED:
6636 case CPU_UP_CANCELED_FROZEN:
6637 case CPU_DOWN_FAILED:
6638 case CPU_DOWN_FAILED_FROZEN:
6640 case CPU_ONLINE_FROZEN:
6642 case CPU_DEAD_FROZEN:
6644 * Fall through and re-initialise the domains.
6651 /* The hotplug lock is already held by cpu_up/cpu_down */
6652 arch_init_sched_domains(&cpu_online_map);
6657 void __init sched_init_smp(void)
6659 cpumask_t non_isolated_cpus;
6661 mutex_lock(&sched_hotcpu_mutex);
6662 arch_init_sched_domains(&cpu_online_map);
6663 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6664 if (cpus_empty(non_isolated_cpus))
6665 cpu_set(smp_processor_id(), non_isolated_cpus);
6666 mutex_unlock(&sched_hotcpu_mutex);
6667 /* XXX: Theoretical race here - CPU may be hotplugged now */
6668 hotcpu_notifier(update_sched_domains, 0);
6670 /* Move init over to a non-isolated CPU */
6671 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6675 void __init sched_init_smp(void)
6678 #endif /* CONFIG_SMP */
6680 int in_sched_functions(unsigned long addr)
6682 /* Linker adds these: start and end of __sched functions */
6683 extern char __sched_text_start[], __sched_text_end[];
6685 return in_lock_functions(addr) ||
6686 (addr >= (unsigned long)__sched_text_start
6687 && addr < (unsigned long)__sched_text_end);
6690 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6692 cfs_rq->tasks_timeline = RB_ROOT;
6693 #ifdef CONFIG_FAIR_GROUP_SCHED
6696 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6699 void __init sched_init(void)
6701 int highest_cpu = 0;
6704 for_each_possible_cpu(i) {
6705 struct rt_prio_array *array;
6709 spin_lock_init(&rq->lock);
6710 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6713 init_cfs_rq(&rq->cfs, rq);
6714 #ifdef CONFIG_FAIR_GROUP_SCHED
6715 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6717 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6718 struct sched_entity *se =
6719 &per_cpu(init_sched_entity, i);
6721 init_cfs_rq_p[i] = cfs_rq;
6722 init_cfs_rq(cfs_rq, rq);
6723 cfs_rq->tg = &init_task_group;
6724 list_add(&cfs_rq->leaf_cfs_rq_list,
6725 &rq->leaf_cfs_rq_list);
6727 init_sched_entity_p[i] = se;
6728 se->cfs_rq = &rq->cfs;
6730 se->load.weight = init_task_group_load;
6731 se->load.inv_weight =
6732 div64_64(1ULL<<32, init_task_group_load);
6735 init_task_group.shares = init_task_group_load;
6736 spin_lock_init(&init_task_group.lock);
6739 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6740 rq->cpu_load[j] = 0;
6743 rq->active_balance = 0;
6744 rq->next_balance = jiffies;
6747 rq->migration_thread = NULL;
6748 INIT_LIST_HEAD(&rq->migration_queue);
6750 atomic_set(&rq->nr_iowait, 0);
6752 array = &rq->rt.active;
6753 for (j = 0; j < MAX_RT_PRIO; j++) {
6754 INIT_LIST_HEAD(array->queue + j);
6755 __clear_bit(j, array->bitmap);
6758 /* delimiter for bitsearch: */
6759 __set_bit(MAX_RT_PRIO, array->bitmap);
6762 set_load_weight(&init_task);
6764 #ifdef CONFIG_PREEMPT_NOTIFIERS
6765 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6769 nr_cpu_ids = highest_cpu + 1;
6770 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6773 #ifdef CONFIG_RT_MUTEXES
6774 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6778 * The boot idle thread does lazy MMU switching as well:
6780 atomic_inc(&init_mm.mm_count);
6781 enter_lazy_tlb(&init_mm, current);
6784 * Make us the idle thread. Technically, schedule() should not be
6785 * called from this thread, however somewhere below it might be,
6786 * but because we are the idle thread, we just pick up running again
6787 * when this runqueue becomes "idle".
6789 init_idle(current, smp_processor_id());
6791 * During early bootup we pretend to be a normal task:
6793 current->sched_class = &fair_sched_class;
6796 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6797 void __might_sleep(char *file, int line)
6800 static unsigned long prev_jiffy; /* ratelimiting */
6802 if ((in_atomic() || irqs_disabled()) &&
6803 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6804 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6806 prev_jiffy = jiffies;
6807 printk(KERN_ERR "BUG: sleeping function called from invalid"
6808 " context at %s:%d\n", file, line);
6809 printk("in_atomic():%d, irqs_disabled():%d\n",
6810 in_atomic(), irqs_disabled());
6811 debug_show_held_locks(current);
6812 if (irqs_disabled())
6813 print_irqtrace_events(current);
6818 EXPORT_SYMBOL(__might_sleep);
6821 #ifdef CONFIG_MAGIC_SYSRQ
6822 static void normalize_task(struct rq *rq, struct task_struct *p)
6825 update_rq_clock(rq);
6826 on_rq = p->se.on_rq;
6828 deactivate_task(rq, p, 0);
6829 __setscheduler(rq, p, SCHED_NORMAL, 0);
6831 activate_task(rq, p, 0);
6832 resched_task(rq->curr);
6836 void normalize_rt_tasks(void)
6838 struct task_struct *g, *p;
6839 unsigned long flags;
6842 read_lock_irq(&tasklist_lock);
6843 do_each_thread(g, p) {
6845 * Only normalize user tasks:
6850 p->se.exec_start = 0;
6851 #ifdef CONFIG_SCHEDSTATS
6852 p->se.wait_start = 0;
6853 p->se.sleep_start = 0;
6854 p->se.block_start = 0;
6856 task_rq(p)->clock = 0;
6860 * Renice negative nice level userspace
6863 if (TASK_NICE(p) < 0 && p->mm)
6864 set_user_nice(p, 0);
6868 spin_lock_irqsave(&p->pi_lock, flags);
6869 rq = __task_rq_lock(p);
6871 normalize_task(rq, p);
6873 __task_rq_unlock(rq);
6874 spin_unlock_irqrestore(&p->pi_lock, flags);
6875 } while_each_thread(g, p);
6877 read_unlock_irq(&tasklist_lock);
6880 #endif /* CONFIG_MAGIC_SYSRQ */
6884 * These functions are only useful for the IA64 MCA handling.
6886 * They can only be called when the whole system has been
6887 * stopped - every CPU needs to be quiescent, and no scheduling
6888 * activity can take place. Using them for anything else would
6889 * be a serious bug, and as a result, they aren't even visible
6890 * under any other configuration.
6894 * curr_task - return the current task for a given cpu.
6895 * @cpu: the processor in question.
6897 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6899 struct task_struct *curr_task(int cpu)
6901 return cpu_curr(cpu);
6905 * set_curr_task - set the current task for a given cpu.
6906 * @cpu: the processor in question.
6907 * @p: the task pointer to set.
6909 * Description: This function must only be used when non-maskable interrupts
6910 * are serviced on a separate stack. It allows the architecture to switch the
6911 * notion of the current task on a cpu in a non-blocking manner. This function
6912 * must be called with all CPU's synchronized, and interrupts disabled, the
6913 * and caller must save the original value of the current task (see
6914 * curr_task() above) and restore that value before reenabling interrupts and
6915 * re-starting the system.
6917 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6919 void set_curr_task(int cpu, struct task_struct *p)
6926 #ifdef CONFIG_FAIR_GROUP_SCHED
6928 /* allocate runqueue etc for a new task group */
6929 struct task_group *sched_create_group(void)
6931 struct task_group *tg;
6932 struct cfs_rq *cfs_rq;
6933 struct sched_entity *se;
6937 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6939 return ERR_PTR(-ENOMEM);
6941 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6944 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6948 for_each_possible_cpu(i) {
6951 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
6956 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
6961 memset(cfs_rq, 0, sizeof(struct cfs_rq));
6962 memset(se, 0, sizeof(struct sched_entity));
6964 tg->cfs_rq[i] = cfs_rq;
6965 init_cfs_rq(cfs_rq, rq);
6969 se->cfs_rq = &rq->cfs;
6971 se->load.weight = NICE_0_LOAD;
6972 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
6976 for_each_possible_cpu(i) {
6978 cfs_rq = tg->cfs_rq[i];
6979 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6982 tg->shares = NICE_0_LOAD;
6983 spin_lock_init(&tg->lock);
6988 for_each_possible_cpu(i) {
6990 kfree(tg->cfs_rq[i]);
6998 return ERR_PTR(-ENOMEM);
7001 /* rcu callback to free various structures associated with a task group */
7002 static void free_sched_group(struct rcu_head *rhp)
7004 struct cfs_rq *cfs_rq = container_of(rhp, struct cfs_rq, rcu);
7005 struct task_group *tg = cfs_rq->tg;
7006 struct sched_entity *se;
7009 /* now it should be safe to free those cfs_rqs */
7010 for_each_possible_cpu(i) {
7011 cfs_rq = tg->cfs_rq[i];
7023 /* Destroy runqueue etc associated with a task group */
7024 void sched_destroy_group(struct task_group *tg)
7026 struct cfs_rq *cfs_rq;
7029 for_each_possible_cpu(i) {
7030 cfs_rq = tg->cfs_rq[i];
7031 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7034 cfs_rq = tg->cfs_rq[0];
7036 /* wait for possible concurrent references to cfs_rqs complete */
7037 call_rcu(&cfs_rq->rcu, free_sched_group);
7040 /* change task's runqueue when it moves between groups.
7041 * The caller of this function should have put the task in its new group
7042 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7043 * reflect its new group.
7045 void sched_move_task(struct task_struct *tsk)
7048 unsigned long flags;
7051 rq = task_rq_lock(tsk, &flags);
7053 if (tsk->sched_class != &fair_sched_class)
7056 update_rq_clock(rq);
7058 running = task_running(rq, tsk);
7059 on_rq = tsk->se.on_rq;
7062 dequeue_task(rq, tsk, 0);
7063 if (unlikely(running))
7064 tsk->sched_class->put_prev_task(rq, tsk);
7067 set_task_cfs_rq(tsk);
7070 if (unlikely(running))
7071 tsk->sched_class->set_curr_task(rq);
7072 enqueue_task(rq, tsk, 0);
7076 task_rq_unlock(rq, &flags);
7079 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7081 struct cfs_rq *cfs_rq = se->cfs_rq;
7082 struct rq *rq = cfs_rq->rq;
7085 spin_lock_irq(&rq->lock);
7089 dequeue_entity(cfs_rq, se, 0);
7091 se->load.weight = shares;
7092 se->load.inv_weight = div64_64((1ULL<<32), shares);
7095 enqueue_entity(cfs_rq, se, 0);
7097 spin_unlock_irq(&rq->lock);
7100 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7104 spin_lock(&tg->lock);
7105 if (tg->shares == shares)
7108 tg->shares = shares;
7109 for_each_possible_cpu(i)
7110 set_se_shares(tg->se[i], shares);
7113 spin_unlock(&tg->lock);
7117 unsigned long sched_group_shares(struct task_group *tg)
7122 #endif /* CONFIG_FAIR_GROUP_SCHED */
7124 #ifdef CONFIG_FAIR_CGROUP_SCHED
7126 /* return corresponding task_group object of a cgroup */
7127 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7129 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7130 struct task_group, css);
7133 static struct cgroup_subsys_state *
7134 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7136 struct task_group *tg;
7138 if (!cgrp->parent) {
7139 /* This is early initialization for the top cgroup */
7140 init_task_group.css.cgroup = cgrp;
7141 return &init_task_group.css;
7144 /* we support only 1-level deep hierarchical scheduler atm */
7145 if (cgrp->parent->parent)
7146 return ERR_PTR(-EINVAL);
7148 tg = sched_create_group();
7150 return ERR_PTR(-ENOMEM);
7152 /* Bind the cgroup to task_group object we just created */
7153 tg->css.cgroup = cgrp;
7158 static void cpu_cgroup_destroy(struct cgroup_subsys *ss,
7159 struct cgroup *cgrp)
7161 struct task_group *tg = cgroup_tg(cgrp);
7163 sched_destroy_group(tg);
7166 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss,
7167 struct cgroup *cgrp, struct task_struct *tsk)
7169 /* We don't support RT-tasks being in separate groups */
7170 if (tsk->sched_class != &fair_sched_class)
7177 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7178 struct cgroup *old_cont, struct task_struct *tsk)
7180 sched_move_task(tsk);
7183 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7186 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7189 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7191 struct task_group *tg = cgroup_tg(cgrp);
7193 return (u64) tg->shares;
7196 static struct cftype cpu_shares = {
7198 .read_uint = cpu_shares_read_uint,
7199 .write_uint = cpu_shares_write_uint,
7202 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7204 return cgroup_add_file(cont, ss, &cpu_shares);
7207 struct cgroup_subsys cpu_cgroup_subsys = {
7209 .create = cpu_cgroup_create,
7210 .destroy = cpu_cgroup_destroy,
7211 .can_attach = cpu_cgroup_can_attach,
7212 .attach = cpu_cgroup_attach,
7213 .populate = cpu_cgroup_populate,
7214 .subsys_id = cpu_cgroup_subsys_id,
7218 #endif /* CONFIG_FAIR_CGROUP_SCHED */