4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/pid_namespace.h>
48 #include <linux/smp.h>
49 #include <linux/threads.h>
50 #include <linux/timer.h>
51 #include <linux/rcupdate.h>
52 #include <linux/cpu.h>
53 #include <linux/cpuset.h>
54 #include <linux/percpu.h>
55 #include <linux/kthread.h>
56 #include <linux/seq_file.h>
57 #include <linux/sysctl.h>
58 #include <linux/syscalls.h>
59 #include <linux/times.h>
60 #include <linux/tsacct_kern.h>
61 #include <linux/kprobes.h>
62 #include <linux/delayacct.h>
63 #include <linux/reciprocal_div.h>
64 #include <linux/unistd.h>
65 #include <linux/pagemap.h>
68 #include <asm/irq_regs.h>
71 * Scheduler clock - returns current time in nanosec units.
72 * This is default implementation.
73 * Architectures and sub-architectures can override this.
75 unsigned long long __attribute__((weak)) sched_clock(void)
77 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Some helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / HZ))
104 #define NICE_0_LOAD SCHED_LOAD_SCALE
105 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
108 * These are the 'tuning knobs' of the scheduler:
110 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
111 * Timeslices get refilled after they expire.
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
122 return reciprocal_divide(load, sg->reciprocal_cpu_power);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
131 sg->__cpu_power += val;
132 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
136 static inline int rt_policy(int policy)
138 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
143 static inline int task_has_rt_policy(struct task_struct *p)
145 return rt_policy(p->policy);
149 * This is the priority-queue data structure of the RT scheduling class:
151 struct rt_prio_array {
152 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
153 struct list_head queue[MAX_RT_PRIO];
156 #ifdef CONFIG_FAIR_GROUP_SCHED
158 #include <linux/cgroup.h>
162 /* task group related information */
164 #ifdef CONFIG_FAIR_CGROUP_SCHED
165 struct cgroup_subsys_state css;
167 /* schedulable entities of this group on each cpu */
168 struct sched_entity **se;
169 /* runqueue "owned" by this group on each cpu */
170 struct cfs_rq **cfs_rq;
171 unsigned long shares;
175 /* Default task group's sched entity on each cpu */
176 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
177 /* Default task group's cfs_rq on each cpu */
178 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
180 static struct sched_entity *init_sched_entity_p[NR_CPUS];
181 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
183 /* task_group_mutex serializes add/remove of task groups and also changes to
184 * a task group's cpu shares.
186 static DEFINE_MUTEX(task_group_mutex);
188 /* Default task group.
189 * Every task in system belong to this group at bootup.
191 struct task_group init_task_group = {
192 .se = init_sched_entity_p,
193 .cfs_rq = init_cfs_rq_p,
196 #ifdef CONFIG_FAIR_USER_SCHED
197 # define INIT_TASK_GROUP_LOAD 2*NICE_0_LOAD
199 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
202 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
204 /* return group to which a task belongs */
205 static inline struct task_group *task_group(struct task_struct *p)
207 struct task_group *tg;
209 #ifdef CONFIG_FAIR_USER_SCHED
211 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
212 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
213 struct task_group, css);
215 tg = &init_task_group;
220 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
221 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu)
223 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
224 p->se.parent = task_group(p)->se[cpu];
227 static inline void lock_task_group_list(void)
229 mutex_lock(&task_group_mutex);
232 static inline void unlock_task_group_list(void)
234 mutex_unlock(&task_group_mutex);
239 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu) { }
240 static inline void lock_task_group_list(void) { }
241 static inline void unlock_task_group_list(void) { }
243 #endif /* CONFIG_FAIR_GROUP_SCHED */
245 /* CFS-related fields in a runqueue */
247 struct load_weight load;
248 unsigned long nr_running;
253 struct rb_root tasks_timeline;
254 struct rb_node *rb_leftmost;
255 struct rb_node *rb_load_balance_curr;
256 /* 'curr' points to currently running entity on this cfs_rq.
257 * It is set to NULL otherwise (i.e when none are currently running).
259 struct sched_entity *curr;
261 unsigned long nr_spread_over;
263 #ifdef CONFIG_FAIR_GROUP_SCHED
264 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
267 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
268 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
269 * (like users, containers etc.)
271 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
272 * list is used during load balance.
274 struct list_head leaf_cfs_rq_list;
275 struct task_group *tg; /* group that "owns" this runqueue */
279 /* Real-Time classes' related field in a runqueue: */
281 struct rt_prio_array active;
282 int rt_load_balance_idx;
283 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
287 * This is the main, per-CPU runqueue data structure.
289 * Locking rule: those places that want to lock multiple runqueues
290 * (such as the load balancing or the thread migration code), lock
291 * acquire operations must be ordered by ascending &runqueue.
298 * nr_running and cpu_load should be in the same cacheline because
299 * remote CPUs use both these fields when doing load calculation.
301 unsigned long nr_running;
302 #define CPU_LOAD_IDX_MAX 5
303 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
304 unsigned char idle_at_tick;
306 unsigned char in_nohz_recently;
308 /* capture load from *all* tasks on this cpu: */
309 struct load_weight load;
310 unsigned long nr_load_updates;
314 #ifdef CONFIG_FAIR_GROUP_SCHED
315 /* list of leaf cfs_rq on this cpu: */
316 struct list_head leaf_cfs_rq_list;
321 * This is part of a global counter where only the total sum
322 * over all CPUs matters. A task can increase this counter on
323 * one CPU and if it got migrated afterwards it may decrease
324 * it on another CPU. Always updated under the runqueue lock:
326 unsigned long nr_uninterruptible;
328 struct task_struct *curr, *idle;
329 unsigned long next_balance;
330 struct mm_struct *prev_mm;
332 u64 clock, prev_clock_raw;
335 unsigned int clock_warps, clock_overflows;
337 unsigned int clock_deep_idle_events;
343 struct sched_domain *sd;
345 /* For active balancing */
348 /* cpu of this runqueue: */
351 struct task_struct *migration_thread;
352 struct list_head migration_queue;
355 #ifdef CONFIG_SCHEDSTATS
357 struct sched_info rq_sched_info;
359 /* sys_sched_yield() stats */
360 unsigned int yld_exp_empty;
361 unsigned int yld_act_empty;
362 unsigned int yld_both_empty;
363 unsigned int yld_count;
365 /* schedule() stats */
366 unsigned int sched_switch;
367 unsigned int sched_count;
368 unsigned int sched_goidle;
370 /* try_to_wake_up() stats */
371 unsigned int ttwu_count;
372 unsigned int ttwu_local;
375 unsigned int bkl_count;
377 struct lock_class_key rq_lock_key;
380 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
381 static DEFINE_MUTEX(sched_hotcpu_mutex);
383 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
385 rq->curr->sched_class->check_preempt_curr(rq, p);
388 static inline int cpu_of(struct rq *rq)
398 * Update the per-runqueue clock, as finegrained as the platform can give
399 * us, but without assuming monotonicity, etc.:
401 static void __update_rq_clock(struct rq *rq)
403 u64 prev_raw = rq->prev_clock_raw;
404 u64 now = sched_clock();
405 s64 delta = now - prev_raw;
406 u64 clock = rq->clock;
408 #ifdef CONFIG_SCHED_DEBUG
409 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
412 * Protect against sched_clock() occasionally going backwards:
414 if (unlikely(delta < 0)) {
419 * Catch too large forward jumps too:
421 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
422 if (clock < rq->tick_timestamp + TICK_NSEC)
423 clock = rq->tick_timestamp + TICK_NSEC;
426 rq->clock_overflows++;
428 if (unlikely(delta > rq->clock_max_delta))
429 rq->clock_max_delta = delta;
434 rq->prev_clock_raw = now;
438 static void update_rq_clock(struct rq *rq)
440 if (likely(smp_processor_id() == cpu_of(rq)))
441 __update_rq_clock(rq);
445 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
446 * See detach_destroy_domains: synchronize_sched for details.
448 * The domain tree of any CPU may only be accessed from within
449 * preempt-disabled sections.
451 #define for_each_domain(cpu, __sd) \
452 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
454 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
455 #define this_rq() (&__get_cpu_var(runqueues))
456 #define task_rq(p) cpu_rq(task_cpu(p))
457 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
460 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
462 #ifdef CONFIG_SCHED_DEBUG
463 # define const_debug __read_mostly
465 # define const_debug static const
469 * Debugging: various feature bits
472 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
473 SCHED_FEAT_WAKEUP_PREEMPT = 2,
474 SCHED_FEAT_START_DEBIT = 4,
475 SCHED_FEAT_TREE_AVG = 8,
476 SCHED_FEAT_APPROX_AVG = 16,
479 const_debug unsigned int sysctl_sched_features =
480 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
481 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
482 SCHED_FEAT_START_DEBIT * 1 |
483 SCHED_FEAT_TREE_AVG * 0 |
484 SCHED_FEAT_APPROX_AVG * 0;
486 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
489 * Number of tasks to iterate in a single balance run.
490 * Limited because this is done with IRQs disabled.
492 const_debug unsigned int sysctl_sched_nr_migrate = 32;
495 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
496 * clock constructed from sched_clock():
498 unsigned long long cpu_clock(int cpu)
500 unsigned long long now;
504 local_irq_save(flags);
507 * Only call sched_clock() if the scheduler has already been
508 * initialized (some code might call cpu_clock() very early):
513 local_irq_restore(flags);
517 EXPORT_SYMBOL_GPL(cpu_clock);
519 #ifndef prepare_arch_switch
520 # define prepare_arch_switch(next) do { } while (0)
522 #ifndef finish_arch_switch
523 # define finish_arch_switch(prev) do { } while (0)
526 static inline int task_current(struct rq *rq, struct task_struct *p)
528 return rq->curr == p;
531 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
532 static inline int task_running(struct rq *rq, struct task_struct *p)
534 return task_current(rq, p);
537 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
541 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
543 #ifdef CONFIG_DEBUG_SPINLOCK
544 /* this is a valid case when another task releases the spinlock */
545 rq->lock.owner = current;
548 * If we are tracking spinlock dependencies then we have to
549 * fix up the runqueue lock - which gets 'carried over' from
552 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
554 spin_unlock_irq(&rq->lock);
557 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
558 static inline int task_running(struct rq *rq, struct task_struct *p)
563 return task_current(rq, p);
567 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
571 * We can optimise this out completely for !SMP, because the
572 * SMP rebalancing from interrupt is the only thing that cares
577 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
578 spin_unlock_irq(&rq->lock);
580 spin_unlock(&rq->lock);
584 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
588 * After ->oncpu is cleared, the task can be moved to a different CPU.
589 * We must ensure this doesn't happen until the switch is completely
595 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
599 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
602 * __task_rq_lock - lock the runqueue a given task resides on.
603 * Must be called interrupts disabled.
605 static inline struct rq *__task_rq_lock(struct task_struct *p)
609 struct rq *rq = task_rq(p);
610 spin_lock(&rq->lock);
611 if (likely(rq == task_rq(p)))
613 spin_unlock(&rq->lock);
618 * task_rq_lock - lock the runqueue a given task resides on and disable
619 * interrupts. Note the ordering: we can safely lookup the task_rq without
620 * explicitly disabling preemption.
622 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
628 local_irq_save(*flags);
630 spin_lock(&rq->lock);
631 if (likely(rq == task_rq(p)))
633 spin_unlock_irqrestore(&rq->lock, *flags);
637 static void __task_rq_unlock(struct rq *rq)
640 spin_unlock(&rq->lock);
643 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
646 spin_unlock_irqrestore(&rq->lock, *flags);
650 * this_rq_lock - lock this runqueue and disable interrupts.
652 static struct rq *this_rq_lock(void)
659 spin_lock(&rq->lock);
665 * We are going deep-idle (irqs are disabled):
667 void sched_clock_idle_sleep_event(void)
669 struct rq *rq = cpu_rq(smp_processor_id());
671 spin_lock(&rq->lock);
672 __update_rq_clock(rq);
673 spin_unlock(&rq->lock);
674 rq->clock_deep_idle_events++;
676 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
679 * We just idled delta nanoseconds (called with irqs disabled):
681 void sched_clock_idle_wakeup_event(u64 delta_ns)
683 struct rq *rq = cpu_rq(smp_processor_id());
684 u64 now = sched_clock();
686 touch_softlockup_watchdog();
687 rq->idle_clock += delta_ns;
689 * Override the previous timestamp and ignore all
690 * sched_clock() deltas that occured while we idled,
691 * and use the PM-provided delta_ns to advance the
694 spin_lock(&rq->lock);
695 rq->prev_clock_raw = now;
696 rq->clock += delta_ns;
697 spin_unlock(&rq->lock);
699 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
702 * resched_task - mark a task 'to be rescheduled now'.
704 * On UP this means the setting of the need_resched flag, on SMP it
705 * might also involve a cross-CPU call to trigger the scheduler on
710 #ifndef tsk_is_polling
711 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
714 static void resched_task(struct task_struct *p)
718 assert_spin_locked(&task_rq(p)->lock);
720 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
723 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
726 if (cpu == smp_processor_id())
729 /* NEED_RESCHED must be visible before we test polling */
731 if (!tsk_is_polling(p))
732 smp_send_reschedule(cpu);
735 static void resched_cpu(int cpu)
737 struct rq *rq = cpu_rq(cpu);
740 if (!spin_trylock_irqsave(&rq->lock, flags))
742 resched_task(cpu_curr(cpu));
743 spin_unlock_irqrestore(&rq->lock, flags);
746 static inline void resched_task(struct task_struct *p)
748 assert_spin_locked(&task_rq(p)->lock);
749 set_tsk_need_resched(p);
753 #if BITS_PER_LONG == 32
754 # define WMULT_CONST (~0UL)
756 # define WMULT_CONST (1UL << 32)
759 #define WMULT_SHIFT 32
762 * Shift right and round:
764 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
767 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
768 struct load_weight *lw)
772 if (unlikely(!lw->inv_weight))
773 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
775 tmp = (u64)delta_exec * weight;
777 * Check whether we'd overflow the 64-bit multiplication:
779 if (unlikely(tmp > WMULT_CONST))
780 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
783 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
785 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
788 static inline unsigned long
789 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
791 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
794 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
799 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
805 * To aid in avoiding the subversion of "niceness" due to uneven distribution
806 * of tasks with abnormal "nice" values across CPUs the contribution that
807 * each task makes to its run queue's load is weighted according to its
808 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
809 * scaled version of the new time slice allocation that they receive on time
813 #define WEIGHT_IDLEPRIO 2
814 #define WMULT_IDLEPRIO (1 << 31)
817 * Nice levels are multiplicative, with a gentle 10% change for every
818 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
819 * nice 1, it will get ~10% less CPU time than another CPU-bound task
820 * that remained on nice 0.
822 * The "10% effect" is relative and cumulative: from _any_ nice level,
823 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
824 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
825 * If a task goes up by ~10% and another task goes down by ~10% then
826 * the relative distance between them is ~25%.)
828 static const int prio_to_weight[40] = {
829 /* -20 */ 88761, 71755, 56483, 46273, 36291,
830 /* -15 */ 29154, 23254, 18705, 14949, 11916,
831 /* -10 */ 9548, 7620, 6100, 4904, 3906,
832 /* -5 */ 3121, 2501, 1991, 1586, 1277,
833 /* 0 */ 1024, 820, 655, 526, 423,
834 /* 5 */ 335, 272, 215, 172, 137,
835 /* 10 */ 110, 87, 70, 56, 45,
836 /* 15 */ 36, 29, 23, 18, 15,
840 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
842 * In cases where the weight does not change often, we can use the
843 * precalculated inverse to speed up arithmetics by turning divisions
844 * into multiplications:
846 static const u32 prio_to_wmult[40] = {
847 /* -20 */ 48388, 59856, 76040, 92818, 118348,
848 /* -15 */ 147320, 184698, 229616, 287308, 360437,
849 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
850 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
851 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
852 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
853 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
854 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
857 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
860 * runqueue iterator, to support SMP load-balancing between different
861 * scheduling classes, without having to expose their internal data
862 * structures to the load-balancing proper:
866 struct task_struct *(*start)(void *);
867 struct task_struct *(*next)(void *);
872 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
873 unsigned long max_load_move, struct sched_domain *sd,
874 enum cpu_idle_type idle, int *all_pinned,
875 int *this_best_prio, struct rq_iterator *iterator);
878 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
879 struct sched_domain *sd, enum cpu_idle_type idle,
880 struct rq_iterator *iterator);
883 #ifdef CONFIG_CGROUP_CPUACCT
884 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
886 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
889 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
891 update_load_add(&rq->load, load);
894 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
896 update_load_sub(&rq->load, load);
899 #include "sched_stats.h"
900 #include "sched_idletask.c"
901 #include "sched_fair.c"
902 #include "sched_rt.c"
903 #ifdef CONFIG_SCHED_DEBUG
904 # include "sched_debug.c"
907 #define sched_class_highest (&rt_sched_class)
909 static void inc_nr_running(struct task_struct *p, struct rq *rq)
914 static void dec_nr_running(struct task_struct *p, struct rq *rq)
919 static void set_load_weight(struct task_struct *p)
921 if (task_has_rt_policy(p)) {
922 p->se.load.weight = prio_to_weight[0] * 2;
923 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
928 * SCHED_IDLE tasks get minimal weight:
930 if (p->policy == SCHED_IDLE) {
931 p->se.load.weight = WEIGHT_IDLEPRIO;
932 p->se.load.inv_weight = WMULT_IDLEPRIO;
936 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
937 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
940 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
942 sched_info_queued(p);
943 p->sched_class->enqueue_task(rq, p, wakeup);
947 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
949 p->sched_class->dequeue_task(rq, p, sleep);
954 * __normal_prio - return the priority that is based on the static prio
956 static inline int __normal_prio(struct task_struct *p)
958 return p->static_prio;
962 * Calculate the expected normal priority: i.e. priority
963 * without taking RT-inheritance into account. Might be
964 * boosted by interactivity modifiers. Changes upon fork,
965 * setprio syscalls, and whenever the interactivity
966 * estimator recalculates.
968 static inline int normal_prio(struct task_struct *p)
972 if (task_has_rt_policy(p))
973 prio = MAX_RT_PRIO-1 - p->rt_priority;
975 prio = __normal_prio(p);
980 * Calculate the current priority, i.e. the priority
981 * taken into account by the scheduler. This value might
982 * be boosted by RT tasks, or might be boosted by
983 * interactivity modifiers. Will be RT if the task got
984 * RT-boosted. If not then it returns p->normal_prio.
986 static int effective_prio(struct task_struct *p)
988 p->normal_prio = normal_prio(p);
990 * If we are RT tasks or we were boosted to RT priority,
991 * keep the priority unchanged. Otherwise, update priority
992 * to the normal priority:
994 if (!rt_prio(p->prio))
995 return p->normal_prio;
1000 * activate_task - move a task to the runqueue.
1002 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1004 if (p->state == TASK_UNINTERRUPTIBLE)
1005 rq->nr_uninterruptible--;
1007 enqueue_task(rq, p, wakeup);
1008 inc_nr_running(p, rq);
1012 * deactivate_task - remove a task from the runqueue.
1014 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1016 if (p->state == TASK_UNINTERRUPTIBLE)
1017 rq->nr_uninterruptible++;
1019 dequeue_task(rq, p, sleep);
1020 dec_nr_running(p, rq);
1024 * task_curr - is this task currently executing on a CPU?
1025 * @p: the task in question.
1027 inline int task_curr(const struct task_struct *p)
1029 return cpu_curr(task_cpu(p)) == p;
1032 /* Used instead of source_load when we know the type == 0 */
1033 unsigned long weighted_cpuload(const int cpu)
1035 return cpu_rq(cpu)->load.weight;
1038 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1040 set_task_cfs_rq(p, cpu);
1043 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1044 * successfuly executed on another CPU. We must ensure that updates of
1045 * per-task data have been completed by this moment.
1048 task_thread_info(p)->cpu = cpu;
1055 * Is this task likely cache-hot:
1058 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1062 if (p->sched_class != &fair_sched_class)
1065 if (sysctl_sched_migration_cost == -1)
1067 if (sysctl_sched_migration_cost == 0)
1070 delta = now - p->se.exec_start;
1072 return delta < (s64)sysctl_sched_migration_cost;
1076 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1078 int old_cpu = task_cpu(p);
1079 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1080 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1081 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1084 clock_offset = old_rq->clock - new_rq->clock;
1086 #ifdef CONFIG_SCHEDSTATS
1087 if (p->se.wait_start)
1088 p->se.wait_start -= clock_offset;
1089 if (p->se.sleep_start)
1090 p->se.sleep_start -= clock_offset;
1091 if (p->se.block_start)
1092 p->se.block_start -= clock_offset;
1093 if (old_cpu != new_cpu) {
1094 schedstat_inc(p, se.nr_migrations);
1095 if (task_hot(p, old_rq->clock, NULL))
1096 schedstat_inc(p, se.nr_forced2_migrations);
1099 p->se.vruntime -= old_cfsrq->min_vruntime -
1100 new_cfsrq->min_vruntime;
1102 __set_task_cpu(p, new_cpu);
1105 struct migration_req {
1106 struct list_head list;
1108 struct task_struct *task;
1111 struct completion done;
1115 * The task's runqueue lock must be held.
1116 * Returns true if you have to wait for migration thread.
1119 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1121 struct rq *rq = task_rq(p);
1124 * If the task is not on a runqueue (and not running), then
1125 * it is sufficient to simply update the task's cpu field.
1127 if (!p->se.on_rq && !task_running(rq, p)) {
1128 set_task_cpu(p, dest_cpu);
1132 init_completion(&req->done);
1134 req->dest_cpu = dest_cpu;
1135 list_add(&req->list, &rq->migration_queue);
1141 * wait_task_inactive - wait for a thread to unschedule.
1143 * The caller must ensure that the task *will* unschedule sometime soon,
1144 * else this function might spin for a *long* time. This function can't
1145 * be called with interrupts off, or it may introduce deadlock with
1146 * smp_call_function() if an IPI is sent by the same process we are
1147 * waiting to become inactive.
1149 void wait_task_inactive(struct task_struct *p)
1151 unsigned long flags;
1157 * We do the initial early heuristics without holding
1158 * any task-queue locks at all. We'll only try to get
1159 * the runqueue lock when things look like they will
1165 * If the task is actively running on another CPU
1166 * still, just relax and busy-wait without holding
1169 * NOTE! Since we don't hold any locks, it's not
1170 * even sure that "rq" stays as the right runqueue!
1171 * But we don't care, since "task_running()" will
1172 * return false if the runqueue has changed and p
1173 * is actually now running somewhere else!
1175 while (task_running(rq, p))
1179 * Ok, time to look more closely! We need the rq
1180 * lock now, to be *sure*. If we're wrong, we'll
1181 * just go back and repeat.
1183 rq = task_rq_lock(p, &flags);
1184 running = task_running(rq, p);
1185 on_rq = p->se.on_rq;
1186 task_rq_unlock(rq, &flags);
1189 * Was it really running after all now that we
1190 * checked with the proper locks actually held?
1192 * Oops. Go back and try again..
1194 if (unlikely(running)) {
1200 * It's not enough that it's not actively running,
1201 * it must be off the runqueue _entirely_, and not
1204 * So if it wa still runnable (but just not actively
1205 * running right now), it's preempted, and we should
1206 * yield - it could be a while.
1208 if (unlikely(on_rq)) {
1209 schedule_timeout_uninterruptible(1);
1214 * Ahh, all good. It wasn't running, and it wasn't
1215 * runnable, which means that it will never become
1216 * running in the future either. We're all done!
1223 * kick_process - kick a running thread to enter/exit the kernel
1224 * @p: the to-be-kicked thread
1226 * Cause a process which is running on another CPU to enter
1227 * kernel-mode, without any delay. (to get signals handled.)
1229 * NOTE: this function doesnt have to take the runqueue lock,
1230 * because all it wants to ensure is that the remote task enters
1231 * the kernel. If the IPI races and the task has been migrated
1232 * to another CPU then no harm is done and the purpose has been
1235 void kick_process(struct task_struct *p)
1241 if ((cpu != smp_processor_id()) && task_curr(p))
1242 smp_send_reschedule(cpu);
1247 * Return a low guess at the load of a migration-source cpu weighted
1248 * according to the scheduling class and "nice" value.
1250 * We want to under-estimate the load of migration sources, to
1251 * balance conservatively.
1253 static unsigned long source_load(int cpu, int type)
1255 struct rq *rq = cpu_rq(cpu);
1256 unsigned long total = weighted_cpuload(cpu);
1261 return min(rq->cpu_load[type-1], total);
1265 * Return a high guess at the load of a migration-target cpu weighted
1266 * according to the scheduling class and "nice" value.
1268 static unsigned long target_load(int cpu, int type)
1270 struct rq *rq = cpu_rq(cpu);
1271 unsigned long total = weighted_cpuload(cpu);
1276 return max(rq->cpu_load[type-1], total);
1280 * Return the average load per task on the cpu's run queue
1282 static inline unsigned long cpu_avg_load_per_task(int cpu)
1284 struct rq *rq = cpu_rq(cpu);
1285 unsigned long total = weighted_cpuload(cpu);
1286 unsigned long n = rq->nr_running;
1288 return n ? total / n : SCHED_LOAD_SCALE;
1292 * find_idlest_group finds and returns the least busy CPU group within the
1295 static struct sched_group *
1296 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1298 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1299 unsigned long min_load = ULONG_MAX, this_load = 0;
1300 int load_idx = sd->forkexec_idx;
1301 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1304 unsigned long load, avg_load;
1308 /* Skip over this group if it has no CPUs allowed */
1309 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1312 local_group = cpu_isset(this_cpu, group->cpumask);
1314 /* Tally up the load of all CPUs in the group */
1317 for_each_cpu_mask(i, group->cpumask) {
1318 /* Bias balancing toward cpus of our domain */
1320 load = source_load(i, load_idx);
1322 load = target_load(i, load_idx);
1327 /* Adjust by relative CPU power of the group */
1328 avg_load = sg_div_cpu_power(group,
1329 avg_load * SCHED_LOAD_SCALE);
1332 this_load = avg_load;
1334 } else if (avg_load < min_load) {
1335 min_load = avg_load;
1338 } while (group = group->next, group != sd->groups);
1340 if (!idlest || 100*this_load < imbalance*min_load)
1346 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1349 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1352 unsigned long load, min_load = ULONG_MAX;
1356 /* Traverse only the allowed CPUs */
1357 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1359 for_each_cpu_mask(i, tmp) {
1360 load = weighted_cpuload(i);
1362 if (load < min_load || (load == min_load && i == this_cpu)) {
1372 * sched_balance_self: balance the current task (running on cpu) in domains
1373 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1376 * Balance, ie. select the least loaded group.
1378 * Returns the target CPU number, or the same CPU if no balancing is needed.
1380 * preempt must be disabled.
1382 static int sched_balance_self(int cpu, int flag)
1384 struct task_struct *t = current;
1385 struct sched_domain *tmp, *sd = NULL;
1387 for_each_domain(cpu, tmp) {
1389 * If power savings logic is enabled for a domain, stop there.
1391 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1393 if (tmp->flags & flag)
1399 struct sched_group *group;
1400 int new_cpu, weight;
1402 if (!(sd->flags & flag)) {
1408 group = find_idlest_group(sd, t, cpu);
1414 new_cpu = find_idlest_cpu(group, t, cpu);
1415 if (new_cpu == -1 || new_cpu == cpu) {
1416 /* Now try balancing at a lower domain level of cpu */
1421 /* Now try balancing at a lower domain level of new_cpu */
1424 weight = cpus_weight(span);
1425 for_each_domain(cpu, tmp) {
1426 if (weight <= cpus_weight(tmp->span))
1428 if (tmp->flags & flag)
1431 /* while loop will break here if sd == NULL */
1437 #endif /* CONFIG_SMP */
1440 * wake_idle() will wake a task on an idle cpu if task->cpu is
1441 * not idle and an idle cpu is available. The span of cpus to
1442 * search starts with cpus closest then further out as needed,
1443 * so we always favor a closer, idle cpu.
1445 * Returns the CPU we should wake onto.
1447 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1448 static int wake_idle(int cpu, struct task_struct *p)
1451 struct sched_domain *sd;
1455 * If it is idle, then it is the best cpu to run this task.
1457 * This cpu is also the best, if it has more than one task already.
1458 * Siblings must be also busy(in most cases) as they didn't already
1459 * pickup the extra load from this cpu and hence we need not check
1460 * sibling runqueue info. This will avoid the checks and cache miss
1461 * penalities associated with that.
1463 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1466 for_each_domain(cpu, sd) {
1467 if (sd->flags & SD_WAKE_IDLE) {
1468 cpus_and(tmp, sd->span, p->cpus_allowed);
1469 for_each_cpu_mask(i, tmp) {
1471 if (i != task_cpu(p)) {
1473 se.nr_wakeups_idle);
1485 static inline int wake_idle(int cpu, struct task_struct *p)
1492 * try_to_wake_up - wake up a thread
1493 * @p: the to-be-woken-up thread
1494 * @state: the mask of task states that can be woken
1495 * @sync: do a synchronous wakeup?
1497 * Put it on the run-queue if it's not already there. The "current"
1498 * thread is always on the run-queue (except when the actual
1499 * re-schedule is in progress), and as such you're allowed to do
1500 * the simpler "current->state = TASK_RUNNING" to mark yourself
1501 * runnable without the overhead of this.
1503 * returns failure only if the task is already active.
1505 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1507 int cpu, orig_cpu, this_cpu, success = 0;
1508 unsigned long flags;
1512 struct sched_domain *sd, *this_sd = NULL;
1513 unsigned long load, this_load;
1517 rq = task_rq_lock(p, &flags);
1518 old_state = p->state;
1519 if (!(old_state & state))
1527 this_cpu = smp_processor_id();
1530 if (unlikely(task_running(rq, p)))
1535 schedstat_inc(rq, ttwu_count);
1536 if (cpu == this_cpu) {
1537 schedstat_inc(rq, ttwu_local);
1541 for_each_domain(this_cpu, sd) {
1542 if (cpu_isset(cpu, sd->span)) {
1543 schedstat_inc(sd, ttwu_wake_remote);
1549 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1553 * Check for affine wakeup and passive balancing possibilities.
1556 int idx = this_sd->wake_idx;
1557 unsigned int imbalance;
1559 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1561 load = source_load(cpu, idx);
1562 this_load = target_load(this_cpu, idx);
1564 new_cpu = this_cpu; /* Wake to this CPU if we can */
1566 if (this_sd->flags & SD_WAKE_AFFINE) {
1567 unsigned long tl = this_load;
1568 unsigned long tl_per_task;
1571 * Attract cache-cold tasks on sync wakeups:
1573 if (sync && !task_hot(p, rq->clock, this_sd))
1576 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1577 tl_per_task = cpu_avg_load_per_task(this_cpu);
1580 * If sync wakeup then subtract the (maximum possible)
1581 * effect of the currently running task from the load
1582 * of the current CPU:
1585 tl -= current->se.load.weight;
1588 tl + target_load(cpu, idx) <= tl_per_task) ||
1589 100*(tl + p->se.load.weight) <= imbalance*load) {
1591 * This domain has SD_WAKE_AFFINE and
1592 * p is cache cold in this domain, and
1593 * there is no bad imbalance.
1595 schedstat_inc(this_sd, ttwu_move_affine);
1596 schedstat_inc(p, se.nr_wakeups_affine);
1602 * Start passive balancing when half the imbalance_pct
1605 if (this_sd->flags & SD_WAKE_BALANCE) {
1606 if (imbalance*this_load <= 100*load) {
1607 schedstat_inc(this_sd, ttwu_move_balance);
1608 schedstat_inc(p, se.nr_wakeups_passive);
1614 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1616 new_cpu = wake_idle(new_cpu, p);
1617 if (new_cpu != cpu) {
1618 set_task_cpu(p, new_cpu);
1619 task_rq_unlock(rq, &flags);
1620 /* might preempt at this point */
1621 rq = task_rq_lock(p, &flags);
1622 old_state = p->state;
1623 if (!(old_state & state))
1628 this_cpu = smp_processor_id();
1633 #endif /* CONFIG_SMP */
1634 schedstat_inc(p, se.nr_wakeups);
1636 schedstat_inc(p, se.nr_wakeups_sync);
1637 if (orig_cpu != cpu)
1638 schedstat_inc(p, se.nr_wakeups_migrate);
1639 if (cpu == this_cpu)
1640 schedstat_inc(p, se.nr_wakeups_local);
1642 schedstat_inc(p, se.nr_wakeups_remote);
1643 update_rq_clock(rq);
1644 activate_task(rq, p, 1);
1645 check_preempt_curr(rq, p);
1649 p->state = TASK_RUNNING;
1651 task_rq_unlock(rq, &flags);
1656 int fastcall wake_up_process(struct task_struct *p)
1658 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1659 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1661 EXPORT_SYMBOL(wake_up_process);
1663 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1665 return try_to_wake_up(p, state, 0);
1669 * Perform scheduler related setup for a newly forked process p.
1670 * p is forked by current.
1672 * __sched_fork() is basic setup used by init_idle() too:
1674 static void __sched_fork(struct task_struct *p)
1676 p->se.exec_start = 0;
1677 p->se.sum_exec_runtime = 0;
1678 p->se.prev_sum_exec_runtime = 0;
1680 #ifdef CONFIG_SCHEDSTATS
1681 p->se.wait_start = 0;
1682 p->se.sum_sleep_runtime = 0;
1683 p->se.sleep_start = 0;
1684 p->se.block_start = 0;
1685 p->se.sleep_max = 0;
1686 p->se.block_max = 0;
1688 p->se.slice_max = 0;
1692 INIT_LIST_HEAD(&p->run_list);
1695 #ifdef CONFIG_PREEMPT_NOTIFIERS
1696 INIT_HLIST_HEAD(&p->preempt_notifiers);
1700 * We mark the process as running here, but have not actually
1701 * inserted it onto the runqueue yet. This guarantees that
1702 * nobody will actually run it, and a signal or other external
1703 * event cannot wake it up and insert it on the runqueue either.
1705 p->state = TASK_RUNNING;
1709 * fork()/clone()-time setup:
1711 void sched_fork(struct task_struct *p, int clone_flags)
1713 int cpu = get_cpu();
1718 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1720 set_task_cpu(p, cpu);
1723 * Make sure we do not leak PI boosting priority to the child:
1725 p->prio = current->normal_prio;
1726 if (!rt_prio(p->prio))
1727 p->sched_class = &fair_sched_class;
1729 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1730 if (likely(sched_info_on()))
1731 memset(&p->sched_info, 0, sizeof(p->sched_info));
1733 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1736 #ifdef CONFIG_PREEMPT
1737 /* Want to start with kernel preemption disabled. */
1738 task_thread_info(p)->preempt_count = 1;
1744 * wake_up_new_task - wake up a newly created task for the first time.
1746 * This function will do some initial scheduler statistics housekeeping
1747 * that must be done for every newly created context, then puts the task
1748 * on the runqueue and wakes it.
1750 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1752 unsigned long flags;
1755 rq = task_rq_lock(p, &flags);
1756 BUG_ON(p->state != TASK_RUNNING);
1757 update_rq_clock(rq);
1759 p->prio = effective_prio(p);
1761 if (!p->sched_class->task_new || !current->se.on_rq) {
1762 activate_task(rq, p, 0);
1765 * Let the scheduling class do new task startup
1766 * management (if any):
1768 p->sched_class->task_new(rq, p);
1769 inc_nr_running(p, rq);
1771 check_preempt_curr(rq, p);
1772 task_rq_unlock(rq, &flags);
1775 #ifdef CONFIG_PREEMPT_NOTIFIERS
1778 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1779 * @notifier: notifier struct to register
1781 void preempt_notifier_register(struct preempt_notifier *notifier)
1783 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1785 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1788 * preempt_notifier_unregister - no longer interested in preemption notifications
1789 * @notifier: notifier struct to unregister
1791 * This is safe to call from within a preemption notifier.
1793 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1795 hlist_del(¬ifier->link);
1797 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1799 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1801 struct preempt_notifier *notifier;
1802 struct hlist_node *node;
1804 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1805 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1809 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1810 struct task_struct *next)
1812 struct preempt_notifier *notifier;
1813 struct hlist_node *node;
1815 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1816 notifier->ops->sched_out(notifier, next);
1821 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1826 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1827 struct task_struct *next)
1834 * prepare_task_switch - prepare to switch tasks
1835 * @rq: the runqueue preparing to switch
1836 * @prev: the current task that is being switched out
1837 * @next: the task we are going to switch to.
1839 * This is called with the rq lock held and interrupts off. It must
1840 * be paired with a subsequent finish_task_switch after the context
1843 * prepare_task_switch sets up locking and calls architecture specific
1847 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1848 struct task_struct *next)
1850 fire_sched_out_preempt_notifiers(prev, next);
1851 prepare_lock_switch(rq, next);
1852 prepare_arch_switch(next);
1856 * finish_task_switch - clean up after a task-switch
1857 * @rq: runqueue associated with task-switch
1858 * @prev: the thread we just switched away from.
1860 * finish_task_switch must be called after the context switch, paired
1861 * with a prepare_task_switch call before the context switch.
1862 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1863 * and do any other architecture-specific cleanup actions.
1865 * Note that we may have delayed dropping an mm in context_switch(). If
1866 * so, we finish that here outside of the runqueue lock. (Doing it
1867 * with the lock held can cause deadlocks; see schedule() for
1870 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1871 __releases(rq->lock)
1873 struct mm_struct *mm = rq->prev_mm;
1879 * A task struct has one reference for the use as "current".
1880 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1881 * schedule one last time. The schedule call will never return, and
1882 * the scheduled task must drop that reference.
1883 * The test for TASK_DEAD must occur while the runqueue locks are
1884 * still held, otherwise prev could be scheduled on another cpu, die
1885 * there before we look at prev->state, and then the reference would
1887 * Manfred Spraul <manfred@colorfullife.com>
1889 prev_state = prev->state;
1890 finish_arch_switch(prev);
1891 finish_lock_switch(rq, prev);
1892 fire_sched_in_preempt_notifiers(current);
1895 if (unlikely(prev_state == TASK_DEAD)) {
1897 * Remove function-return probe instances associated with this
1898 * task and put them back on the free list.
1900 kprobe_flush_task(prev);
1901 put_task_struct(prev);
1906 * schedule_tail - first thing a freshly forked thread must call.
1907 * @prev: the thread we just switched away from.
1909 asmlinkage void schedule_tail(struct task_struct *prev)
1910 __releases(rq->lock)
1912 struct rq *rq = this_rq();
1914 finish_task_switch(rq, prev);
1915 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1916 /* In this case, finish_task_switch does not reenable preemption */
1919 if (current->set_child_tid)
1920 put_user(task_pid_vnr(current), current->set_child_tid);
1924 * context_switch - switch to the new MM and the new
1925 * thread's register state.
1928 context_switch(struct rq *rq, struct task_struct *prev,
1929 struct task_struct *next)
1931 struct mm_struct *mm, *oldmm;
1933 prepare_task_switch(rq, prev, next);
1935 oldmm = prev->active_mm;
1937 * For paravirt, this is coupled with an exit in switch_to to
1938 * combine the page table reload and the switch backend into
1941 arch_enter_lazy_cpu_mode();
1943 if (unlikely(!mm)) {
1944 next->active_mm = oldmm;
1945 atomic_inc(&oldmm->mm_count);
1946 enter_lazy_tlb(oldmm, next);
1948 switch_mm(oldmm, mm, next);
1950 if (unlikely(!prev->mm)) {
1951 prev->active_mm = NULL;
1952 rq->prev_mm = oldmm;
1955 * Since the runqueue lock will be released by the next
1956 * task (which is an invalid locking op but in the case
1957 * of the scheduler it's an obvious special-case), so we
1958 * do an early lockdep release here:
1960 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1961 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1964 /* Here we just switch the register state and the stack. */
1965 switch_to(prev, next, prev);
1969 * this_rq must be evaluated again because prev may have moved
1970 * CPUs since it called schedule(), thus the 'rq' on its stack
1971 * frame will be invalid.
1973 finish_task_switch(this_rq(), prev);
1977 * nr_running, nr_uninterruptible and nr_context_switches:
1979 * externally visible scheduler statistics: current number of runnable
1980 * threads, current number of uninterruptible-sleeping threads, total
1981 * number of context switches performed since bootup.
1983 unsigned long nr_running(void)
1985 unsigned long i, sum = 0;
1987 for_each_online_cpu(i)
1988 sum += cpu_rq(i)->nr_running;
1993 unsigned long nr_uninterruptible(void)
1995 unsigned long i, sum = 0;
1997 for_each_possible_cpu(i)
1998 sum += cpu_rq(i)->nr_uninterruptible;
2001 * Since we read the counters lockless, it might be slightly
2002 * inaccurate. Do not allow it to go below zero though:
2004 if (unlikely((long)sum < 0))
2010 unsigned long long nr_context_switches(void)
2013 unsigned long long sum = 0;
2015 for_each_possible_cpu(i)
2016 sum += cpu_rq(i)->nr_switches;
2021 unsigned long nr_iowait(void)
2023 unsigned long i, sum = 0;
2025 for_each_possible_cpu(i)
2026 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2031 unsigned long nr_active(void)
2033 unsigned long i, running = 0, uninterruptible = 0;
2035 for_each_online_cpu(i) {
2036 running += cpu_rq(i)->nr_running;
2037 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2040 if (unlikely((long)uninterruptible < 0))
2041 uninterruptible = 0;
2043 return running + uninterruptible;
2047 * Update rq->cpu_load[] statistics. This function is usually called every
2048 * scheduler tick (TICK_NSEC).
2050 static void update_cpu_load(struct rq *this_rq)
2052 unsigned long this_load = this_rq->load.weight;
2055 this_rq->nr_load_updates++;
2057 /* Update our load: */
2058 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2059 unsigned long old_load, new_load;
2061 /* scale is effectively 1 << i now, and >> i divides by scale */
2063 old_load = this_rq->cpu_load[i];
2064 new_load = this_load;
2066 * Round up the averaging division if load is increasing. This
2067 * prevents us from getting stuck on 9 if the load is 10, for
2070 if (new_load > old_load)
2071 new_load += scale-1;
2072 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2079 * double_rq_lock - safely lock two runqueues
2081 * Note this does not disable interrupts like task_rq_lock,
2082 * you need to do so manually before calling.
2084 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2085 __acquires(rq1->lock)
2086 __acquires(rq2->lock)
2088 BUG_ON(!irqs_disabled());
2090 spin_lock(&rq1->lock);
2091 __acquire(rq2->lock); /* Fake it out ;) */
2094 spin_lock(&rq1->lock);
2095 spin_lock(&rq2->lock);
2097 spin_lock(&rq2->lock);
2098 spin_lock(&rq1->lock);
2101 update_rq_clock(rq1);
2102 update_rq_clock(rq2);
2106 * double_rq_unlock - safely unlock two runqueues
2108 * Note this does not restore interrupts like task_rq_unlock,
2109 * you need to do so manually after calling.
2111 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2112 __releases(rq1->lock)
2113 __releases(rq2->lock)
2115 spin_unlock(&rq1->lock);
2117 spin_unlock(&rq2->lock);
2119 __release(rq2->lock);
2123 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2125 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2126 __releases(this_rq->lock)
2127 __acquires(busiest->lock)
2128 __acquires(this_rq->lock)
2130 if (unlikely(!irqs_disabled())) {
2131 /* printk() doesn't work good under rq->lock */
2132 spin_unlock(&this_rq->lock);
2135 if (unlikely(!spin_trylock(&busiest->lock))) {
2136 if (busiest < this_rq) {
2137 spin_unlock(&this_rq->lock);
2138 spin_lock(&busiest->lock);
2139 spin_lock(&this_rq->lock);
2141 spin_lock(&busiest->lock);
2146 * If dest_cpu is allowed for this process, migrate the task to it.
2147 * This is accomplished by forcing the cpu_allowed mask to only
2148 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2149 * the cpu_allowed mask is restored.
2151 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2153 struct migration_req req;
2154 unsigned long flags;
2157 rq = task_rq_lock(p, &flags);
2158 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2159 || unlikely(cpu_is_offline(dest_cpu)))
2162 /* force the process onto the specified CPU */
2163 if (migrate_task(p, dest_cpu, &req)) {
2164 /* Need to wait for migration thread (might exit: take ref). */
2165 struct task_struct *mt = rq->migration_thread;
2167 get_task_struct(mt);
2168 task_rq_unlock(rq, &flags);
2169 wake_up_process(mt);
2170 put_task_struct(mt);
2171 wait_for_completion(&req.done);
2176 task_rq_unlock(rq, &flags);
2180 * sched_exec - execve() is a valuable balancing opportunity, because at
2181 * this point the task has the smallest effective memory and cache footprint.
2183 void sched_exec(void)
2185 int new_cpu, this_cpu = get_cpu();
2186 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2188 if (new_cpu != this_cpu)
2189 sched_migrate_task(current, new_cpu);
2193 * pull_task - move a task from a remote runqueue to the local runqueue.
2194 * Both runqueues must be locked.
2196 static void pull_task(struct rq *src_rq, struct task_struct *p,
2197 struct rq *this_rq, int this_cpu)
2199 deactivate_task(src_rq, p, 0);
2200 set_task_cpu(p, this_cpu);
2201 activate_task(this_rq, p, 0);
2203 * Note that idle threads have a prio of MAX_PRIO, for this test
2204 * to be always true for them.
2206 check_preempt_curr(this_rq, p);
2210 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2213 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2214 struct sched_domain *sd, enum cpu_idle_type idle,
2218 * We do not migrate tasks that are:
2219 * 1) running (obviously), or
2220 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2221 * 3) are cache-hot on their current CPU.
2223 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2224 schedstat_inc(p, se.nr_failed_migrations_affine);
2229 if (task_running(rq, p)) {
2230 schedstat_inc(p, se.nr_failed_migrations_running);
2235 * Aggressive migration if:
2236 * 1) task is cache cold, or
2237 * 2) too many balance attempts have failed.
2240 if (!task_hot(p, rq->clock, sd) ||
2241 sd->nr_balance_failed > sd->cache_nice_tries) {
2242 #ifdef CONFIG_SCHEDSTATS
2243 if (task_hot(p, rq->clock, sd)) {
2244 schedstat_inc(sd, lb_hot_gained[idle]);
2245 schedstat_inc(p, se.nr_forced_migrations);
2251 if (task_hot(p, rq->clock, sd)) {
2252 schedstat_inc(p, se.nr_failed_migrations_hot);
2258 static unsigned long
2259 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2260 unsigned long max_load_move, struct sched_domain *sd,
2261 enum cpu_idle_type idle, int *all_pinned,
2262 int *this_best_prio, struct rq_iterator *iterator)
2264 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2265 struct task_struct *p;
2266 long rem_load_move = max_load_move;
2268 if (max_load_move == 0)
2274 * Start the load-balancing iterator:
2276 p = iterator->start(iterator->arg);
2278 if (!p || loops++ > sysctl_sched_nr_migrate)
2281 * To help distribute high priority tasks across CPUs we don't
2282 * skip a task if it will be the highest priority task (i.e. smallest
2283 * prio value) on its new queue regardless of its load weight
2285 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2286 SCHED_LOAD_SCALE_FUZZ;
2287 if ((skip_for_load && p->prio >= *this_best_prio) ||
2288 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2289 p = iterator->next(iterator->arg);
2293 pull_task(busiest, p, this_rq, this_cpu);
2295 rem_load_move -= p->se.load.weight;
2298 * We only want to steal up to the prescribed amount of weighted load.
2300 if (rem_load_move > 0) {
2301 if (p->prio < *this_best_prio)
2302 *this_best_prio = p->prio;
2303 p = iterator->next(iterator->arg);
2308 * Right now, this is one of only two places pull_task() is called,
2309 * so we can safely collect pull_task() stats here rather than
2310 * inside pull_task().
2312 schedstat_add(sd, lb_gained[idle], pulled);
2315 *all_pinned = pinned;
2317 return max_load_move - rem_load_move;
2321 * move_tasks tries to move up to max_load_move weighted load from busiest to
2322 * this_rq, as part of a balancing operation within domain "sd".
2323 * Returns 1 if successful and 0 otherwise.
2325 * Called with both runqueues locked.
2327 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2328 unsigned long max_load_move,
2329 struct sched_domain *sd, enum cpu_idle_type idle,
2332 const struct sched_class *class = sched_class_highest;
2333 unsigned long total_load_moved = 0;
2334 int this_best_prio = this_rq->curr->prio;
2338 class->load_balance(this_rq, this_cpu, busiest,
2339 max_load_move - total_load_moved,
2340 sd, idle, all_pinned, &this_best_prio);
2341 class = class->next;
2342 } while (class && max_load_move > total_load_moved);
2344 return total_load_moved > 0;
2348 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2349 struct sched_domain *sd, enum cpu_idle_type idle,
2350 struct rq_iterator *iterator)
2352 struct task_struct *p = iterator->start(iterator->arg);
2356 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2357 pull_task(busiest, p, this_rq, this_cpu);
2359 * Right now, this is only the second place pull_task()
2360 * is called, so we can safely collect pull_task()
2361 * stats here rather than inside pull_task().
2363 schedstat_inc(sd, lb_gained[idle]);
2367 p = iterator->next(iterator->arg);
2374 * move_one_task tries to move exactly one task from busiest to this_rq, as
2375 * part of active balancing operations within "domain".
2376 * Returns 1 if successful and 0 otherwise.
2378 * Called with both runqueues locked.
2380 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2381 struct sched_domain *sd, enum cpu_idle_type idle)
2383 const struct sched_class *class;
2385 for (class = sched_class_highest; class; class = class->next)
2386 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2393 * find_busiest_group finds and returns the busiest CPU group within the
2394 * domain. It calculates and returns the amount of weighted load which
2395 * should be moved to restore balance via the imbalance parameter.
2397 static struct sched_group *
2398 find_busiest_group(struct sched_domain *sd, int this_cpu,
2399 unsigned long *imbalance, enum cpu_idle_type idle,
2400 int *sd_idle, cpumask_t *cpus, int *balance)
2402 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2403 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2404 unsigned long max_pull;
2405 unsigned long busiest_load_per_task, busiest_nr_running;
2406 unsigned long this_load_per_task, this_nr_running;
2407 int load_idx, group_imb = 0;
2408 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2409 int power_savings_balance = 1;
2410 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2411 unsigned long min_nr_running = ULONG_MAX;
2412 struct sched_group *group_min = NULL, *group_leader = NULL;
2415 max_load = this_load = total_load = total_pwr = 0;
2416 busiest_load_per_task = busiest_nr_running = 0;
2417 this_load_per_task = this_nr_running = 0;
2418 if (idle == CPU_NOT_IDLE)
2419 load_idx = sd->busy_idx;
2420 else if (idle == CPU_NEWLY_IDLE)
2421 load_idx = sd->newidle_idx;
2423 load_idx = sd->idle_idx;
2426 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2429 int __group_imb = 0;
2430 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2431 unsigned long sum_nr_running, sum_weighted_load;
2433 local_group = cpu_isset(this_cpu, group->cpumask);
2436 balance_cpu = first_cpu(group->cpumask);
2438 /* Tally up the load of all CPUs in the group */
2439 sum_weighted_load = sum_nr_running = avg_load = 0;
2441 min_cpu_load = ~0UL;
2443 for_each_cpu_mask(i, group->cpumask) {
2446 if (!cpu_isset(i, *cpus))
2451 if (*sd_idle && rq->nr_running)
2454 /* Bias balancing toward cpus of our domain */
2456 if (idle_cpu(i) && !first_idle_cpu) {
2461 load = target_load(i, load_idx);
2463 load = source_load(i, load_idx);
2464 if (load > max_cpu_load)
2465 max_cpu_load = load;
2466 if (min_cpu_load > load)
2467 min_cpu_load = load;
2471 sum_nr_running += rq->nr_running;
2472 sum_weighted_load += weighted_cpuload(i);
2476 * First idle cpu or the first cpu(busiest) in this sched group
2477 * is eligible for doing load balancing at this and above
2478 * domains. In the newly idle case, we will allow all the cpu's
2479 * to do the newly idle load balance.
2481 if (idle != CPU_NEWLY_IDLE && local_group &&
2482 balance_cpu != this_cpu && balance) {
2487 total_load += avg_load;
2488 total_pwr += group->__cpu_power;
2490 /* Adjust by relative CPU power of the group */
2491 avg_load = sg_div_cpu_power(group,
2492 avg_load * SCHED_LOAD_SCALE);
2494 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2497 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2500 this_load = avg_load;
2502 this_nr_running = sum_nr_running;
2503 this_load_per_task = sum_weighted_load;
2504 } else if (avg_load > max_load &&
2505 (sum_nr_running > group_capacity || __group_imb)) {
2506 max_load = avg_load;
2508 busiest_nr_running = sum_nr_running;
2509 busiest_load_per_task = sum_weighted_load;
2510 group_imb = __group_imb;
2513 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2515 * Busy processors will not participate in power savings
2518 if (idle == CPU_NOT_IDLE ||
2519 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2523 * If the local group is idle or completely loaded
2524 * no need to do power savings balance at this domain
2526 if (local_group && (this_nr_running >= group_capacity ||
2528 power_savings_balance = 0;
2531 * If a group is already running at full capacity or idle,
2532 * don't include that group in power savings calculations
2534 if (!power_savings_balance || sum_nr_running >= group_capacity
2539 * Calculate the group which has the least non-idle load.
2540 * This is the group from where we need to pick up the load
2543 if ((sum_nr_running < min_nr_running) ||
2544 (sum_nr_running == min_nr_running &&
2545 first_cpu(group->cpumask) <
2546 first_cpu(group_min->cpumask))) {
2548 min_nr_running = sum_nr_running;
2549 min_load_per_task = sum_weighted_load /
2554 * Calculate the group which is almost near its
2555 * capacity but still has some space to pick up some load
2556 * from other group and save more power
2558 if (sum_nr_running <= group_capacity - 1) {
2559 if (sum_nr_running > leader_nr_running ||
2560 (sum_nr_running == leader_nr_running &&
2561 first_cpu(group->cpumask) >
2562 first_cpu(group_leader->cpumask))) {
2563 group_leader = group;
2564 leader_nr_running = sum_nr_running;
2569 group = group->next;
2570 } while (group != sd->groups);
2572 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2575 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2577 if (this_load >= avg_load ||
2578 100*max_load <= sd->imbalance_pct*this_load)
2581 busiest_load_per_task /= busiest_nr_running;
2583 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2586 * We're trying to get all the cpus to the average_load, so we don't
2587 * want to push ourselves above the average load, nor do we wish to
2588 * reduce the max loaded cpu below the average load, as either of these
2589 * actions would just result in more rebalancing later, and ping-pong
2590 * tasks around. Thus we look for the minimum possible imbalance.
2591 * Negative imbalances (*we* are more loaded than anyone else) will
2592 * be counted as no imbalance for these purposes -- we can't fix that
2593 * by pulling tasks to us. Be careful of negative numbers as they'll
2594 * appear as very large values with unsigned longs.
2596 if (max_load <= busiest_load_per_task)
2600 * In the presence of smp nice balancing, certain scenarios can have
2601 * max load less than avg load(as we skip the groups at or below
2602 * its cpu_power, while calculating max_load..)
2604 if (max_load < avg_load) {
2606 goto small_imbalance;
2609 /* Don't want to pull so many tasks that a group would go idle */
2610 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2612 /* How much load to actually move to equalise the imbalance */
2613 *imbalance = min(max_pull * busiest->__cpu_power,
2614 (avg_load - this_load) * this->__cpu_power)
2618 * if *imbalance is less than the average load per runnable task
2619 * there is no gaurantee that any tasks will be moved so we'll have
2620 * a think about bumping its value to force at least one task to be
2623 if (*imbalance < busiest_load_per_task) {
2624 unsigned long tmp, pwr_now, pwr_move;
2628 pwr_move = pwr_now = 0;
2630 if (this_nr_running) {
2631 this_load_per_task /= this_nr_running;
2632 if (busiest_load_per_task > this_load_per_task)
2635 this_load_per_task = SCHED_LOAD_SCALE;
2637 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2638 busiest_load_per_task * imbn) {
2639 *imbalance = busiest_load_per_task;
2644 * OK, we don't have enough imbalance to justify moving tasks,
2645 * however we may be able to increase total CPU power used by
2649 pwr_now += busiest->__cpu_power *
2650 min(busiest_load_per_task, max_load);
2651 pwr_now += this->__cpu_power *
2652 min(this_load_per_task, this_load);
2653 pwr_now /= SCHED_LOAD_SCALE;
2655 /* Amount of load we'd subtract */
2656 tmp = sg_div_cpu_power(busiest,
2657 busiest_load_per_task * SCHED_LOAD_SCALE);
2659 pwr_move += busiest->__cpu_power *
2660 min(busiest_load_per_task, max_load - tmp);
2662 /* Amount of load we'd add */
2663 if (max_load * busiest->__cpu_power <
2664 busiest_load_per_task * SCHED_LOAD_SCALE)
2665 tmp = sg_div_cpu_power(this,
2666 max_load * busiest->__cpu_power);
2668 tmp = sg_div_cpu_power(this,
2669 busiest_load_per_task * SCHED_LOAD_SCALE);
2670 pwr_move += this->__cpu_power *
2671 min(this_load_per_task, this_load + tmp);
2672 pwr_move /= SCHED_LOAD_SCALE;
2674 /* Move if we gain throughput */
2675 if (pwr_move > pwr_now)
2676 *imbalance = busiest_load_per_task;
2682 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2683 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2686 if (this == group_leader && group_leader != group_min) {
2687 *imbalance = min_load_per_task;
2697 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2700 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2701 unsigned long imbalance, cpumask_t *cpus)
2703 struct rq *busiest = NULL, *rq;
2704 unsigned long max_load = 0;
2707 for_each_cpu_mask(i, group->cpumask) {
2710 if (!cpu_isset(i, *cpus))
2714 wl = weighted_cpuload(i);
2716 if (rq->nr_running == 1 && wl > imbalance)
2719 if (wl > max_load) {
2729 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2730 * so long as it is large enough.
2732 #define MAX_PINNED_INTERVAL 512
2735 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2736 * tasks if there is an imbalance.
2738 static int load_balance(int this_cpu, struct rq *this_rq,
2739 struct sched_domain *sd, enum cpu_idle_type idle,
2742 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2743 struct sched_group *group;
2744 unsigned long imbalance;
2746 cpumask_t cpus = CPU_MASK_ALL;
2747 unsigned long flags;
2750 * When power savings policy is enabled for the parent domain, idle
2751 * sibling can pick up load irrespective of busy siblings. In this case,
2752 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2753 * portraying it as CPU_NOT_IDLE.
2755 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2756 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2759 schedstat_inc(sd, lb_count[idle]);
2762 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2769 schedstat_inc(sd, lb_nobusyg[idle]);
2773 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2775 schedstat_inc(sd, lb_nobusyq[idle]);
2779 BUG_ON(busiest == this_rq);
2781 schedstat_add(sd, lb_imbalance[idle], imbalance);
2784 if (busiest->nr_running > 1) {
2786 * Attempt to move tasks. If find_busiest_group has found
2787 * an imbalance but busiest->nr_running <= 1, the group is
2788 * still unbalanced. ld_moved simply stays zero, so it is
2789 * correctly treated as an imbalance.
2791 local_irq_save(flags);
2792 double_rq_lock(this_rq, busiest);
2793 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2794 imbalance, sd, idle, &all_pinned);
2795 double_rq_unlock(this_rq, busiest);
2796 local_irq_restore(flags);
2799 * some other cpu did the load balance for us.
2801 if (ld_moved && this_cpu != smp_processor_id())
2802 resched_cpu(this_cpu);
2804 /* All tasks on this runqueue were pinned by CPU affinity */
2805 if (unlikely(all_pinned)) {
2806 cpu_clear(cpu_of(busiest), cpus);
2807 if (!cpus_empty(cpus))
2814 schedstat_inc(sd, lb_failed[idle]);
2815 sd->nr_balance_failed++;
2817 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2819 spin_lock_irqsave(&busiest->lock, flags);
2821 /* don't kick the migration_thread, if the curr
2822 * task on busiest cpu can't be moved to this_cpu
2824 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2825 spin_unlock_irqrestore(&busiest->lock, flags);
2827 goto out_one_pinned;
2830 if (!busiest->active_balance) {
2831 busiest->active_balance = 1;
2832 busiest->push_cpu = this_cpu;
2835 spin_unlock_irqrestore(&busiest->lock, flags);
2837 wake_up_process(busiest->migration_thread);
2840 * We've kicked active balancing, reset the failure
2843 sd->nr_balance_failed = sd->cache_nice_tries+1;
2846 sd->nr_balance_failed = 0;
2848 if (likely(!active_balance)) {
2849 /* We were unbalanced, so reset the balancing interval */
2850 sd->balance_interval = sd->min_interval;
2853 * If we've begun active balancing, start to back off. This
2854 * case may not be covered by the all_pinned logic if there
2855 * is only 1 task on the busy runqueue (because we don't call
2858 if (sd->balance_interval < sd->max_interval)
2859 sd->balance_interval *= 2;
2862 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2863 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2868 schedstat_inc(sd, lb_balanced[idle]);
2870 sd->nr_balance_failed = 0;
2873 /* tune up the balancing interval */
2874 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2875 (sd->balance_interval < sd->max_interval))
2876 sd->balance_interval *= 2;
2878 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2879 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2885 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2886 * tasks if there is an imbalance.
2888 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2889 * this_rq is locked.
2892 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2894 struct sched_group *group;
2895 struct rq *busiest = NULL;
2896 unsigned long imbalance;
2900 cpumask_t cpus = CPU_MASK_ALL;
2903 * When power savings policy is enabled for the parent domain, idle
2904 * sibling can pick up load irrespective of busy siblings. In this case,
2905 * let the state of idle sibling percolate up as IDLE, instead of
2906 * portraying it as CPU_NOT_IDLE.
2908 if (sd->flags & SD_SHARE_CPUPOWER &&
2909 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2912 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2914 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2915 &sd_idle, &cpus, NULL);
2917 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2921 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2924 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2928 BUG_ON(busiest == this_rq);
2930 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2933 if (busiest->nr_running > 1) {
2934 /* Attempt to move tasks */
2935 double_lock_balance(this_rq, busiest);
2936 /* this_rq->clock is already updated */
2937 update_rq_clock(busiest);
2938 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2939 imbalance, sd, CPU_NEWLY_IDLE,
2941 spin_unlock(&busiest->lock);
2943 if (unlikely(all_pinned)) {
2944 cpu_clear(cpu_of(busiest), cpus);
2945 if (!cpus_empty(cpus))
2951 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2952 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2953 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2956 sd->nr_balance_failed = 0;
2961 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2962 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2963 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2965 sd->nr_balance_failed = 0;
2971 * idle_balance is called by schedule() if this_cpu is about to become
2972 * idle. Attempts to pull tasks from other CPUs.
2974 static void idle_balance(int this_cpu, struct rq *this_rq)
2976 struct sched_domain *sd;
2977 int pulled_task = -1;
2978 unsigned long next_balance = jiffies + HZ;
2980 for_each_domain(this_cpu, sd) {
2981 unsigned long interval;
2983 if (!(sd->flags & SD_LOAD_BALANCE))
2986 if (sd->flags & SD_BALANCE_NEWIDLE)
2987 /* If we've pulled tasks over stop searching: */
2988 pulled_task = load_balance_newidle(this_cpu,
2991 interval = msecs_to_jiffies(sd->balance_interval);
2992 if (time_after(next_balance, sd->last_balance + interval))
2993 next_balance = sd->last_balance + interval;
2997 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2999 * We are going idle. next_balance may be set based on
3000 * a busy processor. So reset next_balance.
3002 this_rq->next_balance = next_balance;
3007 * active_load_balance is run by migration threads. It pushes running tasks
3008 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3009 * running on each physical CPU where possible, and avoids physical /
3010 * logical imbalances.
3012 * Called with busiest_rq locked.
3014 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3016 int target_cpu = busiest_rq->push_cpu;
3017 struct sched_domain *sd;
3018 struct rq *target_rq;
3020 /* Is there any task to move? */
3021 if (busiest_rq->nr_running <= 1)
3024 target_rq = cpu_rq(target_cpu);
3027 * This condition is "impossible", if it occurs
3028 * we need to fix it. Originally reported by
3029 * Bjorn Helgaas on a 128-cpu setup.
3031 BUG_ON(busiest_rq == target_rq);
3033 /* move a task from busiest_rq to target_rq */
3034 double_lock_balance(busiest_rq, target_rq);
3035 update_rq_clock(busiest_rq);
3036 update_rq_clock(target_rq);
3038 /* Search for an sd spanning us and the target CPU. */
3039 for_each_domain(target_cpu, sd) {
3040 if ((sd->flags & SD_LOAD_BALANCE) &&
3041 cpu_isset(busiest_cpu, sd->span))
3046 schedstat_inc(sd, alb_count);
3048 if (move_one_task(target_rq, target_cpu, busiest_rq,
3050 schedstat_inc(sd, alb_pushed);
3052 schedstat_inc(sd, alb_failed);
3054 spin_unlock(&target_rq->lock);
3059 atomic_t load_balancer;
3061 } nohz ____cacheline_aligned = {
3062 .load_balancer = ATOMIC_INIT(-1),
3063 .cpu_mask = CPU_MASK_NONE,
3067 * This routine will try to nominate the ilb (idle load balancing)
3068 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3069 * load balancing on behalf of all those cpus. If all the cpus in the system
3070 * go into this tickless mode, then there will be no ilb owner (as there is
3071 * no need for one) and all the cpus will sleep till the next wakeup event
3074 * For the ilb owner, tick is not stopped. And this tick will be used
3075 * for idle load balancing. ilb owner will still be part of
3078 * While stopping the tick, this cpu will become the ilb owner if there
3079 * is no other owner. And will be the owner till that cpu becomes busy
3080 * or if all cpus in the system stop their ticks at which point
3081 * there is no need for ilb owner.
3083 * When the ilb owner becomes busy, it nominates another owner, during the
3084 * next busy scheduler_tick()
3086 int select_nohz_load_balancer(int stop_tick)
3088 int cpu = smp_processor_id();
3091 cpu_set(cpu, nohz.cpu_mask);
3092 cpu_rq(cpu)->in_nohz_recently = 1;
3095 * If we are going offline and still the leader, give up!
3097 if (cpu_is_offline(cpu) &&
3098 atomic_read(&nohz.load_balancer) == cpu) {
3099 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3104 /* time for ilb owner also to sleep */
3105 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3106 if (atomic_read(&nohz.load_balancer) == cpu)
3107 atomic_set(&nohz.load_balancer, -1);
3111 if (atomic_read(&nohz.load_balancer) == -1) {
3112 /* make me the ilb owner */
3113 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3115 } else if (atomic_read(&nohz.load_balancer) == cpu)
3118 if (!cpu_isset(cpu, nohz.cpu_mask))
3121 cpu_clear(cpu, nohz.cpu_mask);
3123 if (atomic_read(&nohz.load_balancer) == cpu)
3124 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3131 static DEFINE_SPINLOCK(balancing);
3134 * It checks each scheduling domain to see if it is due to be balanced,
3135 * and initiates a balancing operation if so.
3137 * Balancing parameters are set up in arch_init_sched_domains.
3139 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3142 struct rq *rq = cpu_rq(cpu);
3143 unsigned long interval;
3144 struct sched_domain *sd;
3145 /* Earliest time when we have to do rebalance again */
3146 unsigned long next_balance = jiffies + 60*HZ;
3147 int update_next_balance = 0;
3149 for_each_domain(cpu, sd) {
3150 if (!(sd->flags & SD_LOAD_BALANCE))
3153 interval = sd->balance_interval;
3154 if (idle != CPU_IDLE)
3155 interval *= sd->busy_factor;
3157 /* scale ms to jiffies */
3158 interval = msecs_to_jiffies(interval);
3159 if (unlikely(!interval))
3161 if (interval > HZ*NR_CPUS/10)
3162 interval = HZ*NR_CPUS/10;
3165 if (sd->flags & SD_SERIALIZE) {
3166 if (!spin_trylock(&balancing))
3170 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3171 if (load_balance(cpu, rq, sd, idle, &balance)) {
3173 * We've pulled tasks over so either we're no
3174 * longer idle, or one of our SMT siblings is
3177 idle = CPU_NOT_IDLE;
3179 sd->last_balance = jiffies;
3181 if (sd->flags & SD_SERIALIZE)
3182 spin_unlock(&balancing);
3184 if (time_after(next_balance, sd->last_balance + interval)) {
3185 next_balance = sd->last_balance + interval;
3186 update_next_balance = 1;
3190 * Stop the load balance at this level. There is another
3191 * CPU in our sched group which is doing load balancing more
3199 * next_balance will be updated only when there is a need.
3200 * When the cpu is attached to null domain for ex, it will not be
3203 if (likely(update_next_balance))
3204 rq->next_balance = next_balance;
3208 * run_rebalance_domains is triggered when needed from the scheduler tick.
3209 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3210 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3212 static void run_rebalance_domains(struct softirq_action *h)
3214 int this_cpu = smp_processor_id();
3215 struct rq *this_rq = cpu_rq(this_cpu);
3216 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3217 CPU_IDLE : CPU_NOT_IDLE;
3219 rebalance_domains(this_cpu, idle);
3223 * If this cpu is the owner for idle load balancing, then do the
3224 * balancing on behalf of the other idle cpus whose ticks are
3227 if (this_rq->idle_at_tick &&
3228 atomic_read(&nohz.load_balancer) == this_cpu) {
3229 cpumask_t cpus = nohz.cpu_mask;
3233 cpu_clear(this_cpu, cpus);
3234 for_each_cpu_mask(balance_cpu, cpus) {
3236 * If this cpu gets work to do, stop the load balancing
3237 * work being done for other cpus. Next load
3238 * balancing owner will pick it up.
3243 rebalance_domains(balance_cpu, CPU_IDLE);
3245 rq = cpu_rq(balance_cpu);
3246 if (time_after(this_rq->next_balance, rq->next_balance))
3247 this_rq->next_balance = rq->next_balance;
3254 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3256 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3257 * idle load balancing owner or decide to stop the periodic load balancing,
3258 * if the whole system is idle.
3260 static inline void trigger_load_balance(struct rq *rq, int cpu)
3264 * If we were in the nohz mode recently and busy at the current
3265 * scheduler tick, then check if we need to nominate new idle
3268 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3269 rq->in_nohz_recently = 0;
3271 if (atomic_read(&nohz.load_balancer) == cpu) {
3272 cpu_clear(cpu, nohz.cpu_mask);
3273 atomic_set(&nohz.load_balancer, -1);
3276 if (atomic_read(&nohz.load_balancer) == -1) {
3278 * simple selection for now: Nominate the
3279 * first cpu in the nohz list to be the next
3282 * TBD: Traverse the sched domains and nominate
3283 * the nearest cpu in the nohz.cpu_mask.
3285 int ilb = first_cpu(nohz.cpu_mask);
3293 * If this cpu is idle and doing idle load balancing for all the
3294 * cpus with ticks stopped, is it time for that to stop?
3296 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3297 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3303 * If this cpu is idle and the idle load balancing is done by
3304 * someone else, then no need raise the SCHED_SOFTIRQ
3306 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3307 cpu_isset(cpu, nohz.cpu_mask))
3310 if (time_after_eq(jiffies, rq->next_balance))
3311 raise_softirq(SCHED_SOFTIRQ);
3314 #else /* CONFIG_SMP */
3317 * on UP we do not need to balance between CPUs:
3319 static inline void idle_balance(int cpu, struct rq *rq)
3325 DEFINE_PER_CPU(struct kernel_stat, kstat);
3327 EXPORT_PER_CPU_SYMBOL(kstat);
3330 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3331 * that have not yet been banked in case the task is currently running.
3333 unsigned long long task_sched_runtime(struct task_struct *p)
3335 unsigned long flags;
3339 rq = task_rq_lock(p, &flags);
3340 ns = p->se.sum_exec_runtime;
3341 if (task_current(rq, p)) {
3342 update_rq_clock(rq);
3343 delta_exec = rq->clock - p->se.exec_start;
3344 if ((s64)delta_exec > 0)
3347 task_rq_unlock(rq, &flags);
3353 * Account user cpu time to a process.
3354 * @p: the process that the cpu time gets accounted to
3355 * @cputime: the cpu time spent in user space since the last update
3357 void account_user_time(struct task_struct *p, cputime_t cputime)
3359 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3362 p->utime = cputime_add(p->utime, cputime);
3364 /* Add user time to cpustat. */
3365 tmp = cputime_to_cputime64(cputime);
3366 if (TASK_NICE(p) > 0)
3367 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3369 cpustat->user = cputime64_add(cpustat->user, tmp);
3373 * Account guest cpu time to a process.
3374 * @p: the process that the cpu time gets accounted to
3375 * @cputime: the cpu time spent in virtual machine since the last update
3377 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3380 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3382 tmp = cputime_to_cputime64(cputime);
3384 p->utime = cputime_add(p->utime, cputime);
3385 p->gtime = cputime_add(p->gtime, cputime);
3387 cpustat->user = cputime64_add(cpustat->user, tmp);
3388 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3392 * Account scaled user cpu time to a process.
3393 * @p: the process that the cpu time gets accounted to
3394 * @cputime: the cpu time spent in user space since the last update
3396 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3398 p->utimescaled = cputime_add(p->utimescaled, cputime);
3402 * Account 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(struct task_struct *p, int hardirq_offset,
3410 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3411 struct rq *rq = this_rq();
3414 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3415 return account_guest_time(p, cputime);
3417 p->stime = cputime_add(p->stime, cputime);
3419 /* Add system time to cpustat. */
3420 tmp = cputime_to_cputime64(cputime);
3421 if (hardirq_count() - hardirq_offset)
3422 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3423 else if (softirq_count())
3424 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3425 else if (p != rq->idle)
3426 cpustat->system = cputime64_add(cpustat->system, tmp);
3427 else if (atomic_read(&rq->nr_iowait) > 0)
3428 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3430 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3431 /* Account for system time used */
3432 acct_update_integrals(p);
3436 * Account scaled system cpu time to a process.
3437 * @p: the process that the cpu time gets accounted to
3438 * @hardirq_offset: the offset to subtract from hardirq_count()
3439 * @cputime: the cpu time spent in kernel space since the last update
3441 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3443 p->stimescaled = cputime_add(p->stimescaled, cputime);
3447 * Account for involuntary wait time.
3448 * @p: the process from which the cpu time has been stolen
3449 * @steal: the cpu time spent in involuntary wait
3451 void account_steal_time(struct task_struct *p, cputime_t steal)
3453 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3454 cputime64_t tmp = cputime_to_cputime64(steal);
3455 struct rq *rq = this_rq();
3457 if (p == rq->idle) {
3458 p->stime = cputime_add(p->stime, steal);
3459 if (atomic_read(&rq->nr_iowait) > 0)
3460 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3462 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3464 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3468 * This function gets called by the timer code, with HZ frequency.
3469 * We call it with interrupts disabled.
3471 * It also gets called by the fork code, when changing the parent's
3474 void scheduler_tick(void)
3476 int cpu = smp_processor_id();
3477 struct rq *rq = cpu_rq(cpu);
3478 struct task_struct *curr = rq->curr;
3479 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3481 spin_lock(&rq->lock);
3482 __update_rq_clock(rq);
3484 * Let rq->clock advance by at least TICK_NSEC:
3486 if (unlikely(rq->clock < next_tick))
3487 rq->clock = next_tick;
3488 rq->tick_timestamp = rq->clock;
3489 update_cpu_load(rq);
3490 if (curr != rq->idle) /* FIXME: needed? */
3491 curr->sched_class->task_tick(rq, curr);
3492 spin_unlock(&rq->lock);
3495 rq->idle_at_tick = idle_cpu(cpu);
3496 trigger_load_balance(rq, cpu);
3500 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3502 void fastcall add_preempt_count(int val)
3507 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3509 preempt_count() += val;
3511 * Spinlock count overflowing soon?
3513 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3516 EXPORT_SYMBOL(add_preempt_count);
3518 void fastcall sub_preempt_count(int val)
3523 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3526 * Is the spinlock portion underflowing?
3528 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3529 !(preempt_count() & PREEMPT_MASK)))
3532 preempt_count() -= val;
3534 EXPORT_SYMBOL(sub_preempt_count);
3539 * Print scheduling while atomic bug:
3541 static noinline void __schedule_bug(struct task_struct *prev)
3543 struct pt_regs *regs = get_irq_regs();
3545 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3546 prev->comm, prev->pid, preempt_count());
3548 debug_show_held_locks(prev);
3549 if (irqs_disabled())
3550 print_irqtrace_events(prev);
3559 * Various schedule()-time debugging checks and statistics:
3561 static inline void schedule_debug(struct task_struct *prev)
3564 * Test if we are atomic. Since do_exit() needs to call into
3565 * schedule() atomically, we ignore that path for now.
3566 * Otherwise, whine if we are scheduling when we should not be.
3568 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3569 __schedule_bug(prev);
3571 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3573 schedstat_inc(this_rq(), sched_count);
3574 #ifdef CONFIG_SCHEDSTATS
3575 if (unlikely(prev->lock_depth >= 0)) {
3576 schedstat_inc(this_rq(), bkl_count);
3577 schedstat_inc(prev, sched_info.bkl_count);
3583 * Pick up the highest-prio task:
3585 static inline struct task_struct *
3586 pick_next_task(struct rq *rq, struct task_struct *prev)
3588 const struct sched_class *class;
3589 struct task_struct *p;
3592 * Optimization: we know that if all tasks are in
3593 * the fair class we can call that function directly:
3595 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3596 p = fair_sched_class.pick_next_task(rq);
3601 class = sched_class_highest;
3603 p = class->pick_next_task(rq);
3607 * Will never be NULL as the idle class always
3608 * returns a non-NULL p:
3610 class = class->next;
3615 * schedule() is the main scheduler function.
3617 asmlinkage void __sched schedule(void)
3619 struct task_struct *prev, *next;
3626 cpu = smp_processor_id();
3630 switch_count = &prev->nivcsw;
3632 release_kernel_lock(prev);
3633 need_resched_nonpreemptible:
3635 schedule_debug(prev);
3638 * Do the rq-clock update outside the rq lock:
3640 local_irq_disable();
3641 __update_rq_clock(rq);
3642 spin_lock(&rq->lock);
3643 clear_tsk_need_resched(prev);
3645 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3646 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3647 unlikely(signal_pending(prev)))) {
3648 prev->state = TASK_RUNNING;
3650 deactivate_task(rq, prev, 1);
3652 switch_count = &prev->nvcsw;
3655 if (unlikely(!rq->nr_running))
3656 idle_balance(cpu, rq);
3658 prev->sched_class->put_prev_task(rq, prev);
3659 next = pick_next_task(rq, prev);
3661 sched_info_switch(prev, next);
3663 if (likely(prev != next)) {
3668 context_switch(rq, prev, next); /* unlocks the rq */
3670 spin_unlock_irq(&rq->lock);
3672 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3673 cpu = smp_processor_id();
3675 goto need_resched_nonpreemptible;
3677 preempt_enable_no_resched();
3678 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3681 EXPORT_SYMBOL(schedule);
3683 #ifdef CONFIG_PREEMPT
3685 * this is the entry point to schedule() from in-kernel preemption
3686 * off of preempt_enable. Kernel preemptions off return from interrupt
3687 * occur there and call schedule directly.
3689 asmlinkage void __sched preempt_schedule(void)
3691 struct thread_info *ti = current_thread_info();
3692 #ifdef CONFIG_PREEMPT_BKL
3693 struct task_struct *task = current;
3694 int saved_lock_depth;
3697 * If there is a non-zero preempt_count or interrupts are disabled,
3698 * we do not want to preempt the current task. Just return..
3700 if (likely(ti->preempt_count || irqs_disabled()))
3704 add_preempt_count(PREEMPT_ACTIVE);
3707 * We keep the big kernel semaphore locked, but we
3708 * clear ->lock_depth so that schedule() doesnt
3709 * auto-release the semaphore:
3711 #ifdef CONFIG_PREEMPT_BKL
3712 saved_lock_depth = task->lock_depth;
3713 task->lock_depth = -1;
3716 #ifdef CONFIG_PREEMPT_BKL
3717 task->lock_depth = saved_lock_depth;
3719 sub_preempt_count(PREEMPT_ACTIVE);
3722 * Check again in case we missed a preemption opportunity
3723 * between schedule and now.
3726 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3728 EXPORT_SYMBOL(preempt_schedule);
3731 * this is the entry point to schedule() from kernel preemption
3732 * off of irq context.
3733 * Note, that this is called and return with irqs disabled. This will
3734 * protect us against recursive calling from irq.
3736 asmlinkage void __sched preempt_schedule_irq(void)
3738 struct thread_info *ti = current_thread_info();
3739 #ifdef CONFIG_PREEMPT_BKL
3740 struct task_struct *task = current;
3741 int saved_lock_depth;
3743 /* Catch callers which need to be fixed */
3744 BUG_ON(ti->preempt_count || !irqs_disabled());
3747 add_preempt_count(PREEMPT_ACTIVE);
3750 * We keep the big kernel semaphore locked, but we
3751 * clear ->lock_depth so that schedule() doesnt
3752 * auto-release the semaphore:
3754 #ifdef CONFIG_PREEMPT_BKL
3755 saved_lock_depth = task->lock_depth;
3756 task->lock_depth = -1;
3760 local_irq_disable();
3761 #ifdef CONFIG_PREEMPT_BKL
3762 task->lock_depth = saved_lock_depth;
3764 sub_preempt_count(PREEMPT_ACTIVE);
3767 * Check again in case we missed a preemption opportunity
3768 * between schedule and now.
3771 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3774 #endif /* CONFIG_PREEMPT */
3776 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3779 return try_to_wake_up(curr->private, mode, sync);
3781 EXPORT_SYMBOL(default_wake_function);
3784 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3785 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3786 * number) then we wake all the non-exclusive tasks and one exclusive task.
3788 * There are circumstances in which we can try to wake a task which has already
3789 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3790 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3792 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3793 int nr_exclusive, int sync, void *key)
3795 wait_queue_t *curr, *next;
3797 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3798 unsigned flags = curr->flags;
3800 if (curr->func(curr, mode, sync, key) &&
3801 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3807 * __wake_up - wake up threads blocked on a waitqueue.
3809 * @mode: which threads
3810 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3811 * @key: is directly passed to the wakeup function
3813 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3814 int nr_exclusive, void *key)
3816 unsigned long flags;
3818 spin_lock_irqsave(&q->lock, flags);
3819 __wake_up_common(q, mode, nr_exclusive, 0, key);
3820 spin_unlock_irqrestore(&q->lock, flags);
3822 EXPORT_SYMBOL(__wake_up);
3825 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3827 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3829 __wake_up_common(q, mode, 1, 0, NULL);
3833 * __wake_up_sync - wake up threads blocked on a waitqueue.
3835 * @mode: which threads
3836 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3838 * The sync wakeup differs that the waker knows that it will schedule
3839 * away soon, so while the target thread will be woken up, it will not
3840 * be migrated to another CPU - ie. the two threads are 'synchronized'
3841 * with each other. This can prevent needless bouncing between CPUs.
3843 * On UP it can prevent extra preemption.
3846 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3848 unsigned long flags;
3854 if (unlikely(!nr_exclusive))
3857 spin_lock_irqsave(&q->lock, flags);
3858 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3859 spin_unlock_irqrestore(&q->lock, flags);
3861 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3863 void complete(struct completion *x)
3865 unsigned long flags;
3867 spin_lock_irqsave(&x->wait.lock, flags);
3869 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3871 spin_unlock_irqrestore(&x->wait.lock, flags);
3873 EXPORT_SYMBOL(complete);
3875 void complete_all(struct completion *x)
3877 unsigned long flags;
3879 spin_lock_irqsave(&x->wait.lock, flags);
3880 x->done += UINT_MAX/2;
3881 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3883 spin_unlock_irqrestore(&x->wait.lock, flags);
3885 EXPORT_SYMBOL(complete_all);
3887 static inline long __sched
3888 do_wait_for_common(struct completion *x, long timeout, int state)
3891 DECLARE_WAITQUEUE(wait, current);
3893 wait.flags |= WQ_FLAG_EXCLUSIVE;
3894 __add_wait_queue_tail(&x->wait, &wait);
3896 if (state == TASK_INTERRUPTIBLE &&
3897 signal_pending(current)) {
3898 __remove_wait_queue(&x->wait, &wait);
3899 return -ERESTARTSYS;
3901 __set_current_state(state);
3902 spin_unlock_irq(&x->wait.lock);
3903 timeout = schedule_timeout(timeout);
3904 spin_lock_irq(&x->wait.lock);
3906 __remove_wait_queue(&x->wait, &wait);
3910 __remove_wait_queue(&x->wait, &wait);
3917 wait_for_common(struct completion *x, long timeout, int state)
3921 spin_lock_irq(&x->wait.lock);
3922 timeout = do_wait_for_common(x, timeout, state);
3923 spin_unlock_irq(&x->wait.lock);
3927 void __sched wait_for_completion(struct completion *x)
3929 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3931 EXPORT_SYMBOL(wait_for_completion);
3933 unsigned long __sched
3934 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3936 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3938 EXPORT_SYMBOL(wait_for_completion_timeout);
3940 int __sched wait_for_completion_interruptible(struct completion *x)
3942 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3943 if (t == -ERESTARTSYS)
3947 EXPORT_SYMBOL(wait_for_completion_interruptible);
3949 unsigned long __sched
3950 wait_for_completion_interruptible_timeout(struct completion *x,
3951 unsigned long timeout)
3953 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3955 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3958 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3960 unsigned long flags;
3963 init_waitqueue_entry(&wait, current);
3965 __set_current_state(state);
3967 spin_lock_irqsave(&q->lock, flags);
3968 __add_wait_queue(q, &wait);
3969 spin_unlock(&q->lock);
3970 timeout = schedule_timeout(timeout);
3971 spin_lock_irq(&q->lock);
3972 __remove_wait_queue(q, &wait);
3973 spin_unlock_irqrestore(&q->lock, flags);
3978 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3980 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3982 EXPORT_SYMBOL(interruptible_sleep_on);
3985 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3987 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3989 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3991 void __sched sleep_on(wait_queue_head_t *q)
3993 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3995 EXPORT_SYMBOL(sleep_on);
3997 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3999 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4001 EXPORT_SYMBOL(sleep_on_timeout);
4003 #ifdef CONFIG_RT_MUTEXES
4006 * rt_mutex_setprio - set the current priority of a task
4008 * @prio: prio value (kernel-internal form)
4010 * This function changes the 'effective' priority of a task. It does
4011 * not touch ->normal_prio like __setscheduler().
4013 * Used by the rt_mutex code to implement priority inheritance logic.
4015 void rt_mutex_setprio(struct task_struct *p, int prio)
4017 unsigned long flags;
4018 int oldprio, on_rq, running;
4021 BUG_ON(prio < 0 || prio > MAX_PRIO);
4023 rq = task_rq_lock(p, &flags);
4024 update_rq_clock(rq);
4027 on_rq = p->se.on_rq;
4028 running = task_current(rq, p);
4030 dequeue_task(rq, p, 0);
4032 p->sched_class->put_prev_task(rq, p);
4036 p->sched_class = &rt_sched_class;
4038 p->sched_class = &fair_sched_class;
4044 p->sched_class->set_curr_task(rq);
4045 enqueue_task(rq, p, 0);
4047 * Reschedule if we are currently running on this runqueue and
4048 * our priority decreased, or if we are not currently running on
4049 * this runqueue and our priority is higher than the current's
4052 if (p->prio > oldprio)
4053 resched_task(rq->curr);
4055 check_preempt_curr(rq, p);
4058 task_rq_unlock(rq, &flags);
4063 void set_user_nice(struct task_struct *p, long nice)
4065 int old_prio, delta, on_rq;
4066 unsigned long flags;
4069 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4072 * We have to be careful, if called from sys_setpriority(),
4073 * the task might be in the middle of scheduling on another CPU.
4075 rq = task_rq_lock(p, &flags);
4076 update_rq_clock(rq);
4078 * The RT priorities are set via sched_setscheduler(), but we still
4079 * allow the 'normal' nice value to be set - but as expected
4080 * it wont have any effect on scheduling until the task is
4081 * SCHED_FIFO/SCHED_RR:
4083 if (task_has_rt_policy(p)) {
4084 p->static_prio = NICE_TO_PRIO(nice);
4087 on_rq = p->se.on_rq;
4089 dequeue_task(rq, p, 0);
4091 p->static_prio = NICE_TO_PRIO(nice);
4094 p->prio = effective_prio(p);
4095 delta = p->prio - old_prio;
4098 enqueue_task(rq, p, 0);
4100 * If the task increased its priority or is running and
4101 * lowered its priority, then reschedule its CPU:
4103 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4104 resched_task(rq->curr);
4107 task_rq_unlock(rq, &flags);
4109 EXPORT_SYMBOL(set_user_nice);
4112 * can_nice - check if a task can reduce its nice value
4116 int can_nice(const struct task_struct *p, const int nice)
4118 /* convert nice value [19,-20] to rlimit style value [1,40] */
4119 int nice_rlim = 20 - nice;
4121 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4122 capable(CAP_SYS_NICE));
4125 #ifdef __ARCH_WANT_SYS_NICE
4128 * sys_nice - change the priority of the current process.
4129 * @increment: priority increment
4131 * sys_setpriority is a more generic, but much slower function that
4132 * does similar things.
4134 asmlinkage long sys_nice(int increment)
4139 * Setpriority might change our priority at the same moment.
4140 * We don't have to worry. Conceptually one call occurs first
4141 * and we have a single winner.
4143 if (increment < -40)
4148 nice = PRIO_TO_NICE(current->static_prio) + increment;
4154 if (increment < 0 && !can_nice(current, nice))
4157 retval = security_task_setnice(current, nice);
4161 set_user_nice(current, nice);
4168 * task_prio - return the priority value of a given task.
4169 * @p: the task in question.
4171 * This is the priority value as seen by users in /proc.
4172 * RT tasks are offset by -200. Normal tasks are centered
4173 * around 0, value goes from -16 to +15.
4175 int task_prio(const struct task_struct *p)
4177 return p->prio - MAX_RT_PRIO;
4181 * task_nice - return the nice value of a given task.
4182 * @p: the task in question.
4184 int task_nice(const struct task_struct *p)
4186 return TASK_NICE(p);
4188 EXPORT_SYMBOL_GPL(task_nice);
4191 * idle_cpu - is a given cpu idle currently?
4192 * @cpu: the processor in question.
4194 int idle_cpu(int cpu)
4196 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4200 * idle_task - return the idle task for a given cpu.
4201 * @cpu: the processor in question.
4203 struct task_struct *idle_task(int cpu)
4205 return cpu_rq(cpu)->idle;
4209 * find_process_by_pid - find a process with a matching PID value.
4210 * @pid: the pid in question.
4212 static struct task_struct *find_process_by_pid(pid_t pid)
4214 return pid ? find_task_by_vpid(pid) : current;
4217 /* Actually do priority change: must hold rq lock. */
4219 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4221 BUG_ON(p->se.on_rq);
4224 switch (p->policy) {
4228 p->sched_class = &fair_sched_class;
4232 p->sched_class = &rt_sched_class;
4236 p->rt_priority = prio;
4237 p->normal_prio = normal_prio(p);
4238 /* we are holding p->pi_lock already */
4239 p->prio = rt_mutex_getprio(p);
4244 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4245 * @p: the task in question.
4246 * @policy: new policy.
4247 * @param: structure containing the new RT priority.
4249 * NOTE that the task may be already dead.
4251 int sched_setscheduler(struct task_struct *p, int policy,
4252 struct sched_param *param)
4254 int retval, oldprio, oldpolicy = -1, on_rq, running;
4255 unsigned long flags;
4258 /* may grab non-irq protected spin_locks */
4259 BUG_ON(in_interrupt());
4261 /* double check policy once rq lock held */
4263 policy = oldpolicy = p->policy;
4264 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4265 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4266 policy != SCHED_IDLE)
4269 * Valid priorities for SCHED_FIFO and SCHED_RR are
4270 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4271 * SCHED_BATCH and SCHED_IDLE is 0.
4273 if (param->sched_priority < 0 ||
4274 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4275 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4277 if (rt_policy(policy) != (param->sched_priority != 0))
4281 * Allow unprivileged RT tasks to decrease priority:
4283 if (!capable(CAP_SYS_NICE)) {
4284 if (rt_policy(policy)) {
4285 unsigned long rlim_rtprio;
4287 if (!lock_task_sighand(p, &flags))
4289 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4290 unlock_task_sighand(p, &flags);
4292 /* can't set/change the rt policy */
4293 if (policy != p->policy && !rlim_rtprio)
4296 /* can't increase priority */
4297 if (param->sched_priority > p->rt_priority &&
4298 param->sched_priority > rlim_rtprio)
4302 * Like positive nice levels, dont allow tasks to
4303 * move out of SCHED_IDLE either:
4305 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4308 /* can't change other user's priorities */
4309 if ((current->euid != p->euid) &&
4310 (current->euid != p->uid))
4314 retval = security_task_setscheduler(p, policy, param);
4318 * make sure no PI-waiters arrive (or leave) while we are
4319 * changing the priority of the task:
4321 spin_lock_irqsave(&p->pi_lock, flags);
4323 * To be able to change p->policy safely, the apropriate
4324 * runqueue lock must be held.
4326 rq = __task_rq_lock(p);
4327 /* recheck policy now with rq lock held */
4328 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4329 policy = oldpolicy = -1;
4330 __task_rq_unlock(rq);
4331 spin_unlock_irqrestore(&p->pi_lock, flags);
4334 update_rq_clock(rq);
4335 on_rq = p->se.on_rq;
4336 running = task_current(rq, p);
4338 deactivate_task(rq, p, 0);
4340 p->sched_class->put_prev_task(rq, p);
4344 __setscheduler(rq, p, policy, param->sched_priority);
4348 p->sched_class->set_curr_task(rq);
4349 activate_task(rq, p, 0);
4351 * Reschedule if we are currently running on this runqueue and
4352 * our priority decreased, or if we are not currently running on
4353 * this runqueue and our priority is higher than the current's
4356 if (p->prio > oldprio)
4357 resched_task(rq->curr);
4359 check_preempt_curr(rq, p);
4362 __task_rq_unlock(rq);
4363 spin_unlock_irqrestore(&p->pi_lock, flags);
4365 rt_mutex_adjust_pi(p);
4369 EXPORT_SYMBOL_GPL(sched_setscheduler);
4372 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4374 struct sched_param lparam;
4375 struct task_struct *p;
4378 if (!param || pid < 0)
4380 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4385 p = find_process_by_pid(pid);
4387 retval = sched_setscheduler(p, policy, &lparam);
4394 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4395 * @pid: the pid in question.
4396 * @policy: new policy.
4397 * @param: structure containing the new RT priority.
4400 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4402 /* negative values for policy are not valid */
4406 return do_sched_setscheduler(pid, policy, param);
4410 * sys_sched_setparam - set/change the RT priority of a thread
4411 * @pid: the pid in question.
4412 * @param: structure containing the new RT priority.
4414 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4416 return do_sched_setscheduler(pid, -1, param);
4420 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4421 * @pid: the pid in question.
4423 asmlinkage long sys_sched_getscheduler(pid_t pid)
4425 struct task_struct *p;
4432 read_lock(&tasklist_lock);
4433 p = find_process_by_pid(pid);
4435 retval = security_task_getscheduler(p);
4439 read_unlock(&tasklist_lock);
4444 * sys_sched_getscheduler - get the RT priority of a thread
4445 * @pid: the pid in question.
4446 * @param: structure containing the RT priority.
4448 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4450 struct sched_param lp;
4451 struct task_struct *p;
4454 if (!param || pid < 0)
4457 read_lock(&tasklist_lock);
4458 p = find_process_by_pid(pid);
4463 retval = security_task_getscheduler(p);
4467 lp.sched_priority = p->rt_priority;
4468 read_unlock(&tasklist_lock);
4471 * This one might sleep, we cannot do it with a spinlock held ...
4473 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4478 read_unlock(&tasklist_lock);
4482 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4484 cpumask_t cpus_allowed;
4485 struct task_struct *p;
4488 mutex_lock(&sched_hotcpu_mutex);
4489 read_lock(&tasklist_lock);
4491 p = find_process_by_pid(pid);
4493 read_unlock(&tasklist_lock);
4494 mutex_unlock(&sched_hotcpu_mutex);
4499 * It is not safe to call set_cpus_allowed with the
4500 * tasklist_lock held. We will bump the task_struct's
4501 * usage count and then drop tasklist_lock.
4504 read_unlock(&tasklist_lock);
4507 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4508 !capable(CAP_SYS_NICE))
4511 retval = security_task_setscheduler(p, 0, NULL);
4515 cpus_allowed = cpuset_cpus_allowed(p);
4516 cpus_and(new_mask, new_mask, cpus_allowed);
4518 retval = set_cpus_allowed(p, new_mask);
4521 cpus_allowed = cpuset_cpus_allowed(p);
4522 if (!cpus_subset(new_mask, cpus_allowed)) {
4524 * We must have raced with a concurrent cpuset
4525 * update. Just reset the cpus_allowed to the
4526 * cpuset's cpus_allowed
4528 new_mask = cpus_allowed;
4534 mutex_unlock(&sched_hotcpu_mutex);
4538 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4539 cpumask_t *new_mask)
4541 if (len < sizeof(cpumask_t)) {
4542 memset(new_mask, 0, sizeof(cpumask_t));
4543 } else if (len > sizeof(cpumask_t)) {
4544 len = sizeof(cpumask_t);
4546 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4550 * sys_sched_setaffinity - set the cpu affinity of a process
4551 * @pid: pid of the process
4552 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4553 * @user_mask_ptr: user-space pointer to the new cpu mask
4555 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4556 unsigned long __user *user_mask_ptr)
4561 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4565 return sched_setaffinity(pid, new_mask);
4569 * Represents all cpu's present in the system
4570 * In systems capable of hotplug, this map could dynamically grow
4571 * as new cpu's are detected in the system via any platform specific
4572 * method, such as ACPI for e.g.
4575 cpumask_t cpu_present_map __read_mostly;
4576 EXPORT_SYMBOL(cpu_present_map);
4579 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4580 EXPORT_SYMBOL(cpu_online_map);
4582 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4583 EXPORT_SYMBOL(cpu_possible_map);
4586 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4588 struct task_struct *p;
4591 mutex_lock(&sched_hotcpu_mutex);
4592 read_lock(&tasklist_lock);
4595 p = find_process_by_pid(pid);
4599 retval = security_task_getscheduler(p);
4603 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4606 read_unlock(&tasklist_lock);
4607 mutex_unlock(&sched_hotcpu_mutex);
4613 * sys_sched_getaffinity - get the cpu affinity of a process
4614 * @pid: pid of the process
4615 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4616 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4618 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4619 unsigned long __user *user_mask_ptr)
4624 if (len < sizeof(cpumask_t))
4627 ret = sched_getaffinity(pid, &mask);
4631 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4634 return sizeof(cpumask_t);
4638 * sys_sched_yield - yield the current processor to other threads.
4640 * This function yields the current CPU to other tasks. If there are no
4641 * other threads running on this CPU then this function will return.
4643 asmlinkage long sys_sched_yield(void)
4645 struct rq *rq = this_rq_lock();
4647 schedstat_inc(rq, yld_count);
4648 current->sched_class->yield_task(rq);
4651 * Since we are going to call schedule() anyway, there's
4652 * no need to preempt or enable interrupts:
4654 __release(rq->lock);
4655 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4656 _raw_spin_unlock(&rq->lock);
4657 preempt_enable_no_resched();
4664 static void __cond_resched(void)
4666 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4667 __might_sleep(__FILE__, __LINE__);
4670 * The BKS might be reacquired before we have dropped
4671 * PREEMPT_ACTIVE, which could trigger a second
4672 * cond_resched() call.
4675 add_preempt_count(PREEMPT_ACTIVE);
4677 sub_preempt_count(PREEMPT_ACTIVE);
4678 } while (need_resched());
4681 int __sched cond_resched(void)
4683 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4684 system_state == SYSTEM_RUNNING) {
4690 EXPORT_SYMBOL(cond_resched);
4693 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4694 * call schedule, and on return reacquire the lock.
4696 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4697 * operations here to prevent schedule() from being called twice (once via
4698 * spin_unlock(), once by hand).
4700 int cond_resched_lock(spinlock_t *lock)
4704 if (need_lockbreak(lock)) {
4710 if (need_resched() && system_state == SYSTEM_RUNNING) {
4711 spin_release(&lock->dep_map, 1, _THIS_IP_);
4712 _raw_spin_unlock(lock);
4713 preempt_enable_no_resched();
4720 EXPORT_SYMBOL(cond_resched_lock);
4722 int __sched cond_resched_softirq(void)
4724 BUG_ON(!in_softirq());
4726 if (need_resched() && system_state == SYSTEM_RUNNING) {
4734 EXPORT_SYMBOL(cond_resched_softirq);
4737 * yield - yield the current processor to other threads.
4739 * This is a shortcut for kernel-space yielding - it marks the
4740 * thread runnable and calls sys_sched_yield().
4742 void __sched yield(void)
4744 set_current_state(TASK_RUNNING);
4747 EXPORT_SYMBOL(yield);
4750 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4751 * that process accounting knows that this is a task in IO wait state.
4753 * But don't do that if it is a deliberate, throttling IO wait (this task
4754 * has set its backing_dev_info: the queue against which it should throttle)
4756 void __sched io_schedule(void)
4758 struct rq *rq = &__raw_get_cpu_var(runqueues);
4760 delayacct_blkio_start();
4761 atomic_inc(&rq->nr_iowait);
4763 atomic_dec(&rq->nr_iowait);
4764 delayacct_blkio_end();
4766 EXPORT_SYMBOL(io_schedule);
4768 long __sched io_schedule_timeout(long timeout)
4770 struct rq *rq = &__raw_get_cpu_var(runqueues);
4773 delayacct_blkio_start();
4774 atomic_inc(&rq->nr_iowait);
4775 ret = schedule_timeout(timeout);
4776 atomic_dec(&rq->nr_iowait);
4777 delayacct_blkio_end();
4782 * sys_sched_get_priority_max - return maximum RT priority.
4783 * @policy: scheduling class.
4785 * this syscall returns the maximum rt_priority that can be used
4786 * by a given scheduling class.
4788 asmlinkage long sys_sched_get_priority_max(int policy)
4795 ret = MAX_USER_RT_PRIO-1;
4807 * sys_sched_get_priority_min - return minimum RT priority.
4808 * @policy: scheduling class.
4810 * this syscall returns the minimum rt_priority that can be used
4811 * by a given scheduling class.
4813 asmlinkage long sys_sched_get_priority_min(int policy)
4831 * sys_sched_rr_get_interval - return the default timeslice of a process.
4832 * @pid: pid of the process.
4833 * @interval: userspace pointer to the timeslice value.
4835 * this syscall writes the default timeslice value of a given process
4836 * into the user-space timespec buffer. A value of '0' means infinity.
4839 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4841 struct task_struct *p;
4842 unsigned int time_slice;
4850 read_lock(&tasklist_lock);
4851 p = find_process_by_pid(pid);
4855 retval = security_task_getscheduler(p);
4860 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4861 * tasks that are on an otherwise idle runqueue:
4864 if (p->policy == SCHED_RR) {
4865 time_slice = DEF_TIMESLICE;
4867 struct sched_entity *se = &p->se;
4868 unsigned long flags;
4871 rq = task_rq_lock(p, &flags);
4872 if (rq->cfs.load.weight)
4873 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4874 task_rq_unlock(rq, &flags);
4876 read_unlock(&tasklist_lock);
4877 jiffies_to_timespec(time_slice, &t);
4878 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4882 read_unlock(&tasklist_lock);
4886 static const char stat_nam[] = "RSDTtZX";
4888 static void show_task(struct task_struct *p)
4890 unsigned long free = 0;
4893 state = p->state ? __ffs(p->state) + 1 : 0;
4894 printk(KERN_INFO "%-13.13s %c", p->comm,
4895 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4896 #if BITS_PER_LONG == 32
4897 if (state == TASK_RUNNING)
4898 printk(KERN_CONT " running ");
4900 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4902 if (state == TASK_RUNNING)
4903 printk(KERN_CONT " running task ");
4905 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4907 #ifdef CONFIG_DEBUG_STACK_USAGE
4909 unsigned long *n = end_of_stack(p);
4912 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4915 printk(KERN_CONT "%5lu %5d %6d\n", free,
4916 task_pid_nr(p), task_pid_nr(p->real_parent));
4918 if (state != TASK_RUNNING)
4919 show_stack(p, NULL);
4922 void show_state_filter(unsigned long state_filter)
4924 struct task_struct *g, *p;
4926 #if BITS_PER_LONG == 32
4928 " task PC stack pid father\n");
4931 " task PC stack pid father\n");
4933 read_lock(&tasklist_lock);
4934 do_each_thread(g, p) {
4936 * reset the NMI-timeout, listing all files on a slow
4937 * console might take alot of time:
4939 touch_nmi_watchdog();
4940 if (!state_filter || (p->state & state_filter))
4942 } while_each_thread(g, p);
4944 touch_all_softlockup_watchdogs();
4946 #ifdef CONFIG_SCHED_DEBUG
4947 sysrq_sched_debug_show();
4949 read_unlock(&tasklist_lock);
4951 * Only show locks if all tasks are dumped:
4953 if (state_filter == -1)
4954 debug_show_all_locks();
4957 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4959 idle->sched_class = &idle_sched_class;
4963 * init_idle - set up an idle thread for a given CPU
4964 * @idle: task in question
4965 * @cpu: cpu the idle task belongs to
4967 * NOTE: this function does not set the idle thread's NEED_RESCHED
4968 * flag, to make booting more robust.
4970 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4972 struct rq *rq = cpu_rq(cpu);
4973 unsigned long flags;
4976 idle->se.exec_start = sched_clock();
4978 idle->prio = idle->normal_prio = MAX_PRIO;
4979 idle->cpus_allowed = cpumask_of_cpu(cpu);
4980 __set_task_cpu(idle, cpu);
4982 spin_lock_irqsave(&rq->lock, flags);
4983 rq->curr = rq->idle = idle;
4984 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4987 spin_unlock_irqrestore(&rq->lock, flags);
4989 /* Set the preempt count _outside_ the spinlocks! */
4990 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4991 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4993 task_thread_info(idle)->preempt_count = 0;
4996 * The idle tasks have their own, simple scheduling class:
4998 idle->sched_class = &idle_sched_class;
5002 * In a system that switches off the HZ timer nohz_cpu_mask
5003 * indicates which cpus entered this state. This is used
5004 * in the rcu update to wait only for active cpus. For system
5005 * which do not switch off the HZ timer nohz_cpu_mask should
5006 * always be CPU_MASK_NONE.
5008 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5011 * Increase the granularity value when there are more CPUs,
5012 * because with more CPUs the 'effective latency' as visible
5013 * to users decreases. But the relationship is not linear,
5014 * so pick a second-best guess by going with the log2 of the
5017 * This idea comes from the SD scheduler of Con Kolivas:
5019 static inline void sched_init_granularity(void)
5021 unsigned int factor = 1 + ilog2(num_online_cpus());
5022 const unsigned long limit = 200000000;
5024 sysctl_sched_min_granularity *= factor;
5025 if (sysctl_sched_min_granularity > limit)
5026 sysctl_sched_min_granularity = limit;
5028 sysctl_sched_latency *= factor;
5029 if (sysctl_sched_latency > limit)
5030 sysctl_sched_latency = limit;
5032 sysctl_sched_wakeup_granularity *= factor;
5033 sysctl_sched_batch_wakeup_granularity *= factor;
5038 * This is how migration works:
5040 * 1) we queue a struct migration_req structure in the source CPU's
5041 * runqueue and wake up that CPU's migration thread.
5042 * 2) we down() the locked semaphore => thread blocks.
5043 * 3) migration thread wakes up (implicitly it forces the migrated
5044 * thread off the CPU)
5045 * 4) it gets the migration request and checks whether the migrated
5046 * task is still in the wrong runqueue.
5047 * 5) if it's in the wrong runqueue then the migration thread removes
5048 * it and puts it into the right queue.
5049 * 6) migration thread up()s the semaphore.
5050 * 7) we wake up and the migration is done.
5054 * Change a given task's CPU affinity. Migrate the thread to a
5055 * proper CPU and schedule it away if the CPU it's executing on
5056 * is removed from the allowed bitmask.
5058 * NOTE: the caller must have a valid reference to the task, the
5059 * task must not exit() & deallocate itself prematurely. The
5060 * call is not atomic; no spinlocks may be held.
5062 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5064 struct migration_req req;
5065 unsigned long flags;
5069 rq = task_rq_lock(p, &flags);
5070 if (!cpus_intersects(new_mask, cpu_online_map)) {
5075 p->cpus_allowed = new_mask;
5076 /* Can the task run on the task's current CPU? If so, we're done */
5077 if (cpu_isset(task_cpu(p), new_mask))
5080 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5081 /* Need help from migration thread: drop lock and wait. */
5082 task_rq_unlock(rq, &flags);
5083 wake_up_process(rq->migration_thread);
5084 wait_for_completion(&req.done);
5085 tlb_migrate_finish(p->mm);
5089 task_rq_unlock(rq, &flags);
5093 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5096 * Move (not current) task off this cpu, onto dest cpu. We're doing
5097 * this because either it can't run here any more (set_cpus_allowed()
5098 * away from this CPU, or CPU going down), or because we're
5099 * attempting to rebalance this task on exec (sched_exec).
5101 * So we race with normal scheduler movements, but that's OK, as long
5102 * as the task is no longer on this CPU.
5104 * Returns non-zero if task was successfully migrated.
5106 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5108 struct rq *rq_dest, *rq_src;
5111 if (unlikely(cpu_is_offline(dest_cpu)))
5114 rq_src = cpu_rq(src_cpu);
5115 rq_dest = cpu_rq(dest_cpu);
5117 double_rq_lock(rq_src, rq_dest);
5118 /* Already moved. */
5119 if (task_cpu(p) != src_cpu)
5121 /* Affinity changed (again). */
5122 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5125 on_rq = p->se.on_rq;
5127 deactivate_task(rq_src, p, 0);
5129 set_task_cpu(p, dest_cpu);
5131 activate_task(rq_dest, p, 0);
5132 check_preempt_curr(rq_dest, p);
5136 double_rq_unlock(rq_src, rq_dest);
5141 * migration_thread - this is a highprio system thread that performs
5142 * thread migration by bumping thread off CPU then 'pushing' onto
5145 static int migration_thread(void *data)
5147 int cpu = (long)data;
5151 BUG_ON(rq->migration_thread != current);
5153 set_current_state(TASK_INTERRUPTIBLE);
5154 while (!kthread_should_stop()) {
5155 struct migration_req *req;
5156 struct list_head *head;
5158 spin_lock_irq(&rq->lock);
5160 if (cpu_is_offline(cpu)) {
5161 spin_unlock_irq(&rq->lock);
5165 if (rq->active_balance) {
5166 active_load_balance(rq, cpu);
5167 rq->active_balance = 0;
5170 head = &rq->migration_queue;
5172 if (list_empty(head)) {
5173 spin_unlock_irq(&rq->lock);
5175 set_current_state(TASK_INTERRUPTIBLE);
5178 req = list_entry(head->next, struct migration_req, list);
5179 list_del_init(head->next);
5181 spin_unlock(&rq->lock);
5182 __migrate_task(req->task, cpu, req->dest_cpu);
5185 complete(&req->done);
5187 __set_current_state(TASK_RUNNING);
5191 /* Wait for kthread_stop */
5192 set_current_state(TASK_INTERRUPTIBLE);
5193 while (!kthread_should_stop()) {
5195 set_current_state(TASK_INTERRUPTIBLE);
5197 __set_current_state(TASK_RUNNING);
5201 #ifdef CONFIG_HOTPLUG_CPU
5203 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5207 local_irq_disable();
5208 ret = __migrate_task(p, src_cpu, dest_cpu);
5214 * Figure out where task on dead CPU should go, use force if necessary.
5215 * NOTE: interrupts should be disabled by the caller
5217 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5219 unsigned long flags;
5226 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5227 cpus_and(mask, mask, p->cpus_allowed);
5228 dest_cpu = any_online_cpu(mask);
5230 /* On any allowed CPU? */
5231 if (dest_cpu == NR_CPUS)
5232 dest_cpu = any_online_cpu(p->cpus_allowed);
5234 /* No more Mr. Nice Guy. */
5235 if (dest_cpu == NR_CPUS) {
5236 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5238 * Try to stay on the same cpuset, where the
5239 * current cpuset may be a subset of all cpus.
5240 * The cpuset_cpus_allowed_locked() variant of
5241 * cpuset_cpus_allowed() will not block. It must be
5242 * called within calls to cpuset_lock/cpuset_unlock.
5244 rq = task_rq_lock(p, &flags);
5245 p->cpus_allowed = cpus_allowed;
5246 dest_cpu = any_online_cpu(p->cpus_allowed);
5247 task_rq_unlock(rq, &flags);
5250 * Don't tell them about moving exiting tasks or
5251 * kernel threads (both mm NULL), since they never
5254 if (p->mm && printk_ratelimit()) {
5255 printk(KERN_INFO "process %d (%s) no "
5256 "longer affine to cpu%d\n",
5257 task_pid_nr(p), p->comm, dead_cpu);
5260 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5264 * While a dead CPU has no uninterruptible tasks queued at this point,
5265 * it might still have a nonzero ->nr_uninterruptible counter, because
5266 * for performance reasons the counter is not stricly tracking tasks to
5267 * their home CPUs. So we just add the counter to another CPU's counter,
5268 * to keep the global sum constant after CPU-down:
5270 static void migrate_nr_uninterruptible(struct rq *rq_src)
5272 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5273 unsigned long flags;
5275 local_irq_save(flags);
5276 double_rq_lock(rq_src, rq_dest);
5277 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5278 rq_src->nr_uninterruptible = 0;
5279 double_rq_unlock(rq_src, rq_dest);
5280 local_irq_restore(flags);
5283 /* Run through task list and migrate tasks from the dead cpu. */
5284 static void migrate_live_tasks(int src_cpu)
5286 struct task_struct *p, *t;
5288 read_lock(&tasklist_lock);
5290 do_each_thread(t, p) {
5294 if (task_cpu(p) == src_cpu)
5295 move_task_off_dead_cpu(src_cpu, p);
5296 } while_each_thread(t, p);
5298 read_unlock(&tasklist_lock);
5302 * Schedules idle task to be the next runnable task on current CPU.
5303 * It does so by boosting its priority to highest possible.
5304 * Used by CPU offline code.
5306 void sched_idle_next(void)
5308 int this_cpu = smp_processor_id();
5309 struct rq *rq = cpu_rq(this_cpu);
5310 struct task_struct *p = rq->idle;
5311 unsigned long flags;
5313 /* cpu has to be offline */
5314 BUG_ON(cpu_online(this_cpu));
5317 * Strictly not necessary since rest of the CPUs are stopped by now
5318 * and interrupts disabled on the current cpu.
5320 spin_lock_irqsave(&rq->lock, flags);
5322 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5324 update_rq_clock(rq);
5325 activate_task(rq, p, 0);
5327 spin_unlock_irqrestore(&rq->lock, flags);
5331 * Ensures that the idle task is using init_mm right before its cpu goes
5334 void idle_task_exit(void)
5336 struct mm_struct *mm = current->active_mm;
5338 BUG_ON(cpu_online(smp_processor_id()));
5341 switch_mm(mm, &init_mm, current);
5345 /* called under rq->lock with disabled interrupts */
5346 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5348 struct rq *rq = cpu_rq(dead_cpu);
5350 /* Must be exiting, otherwise would be on tasklist. */
5351 BUG_ON(!p->exit_state);
5353 /* Cannot have done final schedule yet: would have vanished. */
5354 BUG_ON(p->state == TASK_DEAD);
5359 * Drop lock around migration; if someone else moves it,
5360 * that's OK. No task can be added to this CPU, so iteration is
5363 spin_unlock_irq(&rq->lock);
5364 move_task_off_dead_cpu(dead_cpu, p);
5365 spin_lock_irq(&rq->lock);
5370 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5371 static void migrate_dead_tasks(unsigned int dead_cpu)
5373 struct rq *rq = cpu_rq(dead_cpu);
5374 struct task_struct *next;
5377 if (!rq->nr_running)
5379 update_rq_clock(rq);
5380 next = pick_next_task(rq, rq->curr);
5383 migrate_dead(dead_cpu, next);
5387 #endif /* CONFIG_HOTPLUG_CPU */
5389 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5391 static struct ctl_table sd_ctl_dir[] = {
5393 .procname = "sched_domain",
5399 static struct ctl_table sd_ctl_root[] = {
5401 .ctl_name = CTL_KERN,
5402 .procname = "kernel",
5404 .child = sd_ctl_dir,
5409 static struct ctl_table *sd_alloc_ctl_entry(int n)
5411 struct ctl_table *entry =
5412 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5417 static void sd_free_ctl_entry(struct ctl_table **tablep)
5419 struct ctl_table *entry;
5422 * In the intermediate directories, both the child directory and
5423 * procname are dynamically allocated and could fail but the mode
5424 * will always be set. In the lowest directory the names are
5425 * static strings and all have proc handlers.
5427 for (entry = *tablep; entry->mode; entry++) {
5429 sd_free_ctl_entry(&entry->child);
5430 if (entry->proc_handler == NULL)
5431 kfree(entry->procname);
5439 set_table_entry(struct ctl_table *entry,
5440 const char *procname, void *data, int maxlen,
5441 mode_t mode, proc_handler *proc_handler)
5443 entry->procname = procname;
5445 entry->maxlen = maxlen;
5447 entry->proc_handler = proc_handler;
5450 static struct ctl_table *
5451 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5453 struct ctl_table *table = sd_alloc_ctl_entry(12);
5458 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5459 sizeof(long), 0644, proc_doulongvec_minmax);
5460 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5461 sizeof(long), 0644, proc_doulongvec_minmax);
5462 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5463 sizeof(int), 0644, proc_dointvec_minmax);
5464 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5465 sizeof(int), 0644, proc_dointvec_minmax);
5466 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5467 sizeof(int), 0644, proc_dointvec_minmax);
5468 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5471 sizeof(int), 0644, proc_dointvec_minmax);
5472 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5473 sizeof(int), 0644, proc_dointvec_minmax);
5474 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5475 sizeof(int), 0644, proc_dointvec_minmax);
5476 set_table_entry(&table[9], "cache_nice_tries",
5477 &sd->cache_nice_tries,
5478 sizeof(int), 0644, proc_dointvec_minmax);
5479 set_table_entry(&table[10], "flags", &sd->flags,
5480 sizeof(int), 0644, proc_dointvec_minmax);
5481 /* &table[11] is terminator */
5486 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5488 struct ctl_table *entry, *table;
5489 struct sched_domain *sd;
5490 int domain_num = 0, i;
5493 for_each_domain(cpu, sd)
5495 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5500 for_each_domain(cpu, sd) {
5501 snprintf(buf, 32, "domain%d", i);
5502 entry->procname = kstrdup(buf, GFP_KERNEL);
5504 entry->child = sd_alloc_ctl_domain_table(sd);
5511 static struct ctl_table_header *sd_sysctl_header;
5512 static void register_sched_domain_sysctl(void)
5514 int i, cpu_num = num_online_cpus();
5515 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5518 WARN_ON(sd_ctl_dir[0].child);
5519 sd_ctl_dir[0].child = entry;
5524 for_each_online_cpu(i) {
5525 snprintf(buf, 32, "cpu%d", i);
5526 entry->procname = kstrdup(buf, GFP_KERNEL);
5528 entry->child = sd_alloc_ctl_cpu_table(i);
5532 WARN_ON(sd_sysctl_header);
5533 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5536 /* may be called multiple times per register */
5537 static void unregister_sched_domain_sysctl(void)
5539 if (sd_sysctl_header)
5540 unregister_sysctl_table(sd_sysctl_header);
5541 sd_sysctl_header = NULL;
5542 if (sd_ctl_dir[0].child)
5543 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5546 static void register_sched_domain_sysctl(void)
5549 static void unregister_sched_domain_sysctl(void)
5555 * migration_call - callback that gets triggered when a CPU is added.
5556 * Here we can start up the necessary migration thread for the new CPU.
5558 static int __cpuinit
5559 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5561 struct task_struct *p;
5562 int cpu = (long)hcpu;
5563 unsigned long flags;
5567 case CPU_LOCK_ACQUIRE:
5568 mutex_lock(&sched_hotcpu_mutex);
5571 case CPU_UP_PREPARE:
5572 case CPU_UP_PREPARE_FROZEN:
5573 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5576 kthread_bind(p, cpu);
5577 /* Must be high prio: stop_machine expects to yield to it. */
5578 rq = task_rq_lock(p, &flags);
5579 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5580 task_rq_unlock(rq, &flags);
5581 cpu_rq(cpu)->migration_thread = p;
5585 case CPU_ONLINE_FROZEN:
5586 /* Strictly unnecessary, as first user will wake it. */
5587 wake_up_process(cpu_rq(cpu)->migration_thread);
5590 #ifdef CONFIG_HOTPLUG_CPU
5591 case CPU_UP_CANCELED:
5592 case CPU_UP_CANCELED_FROZEN:
5593 if (!cpu_rq(cpu)->migration_thread)
5595 /* Unbind it from offline cpu so it can run. Fall thru. */
5596 kthread_bind(cpu_rq(cpu)->migration_thread,
5597 any_online_cpu(cpu_online_map));
5598 kthread_stop(cpu_rq(cpu)->migration_thread);
5599 cpu_rq(cpu)->migration_thread = NULL;
5603 case CPU_DEAD_FROZEN:
5604 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5605 migrate_live_tasks(cpu);
5607 kthread_stop(rq->migration_thread);
5608 rq->migration_thread = NULL;
5609 /* Idle task back to normal (off runqueue, low prio) */
5610 spin_lock_irq(&rq->lock);
5611 update_rq_clock(rq);
5612 deactivate_task(rq, rq->idle, 0);
5613 rq->idle->static_prio = MAX_PRIO;
5614 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5615 rq->idle->sched_class = &idle_sched_class;
5616 migrate_dead_tasks(cpu);
5617 spin_unlock_irq(&rq->lock);
5619 migrate_nr_uninterruptible(rq);
5620 BUG_ON(rq->nr_running != 0);
5623 * No need to migrate the tasks: it was best-effort if
5624 * they didn't take sched_hotcpu_mutex. Just wake up
5627 spin_lock_irq(&rq->lock);
5628 while (!list_empty(&rq->migration_queue)) {
5629 struct migration_req *req;
5631 req = list_entry(rq->migration_queue.next,
5632 struct migration_req, list);
5633 list_del_init(&req->list);
5634 complete(&req->done);
5636 spin_unlock_irq(&rq->lock);
5639 case CPU_LOCK_RELEASE:
5640 mutex_unlock(&sched_hotcpu_mutex);
5646 /* Register at highest priority so that task migration (migrate_all_tasks)
5647 * happens before everything else.
5649 static struct notifier_block __cpuinitdata migration_notifier = {
5650 .notifier_call = migration_call,
5654 void __init migration_init(void)
5656 void *cpu = (void *)(long)smp_processor_id();
5659 /* Start one for the boot CPU: */
5660 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5661 BUG_ON(err == NOTIFY_BAD);
5662 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5663 register_cpu_notifier(&migration_notifier);
5669 /* Number of possible processor ids */
5670 int nr_cpu_ids __read_mostly = NR_CPUS;
5671 EXPORT_SYMBOL(nr_cpu_ids);
5673 #ifdef CONFIG_SCHED_DEBUG
5675 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5677 struct sched_group *group = sd->groups;
5678 cpumask_t groupmask;
5681 cpumask_scnprintf(str, NR_CPUS, sd->span);
5682 cpus_clear(groupmask);
5684 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5686 if (!(sd->flags & SD_LOAD_BALANCE)) {
5687 printk("does not load-balance\n");
5689 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5694 printk(KERN_CONT "span %s\n", str);
5696 if (!cpu_isset(cpu, sd->span)) {
5697 printk(KERN_ERR "ERROR: domain->span does not contain "
5700 if (!cpu_isset(cpu, group->cpumask)) {
5701 printk(KERN_ERR "ERROR: domain->groups does not contain"
5705 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5709 printk(KERN_ERR "ERROR: group is NULL\n");
5713 if (!group->__cpu_power) {
5714 printk(KERN_CONT "\n");
5715 printk(KERN_ERR "ERROR: domain->cpu_power not "
5720 if (!cpus_weight(group->cpumask)) {
5721 printk(KERN_CONT "\n");
5722 printk(KERN_ERR "ERROR: empty group\n");
5726 if (cpus_intersects(groupmask, group->cpumask)) {
5727 printk(KERN_CONT "\n");
5728 printk(KERN_ERR "ERROR: repeated CPUs\n");
5732 cpus_or(groupmask, groupmask, group->cpumask);
5734 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5735 printk(KERN_CONT " %s", str);
5737 group = group->next;
5738 } while (group != sd->groups);
5739 printk(KERN_CONT "\n");
5741 if (!cpus_equal(sd->span, groupmask))
5742 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5744 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5745 printk(KERN_ERR "ERROR: parent span is not a superset "
5746 "of domain->span\n");
5750 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5755 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5759 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5762 if (sched_domain_debug_one(sd, cpu, level))
5771 # define sched_domain_debug(sd, cpu) do { } while (0)
5774 static int sd_degenerate(struct sched_domain *sd)
5776 if (cpus_weight(sd->span) == 1)
5779 /* Following flags need at least 2 groups */
5780 if (sd->flags & (SD_LOAD_BALANCE |
5781 SD_BALANCE_NEWIDLE |
5785 SD_SHARE_PKG_RESOURCES)) {
5786 if (sd->groups != sd->groups->next)
5790 /* Following flags don't use groups */
5791 if (sd->flags & (SD_WAKE_IDLE |
5800 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5802 unsigned long cflags = sd->flags, pflags = parent->flags;
5804 if (sd_degenerate(parent))
5807 if (!cpus_equal(sd->span, parent->span))
5810 /* Does parent contain flags not in child? */
5811 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5812 if (cflags & SD_WAKE_AFFINE)
5813 pflags &= ~SD_WAKE_BALANCE;
5814 /* Flags needing groups don't count if only 1 group in parent */
5815 if (parent->groups == parent->groups->next) {
5816 pflags &= ~(SD_LOAD_BALANCE |
5817 SD_BALANCE_NEWIDLE |
5821 SD_SHARE_PKG_RESOURCES);
5823 if (~cflags & pflags)
5830 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5831 * hold the hotplug lock.
5833 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5835 struct rq *rq = cpu_rq(cpu);
5836 struct sched_domain *tmp;
5838 /* Remove the sched domains which do not contribute to scheduling. */
5839 for (tmp = sd; tmp; tmp = tmp->parent) {
5840 struct sched_domain *parent = tmp->parent;
5843 if (sd_parent_degenerate(tmp, parent)) {
5844 tmp->parent = parent->parent;
5846 parent->parent->child = tmp;
5850 if (sd && sd_degenerate(sd)) {
5856 sched_domain_debug(sd, cpu);
5858 rcu_assign_pointer(rq->sd, sd);
5861 /* cpus with isolated domains */
5862 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5864 /* Setup the mask of cpus configured for isolated domains */
5865 static int __init isolated_cpu_setup(char *str)
5867 int ints[NR_CPUS], i;
5869 str = get_options(str, ARRAY_SIZE(ints), ints);
5870 cpus_clear(cpu_isolated_map);
5871 for (i = 1; i <= ints[0]; i++)
5872 if (ints[i] < NR_CPUS)
5873 cpu_set(ints[i], cpu_isolated_map);
5877 __setup("isolcpus=", isolated_cpu_setup);
5880 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5881 * to a function which identifies what group(along with sched group) a CPU
5882 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5883 * (due to the fact that we keep track of groups covered with a cpumask_t).
5885 * init_sched_build_groups will build a circular linked list of the groups
5886 * covered by the given span, and will set each group's ->cpumask correctly,
5887 * and ->cpu_power to 0.
5890 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5891 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5892 struct sched_group **sg))
5894 struct sched_group *first = NULL, *last = NULL;
5895 cpumask_t covered = CPU_MASK_NONE;
5898 for_each_cpu_mask(i, span) {
5899 struct sched_group *sg;
5900 int group = group_fn(i, cpu_map, &sg);
5903 if (cpu_isset(i, covered))
5906 sg->cpumask = CPU_MASK_NONE;
5907 sg->__cpu_power = 0;
5909 for_each_cpu_mask(j, span) {
5910 if (group_fn(j, cpu_map, NULL) != group)
5913 cpu_set(j, covered);
5914 cpu_set(j, sg->cpumask);
5925 #define SD_NODES_PER_DOMAIN 16
5930 * find_next_best_node - find the next node to include in a sched_domain
5931 * @node: node whose sched_domain we're building
5932 * @used_nodes: nodes already in the sched_domain
5934 * Find the next node to include in a given scheduling domain. Simply
5935 * finds the closest node not already in the @used_nodes map.
5937 * Should use nodemask_t.
5939 static int find_next_best_node(int node, unsigned long *used_nodes)
5941 int i, n, val, min_val, best_node = 0;
5945 for (i = 0; i < MAX_NUMNODES; i++) {
5946 /* Start at @node */
5947 n = (node + i) % MAX_NUMNODES;
5949 if (!nr_cpus_node(n))
5952 /* Skip already used nodes */
5953 if (test_bit(n, used_nodes))
5956 /* Simple min distance search */
5957 val = node_distance(node, n);
5959 if (val < min_val) {
5965 set_bit(best_node, used_nodes);
5970 * sched_domain_node_span - get a cpumask for a node's sched_domain
5971 * @node: node whose cpumask we're constructing
5972 * @size: number of nodes to include in this span
5974 * Given a node, construct a good cpumask for its sched_domain to span. It
5975 * should be one that prevents unnecessary balancing, but also spreads tasks
5978 static cpumask_t sched_domain_node_span(int node)
5980 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5981 cpumask_t span, nodemask;
5985 bitmap_zero(used_nodes, MAX_NUMNODES);
5987 nodemask = node_to_cpumask(node);
5988 cpus_or(span, span, nodemask);
5989 set_bit(node, used_nodes);
5991 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5992 int next_node = find_next_best_node(node, used_nodes);
5994 nodemask = node_to_cpumask(next_node);
5995 cpus_or(span, span, nodemask);
6002 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6005 * SMT sched-domains:
6007 #ifdef CONFIG_SCHED_SMT
6008 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6009 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6012 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6015 *sg = &per_cpu(sched_group_cpus, cpu);
6021 * multi-core sched-domains:
6023 #ifdef CONFIG_SCHED_MC
6024 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6025 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6028 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6030 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6033 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6034 cpus_and(mask, mask, *cpu_map);
6035 group = first_cpu(mask);
6037 *sg = &per_cpu(sched_group_core, group);
6040 #elif defined(CONFIG_SCHED_MC)
6042 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6045 *sg = &per_cpu(sched_group_core, cpu);
6050 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6051 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6054 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6057 #ifdef CONFIG_SCHED_MC
6058 cpumask_t mask = cpu_coregroup_map(cpu);
6059 cpus_and(mask, mask, *cpu_map);
6060 group = first_cpu(mask);
6061 #elif defined(CONFIG_SCHED_SMT)
6062 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6063 cpus_and(mask, mask, *cpu_map);
6064 group = first_cpu(mask);
6069 *sg = &per_cpu(sched_group_phys, group);
6075 * The init_sched_build_groups can't handle what we want to do with node
6076 * groups, so roll our own. Now each node has its own list of groups which
6077 * gets dynamically allocated.
6079 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6080 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6082 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6083 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6085 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6086 struct sched_group **sg)
6088 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6091 cpus_and(nodemask, nodemask, *cpu_map);
6092 group = first_cpu(nodemask);
6095 *sg = &per_cpu(sched_group_allnodes, group);
6099 static void init_numa_sched_groups_power(struct sched_group *group_head)
6101 struct sched_group *sg = group_head;
6107 for_each_cpu_mask(j, sg->cpumask) {
6108 struct sched_domain *sd;
6110 sd = &per_cpu(phys_domains, j);
6111 if (j != first_cpu(sd->groups->cpumask)) {
6113 * Only add "power" once for each
6119 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6122 } while (sg != group_head);
6127 /* Free memory allocated for various sched_group structures */
6128 static void free_sched_groups(const cpumask_t *cpu_map)
6132 for_each_cpu_mask(cpu, *cpu_map) {
6133 struct sched_group **sched_group_nodes
6134 = sched_group_nodes_bycpu[cpu];
6136 if (!sched_group_nodes)
6139 for (i = 0; i < MAX_NUMNODES; i++) {
6140 cpumask_t nodemask = node_to_cpumask(i);
6141 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6143 cpus_and(nodemask, nodemask, *cpu_map);
6144 if (cpus_empty(nodemask))
6154 if (oldsg != sched_group_nodes[i])
6157 kfree(sched_group_nodes);
6158 sched_group_nodes_bycpu[cpu] = NULL;
6162 static void free_sched_groups(const cpumask_t *cpu_map)
6168 * Initialize sched groups cpu_power.
6170 * cpu_power indicates the capacity of sched group, which is used while
6171 * distributing the load between different sched groups in a sched domain.
6172 * Typically cpu_power for all the groups in a sched domain will be same unless
6173 * there are asymmetries in the topology. If there are asymmetries, group
6174 * having more cpu_power will pickup more load compared to the group having
6177 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6178 * the maximum number of tasks a group can handle in the presence of other idle
6179 * or lightly loaded groups in the same sched domain.
6181 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6183 struct sched_domain *child;
6184 struct sched_group *group;
6186 WARN_ON(!sd || !sd->groups);
6188 if (cpu != first_cpu(sd->groups->cpumask))
6193 sd->groups->__cpu_power = 0;
6196 * For perf policy, if the groups in child domain share resources
6197 * (for example cores sharing some portions of the cache hierarchy
6198 * or SMT), then set this domain groups cpu_power such that each group
6199 * can handle only one task, when there are other idle groups in the
6200 * same sched domain.
6202 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6204 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6205 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6210 * add cpu_power of each child group to this groups cpu_power
6212 group = child->groups;
6214 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6215 group = group->next;
6216 } while (group != child->groups);
6220 * Build sched domains for a given set of cpus and attach the sched domains
6221 * to the individual cpus
6223 static int build_sched_domains(const cpumask_t *cpu_map)
6227 struct sched_group **sched_group_nodes = NULL;
6228 int sd_allnodes = 0;
6231 * Allocate the per-node list of sched groups
6233 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6235 if (!sched_group_nodes) {
6236 printk(KERN_WARNING "Can not alloc sched group node list\n");
6239 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6243 * Set up domains for cpus specified by the cpu_map.
6245 for_each_cpu_mask(i, *cpu_map) {
6246 struct sched_domain *sd = NULL, *p;
6247 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6249 cpus_and(nodemask, nodemask, *cpu_map);
6252 if (cpus_weight(*cpu_map) >
6253 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6254 sd = &per_cpu(allnodes_domains, i);
6255 *sd = SD_ALLNODES_INIT;
6256 sd->span = *cpu_map;
6257 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6263 sd = &per_cpu(node_domains, i);
6265 sd->span = sched_domain_node_span(cpu_to_node(i));
6269 cpus_and(sd->span, sd->span, *cpu_map);
6273 sd = &per_cpu(phys_domains, i);
6275 sd->span = nodemask;
6279 cpu_to_phys_group(i, cpu_map, &sd->groups);
6281 #ifdef CONFIG_SCHED_MC
6283 sd = &per_cpu(core_domains, i);
6285 sd->span = cpu_coregroup_map(i);
6286 cpus_and(sd->span, sd->span, *cpu_map);
6289 cpu_to_core_group(i, cpu_map, &sd->groups);
6292 #ifdef CONFIG_SCHED_SMT
6294 sd = &per_cpu(cpu_domains, i);
6295 *sd = SD_SIBLING_INIT;
6296 sd->span = per_cpu(cpu_sibling_map, i);
6297 cpus_and(sd->span, sd->span, *cpu_map);
6300 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6304 #ifdef CONFIG_SCHED_SMT
6305 /* Set up CPU (sibling) groups */
6306 for_each_cpu_mask(i, *cpu_map) {
6307 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6308 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6309 if (i != first_cpu(this_sibling_map))
6312 init_sched_build_groups(this_sibling_map, cpu_map,
6317 #ifdef CONFIG_SCHED_MC
6318 /* Set up multi-core groups */
6319 for_each_cpu_mask(i, *cpu_map) {
6320 cpumask_t this_core_map = cpu_coregroup_map(i);
6321 cpus_and(this_core_map, this_core_map, *cpu_map);
6322 if (i != first_cpu(this_core_map))
6324 init_sched_build_groups(this_core_map, cpu_map,
6325 &cpu_to_core_group);
6329 /* Set up physical groups */
6330 for (i = 0; i < MAX_NUMNODES; i++) {
6331 cpumask_t nodemask = node_to_cpumask(i);
6333 cpus_and(nodemask, nodemask, *cpu_map);
6334 if (cpus_empty(nodemask))
6337 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6341 /* Set up node groups */
6343 init_sched_build_groups(*cpu_map, cpu_map,
6344 &cpu_to_allnodes_group);
6346 for (i = 0; i < MAX_NUMNODES; i++) {
6347 /* Set up node groups */
6348 struct sched_group *sg, *prev;
6349 cpumask_t nodemask = node_to_cpumask(i);
6350 cpumask_t domainspan;
6351 cpumask_t covered = CPU_MASK_NONE;
6354 cpus_and(nodemask, nodemask, *cpu_map);
6355 if (cpus_empty(nodemask)) {
6356 sched_group_nodes[i] = NULL;
6360 domainspan = sched_domain_node_span(i);
6361 cpus_and(domainspan, domainspan, *cpu_map);
6363 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6365 printk(KERN_WARNING "Can not alloc domain group for "
6369 sched_group_nodes[i] = sg;
6370 for_each_cpu_mask(j, nodemask) {
6371 struct sched_domain *sd;
6373 sd = &per_cpu(node_domains, j);
6376 sg->__cpu_power = 0;
6377 sg->cpumask = nodemask;
6379 cpus_or(covered, covered, nodemask);
6382 for (j = 0; j < MAX_NUMNODES; j++) {
6383 cpumask_t tmp, notcovered;
6384 int n = (i + j) % MAX_NUMNODES;
6386 cpus_complement(notcovered, covered);
6387 cpus_and(tmp, notcovered, *cpu_map);
6388 cpus_and(tmp, tmp, domainspan);
6389 if (cpus_empty(tmp))
6392 nodemask = node_to_cpumask(n);
6393 cpus_and(tmp, tmp, nodemask);
6394 if (cpus_empty(tmp))
6397 sg = kmalloc_node(sizeof(struct sched_group),
6401 "Can not alloc domain group for node %d\n", j);
6404 sg->__cpu_power = 0;
6406 sg->next = prev->next;
6407 cpus_or(covered, covered, tmp);
6414 /* Calculate CPU power for physical packages and nodes */
6415 #ifdef CONFIG_SCHED_SMT
6416 for_each_cpu_mask(i, *cpu_map) {
6417 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6419 init_sched_groups_power(i, sd);
6422 #ifdef CONFIG_SCHED_MC
6423 for_each_cpu_mask(i, *cpu_map) {
6424 struct sched_domain *sd = &per_cpu(core_domains, i);
6426 init_sched_groups_power(i, sd);
6430 for_each_cpu_mask(i, *cpu_map) {
6431 struct sched_domain *sd = &per_cpu(phys_domains, i);
6433 init_sched_groups_power(i, sd);
6437 for (i = 0; i < MAX_NUMNODES; i++)
6438 init_numa_sched_groups_power(sched_group_nodes[i]);
6441 struct sched_group *sg;
6443 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6444 init_numa_sched_groups_power(sg);
6448 /* Attach the domains */
6449 for_each_cpu_mask(i, *cpu_map) {
6450 struct sched_domain *sd;
6451 #ifdef CONFIG_SCHED_SMT
6452 sd = &per_cpu(cpu_domains, i);
6453 #elif defined(CONFIG_SCHED_MC)
6454 sd = &per_cpu(core_domains, i);
6456 sd = &per_cpu(phys_domains, i);
6458 cpu_attach_domain(sd, i);
6465 free_sched_groups(cpu_map);
6470 static cpumask_t *doms_cur; /* current sched domains */
6471 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6474 * Special case: If a kmalloc of a doms_cur partition (array of
6475 * cpumask_t) fails, then fallback to a single sched domain,
6476 * as determined by the single cpumask_t fallback_doms.
6478 static cpumask_t fallback_doms;
6481 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6482 * For now this just excludes isolated cpus, but could be used to
6483 * exclude other special cases in the future.
6485 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6490 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6492 doms_cur = &fallback_doms;
6493 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6494 err = build_sched_domains(doms_cur);
6495 register_sched_domain_sysctl();
6500 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6502 free_sched_groups(cpu_map);
6506 * Detach sched domains from a group of cpus specified in cpu_map
6507 * These cpus will now be attached to the NULL domain
6509 static void detach_destroy_domains(const cpumask_t *cpu_map)
6513 unregister_sched_domain_sysctl();
6515 for_each_cpu_mask(i, *cpu_map)
6516 cpu_attach_domain(NULL, i);
6517 synchronize_sched();
6518 arch_destroy_sched_domains(cpu_map);
6522 * Partition sched domains as specified by the 'ndoms_new'
6523 * cpumasks in the array doms_new[] of cpumasks. This compares
6524 * doms_new[] to the current sched domain partitioning, doms_cur[].
6525 * It destroys each deleted domain and builds each new domain.
6527 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6528 * The masks don't intersect (don't overlap.) We should setup one
6529 * sched domain for each mask. CPUs not in any of the cpumasks will
6530 * not be load balanced. If the same cpumask appears both in the
6531 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6534 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6535 * ownership of it and will kfree it when done with it. If the caller
6536 * failed the kmalloc call, then it can pass in doms_new == NULL,
6537 * and partition_sched_domains() will fallback to the single partition
6540 * Call with hotplug lock held
6542 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6546 /* always unregister in case we don't destroy any domains */
6547 unregister_sched_domain_sysctl();
6549 if (doms_new == NULL) {
6551 doms_new = &fallback_doms;
6552 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6555 /* Destroy deleted domains */
6556 for (i = 0; i < ndoms_cur; i++) {
6557 for (j = 0; j < ndoms_new; j++) {
6558 if (cpus_equal(doms_cur[i], doms_new[j]))
6561 /* no match - a current sched domain not in new doms_new[] */
6562 detach_destroy_domains(doms_cur + i);
6567 /* Build new domains */
6568 for (i = 0; i < ndoms_new; i++) {
6569 for (j = 0; j < ndoms_cur; j++) {
6570 if (cpus_equal(doms_new[i], doms_cur[j]))
6573 /* no match - add a new doms_new */
6574 build_sched_domains(doms_new + i);
6579 /* Remember the new sched domains */
6580 if (doms_cur != &fallback_doms)
6582 doms_cur = doms_new;
6583 ndoms_cur = ndoms_new;
6585 register_sched_domain_sysctl();
6588 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6589 static int arch_reinit_sched_domains(void)
6593 mutex_lock(&sched_hotcpu_mutex);
6594 detach_destroy_domains(&cpu_online_map);
6595 err = arch_init_sched_domains(&cpu_online_map);
6596 mutex_unlock(&sched_hotcpu_mutex);
6601 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6605 if (buf[0] != '0' && buf[0] != '1')
6609 sched_smt_power_savings = (buf[0] == '1');
6611 sched_mc_power_savings = (buf[0] == '1');
6613 ret = arch_reinit_sched_domains();
6615 return ret ? ret : count;
6618 #ifdef CONFIG_SCHED_MC
6619 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6621 return sprintf(page, "%u\n", sched_mc_power_savings);
6623 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6624 const char *buf, size_t count)
6626 return sched_power_savings_store(buf, count, 0);
6628 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6629 sched_mc_power_savings_store);
6632 #ifdef CONFIG_SCHED_SMT
6633 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6635 return sprintf(page, "%u\n", sched_smt_power_savings);
6637 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6638 const char *buf, size_t count)
6640 return sched_power_savings_store(buf, count, 1);
6642 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6643 sched_smt_power_savings_store);
6646 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6650 #ifdef CONFIG_SCHED_SMT
6652 err = sysfs_create_file(&cls->kset.kobj,
6653 &attr_sched_smt_power_savings.attr);
6655 #ifdef CONFIG_SCHED_MC
6656 if (!err && mc_capable())
6657 err = sysfs_create_file(&cls->kset.kobj,
6658 &attr_sched_mc_power_savings.attr);
6665 * Force a reinitialization of the sched domains hierarchy. The domains
6666 * and groups cannot be updated in place without racing with the balancing
6667 * code, so we temporarily attach all running cpus to the NULL domain
6668 * which will prevent rebalancing while the sched domains are recalculated.
6670 static int update_sched_domains(struct notifier_block *nfb,
6671 unsigned long action, void *hcpu)
6674 case CPU_UP_PREPARE:
6675 case CPU_UP_PREPARE_FROZEN:
6676 case CPU_DOWN_PREPARE:
6677 case CPU_DOWN_PREPARE_FROZEN:
6678 detach_destroy_domains(&cpu_online_map);
6681 case CPU_UP_CANCELED:
6682 case CPU_UP_CANCELED_FROZEN:
6683 case CPU_DOWN_FAILED:
6684 case CPU_DOWN_FAILED_FROZEN:
6686 case CPU_ONLINE_FROZEN:
6688 case CPU_DEAD_FROZEN:
6690 * Fall through and re-initialise the domains.
6697 /* The hotplug lock is already held by cpu_up/cpu_down */
6698 arch_init_sched_domains(&cpu_online_map);
6703 void __init sched_init_smp(void)
6705 cpumask_t non_isolated_cpus;
6707 mutex_lock(&sched_hotcpu_mutex);
6708 arch_init_sched_domains(&cpu_online_map);
6709 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6710 if (cpus_empty(non_isolated_cpus))
6711 cpu_set(smp_processor_id(), non_isolated_cpus);
6712 mutex_unlock(&sched_hotcpu_mutex);
6713 /* XXX: Theoretical race here - CPU may be hotplugged now */
6714 hotcpu_notifier(update_sched_domains, 0);
6716 /* Move init over to a non-isolated CPU */
6717 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6719 sched_init_granularity();
6722 void __init sched_init_smp(void)
6724 sched_init_granularity();
6726 #endif /* CONFIG_SMP */
6728 int in_sched_functions(unsigned long addr)
6730 return in_lock_functions(addr) ||
6731 (addr >= (unsigned long)__sched_text_start
6732 && addr < (unsigned long)__sched_text_end);
6735 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6737 cfs_rq->tasks_timeline = RB_ROOT;
6738 #ifdef CONFIG_FAIR_GROUP_SCHED
6741 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6744 void __init sched_init(void)
6746 int highest_cpu = 0;
6749 for_each_possible_cpu(i) {
6750 struct rt_prio_array *array;
6754 spin_lock_init(&rq->lock);
6755 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6758 init_cfs_rq(&rq->cfs, rq);
6759 #ifdef CONFIG_FAIR_GROUP_SCHED
6760 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6762 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6763 struct sched_entity *se =
6764 &per_cpu(init_sched_entity, i);
6766 init_cfs_rq_p[i] = cfs_rq;
6767 init_cfs_rq(cfs_rq, rq);
6768 cfs_rq->tg = &init_task_group;
6769 list_add(&cfs_rq->leaf_cfs_rq_list,
6770 &rq->leaf_cfs_rq_list);
6772 init_sched_entity_p[i] = se;
6773 se->cfs_rq = &rq->cfs;
6775 se->load.weight = init_task_group_load;
6776 se->load.inv_weight =
6777 div64_64(1ULL<<32, init_task_group_load);
6780 init_task_group.shares = init_task_group_load;
6783 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6784 rq->cpu_load[j] = 0;
6787 rq->active_balance = 0;
6788 rq->next_balance = jiffies;
6791 rq->migration_thread = NULL;
6792 INIT_LIST_HEAD(&rq->migration_queue);
6794 atomic_set(&rq->nr_iowait, 0);
6796 array = &rq->rt.active;
6797 for (j = 0; j < MAX_RT_PRIO; j++) {
6798 INIT_LIST_HEAD(array->queue + j);
6799 __clear_bit(j, array->bitmap);
6802 /* delimiter for bitsearch: */
6803 __set_bit(MAX_RT_PRIO, array->bitmap);
6806 set_load_weight(&init_task);
6808 #ifdef CONFIG_PREEMPT_NOTIFIERS
6809 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6813 nr_cpu_ids = highest_cpu + 1;
6814 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6817 #ifdef CONFIG_RT_MUTEXES
6818 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6822 * The boot idle thread does lazy MMU switching as well:
6824 atomic_inc(&init_mm.mm_count);
6825 enter_lazy_tlb(&init_mm, current);
6828 * Make us the idle thread. Technically, schedule() should not be
6829 * called from this thread, however somewhere below it might be,
6830 * but because we are the idle thread, we just pick up running again
6831 * when this runqueue becomes "idle".
6833 init_idle(current, smp_processor_id());
6835 * During early bootup we pretend to be a normal task:
6837 current->sched_class = &fair_sched_class;
6840 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6841 void __might_sleep(char *file, int line)
6844 static unsigned long prev_jiffy; /* ratelimiting */
6846 if ((in_atomic() || irqs_disabled()) &&
6847 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6848 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6850 prev_jiffy = jiffies;
6851 printk(KERN_ERR "BUG: sleeping function called from invalid"
6852 " context at %s:%d\n", file, line);
6853 printk("in_atomic():%d, irqs_disabled():%d\n",
6854 in_atomic(), irqs_disabled());
6855 debug_show_held_locks(current);
6856 if (irqs_disabled())
6857 print_irqtrace_events(current);
6862 EXPORT_SYMBOL(__might_sleep);
6865 #ifdef CONFIG_MAGIC_SYSRQ
6866 static void normalize_task(struct rq *rq, struct task_struct *p)
6869 update_rq_clock(rq);
6870 on_rq = p->se.on_rq;
6872 deactivate_task(rq, p, 0);
6873 __setscheduler(rq, p, SCHED_NORMAL, 0);
6875 activate_task(rq, p, 0);
6876 resched_task(rq->curr);
6880 void normalize_rt_tasks(void)
6882 struct task_struct *g, *p;
6883 unsigned long flags;
6886 read_lock_irq(&tasklist_lock);
6887 do_each_thread(g, p) {
6889 * Only normalize user tasks:
6894 p->se.exec_start = 0;
6895 #ifdef CONFIG_SCHEDSTATS
6896 p->se.wait_start = 0;
6897 p->se.sleep_start = 0;
6898 p->se.block_start = 0;
6900 task_rq(p)->clock = 0;
6904 * Renice negative nice level userspace
6907 if (TASK_NICE(p) < 0 && p->mm)
6908 set_user_nice(p, 0);
6912 spin_lock_irqsave(&p->pi_lock, flags);
6913 rq = __task_rq_lock(p);
6915 normalize_task(rq, p);
6917 __task_rq_unlock(rq);
6918 spin_unlock_irqrestore(&p->pi_lock, flags);
6919 } while_each_thread(g, p);
6921 read_unlock_irq(&tasklist_lock);
6924 #endif /* CONFIG_MAGIC_SYSRQ */
6928 * These functions are only useful for the IA64 MCA handling.
6930 * They can only be called when the whole system has been
6931 * stopped - every CPU needs to be quiescent, and no scheduling
6932 * activity can take place. Using them for anything else would
6933 * be a serious bug, and as a result, they aren't even visible
6934 * under any other configuration.
6938 * curr_task - return the current task for a given cpu.
6939 * @cpu: the processor in question.
6941 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6943 struct task_struct *curr_task(int cpu)
6945 return cpu_curr(cpu);
6949 * set_curr_task - set the current task for a given cpu.
6950 * @cpu: the processor in question.
6951 * @p: the task pointer to set.
6953 * Description: This function must only be used when non-maskable interrupts
6954 * are serviced on a separate stack. It allows the architecture to switch the
6955 * notion of the current task on a cpu in a non-blocking manner. This function
6956 * must be called with all CPU's synchronized, and interrupts disabled, the
6957 * and caller must save the original value of the current task (see
6958 * curr_task() above) and restore that value before reenabling interrupts and
6959 * re-starting the system.
6961 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6963 void set_curr_task(int cpu, struct task_struct *p)
6970 #ifdef CONFIG_FAIR_GROUP_SCHED
6972 /* allocate runqueue etc for a new task group */
6973 struct task_group *sched_create_group(void)
6975 struct task_group *tg;
6976 struct cfs_rq *cfs_rq;
6977 struct sched_entity *se;
6981 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6983 return ERR_PTR(-ENOMEM);
6985 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6988 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6992 for_each_possible_cpu(i) {
6995 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
7000 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
7005 memset(cfs_rq, 0, sizeof(struct cfs_rq));
7006 memset(se, 0, sizeof(struct sched_entity));
7008 tg->cfs_rq[i] = cfs_rq;
7009 init_cfs_rq(cfs_rq, rq);
7013 se->cfs_rq = &rq->cfs;
7015 se->load.weight = NICE_0_LOAD;
7016 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7020 tg->shares = NICE_0_LOAD;
7022 lock_task_group_list();
7023 for_each_possible_cpu(i) {
7025 cfs_rq = tg->cfs_rq[i];
7026 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7028 unlock_task_group_list();
7033 for_each_possible_cpu(i) {
7035 kfree(tg->cfs_rq[i]);
7043 return ERR_PTR(-ENOMEM);
7046 /* rcu callback to free various structures associated with a task group */
7047 static void free_sched_group(struct rcu_head *rhp)
7049 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7050 struct cfs_rq *cfs_rq;
7051 struct sched_entity *se;
7054 /* now it should be safe to free those cfs_rqs */
7055 for_each_possible_cpu(i) {
7056 cfs_rq = tg->cfs_rq[i];
7068 /* Destroy runqueue etc associated with a task group */
7069 void sched_destroy_group(struct task_group *tg)
7071 struct cfs_rq *cfs_rq = NULL;
7074 lock_task_group_list();
7075 for_each_possible_cpu(i) {
7076 cfs_rq = tg->cfs_rq[i];
7077 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7079 unlock_task_group_list();
7083 /* wait for possible concurrent references to cfs_rqs complete */
7084 call_rcu(&tg->rcu, free_sched_group);
7087 /* change task's runqueue when it moves between groups.
7088 * The caller of this function should have put the task in its new group
7089 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7090 * reflect its new group.
7092 void sched_move_task(struct task_struct *tsk)
7095 unsigned long flags;
7098 rq = task_rq_lock(tsk, &flags);
7100 if (tsk->sched_class != &fair_sched_class) {
7101 set_task_cfs_rq(tsk, task_cpu(tsk));
7105 update_rq_clock(rq);
7107 running = task_current(rq, tsk);
7108 on_rq = tsk->se.on_rq;
7111 dequeue_task(rq, tsk, 0);
7112 if (unlikely(running))
7113 tsk->sched_class->put_prev_task(rq, tsk);
7116 set_task_cfs_rq(tsk, task_cpu(tsk));
7119 if (unlikely(running))
7120 tsk->sched_class->set_curr_task(rq);
7121 enqueue_task(rq, tsk, 0);
7125 task_rq_unlock(rq, &flags);
7128 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7130 struct cfs_rq *cfs_rq = se->cfs_rq;
7131 struct rq *rq = cfs_rq->rq;
7134 spin_lock_irq(&rq->lock);
7138 dequeue_entity(cfs_rq, se, 0);
7140 se->load.weight = shares;
7141 se->load.inv_weight = div64_64((1ULL<<32), shares);
7144 enqueue_entity(cfs_rq, se, 0);
7146 spin_unlock_irq(&rq->lock);
7149 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7154 * A weight of 0 or 1 can cause arithmetics problems.
7155 * (The default weight is 1024 - so there's no practical
7156 * limitation from this.)
7161 lock_task_group_list();
7162 if (tg->shares == shares)
7165 tg->shares = shares;
7166 for_each_possible_cpu(i)
7167 set_se_shares(tg->se[i], shares);
7170 unlock_task_group_list();
7174 unsigned long sched_group_shares(struct task_group *tg)
7179 #endif /* CONFIG_FAIR_GROUP_SCHED */
7181 #ifdef CONFIG_FAIR_CGROUP_SCHED
7183 /* return corresponding task_group object of a cgroup */
7184 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7186 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7187 struct task_group, css);
7190 static struct cgroup_subsys_state *
7191 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7193 struct task_group *tg;
7195 if (!cgrp->parent) {
7196 /* This is early initialization for the top cgroup */
7197 init_task_group.css.cgroup = cgrp;
7198 return &init_task_group.css;
7201 /* we support only 1-level deep hierarchical scheduler atm */
7202 if (cgrp->parent->parent)
7203 return ERR_PTR(-EINVAL);
7205 tg = sched_create_group();
7207 return ERR_PTR(-ENOMEM);
7209 /* Bind the cgroup to task_group object we just created */
7210 tg->css.cgroup = cgrp;
7216 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7218 struct task_group *tg = cgroup_tg(cgrp);
7220 sched_destroy_group(tg);
7224 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7225 struct task_struct *tsk)
7227 /* We don't support RT-tasks being in separate groups */
7228 if (tsk->sched_class != &fair_sched_class)
7235 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7236 struct cgroup *old_cont, struct task_struct *tsk)
7238 sched_move_task(tsk);
7241 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7244 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7247 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7249 struct task_group *tg = cgroup_tg(cgrp);
7251 return (u64) tg->shares;
7254 static struct cftype cpu_files[] = {
7257 .read_uint = cpu_shares_read_uint,
7258 .write_uint = cpu_shares_write_uint,
7262 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7264 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7267 struct cgroup_subsys cpu_cgroup_subsys = {
7269 .create = cpu_cgroup_create,
7270 .destroy = cpu_cgroup_destroy,
7271 .can_attach = cpu_cgroup_can_attach,
7272 .attach = cpu_cgroup_attach,
7273 .populate = cpu_cgroup_populate,
7274 .subsys_id = cpu_cgroup_subsys_id,
7278 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7280 #ifdef CONFIG_CGROUP_CPUACCT
7283 * CPU accounting code for task groups.
7285 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7286 * (balbir@in.ibm.com).
7289 /* track cpu usage of a group of tasks */
7291 struct cgroup_subsys_state css;
7292 /* cpuusage holds pointer to a u64-type object on every cpu */
7296 struct cgroup_subsys cpuacct_subsys;
7298 /* return cpu accounting group corresponding to this container */
7299 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7301 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7302 struct cpuacct, css);
7305 /* return cpu accounting group to which this task belongs */
7306 static inline struct cpuacct *task_ca(struct task_struct *tsk)
7308 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7309 struct cpuacct, css);
7312 /* create a new cpu accounting group */
7313 static struct cgroup_subsys_state *cpuacct_create(
7314 struct cgroup_subsys *ss, struct cgroup *cont)
7316 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7319 return ERR_PTR(-ENOMEM);
7321 ca->cpuusage = alloc_percpu(u64);
7322 if (!ca->cpuusage) {
7324 return ERR_PTR(-ENOMEM);
7330 /* destroy an existing cpu accounting group */
7332 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
7334 struct cpuacct *ca = cgroup_ca(cont);
7336 free_percpu(ca->cpuusage);
7340 /* return total cpu usage (in nanoseconds) of a group */
7341 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
7343 struct cpuacct *ca = cgroup_ca(cont);
7344 u64 totalcpuusage = 0;
7347 for_each_possible_cpu(i) {
7348 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
7351 * Take rq->lock to make 64-bit addition safe on 32-bit
7354 spin_lock_irq(&cpu_rq(i)->lock);
7355 totalcpuusage += *cpuusage;
7356 spin_unlock_irq(&cpu_rq(i)->lock);
7359 return totalcpuusage;
7362 static struct cftype files[] = {
7365 .read_uint = cpuusage_read,
7369 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7371 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
7375 * charge this task's execution time to its accounting group.
7377 * called with rq->lock held.
7379 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
7383 if (!cpuacct_subsys.active)
7388 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
7390 *cpuusage += cputime;
7394 struct cgroup_subsys cpuacct_subsys = {
7396 .create = cpuacct_create,
7397 .destroy = cpuacct_destroy,
7398 .populate = cpuacct_populate,
7399 .subsys_id = cpuacct_subsys_id,
7401 #endif /* CONFIG_CGROUP_CPUACCT */