4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/pid_namespace.h>
48 #include <linux/smp.h>
49 #include <linux/threads.h>
50 #include <linux/timer.h>
51 #include <linux/rcupdate.h>
52 #include <linux/cpu.h>
53 #include <linux/cpuset.h>
54 #include <linux/percpu.h>
55 #include <linux/kthread.h>
56 #include <linux/seq_file.h>
57 #include <linux/sysctl.h>
58 #include <linux/syscalls.h>
59 #include <linux/times.h>
60 #include <linux/tsacct_kern.h>
61 #include <linux/kprobes.h>
62 #include <linux/delayacct.h>
63 #include <linux/reciprocal_div.h>
64 #include <linux/unistd.h>
65 #include <linux/pagemap.h>
68 #include <asm/irq_regs.h>
71 * Scheduler clock - returns current time in nanosec units.
72 * This is default implementation.
73 * Architectures and sub-architectures can override this.
75 unsigned long long __attribute__((weak)) sched_clock(void)
77 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Some helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / HZ))
104 #define NICE_0_LOAD SCHED_LOAD_SCALE
105 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
108 * These are the 'tuning knobs' of the scheduler:
110 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
111 * Timeslices get refilled after they expire.
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
122 return reciprocal_divide(load, sg->reciprocal_cpu_power);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
131 sg->__cpu_power += val;
132 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
136 static inline int rt_policy(int policy)
138 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
143 static inline int task_has_rt_policy(struct task_struct *p)
145 return rt_policy(p->policy);
149 * This is the priority-queue data structure of the RT scheduling class:
151 struct rt_prio_array {
152 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
153 struct list_head queue[MAX_RT_PRIO];
156 #ifdef CONFIG_FAIR_GROUP_SCHED
158 #include <linux/cgroup.h>
162 /* task group related information */
164 #ifdef CONFIG_FAIR_CGROUP_SCHED
165 struct cgroup_subsys_state css;
167 /* schedulable entities of this group on each cpu */
168 struct sched_entity **se;
169 /* runqueue "owned" by this group on each cpu */
170 struct cfs_rq **cfs_rq;
171 unsigned long shares;
172 /* spinlock to serialize modification to shares */
177 /* Default task group's sched entity on each cpu */
178 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
179 /* Default task group's cfs_rq on each cpu */
180 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
182 static struct sched_entity *init_sched_entity_p[NR_CPUS];
183 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
185 /* Default task group.
186 * Every task in system belong to this group at bootup.
188 struct task_group init_task_group = {
189 .se = init_sched_entity_p,
190 .cfs_rq = init_cfs_rq_p,
193 #ifdef CONFIG_FAIR_USER_SCHED
194 # define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
196 # define INIT_TASK_GRP_LOAD NICE_0_LOAD
199 static int init_task_group_load = INIT_TASK_GRP_LOAD;
201 /* return group to which a task belongs */
202 static inline struct task_group *task_group(struct task_struct *p)
204 struct task_group *tg;
206 #ifdef CONFIG_FAIR_USER_SCHED
208 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
209 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
210 struct task_group, css);
212 tg = &init_task_group;
217 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
218 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu)
220 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
221 p->se.parent = task_group(p)->se[cpu];
226 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu) { }
228 #endif /* CONFIG_FAIR_GROUP_SCHED */
230 /* CFS-related fields in a runqueue */
232 struct load_weight load;
233 unsigned long nr_running;
238 struct rb_root tasks_timeline;
239 struct rb_node *rb_leftmost;
240 struct rb_node *rb_load_balance_curr;
241 /* 'curr' points to currently running entity on this cfs_rq.
242 * It is set to NULL otherwise (i.e when none are currently running).
244 struct sched_entity *curr;
246 unsigned long nr_spread_over;
248 #ifdef CONFIG_FAIR_GROUP_SCHED
249 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
252 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
253 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
254 * (like users, containers etc.)
256 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
257 * list is used during load balance.
259 struct list_head leaf_cfs_rq_list;
260 struct task_group *tg; /* group that "owns" this runqueue */
264 /* Real-Time classes' related field in a runqueue: */
266 struct rt_prio_array active;
267 int rt_load_balance_idx;
268 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
272 * This is the main, per-CPU runqueue data structure.
274 * Locking rule: those places that want to lock multiple runqueues
275 * (such as the load balancing or the thread migration code), lock
276 * acquire operations must be ordered by ascending &runqueue.
283 * nr_running and cpu_load should be in the same cacheline because
284 * remote CPUs use both these fields when doing load calculation.
286 unsigned long nr_running;
287 #define CPU_LOAD_IDX_MAX 5
288 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
289 unsigned char idle_at_tick;
291 unsigned char in_nohz_recently;
293 /* capture load from *all* tasks on this cpu: */
294 struct load_weight load;
295 unsigned long nr_load_updates;
299 #ifdef CONFIG_FAIR_GROUP_SCHED
300 /* list of leaf cfs_rq on this cpu: */
301 struct list_head leaf_cfs_rq_list;
306 * This is part of a global counter where only the total sum
307 * over all CPUs matters. A task can increase this counter on
308 * one CPU and if it got migrated afterwards it may decrease
309 * it on another CPU. Always updated under the runqueue lock:
311 unsigned long nr_uninterruptible;
313 struct task_struct *curr, *idle;
314 unsigned long next_balance;
315 struct mm_struct *prev_mm;
317 u64 clock, prev_clock_raw;
320 unsigned int clock_warps, clock_overflows;
322 unsigned int clock_deep_idle_events;
328 struct sched_domain *sd;
330 /* For active balancing */
333 /* cpu of this runqueue: */
336 struct task_struct *migration_thread;
337 struct list_head migration_queue;
340 #ifdef CONFIG_SCHEDSTATS
342 struct sched_info rq_sched_info;
344 /* sys_sched_yield() stats */
345 unsigned int yld_exp_empty;
346 unsigned int yld_act_empty;
347 unsigned int yld_both_empty;
348 unsigned int yld_count;
350 /* schedule() stats */
351 unsigned int sched_switch;
352 unsigned int sched_count;
353 unsigned int sched_goidle;
355 /* try_to_wake_up() stats */
356 unsigned int ttwu_count;
357 unsigned int ttwu_local;
360 unsigned int bkl_count;
362 struct lock_class_key rq_lock_key;
365 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
366 static DEFINE_MUTEX(sched_hotcpu_mutex);
368 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
370 rq->curr->sched_class->check_preempt_curr(rq, p);
373 static inline int cpu_of(struct rq *rq)
383 * Update the per-runqueue clock, as finegrained as the platform can give
384 * us, but without assuming monotonicity, etc.:
386 static void __update_rq_clock(struct rq *rq)
388 u64 prev_raw = rq->prev_clock_raw;
389 u64 now = sched_clock();
390 s64 delta = now - prev_raw;
391 u64 clock = rq->clock;
393 #ifdef CONFIG_SCHED_DEBUG
394 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
397 * Protect against sched_clock() occasionally going backwards:
399 if (unlikely(delta < 0)) {
404 * Catch too large forward jumps too:
406 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
407 if (clock < rq->tick_timestamp + TICK_NSEC)
408 clock = rq->tick_timestamp + TICK_NSEC;
411 rq->clock_overflows++;
413 if (unlikely(delta > rq->clock_max_delta))
414 rq->clock_max_delta = delta;
419 rq->prev_clock_raw = now;
423 static void update_rq_clock(struct rq *rq)
425 if (likely(smp_processor_id() == cpu_of(rq)))
426 __update_rq_clock(rq);
430 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
431 * See detach_destroy_domains: synchronize_sched for details.
433 * The domain tree of any CPU may only be accessed from within
434 * preempt-disabled sections.
436 #define for_each_domain(cpu, __sd) \
437 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
439 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
440 #define this_rq() (&__get_cpu_var(runqueues))
441 #define task_rq(p) cpu_rq(task_cpu(p))
442 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
445 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
447 #ifdef CONFIG_SCHED_DEBUG
448 # define const_debug __read_mostly
450 # define const_debug static const
454 * Debugging: various feature bits
457 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
458 SCHED_FEAT_WAKEUP_PREEMPT = 2,
459 SCHED_FEAT_START_DEBIT = 4,
460 SCHED_FEAT_TREE_AVG = 8,
461 SCHED_FEAT_APPROX_AVG = 16,
464 const_debug unsigned int sysctl_sched_features =
465 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
466 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
467 SCHED_FEAT_START_DEBIT * 1 |
468 SCHED_FEAT_TREE_AVG * 0 |
469 SCHED_FEAT_APPROX_AVG * 0;
471 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
474 * Number of tasks to iterate in a single balance run.
475 * Limited because this is done with IRQs disabled.
477 const_debug unsigned int sysctl_sched_nr_migrate = 32;
480 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
481 * clock constructed from sched_clock():
483 unsigned long long cpu_clock(int cpu)
485 unsigned long long now;
489 local_irq_save(flags);
492 * Only call sched_clock() if the scheduler has already been
493 * initialized (some code might call cpu_clock() very early):
498 local_irq_restore(flags);
502 EXPORT_SYMBOL_GPL(cpu_clock);
504 #ifndef prepare_arch_switch
505 # define prepare_arch_switch(next) do { } while (0)
507 #ifndef finish_arch_switch
508 # define finish_arch_switch(prev) do { } while (0)
511 static inline int task_current(struct rq *rq, struct task_struct *p)
513 return rq->curr == p;
516 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
517 static inline int task_running(struct rq *rq, struct task_struct *p)
519 return task_current(rq, p);
522 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
526 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
528 #ifdef CONFIG_DEBUG_SPINLOCK
529 /* this is a valid case when another task releases the spinlock */
530 rq->lock.owner = current;
533 * If we are tracking spinlock dependencies then we have to
534 * fix up the runqueue lock - which gets 'carried over' from
537 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
539 spin_unlock_irq(&rq->lock);
542 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
543 static inline int task_running(struct rq *rq, struct task_struct *p)
548 return task_current(rq, p);
552 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
556 * We can optimise this out completely for !SMP, because the
557 * SMP rebalancing from interrupt is the only thing that cares
562 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
563 spin_unlock_irq(&rq->lock);
565 spin_unlock(&rq->lock);
569 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
573 * After ->oncpu is cleared, the task can be moved to a different CPU.
574 * We must ensure this doesn't happen until the switch is completely
580 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
584 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
587 * __task_rq_lock - lock the runqueue a given task resides on.
588 * Must be called interrupts disabled.
590 static inline struct rq *__task_rq_lock(struct task_struct *p)
594 struct rq *rq = task_rq(p);
595 spin_lock(&rq->lock);
596 if (likely(rq == task_rq(p)))
598 spin_unlock(&rq->lock);
603 * task_rq_lock - lock the runqueue a given task resides on and disable
604 * interrupts. Note the ordering: we can safely lookup the task_rq without
605 * explicitly disabling preemption.
607 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
613 local_irq_save(*flags);
615 spin_lock(&rq->lock);
616 if (likely(rq == task_rq(p)))
618 spin_unlock_irqrestore(&rq->lock, *flags);
622 static void __task_rq_unlock(struct rq *rq)
625 spin_unlock(&rq->lock);
628 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
631 spin_unlock_irqrestore(&rq->lock, *flags);
635 * this_rq_lock - lock this runqueue and disable interrupts.
637 static struct rq *this_rq_lock(void)
644 spin_lock(&rq->lock);
650 * We are going deep-idle (irqs are disabled):
652 void sched_clock_idle_sleep_event(void)
654 struct rq *rq = cpu_rq(smp_processor_id());
656 spin_lock(&rq->lock);
657 __update_rq_clock(rq);
658 spin_unlock(&rq->lock);
659 rq->clock_deep_idle_events++;
661 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
664 * We just idled delta nanoseconds (called with irqs disabled):
666 void sched_clock_idle_wakeup_event(u64 delta_ns)
668 struct rq *rq = cpu_rq(smp_processor_id());
669 u64 now = sched_clock();
671 rq->idle_clock += delta_ns;
673 * Override the previous timestamp and ignore all
674 * sched_clock() deltas that occured while we idled,
675 * and use the PM-provided delta_ns to advance the
678 spin_lock(&rq->lock);
679 rq->prev_clock_raw = now;
680 rq->clock += delta_ns;
681 spin_unlock(&rq->lock);
683 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
686 * resched_task - mark a task 'to be rescheduled now'.
688 * On UP this means the setting of the need_resched flag, on SMP it
689 * might also involve a cross-CPU call to trigger the scheduler on
694 #ifndef tsk_is_polling
695 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
698 static void resched_task(struct task_struct *p)
702 assert_spin_locked(&task_rq(p)->lock);
704 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
707 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
710 if (cpu == smp_processor_id())
713 /* NEED_RESCHED must be visible before we test polling */
715 if (!tsk_is_polling(p))
716 smp_send_reschedule(cpu);
719 static void resched_cpu(int cpu)
721 struct rq *rq = cpu_rq(cpu);
724 if (!spin_trylock_irqsave(&rq->lock, flags))
726 resched_task(cpu_curr(cpu));
727 spin_unlock_irqrestore(&rq->lock, flags);
730 static inline void resched_task(struct task_struct *p)
732 assert_spin_locked(&task_rq(p)->lock);
733 set_tsk_need_resched(p);
737 #if BITS_PER_LONG == 32
738 # define WMULT_CONST (~0UL)
740 # define WMULT_CONST (1UL << 32)
743 #define WMULT_SHIFT 32
746 * Shift right and round:
748 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
751 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
752 struct load_weight *lw)
756 if (unlikely(!lw->inv_weight))
757 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
759 tmp = (u64)delta_exec * weight;
761 * Check whether we'd overflow the 64-bit multiplication:
763 if (unlikely(tmp > WMULT_CONST))
764 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
767 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
769 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
772 static inline unsigned long
773 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
775 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
778 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
783 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
789 * To aid in avoiding the subversion of "niceness" due to uneven distribution
790 * of tasks with abnormal "nice" values across CPUs the contribution that
791 * each task makes to its run queue's load is weighted according to its
792 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
793 * scaled version of the new time slice allocation that they receive on time
797 #define WEIGHT_IDLEPRIO 2
798 #define WMULT_IDLEPRIO (1 << 31)
801 * Nice levels are multiplicative, with a gentle 10% change for every
802 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
803 * nice 1, it will get ~10% less CPU time than another CPU-bound task
804 * that remained on nice 0.
806 * The "10% effect" is relative and cumulative: from _any_ nice level,
807 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
808 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
809 * If a task goes up by ~10% and another task goes down by ~10% then
810 * the relative distance between them is ~25%.)
812 static const int prio_to_weight[40] = {
813 /* -20 */ 88761, 71755, 56483, 46273, 36291,
814 /* -15 */ 29154, 23254, 18705, 14949, 11916,
815 /* -10 */ 9548, 7620, 6100, 4904, 3906,
816 /* -5 */ 3121, 2501, 1991, 1586, 1277,
817 /* 0 */ 1024, 820, 655, 526, 423,
818 /* 5 */ 335, 272, 215, 172, 137,
819 /* 10 */ 110, 87, 70, 56, 45,
820 /* 15 */ 36, 29, 23, 18, 15,
824 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
826 * In cases where the weight does not change often, we can use the
827 * precalculated inverse to speed up arithmetics by turning divisions
828 * into multiplications:
830 static const u32 prio_to_wmult[40] = {
831 /* -20 */ 48388, 59856, 76040, 92818, 118348,
832 /* -15 */ 147320, 184698, 229616, 287308, 360437,
833 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
834 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
835 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
836 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
837 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
838 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
841 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
844 * runqueue iterator, to support SMP load-balancing between different
845 * scheduling classes, without having to expose their internal data
846 * structures to the load-balancing proper:
850 struct task_struct *(*start)(void *);
851 struct task_struct *(*next)(void *);
856 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
857 unsigned long max_load_move, struct sched_domain *sd,
858 enum cpu_idle_type idle, int *all_pinned,
859 int *this_best_prio, struct rq_iterator *iterator);
862 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
863 struct sched_domain *sd, enum cpu_idle_type idle,
864 struct rq_iterator *iterator);
867 #ifdef CONFIG_CGROUP_CPUACCT
868 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
870 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
873 #include "sched_stats.h"
874 #include "sched_idletask.c"
875 #include "sched_fair.c"
876 #include "sched_rt.c"
877 #ifdef CONFIG_SCHED_DEBUG
878 # include "sched_debug.c"
881 #define sched_class_highest (&rt_sched_class)
884 * Update delta_exec, delta_fair fields for rq.
886 * delta_fair clock advances at a rate inversely proportional to
887 * total load (rq->load.weight) on the runqueue, while
888 * delta_exec advances at the same rate as wall-clock (provided
891 * delta_exec / delta_fair is a measure of the (smoothened) load on this
892 * runqueue over any given interval. This (smoothened) load is used
893 * during load balance.
895 * This function is called /before/ updating rq->load
896 * and when switching tasks.
898 static inline void inc_load(struct rq *rq, const struct task_struct *p)
900 update_load_add(&rq->load, p->se.load.weight);
903 static inline void dec_load(struct rq *rq, const struct task_struct *p)
905 update_load_sub(&rq->load, p->se.load.weight);
908 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)
920 static void set_load_weight(struct task_struct *p)
922 if (task_has_rt_policy(p)) {
923 p->se.load.weight = prio_to_weight[0] * 2;
924 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
929 * SCHED_IDLE tasks get minimal weight:
931 if (p->policy == SCHED_IDLE) {
932 p->se.load.weight = WEIGHT_IDLEPRIO;
933 p->se.load.inv_weight = WMULT_IDLEPRIO;
937 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
938 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
941 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
943 sched_info_queued(p);
944 p->sched_class->enqueue_task(rq, p, wakeup);
948 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
950 p->sched_class->dequeue_task(rq, p, sleep);
955 * __normal_prio - return the priority that is based on the static prio
957 static inline int __normal_prio(struct task_struct *p)
959 return p->static_prio;
963 * Calculate the expected normal priority: i.e. priority
964 * without taking RT-inheritance into account. Might be
965 * boosted by interactivity modifiers. Changes upon fork,
966 * setprio syscalls, and whenever the interactivity
967 * estimator recalculates.
969 static inline int normal_prio(struct task_struct *p)
973 if (task_has_rt_policy(p))
974 prio = MAX_RT_PRIO-1 - p->rt_priority;
976 prio = __normal_prio(p);
981 * Calculate the current priority, i.e. the priority
982 * taken into account by the scheduler. This value might
983 * be boosted by RT tasks, or might be boosted by
984 * interactivity modifiers. Will be RT if the task got
985 * RT-boosted. If not then it returns p->normal_prio.
987 static int effective_prio(struct task_struct *p)
989 p->normal_prio = normal_prio(p);
991 * If we are RT tasks or we were boosted to RT priority,
992 * keep the priority unchanged. Otherwise, update priority
993 * to the normal priority:
995 if (!rt_prio(p->prio))
996 return p->normal_prio;
1001 * activate_task - move a task to the runqueue.
1003 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1005 if (p->state == TASK_UNINTERRUPTIBLE)
1006 rq->nr_uninterruptible--;
1008 enqueue_task(rq, p, wakeup);
1009 inc_nr_running(p, rq);
1013 * deactivate_task - remove a task from the runqueue.
1015 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1017 if (p->state == TASK_UNINTERRUPTIBLE)
1018 rq->nr_uninterruptible++;
1020 dequeue_task(rq, p, sleep);
1021 dec_nr_running(p, rq);
1025 * task_curr - is this task currently executing on a CPU?
1026 * @p: the task in question.
1028 inline int task_curr(const struct task_struct *p)
1030 return cpu_curr(task_cpu(p)) == p;
1033 /* Used instead of source_load when we know the type == 0 */
1034 unsigned long weighted_cpuload(const int cpu)
1036 return cpu_rq(cpu)->load.weight;
1039 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1041 set_task_cfs_rq(p, cpu);
1044 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1045 * successfuly executed on another CPU. We must ensure that updates of
1046 * per-task data have been completed by this moment.
1049 task_thread_info(p)->cpu = cpu;
1056 * Is this task likely cache-hot:
1059 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1063 if (p->sched_class != &fair_sched_class)
1066 if (sysctl_sched_migration_cost == -1)
1068 if (sysctl_sched_migration_cost == 0)
1071 delta = now - p->se.exec_start;
1073 return delta < (s64)sysctl_sched_migration_cost;
1077 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1079 int old_cpu = task_cpu(p);
1080 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1081 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1082 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1085 clock_offset = old_rq->clock - new_rq->clock;
1087 #ifdef CONFIG_SCHEDSTATS
1088 if (p->se.wait_start)
1089 p->se.wait_start -= clock_offset;
1090 if (p->se.sleep_start)
1091 p->se.sleep_start -= clock_offset;
1092 if (p->se.block_start)
1093 p->se.block_start -= clock_offset;
1094 if (old_cpu != new_cpu) {
1095 schedstat_inc(p, se.nr_migrations);
1096 if (task_hot(p, old_rq->clock, NULL))
1097 schedstat_inc(p, se.nr_forced2_migrations);
1100 p->se.vruntime -= old_cfsrq->min_vruntime -
1101 new_cfsrq->min_vruntime;
1103 __set_task_cpu(p, new_cpu);
1106 struct migration_req {
1107 struct list_head list;
1109 struct task_struct *task;
1112 struct completion done;
1116 * The task's runqueue lock must be held.
1117 * Returns true if you have to wait for migration thread.
1120 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1122 struct rq *rq = task_rq(p);
1125 * If the task is not on a runqueue (and not running), then
1126 * it is sufficient to simply update the task's cpu field.
1128 if (!p->se.on_rq && !task_running(rq, p)) {
1129 set_task_cpu(p, dest_cpu);
1133 init_completion(&req->done);
1135 req->dest_cpu = dest_cpu;
1136 list_add(&req->list, &rq->migration_queue);
1142 * wait_task_inactive - wait for a thread to unschedule.
1144 * The caller must ensure that the task *will* unschedule sometime soon,
1145 * else this function might spin for a *long* time. This function can't
1146 * be called with interrupts off, or it may introduce deadlock with
1147 * smp_call_function() if an IPI is sent by the same process we are
1148 * waiting to become inactive.
1150 void wait_task_inactive(struct task_struct *p)
1152 unsigned long flags;
1158 * We do the initial early heuristics without holding
1159 * any task-queue locks at all. We'll only try to get
1160 * the runqueue lock when things look like they will
1166 * If the task is actively running on another CPU
1167 * still, just relax and busy-wait without holding
1170 * NOTE! Since we don't hold any locks, it's not
1171 * even sure that "rq" stays as the right runqueue!
1172 * But we don't care, since "task_running()" will
1173 * return false if the runqueue has changed and p
1174 * is actually now running somewhere else!
1176 while (task_running(rq, p))
1180 * Ok, time to look more closely! We need the rq
1181 * lock now, to be *sure*. If we're wrong, we'll
1182 * just go back and repeat.
1184 rq = task_rq_lock(p, &flags);
1185 running = task_running(rq, p);
1186 on_rq = p->se.on_rq;
1187 task_rq_unlock(rq, &flags);
1190 * Was it really running after all now that we
1191 * checked with the proper locks actually held?
1193 * Oops. Go back and try again..
1195 if (unlikely(running)) {
1201 * It's not enough that it's not actively running,
1202 * it must be off the runqueue _entirely_, and not
1205 * So if it wa still runnable (but just not actively
1206 * running right now), it's preempted, and we should
1207 * yield - it could be a while.
1209 if (unlikely(on_rq)) {
1210 schedule_timeout_uninterruptible(1);
1215 * Ahh, all good. It wasn't running, and it wasn't
1216 * runnable, which means that it will never become
1217 * running in the future either. We're all done!
1224 * kick_process - kick a running thread to enter/exit the kernel
1225 * @p: the to-be-kicked thread
1227 * Cause a process which is running on another CPU to enter
1228 * kernel-mode, without any delay. (to get signals handled.)
1230 * NOTE: this function doesnt have to take the runqueue lock,
1231 * because all it wants to ensure is that the remote task enters
1232 * the kernel. If the IPI races and the task has been migrated
1233 * to another CPU then no harm is done and the purpose has been
1236 void kick_process(struct task_struct *p)
1242 if ((cpu != smp_processor_id()) && task_curr(p))
1243 smp_send_reschedule(cpu);
1248 * Return a low guess at the load of a migration-source cpu weighted
1249 * according to the scheduling class and "nice" value.
1251 * We want to under-estimate the load of migration sources, to
1252 * balance conservatively.
1254 static unsigned long source_load(int cpu, int type)
1256 struct rq *rq = cpu_rq(cpu);
1257 unsigned long total = weighted_cpuload(cpu);
1262 return min(rq->cpu_load[type-1], total);
1266 * Return a high guess at the load of a migration-target cpu weighted
1267 * according to the scheduling class and "nice" value.
1269 static unsigned long target_load(int cpu, int type)
1271 struct rq *rq = cpu_rq(cpu);
1272 unsigned long total = weighted_cpuload(cpu);
1277 return max(rq->cpu_load[type-1], total);
1281 * Return the average load per task on the cpu's run queue
1283 static inline unsigned long cpu_avg_load_per_task(int cpu)
1285 struct rq *rq = cpu_rq(cpu);
1286 unsigned long total = weighted_cpuload(cpu);
1287 unsigned long n = rq->nr_running;
1289 return n ? total / n : SCHED_LOAD_SCALE;
1293 * find_idlest_group finds and returns the least busy CPU group within the
1296 static struct sched_group *
1297 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1299 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1300 unsigned long min_load = ULONG_MAX, this_load = 0;
1301 int load_idx = sd->forkexec_idx;
1302 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1305 unsigned long load, avg_load;
1309 /* Skip over this group if it has no CPUs allowed */
1310 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1313 local_group = cpu_isset(this_cpu, group->cpumask);
1315 /* Tally up the load of all CPUs in the group */
1318 for_each_cpu_mask(i, group->cpumask) {
1319 /* Bias balancing toward cpus of our domain */
1321 load = source_load(i, load_idx);
1323 load = target_load(i, load_idx);
1328 /* Adjust by relative CPU power of the group */
1329 avg_load = sg_div_cpu_power(group,
1330 avg_load * SCHED_LOAD_SCALE);
1333 this_load = avg_load;
1335 } else if (avg_load < min_load) {
1336 min_load = avg_load;
1339 } while (group = group->next, group != sd->groups);
1341 if (!idlest || 100*this_load < imbalance*min_load)
1347 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1350 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1353 unsigned long load, min_load = ULONG_MAX;
1357 /* Traverse only the allowed CPUs */
1358 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1360 for_each_cpu_mask(i, tmp) {
1361 load = weighted_cpuload(i);
1363 if (load < min_load || (load == min_load && i == this_cpu)) {
1373 * sched_balance_self: balance the current task (running on cpu) in domains
1374 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1377 * Balance, ie. select the least loaded group.
1379 * Returns the target CPU number, or the same CPU if no balancing is needed.
1381 * preempt must be disabled.
1383 static int sched_balance_self(int cpu, int flag)
1385 struct task_struct *t = current;
1386 struct sched_domain *tmp, *sd = NULL;
1388 for_each_domain(cpu, tmp) {
1390 * If power savings logic is enabled for a domain, stop there.
1392 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1394 if (tmp->flags & flag)
1400 struct sched_group *group;
1401 int new_cpu, weight;
1403 if (!(sd->flags & flag)) {
1409 group = find_idlest_group(sd, t, cpu);
1415 new_cpu = find_idlest_cpu(group, t, cpu);
1416 if (new_cpu == -1 || new_cpu == cpu) {
1417 /* Now try balancing at a lower domain level of cpu */
1422 /* Now try balancing at a lower domain level of new_cpu */
1425 weight = cpus_weight(span);
1426 for_each_domain(cpu, tmp) {
1427 if (weight <= cpus_weight(tmp->span))
1429 if (tmp->flags & flag)
1432 /* while loop will break here if sd == NULL */
1438 #endif /* CONFIG_SMP */
1441 * wake_idle() will wake a task on an idle cpu if task->cpu is
1442 * not idle and an idle cpu is available. The span of cpus to
1443 * search starts with cpus closest then further out as needed,
1444 * so we always favor a closer, idle cpu.
1446 * Returns the CPU we should wake onto.
1448 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1449 static int wake_idle(int cpu, struct task_struct *p)
1452 struct sched_domain *sd;
1456 * If it is idle, then it is the best cpu to run this task.
1458 * This cpu is also the best, if it has more than one task already.
1459 * Siblings must be also busy(in most cases) as they didn't already
1460 * pickup the extra load from this cpu and hence we need not check
1461 * sibling runqueue info. This will avoid the checks and cache miss
1462 * penalities associated with that.
1464 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1467 for_each_domain(cpu, sd) {
1468 if (sd->flags & SD_WAKE_IDLE) {
1469 cpus_and(tmp, sd->span, p->cpus_allowed);
1470 for_each_cpu_mask(i, tmp) {
1472 if (i != task_cpu(p)) {
1474 se.nr_wakeups_idle);
1486 static inline int wake_idle(int cpu, struct task_struct *p)
1493 * try_to_wake_up - wake up a thread
1494 * @p: the to-be-woken-up thread
1495 * @state: the mask of task states that can be woken
1496 * @sync: do a synchronous wakeup?
1498 * Put it on the run-queue if it's not already there. The "current"
1499 * thread is always on the run-queue (except when the actual
1500 * re-schedule is in progress), and as such you're allowed to do
1501 * the simpler "current->state = TASK_RUNNING" to mark yourself
1502 * runnable without the overhead of this.
1504 * returns failure only if the task is already active.
1506 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1508 int cpu, orig_cpu, this_cpu, success = 0;
1509 unsigned long flags;
1513 struct sched_domain *sd, *this_sd = NULL;
1514 unsigned long load, this_load;
1518 rq = task_rq_lock(p, &flags);
1519 old_state = p->state;
1520 if (!(old_state & state))
1528 this_cpu = smp_processor_id();
1531 if (unlikely(task_running(rq, p)))
1536 schedstat_inc(rq, ttwu_count);
1537 if (cpu == this_cpu) {
1538 schedstat_inc(rq, ttwu_local);
1542 for_each_domain(this_cpu, sd) {
1543 if (cpu_isset(cpu, sd->span)) {
1544 schedstat_inc(sd, ttwu_wake_remote);
1550 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1554 * Check for affine wakeup and passive balancing possibilities.
1557 int idx = this_sd->wake_idx;
1558 unsigned int imbalance;
1560 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1562 load = source_load(cpu, idx);
1563 this_load = target_load(this_cpu, idx);
1565 new_cpu = this_cpu; /* Wake to this CPU if we can */
1567 if (this_sd->flags & SD_WAKE_AFFINE) {
1568 unsigned long tl = this_load;
1569 unsigned long tl_per_task;
1572 * Attract cache-cold tasks on sync wakeups:
1574 if (sync && !task_hot(p, rq->clock, this_sd))
1577 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1578 tl_per_task = cpu_avg_load_per_task(this_cpu);
1581 * If sync wakeup then subtract the (maximum possible)
1582 * effect of the currently running task from the load
1583 * of the current CPU:
1586 tl -= current->se.load.weight;
1589 tl + target_load(cpu, idx) <= tl_per_task) ||
1590 100*(tl + p->se.load.weight) <= imbalance*load) {
1592 * This domain has SD_WAKE_AFFINE and
1593 * p is cache cold in this domain, and
1594 * there is no bad imbalance.
1596 schedstat_inc(this_sd, ttwu_move_affine);
1597 schedstat_inc(p, se.nr_wakeups_affine);
1603 * Start passive balancing when half the imbalance_pct
1606 if (this_sd->flags & SD_WAKE_BALANCE) {
1607 if (imbalance*this_load <= 100*load) {
1608 schedstat_inc(this_sd, ttwu_move_balance);
1609 schedstat_inc(p, se.nr_wakeups_passive);
1615 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1617 new_cpu = wake_idle(new_cpu, p);
1618 if (new_cpu != cpu) {
1619 set_task_cpu(p, new_cpu);
1620 task_rq_unlock(rq, &flags);
1621 /* might preempt at this point */
1622 rq = task_rq_lock(p, &flags);
1623 old_state = p->state;
1624 if (!(old_state & state))
1629 this_cpu = smp_processor_id();
1634 #endif /* CONFIG_SMP */
1635 schedstat_inc(p, se.nr_wakeups);
1637 schedstat_inc(p, se.nr_wakeups_sync);
1638 if (orig_cpu != cpu)
1639 schedstat_inc(p, se.nr_wakeups_migrate);
1640 if (cpu == this_cpu)
1641 schedstat_inc(p, se.nr_wakeups_local);
1643 schedstat_inc(p, se.nr_wakeups_remote);
1644 update_rq_clock(rq);
1645 activate_task(rq, p, 1);
1646 check_preempt_curr(rq, p);
1650 p->state = TASK_RUNNING;
1652 task_rq_unlock(rq, &flags);
1657 int fastcall wake_up_process(struct task_struct *p)
1659 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1660 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1662 EXPORT_SYMBOL(wake_up_process);
1664 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1666 return try_to_wake_up(p, state, 0);
1670 * Perform scheduler related setup for a newly forked process p.
1671 * p is forked by current.
1673 * __sched_fork() is basic setup used by init_idle() too:
1675 static void __sched_fork(struct task_struct *p)
1677 p->se.exec_start = 0;
1678 p->se.sum_exec_runtime = 0;
1679 p->se.prev_sum_exec_runtime = 0;
1681 #ifdef CONFIG_SCHEDSTATS
1682 p->se.wait_start = 0;
1683 p->se.sum_sleep_runtime = 0;
1684 p->se.sleep_start = 0;
1685 p->se.block_start = 0;
1686 p->se.sleep_max = 0;
1687 p->se.block_max = 0;
1689 p->se.slice_max = 0;
1693 INIT_LIST_HEAD(&p->run_list);
1696 #ifdef CONFIG_PREEMPT_NOTIFIERS
1697 INIT_HLIST_HEAD(&p->preempt_notifiers);
1701 * We mark the process as running here, but have not actually
1702 * inserted it onto the runqueue yet. This guarantees that
1703 * nobody will actually run it, and a signal or other external
1704 * event cannot wake it up and insert it on the runqueue either.
1706 p->state = TASK_RUNNING;
1710 * fork()/clone()-time setup:
1712 void sched_fork(struct task_struct *p, int clone_flags)
1714 int cpu = get_cpu();
1719 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1721 set_task_cpu(p, cpu);
1724 * Make sure we do not leak PI boosting priority to the child:
1726 p->prio = current->normal_prio;
1727 if (!rt_prio(p->prio))
1728 p->sched_class = &fair_sched_class;
1730 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1731 if (likely(sched_info_on()))
1732 memset(&p->sched_info, 0, sizeof(p->sched_info));
1734 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1737 #ifdef CONFIG_PREEMPT
1738 /* Want to start with kernel preemption disabled. */
1739 task_thread_info(p)->preempt_count = 1;
1745 * wake_up_new_task - wake up a newly created task for the first time.
1747 * This function will do some initial scheduler statistics housekeeping
1748 * that must be done for every newly created context, then puts the task
1749 * on the runqueue and wakes it.
1751 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1753 unsigned long flags;
1756 rq = task_rq_lock(p, &flags);
1757 BUG_ON(p->state != TASK_RUNNING);
1758 update_rq_clock(rq);
1760 p->prio = effective_prio(p);
1762 if (!p->sched_class->task_new || !current->se.on_rq) {
1763 activate_task(rq, p, 0);
1766 * Let the scheduling class do new task startup
1767 * management (if any):
1769 p->sched_class->task_new(rq, p);
1770 inc_nr_running(p, rq);
1772 check_preempt_curr(rq, p);
1773 task_rq_unlock(rq, &flags);
1776 #ifdef CONFIG_PREEMPT_NOTIFIERS
1779 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1780 * @notifier: notifier struct to register
1782 void preempt_notifier_register(struct preempt_notifier *notifier)
1784 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1786 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1789 * preempt_notifier_unregister - no longer interested in preemption notifications
1790 * @notifier: notifier struct to unregister
1792 * This is safe to call from within a preemption notifier.
1794 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1796 hlist_del(¬ifier->link);
1798 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1800 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1802 struct preempt_notifier *notifier;
1803 struct hlist_node *node;
1805 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1806 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1810 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1811 struct task_struct *next)
1813 struct preempt_notifier *notifier;
1814 struct hlist_node *node;
1816 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1817 notifier->ops->sched_out(notifier, next);
1822 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1827 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1828 struct task_struct *next)
1835 * prepare_task_switch - prepare to switch tasks
1836 * @rq: the runqueue preparing to switch
1837 * @prev: the current task that is being switched out
1838 * @next: the task we are going to switch to.
1840 * This is called with the rq lock held and interrupts off. It must
1841 * be paired with a subsequent finish_task_switch after the context
1844 * prepare_task_switch sets up locking and calls architecture specific
1848 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1849 struct task_struct *next)
1851 fire_sched_out_preempt_notifiers(prev, next);
1852 prepare_lock_switch(rq, next);
1853 prepare_arch_switch(next);
1857 * finish_task_switch - clean up after a task-switch
1858 * @rq: runqueue associated with task-switch
1859 * @prev: the thread we just switched away from.
1861 * finish_task_switch must be called after the context switch, paired
1862 * with a prepare_task_switch call before the context switch.
1863 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1864 * and do any other architecture-specific cleanup actions.
1866 * Note that we may have delayed dropping an mm in context_switch(). If
1867 * so, we finish that here outside of the runqueue lock. (Doing it
1868 * with the lock held can cause deadlocks; see schedule() for
1871 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1872 __releases(rq->lock)
1874 struct mm_struct *mm = rq->prev_mm;
1880 * A task struct has one reference for the use as "current".
1881 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1882 * schedule one last time. The schedule call will never return, and
1883 * the scheduled task must drop that reference.
1884 * The test for TASK_DEAD must occur while the runqueue locks are
1885 * still held, otherwise prev could be scheduled on another cpu, die
1886 * there before we look at prev->state, and then the reference would
1888 * Manfred Spraul <manfred@colorfullife.com>
1890 prev_state = prev->state;
1891 finish_arch_switch(prev);
1892 finish_lock_switch(rq, prev);
1893 fire_sched_in_preempt_notifiers(current);
1896 if (unlikely(prev_state == TASK_DEAD)) {
1898 * Remove function-return probe instances associated with this
1899 * task and put them back on the free list.
1901 kprobe_flush_task(prev);
1902 put_task_struct(prev);
1907 * schedule_tail - first thing a freshly forked thread must call.
1908 * @prev: the thread we just switched away from.
1910 asmlinkage void schedule_tail(struct task_struct *prev)
1911 __releases(rq->lock)
1913 struct rq *rq = this_rq();
1915 finish_task_switch(rq, prev);
1916 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1917 /* In this case, finish_task_switch does not reenable preemption */
1920 if (current->set_child_tid)
1921 put_user(task_pid_vnr(current), current->set_child_tid);
1925 * context_switch - switch to the new MM and the new
1926 * thread's register state.
1929 context_switch(struct rq *rq, struct task_struct *prev,
1930 struct task_struct *next)
1932 struct mm_struct *mm, *oldmm;
1934 prepare_task_switch(rq, prev, next);
1936 oldmm = prev->active_mm;
1938 * For paravirt, this is coupled with an exit in switch_to to
1939 * combine the page table reload and the switch backend into
1942 arch_enter_lazy_cpu_mode();
1944 if (unlikely(!mm)) {
1945 next->active_mm = oldmm;
1946 atomic_inc(&oldmm->mm_count);
1947 enter_lazy_tlb(oldmm, next);
1949 switch_mm(oldmm, mm, next);
1951 if (unlikely(!prev->mm)) {
1952 prev->active_mm = NULL;
1953 rq->prev_mm = oldmm;
1956 * Since the runqueue lock will be released by the next
1957 * task (which is an invalid locking op but in the case
1958 * of the scheduler it's an obvious special-case), so we
1959 * do an early lockdep release here:
1961 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1962 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1965 /* Here we just switch the register state and the stack. */
1966 switch_to(prev, next, prev);
1970 * this_rq must be evaluated again because prev may have moved
1971 * CPUs since it called schedule(), thus the 'rq' on its stack
1972 * frame will be invalid.
1974 finish_task_switch(this_rq(), prev);
1978 * nr_running, nr_uninterruptible and nr_context_switches:
1980 * externally visible scheduler statistics: current number of runnable
1981 * threads, current number of uninterruptible-sleeping threads, total
1982 * number of context switches performed since bootup.
1984 unsigned long nr_running(void)
1986 unsigned long i, sum = 0;
1988 for_each_online_cpu(i)
1989 sum += cpu_rq(i)->nr_running;
1994 unsigned long nr_uninterruptible(void)
1996 unsigned long i, sum = 0;
1998 for_each_possible_cpu(i)
1999 sum += cpu_rq(i)->nr_uninterruptible;
2002 * Since we read the counters lockless, it might be slightly
2003 * inaccurate. Do not allow it to go below zero though:
2005 if (unlikely((long)sum < 0))
2011 unsigned long long nr_context_switches(void)
2014 unsigned long long sum = 0;
2016 for_each_possible_cpu(i)
2017 sum += cpu_rq(i)->nr_switches;
2022 unsigned long nr_iowait(void)
2024 unsigned long i, sum = 0;
2026 for_each_possible_cpu(i)
2027 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2032 unsigned long nr_active(void)
2034 unsigned long i, running = 0, uninterruptible = 0;
2036 for_each_online_cpu(i) {
2037 running += cpu_rq(i)->nr_running;
2038 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2041 if (unlikely((long)uninterruptible < 0))
2042 uninterruptible = 0;
2044 return running + uninterruptible;
2048 * Update rq->cpu_load[] statistics. This function is usually called every
2049 * scheduler tick (TICK_NSEC).
2051 static void update_cpu_load(struct rq *this_rq)
2053 unsigned long this_load = this_rq->load.weight;
2056 this_rq->nr_load_updates++;
2058 /* Update our load: */
2059 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2060 unsigned long old_load, new_load;
2062 /* scale is effectively 1 << i now, and >> i divides by scale */
2064 old_load = this_rq->cpu_load[i];
2065 new_load = this_load;
2067 * Round up the averaging division if load is increasing. This
2068 * prevents us from getting stuck on 9 if the load is 10, for
2071 if (new_load > old_load)
2072 new_load += scale-1;
2073 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2080 * double_rq_lock - safely lock two runqueues
2082 * Note this does not disable interrupts like task_rq_lock,
2083 * you need to do so manually before calling.
2085 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2086 __acquires(rq1->lock)
2087 __acquires(rq2->lock)
2089 BUG_ON(!irqs_disabled());
2091 spin_lock(&rq1->lock);
2092 __acquire(rq2->lock); /* Fake it out ;) */
2095 spin_lock(&rq1->lock);
2096 spin_lock(&rq2->lock);
2098 spin_lock(&rq2->lock);
2099 spin_lock(&rq1->lock);
2102 update_rq_clock(rq1);
2103 update_rq_clock(rq2);
2107 * double_rq_unlock - safely unlock two runqueues
2109 * Note this does not restore interrupts like task_rq_unlock,
2110 * you need to do so manually after calling.
2112 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2113 __releases(rq1->lock)
2114 __releases(rq2->lock)
2116 spin_unlock(&rq1->lock);
2118 spin_unlock(&rq2->lock);
2120 __release(rq2->lock);
2124 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2126 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2127 __releases(this_rq->lock)
2128 __acquires(busiest->lock)
2129 __acquires(this_rq->lock)
2131 if (unlikely(!irqs_disabled())) {
2132 /* printk() doesn't work good under rq->lock */
2133 spin_unlock(&this_rq->lock);
2136 if (unlikely(!spin_trylock(&busiest->lock))) {
2137 if (busiest < this_rq) {
2138 spin_unlock(&this_rq->lock);
2139 spin_lock(&busiest->lock);
2140 spin_lock(&this_rq->lock);
2142 spin_lock(&busiest->lock);
2147 * If dest_cpu is allowed for this process, migrate the task to it.
2148 * This is accomplished by forcing the cpu_allowed mask to only
2149 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2150 * the cpu_allowed mask is restored.
2152 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2154 struct migration_req req;
2155 unsigned long flags;
2158 rq = task_rq_lock(p, &flags);
2159 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2160 || unlikely(cpu_is_offline(dest_cpu)))
2163 /* force the process onto the specified CPU */
2164 if (migrate_task(p, dest_cpu, &req)) {
2165 /* Need to wait for migration thread (might exit: take ref). */
2166 struct task_struct *mt = rq->migration_thread;
2168 get_task_struct(mt);
2169 task_rq_unlock(rq, &flags);
2170 wake_up_process(mt);
2171 put_task_struct(mt);
2172 wait_for_completion(&req.done);
2177 task_rq_unlock(rq, &flags);
2181 * sched_exec - execve() is a valuable balancing opportunity, because at
2182 * this point the task has the smallest effective memory and cache footprint.
2184 void sched_exec(void)
2186 int new_cpu, this_cpu = get_cpu();
2187 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2189 if (new_cpu != this_cpu)
2190 sched_migrate_task(current, new_cpu);
2194 * pull_task - move a task from a remote runqueue to the local runqueue.
2195 * Both runqueues must be locked.
2197 static void pull_task(struct rq *src_rq, struct task_struct *p,
2198 struct rq *this_rq, int this_cpu)
2200 deactivate_task(src_rq, p, 0);
2201 set_task_cpu(p, this_cpu);
2202 activate_task(this_rq, p, 0);
2204 * Note that idle threads have a prio of MAX_PRIO, for this test
2205 * to be always true for them.
2207 check_preempt_curr(this_rq, p);
2211 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2214 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2215 struct sched_domain *sd, enum cpu_idle_type idle,
2219 * We do not migrate tasks that are:
2220 * 1) running (obviously), or
2221 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2222 * 3) are cache-hot on their current CPU.
2224 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2225 schedstat_inc(p, se.nr_failed_migrations_affine);
2230 if (task_running(rq, p)) {
2231 schedstat_inc(p, se.nr_failed_migrations_running);
2236 * Aggressive migration if:
2237 * 1) task is cache cold, or
2238 * 2) too many balance attempts have failed.
2241 if (!task_hot(p, rq->clock, sd) ||
2242 sd->nr_balance_failed > sd->cache_nice_tries) {
2243 #ifdef CONFIG_SCHEDSTATS
2244 if (task_hot(p, rq->clock, sd)) {
2245 schedstat_inc(sd, lb_hot_gained[idle]);
2246 schedstat_inc(p, se.nr_forced_migrations);
2252 if (task_hot(p, rq->clock, sd)) {
2253 schedstat_inc(p, se.nr_failed_migrations_hot);
2259 static unsigned long
2260 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2261 unsigned long max_load_move, struct sched_domain *sd,
2262 enum cpu_idle_type idle, int *all_pinned,
2263 int *this_best_prio, struct rq_iterator *iterator)
2265 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2266 struct task_struct *p;
2267 long rem_load_move = max_load_move;
2269 if (max_load_move == 0)
2275 * Start the load-balancing iterator:
2277 p = iterator->start(iterator->arg);
2279 if (!p || loops++ > sysctl_sched_nr_migrate)
2282 * To help distribute high priority tasks across CPUs we don't
2283 * skip a task if it will be the highest priority task (i.e. smallest
2284 * prio value) on its new queue regardless of its load weight
2286 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2287 SCHED_LOAD_SCALE_FUZZ;
2288 if ((skip_for_load && p->prio >= *this_best_prio) ||
2289 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2290 p = iterator->next(iterator->arg);
2294 pull_task(busiest, p, this_rq, this_cpu);
2296 rem_load_move -= p->se.load.weight;
2299 * We only want to steal up to the prescribed amount of weighted load.
2301 if (rem_load_move > 0) {
2302 if (p->prio < *this_best_prio)
2303 *this_best_prio = p->prio;
2304 p = iterator->next(iterator->arg);
2309 * Right now, this is one of only two places pull_task() is called,
2310 * so we can safely collect pull_task() stats here rather than
2311 * inside pull_task().
2313 schedstat_add(sd, lb_gained[idle], pulled);
2316 *all_pinned = pinned;
2318 return max_load_move - rem_load_move;
2322 * move_tasks tries to move up to max_load_move weighted load from busiest to
2323 * this_rq, as part of a balancing operation within domain "sd".
2324 * Returns 1 if successful and 0 otherwise.
2326 * Called with both runqueues locked.
2328 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2329 unsigned long max_load_move,
2330 struct sched_domain *sd, enum cpu_idle_type idle,
2333 const struct sched_class *class = sched_class_highest;
2334 unsigned long total_load_moved = 0;
2335 int this_best_prio = this_rq->curr->prio;
2339 class->load_balance(this_rq, this_cpu, busiest,
2340 max_load_move - total_load_moved,
2341 sd, idle, all_pinned, &this_best_prio);
2342 class = class->next;
2343 } while (class && max_load_move > total_load_moved);
2345 return total_load_moved > 0;
2349 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2350 struct sched_domain *sd, enum cpu_idle_type idle,
2351 struct rq_iterator *iterator)
2353 struct task_struct *p = iterator->start(iterator->arg);
2357 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2358 pull_task(busiest, p, this_rq, this_cpu);
2360 * Right now, this is only the second place pull_task()
2361 * is called, so we can safely collect pull_task()
2362 * stats here rather than inside pull_task().
2364 schedstat_inc(sd, lb_gained[idle]);
2368 p = iterator->next(iterator->arg);
2375 * move_one_task tries to move exactly one task from busiest to this_rq, as
2376 * part of active balancing operations within "domain".
2377 * Returns 1 if successful and 0 otherwise.
2379 * Called with both runqueues locked.
2381 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2382 struct sched_domain *sd, enum cpu_idle_type idle)
2384 const struct sched_class *class;
2386 for (class = sched_class_highest; class; class = class->next)
2387 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2394 * find_busiest_group finds and returns the busiest CPU group within the
2395 * domain. It calculates and returns the amount of weighted load which
2396 * should be moved to restore balance via the imbalance parameter.
2398 static struct sched_group *
2399 find_busiest_group(struct sched_domain *sd, int this_cpu,
2400 unsigned long *imbalance, enum cpu_idle_type idle,
2401 int *sd_idle, cpumask_t *cpus, int *balance)
2403 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2404 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2405 unsigned long max_pull;
2406 unsigned long busiest_load_per_task, busiest_nr_running;
2407 unsigned long this_load_per_task, this_nr_running;
2408 int load_idx, group_imb = 0;
2409 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2410 int power_savings_balance = 1;
2411 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2412 unsigned long min_nr_running = ULONG_MAX;
2413 struct sched_group *group_min = NULL, *group_leader = NULL;
2416 max_load = this_load = total_load = total_pwr = 0;
2417 busiest_load_per_task = busiest_nr_running = 0;
2418 this_load_per_task = this_nr_running = 0;
2419 if (idle == CPU_NOT_IDLE)
2420 load_idx = sd->busy_idx;
2421 else if (idle == CPU_NEWLY_IDLE)
2422 load_idx = sd->newidle_idx;
2424 load_idx = sd->idle_idx;
2427 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2430 int __group_imb = 0;
2431 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2432 unsigned long sum_nr_running, sum_weighted_load;
2434 local_group = cpu_isset(this_cpu, group->cpumask);
2437 balance_cpu = first_cpu(group->cpumask);
2439 /* Tally up the load of all CPUs in the group */
2440 sum_weighted_load = sum_nr_running = avg_load = 0;
2442 min_cpu_load = ~0UL;
2444 for_each_cpu_mask(i, group->cpumask) {
2447 if (!cpu_isset(i, *cpus))
2452 if (*sd_idle && rq->nr_running)
2455 /* Bias balancing toward cpus of our domain */
2457 if (idle_cpu(i) && !first_idle_cpu) {
2462 load = target_load(i, load_idx);
2464 load = source_load(i, load_idx);
2465 if (load > max_cpu_load)
2466 max_cpu_load = load;
2467 if (min_cpu_load > load)
2468 min_cpu_load = load;
2472 sum_nr_running += rq->nr_running;
2473 sum_weighted_load += weighted_cpuload(i);
2477 * First idle cpu or the first cpu(busiest) in this sched group
2478 * is eligible for doing load balancing at this and above
2479 * domains. In the newly idle case, we will allow all the cpu's
2480 * to do the newly idle load balance.
2482 if (idle != CPU_NEWLY_IDLE && local_group &&
2483 balance_cpu != this_cpu && balance) {
2488 total_load += avg_load;
2489 total_pwr += group->__cpu_power;
2491 /* Adjust by relative CPU power of the group */
2492 avg_load = sg_div_cpu_power(group,
2493 avg_load * SCHED_LOAD_SCALE);
2495 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2498 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2501 this_load = avg_load;
2503 this_nr_running = sum_nr_running;
2504 this_load_per_task = sum_weighted_load;
2505 } else if (avg_load > max_load &&
2506 (sum_nr_running > group_capacity || __group_imb)) {
2507 max_load = avg_load;
2509 busiest_nr_running = sum_nr_running;
2510 busiest_load_per_task = sum_weighted_load;
2511 group_imb = __group_imb;
2514 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2516 * Busy processors will not participate in power savings
2519 if (idle == CPU_NOT_IDLE ||
2520 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2524 * If the local group is idle or completely loaded
2525 * no need to do power savings balance at this domain
2527 if (local_group && (this_nr_running >= group_capacity ||
2529 power_savings_balance = 0;
2532 * If a group is already running at full capacity or idle,
2533 * don't include that group in power savings calculations
2535 if (!power_savings_balance || sum_nr_running >= group_capacity
2540 * Calculate the group which has the least non-idle load.
2541 * This is the group from where we need to pick up the load
2544 if ((sum_nr_running < min_nr_running) ||
2545 (sum_nr_running == min_nr_running &&
2546 first_cpu(group->cpumask) <
2547 first_cpu(group_min->cpumask))) {
2549 min_nr_running = sum_nr_running;
2550 min_load_per_task = sum_weighted_load /
2555 * Calculate the group which is almost near its
2556 * capacity but still has some space to pick up some load
2557 * from other group and save more power
2559 if (sum_nr_running <= group_capacity - 1) {
2560 if (sum_nr_running > leader_nr_running ||
2561 (sum_nr_running == leader_nr_running &&
2562 first_cpu(group->cpumask) >
2563 first_cpu(group_leader->cpumask))) {
2564 group_leader = group;
2565 leader_nr_running = sum_nr_running;
2570 group = group->next;
2571 } while (group != sd->groups);
2573 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2576 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2578 if (this_load >= avg_load ||
2579 100*max_load <= sd->imbalance_pct*this_load)
2582 busiest_load_per_task /= busiest_nr_running;
2584 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2587 * We're trying to get all the cpus to the average_load, so we don't
2588 * want to push ourselves above the average load, nor do we wish to
2589 * reduce the max loaded cpu below the average load, as either of these
2590 * actions would just result in more rebalancing later, and ping-pong
2591 * tasks around. Thus we look for the minimum possible imbalance.
2592 * Negative imbalances (*we* are more loaded than anyone else) will
2593 * be counted as no imbalance for these purposes -- we can't fix that
2594 * by pulling tasks to us. Be careful of negative numbers as they'll
2595 * appear as very large values with unsigned longs.
2597 if (max_load <= busiest_load_per_task)
2601 * In the presence of smp nice balancing, certain scenarios can have
2602 * max load less than avg load(as we skip the groups at or below
2603 * its cpu_power, while calculating max_load..)
2605 if (max_load < avg_load) {
2607 goto small_imbalance;
2610 /* Don't want to pull so many tasks that a group would go idle */
2611 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2613 /* How much load to actually move to equalise the imbalance */
2614 *imbalance = min(max_pull * busiest->__cpu_power,
2615 (avg_load - this_load) * this->__cpu_power)
2619 * if *imbalance is less than the average load per runnable task
2620 * there is no gaurantee that any tasks will be moved so we'll have
2621 * a think about bumping its value to force at least one task to be
2624 if (*imbalance < busiest_load_per_task) {
2625 unsigned long tmp, pwr_now, pwr_move;
2629 pwr_move = pwr_now = 0;
2631 if (this_nr_running) {
2632 this_load_per_task /= this_nr_running;
2633 if (busiest_load_per_task > this_load_per_task)
2636 this_load_per_task = SCHED_LOAD_SCALE;
2638 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2639 busiest_load_per_task * imbn) {
2640 *imbalance = busiest_load_per_task;
2645 * OK, we don't have enough imbalance to justify moving tasks,
2646 * however we may be able to increase total CPU power used by
2650 pwr_now += busiest->__cpu_power *
2651 min(busiest_load_per_task, max_load);
2652 pwr_now += this->__cpu_power *
2653 min(this_load_per_task, this_load);
2654 pwr_now /= SCHED_LOAD_SCALE;
2656 /* Amount of load we'd subtract */
2657 tmp = sg_div_cpu_power(busiest,
2658 busiest_load_per_task * SCHED_LOAD_SCALE);
2660 pwr_move += busiest->__cpu_power *
2661 min(busiest_load_per_task, max_load - tmp);
2663 /* Amount of load we'd add */
2664 if (max_load * busiest->__cpu_power <
2665 busiest_load_per_task * SCHED_LOAD_SCALE)
2666 tmp = sg_div_cpu_power(this,
2667 max_load * busiest->__cpu_power);
2669 tmp = sg_div_cpu_power(this,
2670 busiest_load_per_task * SCHED_LOAD_SCALE);
2671 pwr_move += this->__cpu_power *
2672 min(this_load_per_task, this_load + tmp);
2673 pwr_move /= SCHED_LOAD_SCALE;
2675 /* Move if we gain throughput */
2676 if (pwr_move > pwr_now)
2677 *imbalance = busiest_load_per_task;
2683 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2684 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2687 if (this == group_leader && group_leader != group_min) {
2688 *imbalance = min_load_per_task;
2698 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2701 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2702 unsigned long imbalance, cpumask_t *cpus)
2704 struct rq *busiest = NULL, *rq;
2705 unsigned long max_load = 0;
2708 for_each_cpu_mask(i, group->cpumask) {
2711 if (!cpu_isset(i, *cpus))
2715 wl = weighted_cpuload(i);
2717 if (rq->nr_running == 1 && wl > imbalance)
2720 if (wl > max_load) {
2730 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2731 * so long as it is large enough.
2733 #define MAX_PINNED_INTERVAL 512
2736 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2737 * tasks if there is an imbalance.
2739 static int load_balance(int this_cpu, struct rq *this_rq,
2740 struct sched_domain *sd, enum cpu_idle_type idle,
2743 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2744 struct sched_group *group;
2745 unsigned long imbalance;
2747 cpumask_t cpus = CPU_MASK_ALL;
2748 unsigned long flags;
2751 * When power savings policy is enabled for the parent domain, idle
2752 * sibling can pick up load irrespective of busy siblings. In this case,
2753 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2754 * portraying it as CPU_NOT_IDLE.
2756 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2757 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2760 schedstat_inc(sd, lb_count[idle]);
2763 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2770 schedstat_inc(sd, lb_nobusyg[idle]);
2774 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2776 schedstat_inc(sd, lb_nobusyq[idle]);
2780 BUG_ON(busiest == this_rq);
2782 schedstat_add(sd, lb_imbalance[idle], imbalance);
2785 if (busiest->nr_running > 1) {
2787 * Attempt to move tasks. If find_busiest_group has found
2788 * an imbalance but busiest->nr_running <= 1, the group is
2789 * still unbalanced. ld_moved simply stays zero, so it is
2790 * correctly treated as an imbalance.
2792 local_irq_save(flags);
2793 double_rq_lock(this_rq, busiest);
2794 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2795 imbalance, sd, idle, &all_pinned);
2796 double_rq_unlock(this_rq, busiest);
2797 local_irq_restore(flags);
2800 * some other cpu did the load balance for us.
2802 if (ld_moved && this_cpu != smp_processor_id())
2803 resched_cpu(this_cpu);
2805 /* All tasks on this runqueue were pinned by CPU affinity */
2806 if (unlikely(all_pinned)) {
2807 cpu_clear(cpu_of(busiest), cpus);
2808 if (!cpus_empty(cpus))
2815 schedstat_inc(sd, lb_failed[idle]);
2816 sd->nr_balance_failed++;
2818 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2820 spin_lock_irqsave(&busiest->lock, flags);
2822 /* don't kick the migration_thread, if the curr
2823 * task on busiest cpu can't be moved to this_cpu
2825 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2826 spin_unlock_irqrestore(&busiest->lock, flags);
2828 goto out_one_pinned;
2831 if (!busiest->active_balance) {
2832 busiest->active_balance = 1;
2833 busiest->push_cpu = this_cpu;
2836 spin_unlock_irqrestore(&busiest->lock, flags);
2838 wake_up_process(busiest->migration_thread);
2841 * We've kicked active balancing, reset the failure
2844 sd->nr_balance_failed = sd->cache_nice_tries+1;
2847 sd->nr_balance_failed = 0;
2849 if (likely(!active_balance)) {
2850 /* We were unbalanced, so reset the balancing interval */
2851 sd->balance_interval = sd->min_interval;
2854 * If we've begun active balancing, start to back off. This
2855 * case may not be covered by the all_pinned logic if there
2856 * is only 1 task on the busy runqueue (because we don't call
2859 if (sd->balance_interval < sd->max_interval)
2860 sd->balance_interval *= 2;
2863 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2864 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2869 schedstat_inc(sd, lb_balanced[idle]);
2871 sd->nr_balance_failed = 0;
2874 /* tune up the balancing interval */
2875 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2876 (sd->balance_interval < sd->max_interval))
2877 sd->balance_interval *= 2;
2879 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2880 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2886 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2887 * tasks if there is an imbalance.
2889 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2890 * this_rq is locked.
2893 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2895 struct sched_group *group;
2896 struct rq *busiest = NULL;
2897 unsigned long imbalance;
2901 cpumask_t cpus = CPU_MASK_ALL;
2904 * When power savings policy is enabled for the parent domain, idle
2905 * sibling can pick up load irrespective of busy siblings. In this case,
2906 * let the state of idle sibling percolate up as IDLE, instead of
2907 * portraying it as CPU_NOT_IDLE.
2909 if (sd->flags & SD_SHARE_CPUPOWER &&
2910 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2913 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2915 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2916 &sd_idle, &cpus, NULL);
2918 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2922 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2925 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2929 BUG_ON(busiest == this_rq);
2931 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2934 if (busiest->nr_running > 1) {
2935 /* Attempt to move tasks */
2936 double_lock_balance(this_rq, busiest);
2937 /* this_rq->clock is already updated */
2938 update_rq_clock(busiest);
2939 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2940 imbalance, sd, CPU_NEWLY_IDLE,
2942 spin_unlock(&busiest->lock);
2944 if (unlikely(all_pinned)) {
2945 cpu_clear(cpu_of(busiest), cpus);
2946 if (!cpus_empty(cpus))
2952 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2953 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2954 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2957 sd->nr_balance_failed = 0;
2962 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2963 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2964 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2966 sd->nr_balance_failed = 0;
2972 * idle_balance is called by schedule() if this_cpu is about to become
2973 * idle. Attempts to pull tasks from other CPUs.
2975 static void idle_balance(int this_cpu, struct rq *this_rq)
2977 struct sched_domain *sd;
2978 int pulled_task = -1;
2979 unsigned long next_balance = jiffies + HZ;
2981 for_each_domain(this_cpu, sd) {
2982 unsigned long interval;
2984 if (!(sd->flags & SD_LOAD_BALANCE))
2987 if (sd->flags & SD_BALANCE_NEWIDLE)
2988 /* If we've pulled tasks over stop searching: */
2989 pulled_task = load_balance_newidle(this_cpu,
2992 interval = msecs_to_jiffies(sd->balance_interval);
2993 if (time_after(next_balance, sd->last_balance + interval))
2994 next_balance = sd->last_balance + interval;
2998 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3000 * We are going idle. next_balance may be set based on
3001 * a busy processor. So reset next_balance.
3003 this_rq->next_balance = next_balance;
3008 * active_load_balance is run by migration threads. It pushes running tasks
3009 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3010 * running on each physical CPU where possible, and avoids physical /
3011 * logical imbalances.
3013 * Called with busiest_rq locked.
3015 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3017 int target_cpu = busiest_rq->push_cpu;
3018 struct sched_domain *sd;
3019 struct rq *target_rq;
3021 /* Is there any task to move? */
3022 if (busiest_rq->nr_running <= 1)
3025 target_rq = cpu_rq(target_cpu);
3028 * This condition is "impossible", if it occurs
3029 * we need to fix it. Originally reported by
3030 * Bjorn Helgaas on a 128-cpu setup.
3032 BUG_ON(busiest_rq == target_rq);
3034 /* move a task from busiest_rq to target_rq */
3035 double_lock_balance(busiest_rq, target_rq);
3036 update_rq_clock(busiest_rq);
3037 update_rq_clock(target_rq);
3039 /* Search for an sd spanning us and the target CPU. */
3040 for_each_domain(target_cpu, sd) {
3041 if ((sd->flags & SD_LOAD_BALANCE) &&
3042 cpu_isset(busiest_cpu, sd->span))
3047 schedstat_inc(sd, alb_count);
3049 if (move_one_task(target_rq, target_cpu, busiest_rq,
3051 schedstat_inc(sd, alb_pushed);
3053 schedstat_inc(sd, alb_failed);
3055 spin_unlock(&target_rq->lock);
3060 atomic_t load_balancer;
3062 } nohz ____cacheline_aligned = {
3063 .load_balancer = ATOMIC_INIT(-1),
3064 .cpu_mask = CPU_MASK_NONE,
3068 * This routine will try to nominate the ilb (idle load balancing)
3069 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3070 * load balancing on behalf of all those cpus. If all the cpus in the system
3071 * go into this tickless mode, then there will be no ilb owner (as there is
3072 * no need for one) and all the cpus will sleep till the next wakeup event
3075 * For the ilb owner, tick is not stopped. And this tick will be used
3076 * for idle load balancing. ilb owner will still be part of
3079 * While stopping the tick, this cpu will become the ilb owner if there
3080 * is no other owner. And will be the owner till that cpu becomes busy
3081 * or if all cpus in the system stop their ticks at which point
3082 * there is no need for ilb owner.
3084 * When the ilb owner becomes busy, it nominates another owner, during the
3085 * next busy scheduler_tick()
3087 int select_nohz_load_balancer(int stop_tick)
3089 int cpu = smp_processor_id();
3092 cpu_set(cpu, nohz.cpu_mask);
3093 cpu_rq(cpu)->in_nohz_recently = 1;
3096 * If we are going offline and still the leader, give up!
3098 if (cpu_is_offline(cpu) &&
3099 atomic_read(&nohz.load_balancer) == cpu) {
3100 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3105 /* time for ilb owner also to sleep */
3106 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3107 if (atomic_read(&nohz.load_balancer) == cpu)
3108 atomic_set(&nohz.load_balancer, -1);
3112 if (atomic_read(&nohz.load_balancer) == -1) {
3113 /* make me the ilb owner */
3114 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3116 } else if (atomic_read(&nohz.load_balancer) == cpu)
3119 if (!cpu_isset(cpu, nohz.cpu_mask))
3122 cpu_clear(cpu, nohz.cpu_mask);
3124 if (atomic_read(&nohz.load_balancer) == cpu)
3125 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3132 static DEFINE_SPINLOCK(balancing);
3135 * It checks each scheduling domain to see if it is due to be balanced,
3136 * and initiates a balancing operation if so.
3138 * Balancing parameters are set up in arch_init_sched_domains.
3140 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3143 struct rq *rq = cpu_rq(cpu);
3144 unsigned long interval;
3145 struct sched_domain *sd;
3146 /* Earliest time when we have to do rebalance again */
3147 unsigned long next_balance = jiffies + 60*HZ;
3148 int update_next_balance = 0;
3150 for_each_domain(cpu, sd) {
3151 if (!(sd->flags & SD_LOAD_BALANCE))
3154 interval = sd->balance_interval;
3155 if (idle != CPU_IDLE)
3156 interval *= sd->busy_factor;
3158 /* scale ms to jiffies */
3159 interval = msecs_to_jiffies(interval);
3160 if (unlikely(!interval))
3162 if (interval > HZ*NR_CPUS/10)
3163 interval = HZ*NR_CPUS/10;
3166 if (sd->flags & SD_SERIALIZE) {
3167 if (!spin_trylock(&balancing))
3171 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3172 if (load_balance(cpu, rq, sd, idle, &balance)) {
3174 * We've pulled tasks over so either we're no
3175 * longer idle, or one of our SMT siblings is
3178 idle = CPU_NOT_IDLE;
3180 sd->last_balance = jiffies;
3182 if (sd->flags & SD_SERIALIZE)
3183 spin_unlock(&balancing);
3185 if (time_after(next_balance, sd->last_balance + interval)) {
3186 next_balance = sd->last_balance + interval;
3187 update_next_balance = 1;
3191 * Stop the load balance at this level. There is another
3192 * CPU in our sched group which is doing load balancing more
3200 * next_balance will be updated only when there is a need.
3201 * When the cpu is attached to null domain for ex, it will not be
3204 if (likely(update_next_balance))
3205 rq->next_balance = next_balance;
3209 * run_rebalance_domains is triggered when needed from the scheduler tick.
3210 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3211 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3213 static void run_rebalance_domains(struct softirq_action *h)
3215 int this_cpu = smp_processor_id();
3216 struct rq *this_rq = cpu_rq(this_cpu);
3217 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3218 CPU_IDLE : CPU_NOT_IDLE;
3220 rebalance_domains(this_cpu, idle);
3224 * If this cpu is the owner for idle load balancing, then do the
3225 * balancing on behalf of the other idle cpus whose ticks are
3228 if (this_rq->idle_at_tick &&
3229 atomic_read(&nohz.load_balancer) == this_cpu) {
3230 cpumask_t cpus = nohz.cpu_mask;
3234 cpu_clear(this_cpu, cpus);
3235 for_each_cpu_mask(balance_cpu, cpus) {
3237 * If this cpu gets work to do, stop the load balancing
3238 * work being done for other cpus. Next load
3239 * balancing owner will pick it up.
3244 rebalance_domains(balance_cpu, CPU_IDLE);
3246 rq = cpu_rq(balance_cpu);
3247 if (time_after(this_rq->next_balance, rq->next_balance))
3248 this_rq->next_balance = rq->next_balance;
3255 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3257 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3258 * idle load balancing owner or decide to stop the periodic load balancing,
3259 * if the whole system is idle.
3261 static inline void trigger_load_balance(struct rq *rq, int cpu)
3265 * If we were in the nohz mode recently and busy at the current
3266 * scheduler tick, then check if we need to nominate new idle
3269 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3270 rq->in_nohz_recently = 0;
3272 if (atomic_read(&nohz.load_balancer) == cpu) {
3273 cpu_clear(cpu, nohz.cpu_mask);
3274 atomic_set(&nohz.load_balancer, -1);
3277 if (atomic_read(&nohz.load_balancer) == -1) {
3279 * simple selection for now: Nominate the
3280 * first cpu in the nohz list to be the next
3283 * TBD: Traverse the sched domains and nominate
3284 * the nearest cpu in the nohz.cpu_mask.
3286 int ilb = first_cpu(nohz.cpu_mask);
3294 * If this cpu is idle and doing idle load balancing for all the
3295 * cpus with ticks stopped, is it time for that to stop?
3297 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3298 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3304 * If this cpu is idle and the idle load balancing is done by
3305 * someone else, then no need raise the SCHED_SOFTIRQ
3307 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3308 cpu_isset(cpu, nohz.cpu_mask))
3311 if (time_after_eq(jiffies, rq->next_balance))
3312 raise_softirq(SCHED_SOFTIRQ);
3315 #else /* CONFIG_SMP */
3318 * on UP we do not need to balance between CPUs:
3320 static inline void idle_balance(int cpu, struct rq *rq)
3326 DEFINE_PER_CPU(struct kernel_stat, kstat);
3328 EXPORT_PER_CPU_SYMBOL(kstat);
3331 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3332 * that have not yet been banked in case the task is currently running.
3334 unsigned long long task_sched_runtime(struct task_struct *p)
3336 unsigned long flags;
3340 rq = task_rq_lock(p, &flags);
3341 ns = p->se.sum_exec_runtime;
3342 if (task_current(rq, p)) {
3343 update_rq_clock(rq);
3344 delta_exec = rq->clock - p->se.exec_start;
3345 if ((s64)delta_exec > 0)
3348 task_rq_unlock(rq, &flags);
3354 * Account user cpu time to a process.
3355 * @p: the process that the cpu time gets accounted to
3356 * @cputime: the cpu time spent in user space since the last update
3358 void account_user_time(struct task_struct *p, cputime_t cputime)
3360 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3363 p->utime = cputime_add(p->utime, cputime);
3365 /* Add user time to cpustat. */
3366 tmp = cputime_to_cputime64(cputime);
3367 if (TASK_NICE(p) > 0)
3368 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3370 cpustat->user = cputime64_add(cpustat->user, tmp);
3374 * Account guest cpu time to a process.
3375 * @p: the process that the cpu time gets accounted to
3376 * @cputime: the cpu time spent in virtual machine since the last update
3378 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3381 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3383 tmp = cputime_to_cputime64(cputime);
3385 p->utime = cputime_add(p->utime, cputime);
3386 p->gtime = cputime_add(p->gtime, cputime);
3388 cpustat->user = cputime64_add(cpustat->user, tmp);
3389 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3393 * Account scaled user cpu time to a process.
3394 * @p: the process that the cpu time gets accounted to
3395 * @cputime: the cpu time spent in user space since the last update
3397 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3399 p->utimescaled = cputime_add(p->utimescaled, cputime);
3403 * Account system cpu time to a process.
3404 * @p: the process that the cpu time gets accounted to
3405 * @hardirq_offset: the offset to subtract from hardirq_count()
3406 * @cputime: the cpu time spent in kernel space since the last update
3408 void account_system_time(struct task_struct *p, int hardirq_offset,
3411 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3412 struct rq *rq = this_rq();
3415 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3416 return account_guest_time(p, cputime);
3418 p->stime = cputime_add(p->stime, cputime);
3420 /* Add system time to cpustat. */
3421 tmp = cputime_to_cputime64(cputime);
3422 if (hardirq_count() - hardirq_offset)
3423 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3424 else if (softirq_count())
3425 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3426 else if (p != rq->idle)
3427 cpustat->system = cputime64_add(cpustat->system, tmp);
3428 else if (atomic_read(&rq->nr_iowait) > 0)
3429 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3431 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3432 /* Account for system time used */
3433 acct_update_integrals(p);
3437 * Account scaled system cpu time to a process.
3438 * @p: the process that the cpu time gets accounted to
3439 * @hardirq_offset: the offset to subtract from hardirq_count()
3440 * @cputime: the cpu time spent in kernel space since the last update
3442 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3444 p->stimescaled = cputime_add(p->stimescaled, cputime);
3448 * Account for involuntary wait time.
3449 * @p: the process from which the cpu time has been stolen
3450 * @steal: the cpu time spent in involuntary wait
3452 void account_steal_time(struct task_struct *p, cputime_t steal)
3454 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3455 cputime64_t tmp = cputime_to_cputime64(steal);
3456 struct rq *rq = this_rq();
3458 if (p == rq->idle) {
3459 p->stime = cputime_add(p->stime, steal);
3460 if (atomic_read(&rq->nr_iowait) > 0)
3461 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3463 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3465 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3469 * This function gets called by the timer code, with HZ frequency.
3470 * We call it with interrupts disabled.
3472 * It also gets called by the fork code, when changing the parent's
3475 void scheduler_tick(void)
3477 int cpu = smp_processor_id();
3478 struct rq *rq = cpu_rq(cpu);
3479 struct task_struct *curr = rq->curr;
3480 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3482 spin_lock(&rq->lock);
3483 __update_rq_clock(rq);
3485 * Let rq->clock advance by at least TICK_NSEC:
3487 if (unlikely(rq->clock < next_tick))
3488 rq->clock = next_tick;
3489 rq->tick_timestamp = rq->clock;
3490 update_cpu_load(rq);
3491 if (curr != rq->idle) /* FIXME: needed? */
3492 curr->sched_class->task_tick(rq, curr);
3493 spin_unlock(&rq->lock);
3496 rq->idle_at_tick = idle_cpu(cpu);
3497 trigger_load_balance(rq, cpu);
3501 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3503 void fastcall add_preempt_count(int val)
3508 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3510 preempt_count() += val;
3512 * Spinlock count overflowing soon?
3514 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3517 EXPORT_SYMBOL(add_preempt_count);
3519 void fastcall sub_preempt_count(int val)
3524 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3527 * Is the spinlock portion underflowing?
3529 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3530 !(preempt_count() & PREEMPT_MASK)))
3533 preempt_count() -= val;
3535 EXPORT_SYMBOL(sub_preempt_count);
3540 * Print scheduling while atomic bug:
3542 static noinline void __schedule_bug(struct task_struct *prev)
3544 struct pt_regs *regs = get_irq_regs();
3546 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3547 prev->comm, prev->pid, preempt_count());
3549 debug_show_held_locks(prev);
3550 if (irqs_disabled())
3551 print_irqtrace_events(prev);
3560 * Various schedule()-time debugging checks and statistics:
3562 static inline void schedule_debug(struct task_struct *prev)
3565 * Test if we are atomic. Since do_exit() needs to call into
3566 * schedule() atomically, we ignore that path for now.
3567 * Otherwise, whine if we are scheduling when we should not be.
3569 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3570 __schedule_bug(prev);
3572 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3574 schedstat_inc(this_rq(), sched_count);
3575 #ifdef CONFIG_SCHEDSTATS
3576 if (unlikely(prev->lock_depth >= 0)) {
3577 schedstat_inc(this_rq(), bkl_count);
3578 schedstat_inc(prev, sched_info.bkl_count);
3584 * Pick up the highest-prio task:
3586 static inline struct task_struct *
3587 pick_next_task(struct rq *rq, struct task_struct *prev)
3589 const struct sched_class *class;
3590 struct task_struct *p;
3593 * Optimization: we know that if all tasks are in
3594 * the fair class we can call that function directly:
3596 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3597 p = fair_sched_class.pick_next_task(rq);
3602 class = sched_class_highest;
3604 p = class->pick_next_task(rq);
3608 * Will never be NULL as the idle class always
3609 * returns a non-NULL p:
3611 class = class->next;
3616 * schedule() is the main scheduler function.
3618 asmlinkage void __sched schedule(void)
3620 struct task_struct *prev, *next;
3627 cpu = smp_processor_id();
3631 switch_count = &prev->nivcsw;
3633 release_kernel_lock(prev);
3634 need_resched_nonpreemptible:
3636 schedule_debug(prev);
3639 * Do the rq-clock update outside the rq lock:
3641 local_irq_disable();
3642 __update_rq_clock(rq);
3643 spin_lock(&rq->lock);
3644 clear_tsk_need_resched(prev);
3646 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3647 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3648 unlikely(signal_pending(prev)))) {
3649 prev->state = TASK_RUNNING;
3651 deactivate_task(rq, prev, 1);
3653 switch_count = &prev->nvcsw;
3656 if (unlikely(!rq->nr_running))
3657 idle_balance(cpu, rq);
3659 prev->sched_class->put_prev_task(rq, prev);
3660 next = pick_next_task(rq, prev);
3662 sched_info_switch(prev, next);
3664 if (likely(prev != next)) {
3669 context_switch(rq, prev, next); /* unlocks the rq */
3671 spin_unlock_irq(&rq->lock);
3673 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3674 cpu = smp_processor_id();
3676 goto need_resched_nonpreemptible;
3678 preempt_enable_no_resched();
3679 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3682 EXPORT_SYMBOL(schedule);
3684 #ifdef CONFIG_PREEMPT
3686 * this is the entry point to schedule() from in-kernel preemption
3687 * off of preempt_enable. Kernel preemptions off return from interrupt
3688 * occur there and call schedule directly.
3690 asmlinkage void __sched preempt_schedule(void)
3692 struct thread_info *ti = current_thread_info();
3693 #ifdef CONFIG_PREEMPT_BKL
3694 struct task_struct *task = current;
3695 int saved_lock_depth;
3698 * If there is a non-zero preempt_count or interrupts are disabled,
3699 * we do not want to preempt the current task. Just return..
3701 if (likely(ti->preempt_count || irqs_disabled()))
3705 add_preempt_count(PREEMPT_ACTIVE);
3708 * We keep the big kernel semaphore locked, but we
3709 * clear ->lock_depth so that schedule() doesnt
3710 * auto-release the semaphore:
3712 #ifdef CONFIG_PREEMPT_BKL
3713 saved_lock_depth = task->lock_depth;
3714 task->lock_depth = -1;
3717 #ifdef CONFIG_PREEMPT_BKL
3718 task->lock_depth = saved_lock_depth;
3720 sub_preempt_count(PREEMPT_ACTIVE);
3723 * Check again in case we missed a preemption opportunity
3724 * between schedule and now.
3727 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3729 EXPORT_SYMBOL(preempt_schedule);
3732 * this is the entry point to schedule() from kernel preemption
3733 * off of irq context.
3734 * Note, that this is called and return with irqs disabled. This will
3735 * protect us against recursive calling from irq.
3737 asmlinkage void __sched preempt_schedule_irq(void)
3739 struct thread_info *ti = current_thread_info();
3740 #ifdef CONFIG_PREEMPT_BKL
3741 struct task_struct *task = current;
3742 int saved_lock_depth;
3744 /* Catch callers which need to be fixed */
3745 BUG_ON(ti->preempt_count || !irqs_disabled());
3748 add_preempt_count(PREEMPT_ACTIVE);
3751 * We keep the big kernel semaphore locked, but we
3752 * clear ->lock_depth so that schedule() doesnt
3753 * auto-release the semaphore:
3755 #ifdef CONFIG_PREEMPT_BKL
3756 saved_lock_depth = task->lock_depth;
3757 task->lock_depth = -1;
3761 local_irq_disable();
3762 #ifdef CONFIG_PREEMPT_BKL
3763 task->lock_depth = saved_lock_depth;
3765 sub_preempt_count(PREEMPT_ACTIVE);
3768 * Check again in case we missed a preemption opportunity
3769 * between schedule and now.
3772 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3775 #endif /* CONFIG_PREEMPT */
3777 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3780 return try_to_wake_up(curr->private, mode, sync);
3782 EXPORT_SYMBOL(default_wake_function);
3785 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3786 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3787 * number) then we wake all the non-exclusive tasks and one exclusive task.
3789 * There are circumstances in which we can try to wake a task which has already
3790 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3791 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3793 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3794 int nr_exclusive, int sync, void *key)
3796 wait_queue_t *curr, *next;
3798 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3799 unsigned flags = curr->flags;
3801 if (curr->func(curr, mode, sync, key) &&
3802 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3808 * __wake_up - wake up threads blocked on a waitqueue.
3810 * @mode: which threads
3811 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3812 * @key: is directly passed to the wakeup function
3814 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3815 int nr_exclusive, void *key)
3817 unsigned long flags;
3819 spin_lock_irqsave(&q->lock, flags);
3820 __wake_up_common(q, mode, nr_exclusive, 0, key);
3821 spin_unlock_irqrestore(&q->lock, flags);
3823 EXPORT_SYMBOL(__wake_up);
3826 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3828 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3830 __wake_up_common(q, mode, 1, 0, NULL);
3834 * __wake_up_sync - wake up threads blocked on a waitqueue.
3836 * @mode: which threads
3837 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3839 * The sync wakeup differs that the waker knows that it will schedule
3840 * away soon, so while the target thread will be woken up, it will not
3841 * be migrated to another CPU - ie. the two threads are 'synchronized'
3842 * with each other. This can prevent needless bouncing between CPUs.
3844 * On UP it can prevent extra preemption.
3847 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3849 unsigned long flags;
3855 if (unlikely(!nr_exclusive))
3858 spin_lock_irqsave(&q->lock, flags);
3859 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3860 spin_unlock_irqrestore(&q->lock, flags);
3862 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3864 void complete(struct completion *x)
3866 unsigned long flags;
3868 spin_lock_irqsave(&x->wait.lock, flags);
3870 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3872 spin_unlock_irqrestore(&x->wait.lock, flags);
3874 EXPORT_SYMBOL(complete);
3876 void complete_all(struct completion *x)
3878 unsigned long flags;
3880 spin_lock_irqsave(&x->wait.lock, flags);
3881 x->done += UINT_MAX/2;
3882 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3884 spin_unlock_irqrestore(&x->wait.lock, flags);
3886 EXPORT_SYMBOL(complete_all);
3888 static inline long __sched
3889 do_wait_for_common(struct completion *x, long timeout, int state)
3892 DECLARE_WAITQUEUE(wait, current);
3894 wait.flags |= WQ_FLAG_EXCLUSIVE;
3895 __add_wait_queue_tail(&x->wait, &wait);
3897 if (state == TASK_INTERRUPTIBLE &&
3898 signal_pending(current)) {
3899 __remove_wait_queue(&x->wait, &wait);
3900 return -ERESTARTSYS;
3902 __set_current_state(state);
3903 spin_unlock_irq(&x->wait.lock);
3904 timeout = schedule_timeout(timeout);
3905 spin_lock_irq(&x->wait.lock);
3907 __remove_wait_queue(&x->wait, &wait);
3911 __remove_wait_queue(&x->wait, &wait);
3918 wait_for_common(struct completion *x, long timeout, int state)
3922 spin_lock_irq(&x->wait.lock);
3923 timeout = do_wait_for_common(x, timeout, state);
3924 spin_unlock_irq(&x->wait.lock);
3928 void __sched wait_for_completion(struct completion *x)
3930 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3932 EXPORT_SYMBOL(wait_for_completion);
3934 unsigned long __sched
3935 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3937 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3939 EXPORT_SYMBOL(wait_for_completion_timeout);
3941 int __sched wait_for_completion_interruptible(struct completion *x)
3943 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3944 if (t == -ERESTARTSYS)
3948 EXPORT_SYMBOL(wait_for_completion_interruptible);
3950 unsigned long __sched
3951 wait_for_completion_interruptible_timeout(struct completion *x,
3952 unsigned long timeout)
3954 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3956 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3959 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3961 unsigned long flags;
3964 init_waitqueue_entry(&wait, current);
3966 __set_current_state(state);
3968 spin_lock_irqsave(&q->lock, flags);
3969 __add_wait_queue(q, &wait);
3970 spin_unlock(&q->lock);
3971 timeout = schedule_timeout(timeout);
3972 spin_lock_irq(&q->lock);
3973 __remove_wait_queue(q, &wait);
3974 spin_unlock_irqrestore(&q->lock, flags);
3979 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3981 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3983 EXPORT_SYMBOL(interruptible_sleep_on);
3986 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3988 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3990 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3992 void __sched sleep_on(wait_queue_head_t *q)
3994 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3996 EXPORT_SYMBOL(sleep_on);
3998 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4000 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4002 EXPORT_SYMBOL(sleep_on_timeout);
4004 #ifdef CONFIG_RT_MUTEXES
4007 * rt_mutex_setprio - set the current priority of a task
4009 * @prio: prio value (kernel-internal form)
4011 * This function changes the 'effective' priority of a task. It does
4012 * not touch ->normal_prio like __setscheduler().
4014 * Used by the rt_mutex code to implement priority inheritance logic.
4016 void rt_mutex_setprio(struct task_struct *p, int prio)
4018 unsigned long flags;
4019 int oldprio, on_rq, running;
4022 BUG_ON(prio < 0 || prio > MAX_PRIO);
4024 rq = task_rq_lock(p, &flags);
4025 update_rq_clock(rq);
4028 on_rq = p->se.on_rq;
4029 running = task_current(rq, p);
4031 dequeue_task(rq, p, 0);
4033 p->sched_class->put_prev_task(rq, p);
4037 p->sched_class = &rt_sched_class;
4039 p->sched_class = &fair_sched_class;
4045 p->sched_class->set_curr_task(rq);
4046 enqueue_task(rq, p, 0);
4048 * Reschedule if we are currently running on this runqueue and
4049 * our priority decreased, or if we are not currently running on
4050 * this runqueue and our priority is higher than the current's
4053 if (p->prio > oldprio)
4054 resched_task(rq->curr);
4056 check_preempt_curr(rq, p);
4059 task_rq_unlock(rq, &flags);
4064 void set_user_nice(struct task_struct *p, long nice)
4066 int old_prio, delta, on_rq;
4067 unsigned long flags;
4070 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4073 * We have to be careful, if called from sys_setpriority(),
4074 * the task might be in the middle of scheduling on another CPU.
4076 rq = task_rq_lock(p, &flags);
4077 update_rq_clock(rq);
4079 * The RT priorities are set via sched_setscheduler(), but we still
4080 * allow the 'normal' nice value to be set - but as expected
4081 * it wont have any effect on scheduling until the task is
4082 * SCHED_FIFO/SCHED_RR:
4084 if (task_has_rt_policy(p)) {
4085 p->static_prio = NICE_TO_PRIO(nice);
4088 on_rq = p->se.on_rq;
4090 dequeue_task(rq, p, 0);
4094 p->static_prio = NICE_TO_PRIO(nice);
4097 p->prio = effective_prio(p);
4098 delta = p->prio - old_prio;
4101 enqueue_task(rq, p, 0);
4104 * If the task increased its priority or is running and
4105 * lowered its priority, then reschedule its CPU:
4107 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4108 resched_task(rq->curr);
4111 task_rq_unlock(rq, &flags);
4113 EXPORT_SYMBOL(set_user_nice);
4116 * can_nice - check if a task can reduce its nice value
4120 int can_nice(const struct task_struct *p, const int nice)
4122 /* convert nice value [19,-20] to rlimit style value [1,40] */
4123 int nice_rlim = 20 - nice;
4125 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4126 capable(CAP_SYS_NICE));
4129 #ifdef __ARCH_WANT_SYS_NICE
4132 * sys_nice - change the priority of the current process.
4133 * @increment: priority increment
4135 * sys_setpriority is a more generic, but much slower function that
4136 * does similar things.
4138 asmlinkage long sys_nice(int increment)
4143 * Setpriority might change our priority at the same moment.
4144 * We don't have to worry. Conceptually one call occurs first
4145 * and we have a single winner.
4147 if (increment < -40)
4152 nice = PRIO_TO_NICE(current->static_prio) + increment;
4158 if (increment < 0 && !can_nice(current, nice))
4161 retval = security_task_setnice(current, nice);
4165 set_user_nice(current, nice);
4172 * task_prio - return the priority value of a given task.
4173 * @p: the task in question.
4175 * This is the priority value as seen by users in /proc.
4176 * RT tasks are offset by -200. Normal tasks are centered
4177 * around 0, value goes from -16 to +15.
4179 int task_prio(const struct task_struct *p)
4181 return p->prio - MAX_RT_PRIO;
4185 * task_nice - return the nice value of a given task.
4186 * @p: the task in question.
4188 int task_nice(const struct task_struct *p)
4190 return TASK_NICE(p);
4192 EXPORT_SYMBOL_GPL(task_nice);
4195 * idle_cpu - is a given cpu idle currently?
4196 * @cpu: the processor in question.
4198 int idle_cpu(int cpu)
4200 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4204 * idle_task - return the idle task for a given cpu.
4205 * @cpu: the processor in question.
4207 struct task_struct *idle_task(int cpu)
4209 return cpu_rq(cpu)->idle;
4213 * find_process_by_pid - find a process with a matching PID value.
4214 * @pid: the pid in question.
4216 static struct task_struct *find_process_by_pid(pid_t pid)
4218 return pid ? find_task_by_vpid(pid) : current;
4221 /* Actually do priority change: must hold rq lock. */
4223 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4225 BUG_ON(p->se.on_rq);
4228 switch (p->policy) {
4232 p->sched_class = &fair_sched_class;
4236 p->sched_class = &rt_sched_class;
4240 p->rt_priority = prio;
4241 p->normal_prio = normal_prio(p);
4242 /* we are holding p->pi_lock already */
4243 p->prio = rt_mutex_getprio(p);
4248 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4249 * @p: the task in question.
4250 * @policy: new policy.
4251 * @param: structure containing the new RT priority.
4253 * NOTE that the task may be already dead.
4255 int sched_setscheduler(struct task_struct *p, int policy,
4256 struct sched_param *param)
4258 int retval, oldprio, oldpolicy = -1, on_rq, running;
4259 unsigned long flags;
4262 /* may grab non-irq protected spin_locks */
4263 BUG_ON(in_interrupt());
4265 /* double check policy once rq lock held */
4267 policy = oldpolicy = p->policy;
4268 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4269 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4270 policy != SCHED_IDLE)
4273 * Valid priorities for SCHED_FIFO and SCHED_RR are
4274 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4275 * SCHED_BATCH and SCHED_IDLE is 0.
4277 if (param->sched_priority < 0 ||
4278 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4279 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4281 if (rt_policy(policy) != (param->sched_priority != 0))
4285 * Allow unprivileged RT tasks to decrease priority:
4287 if (!capable(CAP_SYS_NICE)) {
4288 if (rt_policy(policy)) {
4289 unsigned long rlim_rtprio;
4291 if (!lock_task_sighand(p, &flags))
4293 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4294 unlock_task_sighand(p, &flags);
4296 /* can't set/change the rt policy */
4297 if (policy != p->policy && !rlim_rtprio)
4300 /* can't increase priority */
4301 if (param->sched_priority > p->rt_priority &&
4302 param->sched_priority > rlim_rtprio)
4306 * Like positive nice levels, dont allow tasks to
4307 * move out of SCHED_IDLE either:
4309 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4312 /* can't change other user's priorities */
4313 if ((current->euid != p->euid) &&
4314 (current->euid != p->uid))
4318 retval = security_task_setscheduler(p, policy, param);
4322 * make sure no PI-waiters arrive (or leave) while we are
4323 * changing the priority of the task:
4325 spin_lock_irqsave(&p->pi_lock, flags);
4327 * To be able to change p->policy safely, the apropriate
4328 * runqueue lock must be held.
4330 rq = __task_rq_lock(p);
4331 /* recheck policy now with rq lock held */
4332 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4333 policy = oldpolicy = -1;
4334 __task_rq_unlock(rq);
4335 spin_unlock_irqrestore(&p->pi_lock, flags);
4338 update_rq_clock(rq);
4339 on_rq = p->se.on_rq;
4340 running = task_current(rq, p);
4342 deactivate_task(rq, p, 0);
4344 p->sched_class->put_prev_task(rq, p);
4348 __setscheduler(rq, p, policy, param->sched_priority);
4352 p->sched_class->set_curr_task(rq);
4353 activate_task(rq, p, 0);
4355 * Reschedule if we are currently running on this runqueue and
4356 * our priority decreased, or if we are not currently running on
4357 * this runqueue and our priority is higher than the current's
4360 if (p->prio > oldprio)
4361 resched_task(rq->curr);
4363 check_preempt_curr(rq, p);
4366 __task_rq_unlock(rq);
4367 spin_unlock_irqrestore(&p->pi_lock, flags);
4369 rt_mutex_adjust_pi(p);
4373 EXPORT_SYMBOL_GPL(sched_setscheduler);
4376 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4378 struct sched_param lparam;
4379 struct task_struct *p;
4382 if (!param || pid < 0)
4384 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4389 p = find_process_by_pid(pid);
4391 retval = sched_setscheduler(p, policy, &lparam);
4398 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4399 * @pid: the pid in question.
4400 * @policy: new policy.
4401 * @param: structure containing the new RT priority.
4404 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4406 /* negative values for policy are not valid */
4410 return do_sched_setscheduler(pid, policy, param);
4414 * sys_sched_setparam - set/change the RT priority of a thread
4415 * @pid: the pid in question.
4416 * @param: structure containing the new RT priority.
4418 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4420 return do_sched_setscheduler(pid, -1, param);
4424 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4425 * @pid: the pid in question.
4427 asmlinkage long sys_sched_getscheduler(pid_t pid)
4429 struct task_struct *p;
4436 read_lock(&tasklist_lock);
4437 p = find_process_by_pid(pid);
4439 retval = security_task_getscheduler(p);
4443 read_unlock(&tasklist_lock);
4448 * sys_sched_getscheduler - get the RT priority of a thread
4449 * @pid: the pid in question.
4450 * @param: structure containing the RT priority.
4452 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4454 struct sched_param lp;
4455 struct task_struct *p;
4458 if (!param || pid < 0)
4461 read_lock(&tasklist_lock);
4462 p = find_process_by_pid(pid);
4467 retval = security_task_getscheduler(p);
4471 lp.sched_priority = p->rt_priority;
4472 read_unlock(&tasklist_lock);
4475 * This one might sleep, we cannot do it with a spinlock held ...
4477 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4482 read_unlock(&tasklist_lock);
4486 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4488 cpumask_t cpus_allowed;
4489 struct task_struct *p;
4492 mutex_lock(&sched_hotcpu_mutex);
4493 read_lock(&tasklist_lock);
4495 p = find_process_by_pid(pid);
4497 read_unlock(&tasklist_lock);
4498 mutex_unlock(&sched_hotcpu_mutex);
4503 * It is not safe to call set_cpus_allowed with the
4504 * tasklist_lock held. We will bump the task_struct's
4505 * usage count and then drop tasklist_lock.
4508 read_unlock(&tasklist_lock);
4511 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4512 !capable(CAP_SYS_NICE))
4515 retval = security_task_setscheduler(p, 0, NULL);
4519 cpus_allowed = cpuset_cpus_allowed(p);
4520 cpus_and(new_mask, new_mask, cpus_allowed);
4522 retval = set_cpus_allowed(p, new_mask);
4525 cpus_allowed = cpuset_cpus_allowed(p);
4526 if (!cpus_subset(new_mask, cpus_allowed)) {
4528 * We must have raced with a concurrent cpuset
4529 * update. Just reset the cpus_allowed to the
4530 * cpuset's cpus_allowed
4532 new_mask = cpus_allowed;
4538 mutex_unlock(&sched_hotcpu_mutex);
4542 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4543 cpumask_t *new_mask)
4545 if (len < sizeof(cpumask_t)) {
4546 memset(new_mask, 0, sizeof(cpumask_t));
4547 } else if (len > sizeof(cpumask_t)) {
4548 len = sizeof(cpumask_t);
4550 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4554 * sys_sched_setaffinity - set the cpu affinity of a process
4555 * @pid: pid of the process
4556 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4557 * @user_mask_ptr: user-space pointer to the new cpu mask
4559 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4560 unsigned long __user *user_mask_ptr)
4565 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4569 return sched_setaffinity(pid, new_mask);
4573 * Represents all cpu's present in the system
4574 * In systems capable of hotplug, this map could dynamically grow
4575 * as new cpu's are detected in the system via any platform specific
4576 * method, such as ACPI for e.g.
4579 cpumask_t cpu_present_map __read_mostly;
4580 EXPORT_SYMBOL(cpu_present_map);
4583 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4584 EXPORT_SYMBOL(cpu_online_map);
4586 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4587 EXPORT_SYMBOL(cpu_possible_map);
4590 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4592 struct task_struct *p;
4595 mutex_lock(&sched_hotcpu_mutex);
4596 read_lock(&tasklist_lock);
4599 p = find_process_by_pid(pid);
4603 retval = security_task_getscheduler(p);
4607 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4610 read_unlock(&tasklist_lock);
4611 mutex_unlock(&sched_hotcpu_mutex);
4617 * sys_sched_getaffinity - get the cpu affinity of a process
4618 * @pid: pid of the process
4619 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4620 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4622 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4623 unsigned long __user *user_mask_ptr)
4628 if (len < sizeof(cpumask_t))
4631 ret = sched_getaffinity(pid, &mask);
4635 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4638 return sizeof(cpumask_t);
4642 * sys_sched_yield - yield the current processor to other threads.
4644 * This function yields the current CPU to other tasks. If there are no
4645 * other threads running on this CPU then this function will return.
4647 asmlinkage long sys_sched_yield(void)
4649 struct rq *rq = this_rq_lock();
4651 schedstat_inc(rq, yld_count);
4652 current->sched_class->yield_task(rq);
4655 * Since we are going to call schedule() anyway, there's
4656 * no need to preempt or enable interrupts:
4658 __release(rq->lock);
4659 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4660 _raw_spin_unlock(&rq->lock);
4661 preempt_enable_no_resched();
4668 static void __cond_resched(void)
4670 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4671 __might_sleep(__FILE__, __LINE__);
4674 * The BKS might be reacquired before we have dropped
4675 * PREEMPT_ACTIVE, which could trigger a second
4676 * cond_resched() call.
4679 add_preempt_count(PREEMPT_ACTIVE);
4681 sub_preempt_count(PREEMPT_ACTIVE);
4682 } while (need_resched());
4685 int __sched cond_resched(void)
4687 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4688 system_state == SYSTEM_RUNNING) {
4694 EXPORT_SYMBOL(cond_resched);
4697 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4698 * call schedule, and on return reacquire the lock.
4700 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4701 * operations here to prevent schedule() from being called twice (once via
4702 * spin_unlock(), once by hand).
4704 int cond_resched_lock(spinlock_t *lock)
4708 if (need_lockbreak(lock)) {
4714 if (need_resched() && system_state == SYSTEM_RUNNING) {
4715 spin_release(&lock->dep_map, 1, _THIS_IP_);
4716 _raw_spin_unlock(lock);
4717 preempt_enable_no_resched();
4724 EXPORT_SYMBOL(cond_resched_lock);
4726 int __sched cond_resched_softirq(void)
4728 BUG_ON(!in_softirq());
4730 if (need_resched() && system_state == SYSTEM_RUNNING) {
4738 EXPORT_SYMBOL(cond_resched_softirq);
4741 * yield - yield the current processor to other threads.
4743 * This is a shortcut for kernel-space yielding - it marks the
4744 * thread runnable and calls sys_sched_yield().
4746 void __sched yield(void)
4748 set_current_state(TASK_RUNNING);
4751 EXPORT_SYMBOL(yield);
4754 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4755 * that process accounting knows that this is a task in IO wait state.
4757 * But don't do that if it is a deliberate, throttling IO wait (this task
4758 * has set its backing_dev_info: the queue against which it should throttle)
4760 void __sched io_schedule(void)
4762 struct rq *rq = &__raw_get_cpu_var(runqueues);
4764 delayacct_blkio_start();
4765 atomic_inc(&rq->nr_iowait);
4767 atomic_dec(&rq->nr_iowait);
4768 delayacct_blkio_end();
4770 EXPORT_SYMBOL(io_schedule);
4772 long __sched io_schedule_timeout(long timeout)
4774 struct rq *rq = &__raw_get_cpu_var(runqueues);
4777 delayacct_blkio_start();
4778 atomic_inc(&rq->nr_iowait);
4779 ret = schedule_timeout(timeout);
4780 atomic_dec(&rq->nr_iowait);
4781 delayacct_blkio_end();
4786 * sys_sched_get_priority_max - return maximum RT priority.
4787 * @policy: scheduling class.
4789 * this syscall returns the maximum rt_priority that can be used
4790 * by a given scheduling class.
4792 asmlinkage long sys_sched_get_priority_max(int policy)
4799 ret = MAX_USER_RT_PRIO-1;
4811 * sys_sched_get_priority_min - return minimum RT priority.
4812 * @policy: scheduling class.
4814 * this syscall returns the minimum rt_priority that can be used
4815 * by a given scheduling class.
4817 asmlinkage long sys_sched_get_priority_min(int policy)
4835 * sys_sched_rr_get_interval - return the default timeslice of a process.
4836 * @pid: pid of the process.
4837 * @interval: userspace pointer to the timeslice value.
4839 * this syscall writes the default timeslice value of a given process
4840 * into the user-space timespec buffer. A value of '0' means infinity.
4843 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4845 struct task_struct *p;
4846 unsigned int time_slice;
4854 read_lock(&tasklist_lock);
4855 p = find_process_by_pid(pid);
4859 retval = security_task_getscheduler(p);
4864 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4865 * tasks that are on an otherwise idle runqueue:
4868 if (p->policy == SCHED_RR) {
4869 time_slice = DEF_TIMESLICE;
4871 struct sched_entity *se = &p->se;
4872 unsigned long flags;
4875 rq = task_rq_lock(p, &flags);
4876 if (rq->cfs.load.weight)
4877 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4878 task_rq_unlock(rq, &flags);
4880 read_unlock(&tasklist_lock);
4881 jiffies_to_timespec(time_slice, &t);
4882 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4886 read_unlock(&tasklist_lock);
4890 static const char stat_nam[] = "RSDTtZX";
4892 static void show_task(struct task_struct *p)
4894 unsigned long free = 0;
4897 state = p->state ? __ffs(p->state) + 1 : 0;
4898 printk(KERN_INFO "%-13.13s %c", p->comm,
4899 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4900 #if BITS_PER_LONG == 32
4901 if (state == TASK_RUNNING)
4902 printk(KERN_CONT " running ");
4904 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4906 if (state == TASK_RUNNING)
4907 printk(KERN_CONT " running task ");
4909 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4911 #ifdef CONFIG_DEBUG_STACK_USAGE
4913 unsigned long *n = end_of_stack(p);
4916 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4919 printk(KERN_CONT "%5lu %5d %6d\n", free,
4920 task_pid_nr(p), task_pid_nr(p->parent));
4922 if (state != TASK_RUNNING)
4923 show_stack(p, NULL);
4926 void show_state_filter(unsigned long state_filter)
4928 struct task_struct *g, *p;
4930 #if BITS_PER_LONG == 32
4932 " task PC stack pid father\n");
4935 " task PC stack pid father\n");
4937 read_lock(&tasklist_lock);
4938 do_each_thread(g, p) {
4940 * reset the NMI-timeout, listing all files on a slow
4941 * console might take alot of time:
4943 touch_nmi_watchdog();
4944 if (!state_filter || (p->state & state_filter))
4946 } while_each_thread(g, p);
4948 touch_all_softlockup_watchdogs();
4950 #ifdef CONFIG_SCHED_DEBUG
4951 sysrq_sched_debug_show();
4953 read_unlock(&tasklist_lock);
4955 * Only show locks if all tasks are dumped:
4957 if (state_filter == -1)
4958 debug_show_all_locks();
4961 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4963 idle->sched_class = &idle_sched_class;
4967 * init_idle - set up an idle thread for a given CPU
4968 * @idle: task in question
4969 * @cpu: cpu the idle task belongs to
4971 * NOTE: this function does not set the idle thread's NEED_RESCHED
4972 * flag, to make booting more robust.
4974 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4976 struct rq *rq = cpu_rq(cpu);
4977 unsigned long flags;
4980 idle->se.exec_start = sched_clock();
4982 idle->prio = idle->normal_prio = MAX_PRIO;
4983 idle->cpus_allowed = cpumask_of_cpu(cpu);
4984 __set_task_cpu(idle, cpu);
4986 spin_lock_irqsave(&rq->lock, flags);
4987 rq->curr = rq->idle = idle;
4988 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4991 spin_unlock_irqrestore(&rq->lock, flags);
4993 /* Set the preempt count _outside_ the spinlocks! */
4994 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4995 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4997 task_thread_info(idle)->preempt_count = 0;
5000 * The idle tasks have their own, simple scheduling class:
5002 idle->sched_class = &idle_sched_class;
5006 * In a system that switches off the HZ timer nohz_cpu_mask
5007 * indicates which cpus entered this state. This is used
5008 * in the rcu update to wait only for active cpus. For system
5009 * which do not switch off the HZ timer nohz_cpu_mask should
5010 * always be CPU_MASK_NONE.
5012 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5015 * Increase the granularity value when there are more CPUs,
5016 * because with more CPUs the 'effective latency' as visible
5017 * to users decreases. But the relationship is not linear,
5018 * so pick a second-best guess by going with the log2 of the
5021 * This idea comes from the SD scheduler of Con Kolivas:
5023 static inline void sched_init_granularity(void)
5025 unsigned int factor = 1 + ilog2(num_online_cpus());
5026 const unsigned long limit = 200000000;
5028 sysctl_sched_min_granularity *= factor;
5029 if (sysctl_sched_min_granularity > limit)
5030 sysctl_sched_min_granularity = limit;
5032 sysctl_sched_latency *= factor;
5033 if (sysctl_sched_latency > limit)
5034 sysctl_sched_latency = limit;
5036 sysctl_sched_wakeup_granularity *= factor;
5037 sysctl_sched_batch_wakeup_granularity *= factor;
5042 * This is how migration works:
5044 * 1) we queue a struct migration_req structure in the source CPU's
5045 * runqueue and wake up that CPU's migration thread.
5046 * 2) we down() the locked semaphore => thread blocks.
5047 * 3) migration thread wakes up (implicitly it forces the migrated
5048 * thread off the CPU)
5049 * 4) it gets the migration request and checks whether the migrated
5050 * task is still in the wrong runqueue.
5051 * 5) if it's in the wrong runqueue then the migration thread removes
5052 * it and puts it into the right queue.
5053 * 6) migration thread up()s the semaphore.
5054 * 7) we wake up and the migration is done.
5058 * Change a given task's CPU affinity. Migrate the thread to a
5059 * proper CPU and schedule it away if the CPU it's executing on
5060 * is removed from the allowed bitmask.
5062 * NOTE: the caller must have a valid reference to the task, the
5063 * task must not exit() & deallocate itself prematurely. The
5064 * call is not atomic; no spinlocks may be held.
5066 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5068 struct migration_req req;
5069 unsigned long flags;
5073 rq = task_rq_lock(p, &flags);
5074 if (!cpus_intersects(new_mask, cpu_online_map)) {
5079 p->cpus_allowed = new_mask;
5080 /* Can the task run on the task's current CPU? If so, we're done */
5081 if (cpu_isset(task_cpu(p), new_mask))
5084 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5085 /* Need help from migration thread: drop lock and wait. */
5086 task_rq_unlock(rq, &flags);
5087 wake_up_process(rq->migration_thread);
5088 wait_for_completion(&req.done);
5089 tlb_migrate_finish(p->mm);
5093 task_rq_unlock(rq, &flags);
5097 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5100 * Move (not current) task off this cpu, onto dest cpu. We're doing
5101 * this because either it can't run here any more (set_cpus_allowed()
5102 * away from this CPU, or CPU going down), or because we're
5103 * attempting to rebalance this task on exec (sched_exec).
5105 * So we race with normal scheduler movements, but that's OK, as long
5106 * as the task is no longer on this CPU.
5108 * Returns non-zero if task was successfully migrated.
5110 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5112 struct rq *rq_dest, *rq_src;
5115 if (unlikely(cpu_is_offline(dest_cpu)))
5118 rq_src = cpu_rq(src_cpu);
5119 rq_dest = cpu_rq(dest_cpu);
5121 double_rq_lock(rq_src, rq_dest);
5122 /* Already moved. */
5123 if (task_cpu(p) != src_cpu)
5125 /* Affinity changed (again). */
5126 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5129 on_rq = p->se.on_rq;
5131 deactivate_task(rq_src, p, 0);
5133 set_task_cpu(p, dest_cpu);
5135 activate_task(rq_dest, p, 0);
5136 check_preempt_curr(rq_dest, p);
5140 double_rq_unlock(rq_src, rq_dest);
5145 * migration_thread - this is a highprio system thread that performs
5146 * thread migration by bumping thread off CPU then 'pushing' onto
5149 static int migration_thread(void *data)
5151 int cpu = (long)data;
5155 BUG_ON(rq->migration_thread != current);
5157 set_current_state(TASK_INTERRUPTIBLE);
5158 while (!kthread_should_stop()) {
5159 struct migration_req *req;
5160 struct list_head *head;
5162 spin_lock_irq(&rq->lock);
5164 if (cpu_is_offline(cpu)) {
5165 spin_unlock_irq(&rq->lock);
5169 if (rq->active_balance) {
5170 active_load_balance(rq, cpu);
5171 rq->active_balance = 0;
5174 head = &rq->migration_queue;
5176 if (list_empty(head)) {
5177 spin_unlock_irq(&rq->lock);
5179 set_current_state(TASK_INTERRUPTIBLE);
5182 req = list_entry(head->next, struct migration_req, list);
5183 list_del_init(head->next);
5185 spin_unlock(&rq->lock);
5186 __migrate_task(req->task, cpu, req->dest_cpu);
5189 complete(&req->done);
5191 __set_current_state(TASK_RUNNING);
5195 /* Wait for kthread_stop */
5196 set_current_state(TASK_INTERRUPTIBLE);
5197 while (!kthread_should_stop()) {
5199 set_current_state(TASK_INTERRUPTIBLE);
5201 __set_current_state(TASK_RUNNING);
5205 #ifdef CONFIG_HOTPLUG_CPU
5207 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5211 local_irq_disable();
5212 ret = __migrate_task(p, src_cpu, dest_cpu);
5218 * Figure out where task on dead CPU should go, use force if necessary.
5219 * NOTE: interrupts should be disabled by the caller
5221 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5223 unsigned long flags;
5230 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5231 cpus_and(mask, mask, p->cpus_allowed);
5232 dest_cpu = any_online_cpu(mask);
5234 /* On any allowed CPU? */
5235 if (dest_cpu == NR_CPUS)
5236 dest_cpu = any_online_cpu(p->cpus_allowed);
5238 /* No more Mr. Nice Guy. */
5239 if (dest_cpu == NR_CPUS) {
5240 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5242 * Try to stay on the same cpuset, where the
5243 * current cpuset may be a subset of all cpus.
5244 * The cpuset_cpus_allowed_locked() variant of
5245 * cpuset_cpus_allowed() will not block. It must be
5246 * called within calls to cpuset_lock/cpuset_unlock.
5248 rq = task_rq_lock(p, &flags);
5249 p->cpus_allowed = cpus_allowed;
5250 dest_cpu = any_online_cpu(p->cpus_allowed);
5251 task_rq_unlock(rq, &flags);
5254 * Don't tell them about moving exiting tasks or
5255 * kernel threads (both mm NULL), since they never
5258 if (p->mm && printk_ratelimit()) {
5259 printk(KERN_INFO "process %d (%s) no "
5260 "longer affine to cpu%d\n",
5261 task_pid_nr(p), p->comm, dead_cpu);
5264 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5268 * While a dead CPU has no uninterruptible tasks queued at this point,
5269 * it might still have a nonzero ->nr_uninterruptible counter, because
5270 * for performance reasons the counter is not stricly tracking tasks to
5271 * their home CPUs. So we just add the counter to another CPU's counter,
5272 * to keep the global sum constant after CPU-down:
5274 static void migrate_nr_uninterruptible(struct rq *rq_src)
5276 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5277 unsigned long flags;
5279 local_irq_save(flags);
5280 double_rq_lock(rq_src, rq_dest);
5281 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5282 rq_src->nr_uninterruptible = 0;
5283 double_rq_unlock(rq_src, rq_dest);
5284 local_irq_restore(flags);
5287 /* Run through task list and migrate tasks from the dead cpu. */
5288 static void migrate_live_tasks(int src_cpu)
5290 struct task_struct *p, *t;
5292 read_lock(&tasklist_lock);
5294 do_each_thread(t, p) {
5298 if (task_cpu(p) == src_cpu)
5299 move_task_off_dead_cpu(src_cpu, p);
5300 } while_each_thread(t, p);
5302 read_unlock(&tasklist_lock);
5306 * Schedules idle task to be the next runnable task on current CPU.
5307 * It does so by boosting its priority to highest possible.
5308 * Used by CPU offline code.
5310 void sched_idle_next(void)
5312 int this_cpu = smp_processor_id();
5313 struct rq *rq = cpu_rq(this_cpu);
5314 struct task_struct *p = rq->idle;
5315 unsigned long flags;
5317 /* cpu has to be offline */
5318 BUG_ON(cpu_online(this_cpu));
5321 * Strictly not necessary since rest of the CPUs are stopped by now
5322 * and interrupts disabled on the current cpu.
5324 spin_lock_irqsave(&rq->lock, flags);
5326 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5328 update_rq_clock(rq);
5329 activate_task(rq, p, 0);
5331 spin_unlock_irqrestore(&rq->lock, flags);
5335 * Ensures that the idle task is using init_mm right before its cpu goes
5338 void idle_task_exit(void)
5340 struct mm_struct *mm = current->active_mm;
5342 BUG_ON(cpu_online(smp_processor_id()));
5345 switch_mm(mm, &init_mm, current);
5349 /* called under rq->lock with disabled interrupts */
5350 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5352 struct rq *rq = cpu_rq(dead_cpu);
5354 /* Must be exiting, otherwise would be on tasklist. */
5355 BUG_ON(!p->exit_state);
5357 /* Cannot have done final schedule yet: would have vanished. */
5358 BUG_ON(p->state == TASK_DEAD);
5363 * Drop lock around migration; if someone else moves it,
5364 * that's OK. No task can be added to this CPU, so iteration is
5367 spin_unlock_irq(&rq->lock);
5368 move_task_off_dead_cpu(dead_cpu, p);
5369 spin_lock_irq(&rq->lock);
5374 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5375 static void migrate_dead_tasks(unsigned int dead_cpu)
5377 struct rq *rq = cpu_rq(dead_cpu);
5378 struct task_struct *next;
5381 if (!rq->nr_running)
5383 update_rq_clock(rq);
5384 next = pick_next_task(rq, rq->curr);
5387 migrate_dead(dead_cpu, next);
5391 #endif /* CONFIG_HOTPLUG_CPU */
5393 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5395 static struct ctl_table sd_ctl_dir[] = {
5397 .procname = "sched_domain",
5403 static struct ctl_table sd_ctl_root[] = {
5405 .ctl_name = CTL_KERN,
5406 .procname = "kernel",
5408 .child = sd_ctl_dir,
5413 static struct ctl_table *sd_alloc_ctl_entry(int n)
5415 struct ctl_table *entry =
5416 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5421 static void sd_free_ctl_entry(struct ctl_table **tablep)
5423 struct ctl_table *entry;
5426 * In the intermediate directories, both the child directory and
5427 * procname are dynamically allocated and could fail but the mode
5428 * will always be set. In the lowest directory the names are
5429 * static strings and all have proc handlers.
5431 for (entry = *tablep; entry->mode; entry++) {
5433 sd_free_ctl_entry(&entry->child);
5434 if (entry->proc_handler == NULL)
5435 kfree(entry->procname);
5443 set_table_entry(struct ctl_table *entry,
5444 const char *procname, void *data, int maxlen,
5445 mode_t mode, proc_handler *proc_handler)
5447 entry->procname = procname;
5449 entry->maxlen = maxlen;
5451 entry->proc_handler = proc_handler;
5454 static struct ctl_table *
5455 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5457 struct ctl_table *table = sd_alloc_ctl_entry(12);
5462 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5463 sizeof(long), 0644, proc_doulongvec_minmax);
5464 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5465 sizeof(long), 0644, proc_doulongvec_minmax);
5466 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5467 sizeof(int), 0644, proc_dointvec_minmax);
5468 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5471 sizeof(int), 0644, proc_dointvec_minmax);
5472 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5473 sizeof(int), 0644, proc_dointvec_minmax);
5474 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5475 sizeof(int), 0644, proc_dointvec_minmax);
5476 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5477 sizeof(int), 0644, proc_dointvec_minmax);
5478 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5479 sizeof(int), 0644, proc_dointvec_minmax);
5480 set_table_entry(&table[9], "cache_nice_tries",
5481 &sd->cache_nice_tries,
5482 sizeof(int), 0644, proc_dointvec_minmax);
5483 set_table_entry(&table[10], "flags", &sd->flags,
5484 sizeof(int), 0644, proc_dointvec_minmax);
5485 /* &table[11] is terminator */
5490 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5492 struct ctl_table *entry, *table;
5493 struct sched_domain *sd;
5494 int domain_num = 0, i;
5497 for_each_domain(cpu, sd)
5499 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5504 for_each_domain(cpu, sd) {
5505 snprintf(buf, 32, "domain%d", i);
5506 entry->procname = kstrdup(buf, GFP_KERNEL);
5508 entry->child = sd_alloc_ctl_domain_table(sd);
5515 static struct ctl_table_header *sd_sysctl_header;
5516 static void register_sched_domain_sysctl(void)
5518 int i, cpu_num = num_online_cpus();
5519 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5522 WARN_ON(sd_ctl_dir[0].child);
5523 sd_ctl_dir[0].child = entry;
5528 for_each_online_cpu(i) {
5529 snprintf(buf, 32, "cpu%d", i);
5530 entry->procname = kstrdup(buf, GFP_KERNEL);
5532 entry->child = sd_alloc_ctl_cpu_table(i);
5536 WARN_ON(sd_sysctl_header);
5537 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5540 /* may be called multiple times per register */
5541 static void unregister_sched_domain_sysctl(void)
5543 if (sd_sysctl_header)
5544 unregister_sysctl_table(sd_sysctl_header);
5545 sd_sysctl_header = NULL;
5546 if (sd_ctl_dir[0].child)
5547 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5550 static void register_sched_domain_sysctl(void)
5553 static void unregister_sched_domain_sysctl(void)
5559 * migration_call - callback that gets triggered when a CPU is added.
5560 * Here we can start up the necessary migration thread for the new CPU.
5562 static int __cpuinit
5563 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5565 struct task_struct *p;
5566 int cpu = (long)hcpu;
5567 unsigned long flags;
5571 case CPU_LOCK_ACQUIRE:
5572 mutex_lock(&sched_hotcpu_mutex);
5575 case CPU_UP_PREPARE:
5576 case CPU_UP_PREPARE_FROZEN:
5577 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5580 kthread_bind(p, cpu);
5581 /* Must be high prio: stop_machine expects to yield to it. */
5582 rq = task_rq_lock(p, &flags);
5583 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5584 task_rq_unlock(rq, &flags);
5585 cpu_rq(cpu)->migration_thread = p;
5589 case CPU_ONLINE_FROZEN:
5590 /* Strictly unnecessary, as first user will wake it. */
5591 wake_up_process(cpu_rq(cpu)->migration_thread);
5594 #ifdef CONFIG_HOTPLUG_CPU
5595 case CPU_UP_CANCELED:
5596 case CPU_UP_CANCELED_FROZEN:
5597 if (!cpu_rq(cpu)->migration_thread)
5599 /* Unbind it from offline cpu so it can run. Fall thru. */
5600 kthread_bind(cpu_rq(cpu)->migration_thread,
5601 any_online_cpu(cpu_online_map));
5602 kthread_stop(cpu_rq(cpu)->migration_thread);
5603 cpu_rq(cpu)->migration_thread = NULL;
5607 case CPU_DEAD_FROZEN:
5608 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5609 migrate_live_tasks(cpu);
5611 kthread_stop(rq->migration_thread);
5612 rq->migration_thread = NULL;
5613 /* Idle task back to normal (off runqueue, low prio) */
5614 spin_lock_irq(&rq->lock);
5615 update_rq_clock(rq);
5616 deactivate_task(rq, rq->idle, 0);
5617 rq->idle->static_prio = MAX_PRIO;
5618 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5619 rq->idle->sched_class = &idle_sched_class;
5620 migrate_dead_tasks(cpu);
5621 spin_unlock_irq(&rq->lock);
5623 migrate_nr_uninterruptible(rq);
5624 BUG_ON(rq->nr_running != 0);
5627 * No need to migrate the tasks: it was best-effort if
5628 * they didn't take sched_hotcpu_mutex. Just wake up
5631 spin_lock_irq(&rq->lock);
5632 while (!list_empty(&rq->migration_queue)) {
5633 struct migration_req *req;
5635 req = list_entry(rq->migration_queue.next,
5636 struct migration_req, list);
5637 list_del_init(&req->list);
5638 complete(&req->done);
5640 spin_unlock_irq(&rq->lock);
5643 case CPU_LOCK_RELEASE:
5644 mutex_unlock(&sched_hotcpu_mutex);
5650 /* Register at highest priority so that task migration (migrate_all_tasks)
5651 * happens before everything else.
5653 static struct notifier_block __cpuinitdata migration_notifier = {
5654 .notifier_call = migration_call,
5658 void __init migration_init(void)
5660 void *cpu = (void *)(long)smp_processor_id();
5663 /* Start one for the boot CPU: */
5664 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5665 BUG_ON(err == NOTIFY_BAD);
5666 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5667 register_cpu_notifier(&migration_notifier);
5673 /* Number of possible processor ids */
5674 int nr_cpu_ids __read_mostly = NR_CPUS;
5675 EXPORT_SYMBOL(nr_cpu_ids);
5677 #ifdef CONFIG_SCHED_DEBUG
5679 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5681 struct sched_group *group = sd->groups;
5682 cpumask_t groupmask;
5685 cpumask_scnprintf(str, NR_CPUS, sd->span);
5686 cpus_clear(groupmask);
5688 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5690 if (!(sd->flags & SD_LOAD_BALANCE)) {
5691 printk("does not load-balance\n");
5693 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5698 printk(KERN_CONT "span %s\n", str);
5700 if (!cpu_isset(cpu, sd->span)) {
5701 printk(KERN_ERR "ERROR: domain->span does not contain "
5704 if (!cpu_isset(cpu, group->cpumask)) {
5705 printk(KERN_ERR "ERROR: domain->groups does not contain"
5709 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5713 printk(KERN_ERR "ERROR: group is NULL\n");
5717 if (!group->__cpu_power) {
5718 printk(KERN_CONT "\n");
5719 printk(KERN_ERR "ERROR: domain->cpu_power not "
5724 if (!cpus_weight(group->cpumask)) {
5725 printk(KERN_CONT "\n");
5726 printk(KERN_ERR "ERROR: empty group\n");
5730 if (cpus_intersects(groupmask, group->cpumask)) {
5731 printk(KERN_CONT "\n");
5732 printk(KERN_ERR "ERROR: repeated CPUs\n");
5736 cpus_or(groupmask, groupmask, group->cpumask);
5738 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5739 printk(KERN_CONT " %s", str);
5741 group = group->next;
5742 } while (group != sd->groups);
5743 printk(KERN_CONT "\n");
5745 if (!cpus_equal(sd->span, groupmask))
5746 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5748 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5749 printk(KERN_ERR "ERROR: parent span is not a superset "
5750 "of domain->span\n");
5754 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5759 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5763 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5766 if (sched_domain_debug_one(sd, cpu, level))
5775 # define sched_domain_debug(sd, cpu) do { } while (0)
5778 static int sd_degenerate(struct sched_domain *sd)
5780 if (cpus_weight(sd->span) == 1)
5783 /* Following flags need at least 2 groups */
5784 if (sd->flags & (SD_LOAD_BALANCE |
5785 SD_BALANCE_NEWIDLE |
5789 SD_SHARE_PKG_RESOURCES)) {
5790 if (sd->groups != sd->groups->next)
5794 /* Following flags don't use groups */
5795 if (sd->flags & (SD_WAKE_IDLE |
5804 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5806 unsigned long cflags = sd->flags, pflags = parent->flags;
5808 if (sd_degenerate(parent))
5811 if (!cpus_equal(sd->span, parent->span))
5814 /* Does parent contain flags not in child? */
5815 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5816 if (cflags & SD_WAKE_AFFINE)
5817 pflags &= ~SD_WAKE_BALANCE;
5818 /* Flags needing groups don't count if only 1 group in parent */
5819 if (parent->groups == parent->groups->next) {
5820 pflags &= ~(SD_LOAD_BALANCE |
5821 SD_BALANCE_NEWIDLE |
5825 SD_SHARE_PKG_RESOURCES);
5827 if (~cflags & pflags)
5834 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5835 * hold the hotplug lock.
5837 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5839 struct rq *rq = cpu_rq(cpu);
5840 struct sched_domain *tmp;
5842 /* Remove the sched domains which do not contribute to scheduling. */
5843 for (tmp = sd; tmp; tmp = tmp->parent) {
5844 struct sched_domain *parent = tmp->parent;
5847 if (sd_parent_degenerate(tmp, parent)) {
5848 tmp->parent = parent->parent;
5850 parent->parent->child = tmp;
5854 if (sd && sd_degenerate(sd)) {
5860 sched_domain_debug(sd, cpu);
5862 rcu_assign_pointer(rq->sd, sd);
5865 /* cpus with isolated domains */
5866 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5868 /* Setup the mask of cpus configured for isolated domains */
5869 static int __init isolated_cpu_setup(char *str)
5871 int ints[NR_CPUS], i;
5873 str = get_options(str, ARRAY_SIZE(ints), ints);
5874 cpus_clear(cpu_isolated_map);
5875 for (i = 1; i <= ints[0]; i++)
5876 if (ints[i] < NR_CPUS)
5877 cpu_set(ints[i], cpu_isolated_map);
5881 __setup("isolcpus=", isolated_cpu_setup);
5884 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5885 * to a function which identifies what group(along with sched group) a CPU
5886 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5887 * (due to the fact that we keep track of groups covered with a cpumask_t).
5889 * init_sched_build_groups will build a circular linked list of the groups
5890 * covered by the given span, and will set each group's ->cpumask correctly,
5891 * and ->cpu_power to 0.
5894 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5895 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5896 struct sched_group **sg))
5898 struct sched_group *first = NULL, *last = NULL;
5899 cpumask_t covered = CPU_MASK_NONE;
5902 for_each_cpu_mask(i, span) {
5903 struct sched_group *sg;
5904 int group = group_fn(i, cpu_map, &sg);
5907 if (cpu_isset(i, covered))
5910 sg->cpumask = CPU_MASK_NONE;
5911 sg->__cpu_power = 0;
5913 for_each_cpu_mask(j, span) {
5914 if (group_fn(j, cpu_map, NULL) != group)
5917 cpu_set(j, covered);
5918 cpu_set(j, sg->cpumask);
5929 #define SD_NODES_PER_DOMAIN 16
5934 * find_next_best_node - find the next node to include in a sched_domain
5935 * @node: node whose sched_domain we're building
5936 * @used_nodes: nodes already in the sched_domain
5938 * Find the next node to include in a given scheduling domain. Simply
5939 * finds the closest node not already in the @used_nodes map.
5941 * Should use nodemask_t.
5943 static int find_next_best_node(int node, unsigned long *used_nodes)
5945 int i, n, val, min_val, best_node = 0;
5949 for (i = 0; i < MAX_NUMNODES; i++) {
5950 /* Start at @node */
5951 n = (node + i) % MAX_NUMNODES;
5953 if (!nr_cpus_node(n))
5956 /* Skip already used nodes */
5957 if (test_bit(n, used_nodes))
5960 /* Simple min distance search */
5961 val = node_distance(node, n);
5963 if (val < min_val) {
5969 set_bit(best_node, used_nodes);
5974 * sched_domain_node_span - get a cpumask for a node's sched_domain
5975 * @node: node whose cpumask we're constructing
5976 * @size: number of nodes to include in this span
5978 * Given a node, construct a good cpumask for its sched_domain to span. It
5979 * should be one that prevents unnecessary balancing, but also spreads tasks
5982 static cpumask_t sched_domain_node_span(int node)
5984 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5985 cpumask_t span, nodemask;
5989 bitmap_zero(used_nodes, MAX_NUMNODES);
5991 nodemask = node_to_cpumask(node);
5992 cpus_or(span, span, nodemask);
5993 set_bit(node, used_nodes);
5995 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5996 int next_node = find_next_best_node(node, used_nodes);
5998 nodemask = node_to_cpumask(next_node);
5999 cpus_or(span, span, nodemask);
6006 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6009 * SMT sched-domains:
6011 #ifdef CONFIG_SCHED_SMT
6012 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6013 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6016 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6019 *sg = &per_cpu(sched_group_cpus, cpu);
6025 * multi-core sched-domains:
6027 #ifdef CONFIG_SCHED_MC
6028 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6029 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6032 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6034 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6037 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6038 cpus_and(mask, mask, *cpu_map);
6039 group = first_cpu(mask);
6041 *sg = &per_cpu(sched_group_core, group);
6044 #elif defined(CONFIG_SCHED_MC)
6046 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6049 *sg = &per_cpu(sched_group_core, cpu);
6054 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6055 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6058 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6061 #ifdef CONFIG_SCHED_MC
6062 cpumask_t mask = cpu_coregroup_map(cpu);
6063 cpus_and(mask, mask, *cpu_map);
6064 group = first_cpu(mask);
6065 #elif defined(CONFIG_SCHED_SMT)
6066 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6067 cpus_and(mask, mask, *cpu_map);
6068 group = first_cpu(mask);
6073 *sg = &per_cpu(sched_group_phys, group);
6079 * The init_sched_build_groups can't handle what we want to do with node
6080 * groups, so roll our own. Now each node has its own list of groups which
6081 * gets dynamically allocated.
6083 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6084 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6086 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6087 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6089 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6090 struct sched_group **sg)
6092 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6095 cpus_and(nodemask, nodemask, *cpu_map);
6096 group = first_cpu(nodemask);
6099 *sg = &per_cpu(sched_group_allnodes, group);
6103 static void init_numa_sched_groups_power(struct sched_group *group_head)
6105 struct sched_group *sg = group_head;
6111 for_each_cpu_mask(j, sg->cpumask) {
6112 struct sched_domain *sd;
6114 sd = &per_cpu(phys_domains, j);
6115 if (j != first_cpu(sd->groups->cpumask)) {
6117 * Only add "power" once for each
6123 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6126 } while (sg != group_head);
6131 /* Free memory allocated for various sched_group structures */
6132 static void free_sched_groups(const cpumask_t *cpu_map)
6136 for_each_cpu_mask(cpu, *cpu_map) {
6137 struct sched_group **sched_group_nodes
6138 = sched_group_nodes_bycpu[cpu];
6140 if (!sched_group_nodes)
6143 for (i = 0; i < MAX_NUMNODES; i++) {
6144 cpumask_t nodemask = node_to_cpumask(i);
6145 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6147 cpus_and(nodemask, nodemask, *cpu_map);
6148 if (cpus_empty(nodemask))
6158 if (oldsg != sched_group_nodes[i])
6161 kfree(sched_group_nodes);
6162 sched_group_nodes_bycpu[cpu] = NULL;
6166 static void free_sched_groups(const cpumask_t *cpu_map)
6172 * Initialize sched groups cpu_power.
6174 * cpu_power indicates the capacity of sched group, which is used while
6175 * distributing the load between different sched groups in a sched domain.
6176 * Typically cpu_power for all the groups in a sched domain will be same unless
6177 * there are asymmetries in the topology. If there are asymmetries, group
6178 * having more cpu_power will pickup more load compared to the group having
6181 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6182 * the maximum number of tasks a group can handle in the presence of other idle
6183 * or lightly loaded groups in the same sched domain.
6185 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6187 struct sched_domain *child;
6188 struct sched_group *group;
6190 WARN_ON(!sd || !sd->groups);
6192 if (cpu != first_cpu(sd->groups->cpumask))
6197 sd->groups->__cpu_power = 0;
6200 * For perf policy, if the groups in child domain share resources
6201 * (for example cores sharing some portions of the cache hierarchy
6202 * or SMT), then set this domain groups cpu_power such that each group
6203 * can handle only one task, when there are other idle groups in the
6204 * same sched domain.
6206 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6208 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6209 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6214 * add cpu_power of each child group to this groups cpu_power
6216 group = child->groups;
6218 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6219 group = group->next;
6220 } while (group != child->groups);
6224 * Build sched domains for a given set of cpus and attach the sched domains
6225 * to the individual cpus
6227 static int build_sched_domains(const cpumask_t *cpu_map)
6231 struct sched_group **sched_group_nodes = NULL;
6232 int sd_allnodes = 0;
6235 * Allocate the per-node list of sched groups
6237 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6239 if (!sched_group_nodes) {
6240 printk(KERN_WARNING "Can not alloc sched group node list\n");
6243 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6247 * Set up domains for cpus specified by the cpu_map.
6249 for_each_cpu_mask(i, *cpu_map) {
6250 struct sched_domain *sd = NULL, *p;
6251 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6253 cpus_and(nodemask, nodemask, *cpu_map);
6256 if (cpus_weight(*cpu_map) >
6257 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6258 sd = &per_cpu(allnodes_domains, i);
6259 *sd = SD_ALLNODES_INIT;
6260 sd->span = *cpu_map;
6261 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6267 sd = &per_cpu(node_domains, i);
6269 sd->span = sched_domain_node_span(cpu_to_node(i));
6273 cpus_and(sd->span, sd->span, *cpu_map);
6277 sd = &per_cpu(phys_domains, i);
6279 sd->span = nodemask;
6283 cpu_to_phys_group(i, cpu_map, &sd->groups);
6285 #ifdef CONFIG_SCHED_MC
6287 sd = &per_cpu(core_domains, i);
6289 sd->span = cpu_coregroup_map(i);
6290 cpus_and(sd->span, sd->span, *cpu_map);
6293 cpu_to_core_group(i, cpu_map, &sd->groups);
6296 #ifdef CONFIG_SCHED_SMT
6298 sd = &per_cpu(cpu_domains, i);
6299 *sd = SD_SIBLING_INIT;
6300 sd->span = per_cpu(cpu_sibling_map, i);
6301 cpus_and(sd->span, sd->span, *cpu_map);
6304 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6308 #ifdef CONFIG_SCHED_SMT
6309 /* Set up CPU (sibling) groups */
6310 for_each_cpu_mask(i, *cpu_map) {
6311 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6312 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6313 if (i != first_cpu(this_sibling_map))
6316 init_sched_build_groups(this_sibling_map, cpu_map,
6321 #ifdef CONFIG_SCHED_MC
6322 /* Set up multi-core groups */
6323 for_each_cpu_mask(i, *cpu_map) {
6324 cpumask_t this_core_map = cpu_coregroup_map(i);
6325 cpus_and(this_core_map, this_core_map, *cpu_map);
6326 if (i != first_cpu(this_core_map))
6328 init_sched_build_groups(this_core_map, cpu_map,
6329 &cpu_to_core_group);
6333 /* Set up physical groups */
6334 for (i = 0; i < MAX_NUMNODES; i++) {
6335 cpumask_t nodemask = node_to_cpumask(i);
6337 cpus_and(nodemask, nodemask, *cpu_map);
6338 if (cpus_empty(nodemask))
6341 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6345 /* Set up node groups */
6347 init_sched_build_groups(*cpu_map, cpu_map,
6348 &cpu_to_allnodes_group);
6350 for (i = 0; i < MAX_NUMNODES; i++) {
6351 /* Set up node groups */
6352 struct sched_group *sg, *prev;
6353 cpumask_t nodemask = node_to_cpumask(i);
6354 cpumask_t domainspan;
6355 cpumask_t covered = CPU_MASK_NONE;
6358 cpus_and(nodemask, nodemask, *cpu_map);
6359 if (cpus_empty(nodemask)) {
6360 sched_group_nodes[i] = NULL;
6364 domainspan = sched_domain_node_span(i);
6365 cpus_and(domainspan, domainspan, *cpu_map);
6367 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6369 printk(KERN_WARNING "Can not alloc domain group for "
6373 sched_group_nodes[i] = sg;
6374 for_each_cpu_mask(j, nodemask) {
6375 struct sched_domain *sd;
6377 sd = &per_cpu(node_domains, j);
6380 sg->__cpu_power = 0;
6381 sg->cpumask = nodemask;
6383 cpus_or(covered, covered, nodemask);
6386 for (j = 0; j < MAX_NUMNODES; j++) {
6387 cpumask_t tmp, notcovered;
6388 int n = (i + j) % MAX_NUMNODES;
6390 cpus_complement(notcovered, covered);
6391 cpus_and(tmp, notcovered, *cpu_map);
6392 cpus_and(tmp, tmp, domainspan);
6393 if (cpus_empty(tmp))
6396 nodemask = node_to_cpumask(n);
6397 cpus_and(tmp, tmp, nodemask);
6398 if (cpus_empty(tmp))
6401 sg = kmalloc_node(sizeof(struct sched_group),
6405 "Can not alloc domain group for node %d\n", j);
6408 sg->__cpu_power = 0;
6410 sg->next = prev->next;
6411 cpus_or(covered, covered, tmp);
6418 /* Calculate CPU power for physical packages and nodes */
6419 #ifdef CONFIG_SCHED_SMT
6420 for_each_cpu_mask(i, *cpu_map) {
6421 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6423 init_sched_groups_power(i, sd);
6426 #ifdef CONFIG_SCHED_MC
6427 for_each_cpu_mask(i, *cpu_map) {
6428 struct sched_domain *sd = &per_cpu(core_domains, i);
6430 init_sched_groups_power(i, sd);
6434 for_each_cpu_mask(i, *cpu_map) {
6435 struct sched_domain *sd = &per_cpu(phys_domains, i);
6437 init_sched_groups_power(i, sd);
6441 for (i = 0; i < MAX_NUMNODES; i++)
6442 init_numa_sched_groups_power(sched_group_nodes[i]);
6445 struct sched_group *sg;
6447 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6448 init_numa_sched_groups_power(sg);
6452 /* Attach the domains */
6453 for_each_cpu_mask(i, *cpu_map) {
6454 struct sched_domain *sd;
6455 #ifdef CONFIG_SCHED_SMT
6456 sd = &per_cpu(cpu_domains, i);
6457 #elif defined(CONFIG_SCHED_MC)
6458 sd = &per_cpu(core_domains, i);
6460 sd = &per_cpu(phys_domains, i);
6462 cpu_attach_domain(sd, i);
6469 free_sched_groups(cpu_map);
6474 static cpumask_t *doms_cur; /* current sched domains */
6475 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6478 * Special case: If a kmalloc of a doms_cur partition (array of
6479 * cpumask_t) fails, then fallback to a single sched domain,
6480 * as determined by the single cpumask_t fallback_doms.
6482 static cpumask_t fallback_doms;
6485 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6486 * For now this just excludes isolated cpus, but could be used to
6487 * exclude other special cases in the future.
6489 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6494 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6496 doms_cur = &fallback_doms;
6497 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6498 err = build_sched_domains(doms_cur);
6499 register_sched_domain_sysctl();
6504 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6506 free_sched_groups(cpu_map);
6510 * Detach sched domains from a group of cpus specified in cpu_map
6511 * These cpus will now be attached to the NULL domain
6513 static void detach_destroy_domains(const cpumask_t *cpu_map)
6517 unregister_sched_domain_sysctl();
6519 for_each_cpu_mask(i, *cpu_map)
6520 cpu_attach_domain(NULL, i);
6521 synchronize_sched();
6522 arch_destroy_sched_domains(cpu_map);
6526 * Partition sched domains as specified by the 'ndoms_new'
6527 * cpumasks in the array doms_new[] of cpumasks. This compares
6528 * doms_new[] to the current sched domain partitioning, doms_cur[].
6529 * It destroys each deleted domain and builds each new domain.
6531 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6532 * The masks don't intersect (don't overlap.) We should setup one
6533 * sched domain for each mask. CPUs not in any of the cpumasks will
6534 * not be load balanced. If the same cpumask appears both in the
6535 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6538 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6539 * ownership of it and will kfree it when done with it. If the caller
6540 * failed the kmalloc call, then it can pass in doms_new == NULL,
6541 * and partition_sched_domains() will fallback to the single partition
6544 * Call with hotplug lock held
6546 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6550 /* always unregister in case we don't destroy any domains */
6551 unregister_sched_domain_sysctl();
6553 if (doms_new == NULL) {
6555 doms_new = &fallback_doms;
6556 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6559 /* Destroy deleted domains */
6560 for (i = 0; i < ndoms_cur; i++) {
6561 for (j = 0; j < ndoms_new; j++) {
6562 if (cpus_equal(doms_cur[i], doms_new[j]))
6565 /* no match - a current sched domain not in new doms_new[] */
6566 detach_destroy_domains(doms_cur + i);
6571 /* Build new domains */
6572 for (i = 0; i < ndoms_new; i++) {
6573 for (j = 0; j < ndoms_cur; j++) {
6574 if (cpus_equal(doms_new[i], doms_cur[j]))
6577 /* no match - add a new doms_new */
6578 build_sched_domains(doms_new + i);
6583 /* Remember the new sched domains */
6584 if (doms_cur != &fallback_doms)
6586 doms_cur = doms_new;
6587 ndoms_cur = ndoms_new;
6589 register_sched_domain_sysctl();
6592 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6593 static int arch_reinit_sched_domains(void)
6597 mutex_lock(&sched_hotcpu_mutex);
6598 detach_destroy_domains(&cpu_online_map);
6599 err = arch_init_sched_domains(&cpu_online_map);
6600 mutex_unlock(&sched_hotcpu_mutex);
6605 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6609 if (buf[0] != '0' && buf[0] != '1')
6613 sched_smt_power_savings = (buf[0] == '1');
6615 sched_mc_power_savings = (buf[0] == '1');
6617 ret = arch_reinit_sched_domains();
6619 return ret ? ret : count;
6622 #ifdef CONFIG_SCHED_MC
6623 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6625 return sprintf(page, "%u\n", sched_mc_power_savings);
6627 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6628 const char *buf, size_t count)
6630 return sched_power_savings_store(buf, count, 0);
6632 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6633 sched_mc_power_savings_store);
6636 #ifdef CONFIG_SCHED_SMT
6637 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6639 return sprintf(page, "%u\n", sched_smt_power_savings);
6641 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6642 const char *buf, size_t count)
6644 return sched_power_savings_store(buf, count, 1);
6646 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6647 sched_smt_power_savings_store);
6650 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6654 #ifdef CONFIG_SCHED_SMT
6656 err = sysfs_create_file(&cls->kset.kobj,
6657 &attr_sched_smt_power_savings.attr);
6659 #ifdef CONFIG_SCHED_MC
6660 if (!err && mc_capable())
6661 err = sysfs_create_file(&cls->kset.kobj,
6662 &attr_sched_mc_power_savings.attr);
6669 * Force a reinitialization of the sched domains hierarchy. The domains
6670 * and groups cannot be updated in place without racing with the balancing
6671 * code, so we temporarily attach all running cpus to the NULL domain
6672 * which will prevent rebalancing while the sched domains are recalculated.
6674 static int update_sched_domains(struct notifier_block *nfb,
6675 unsigned long action, void *hcpu)
6678 case CPU_UP_PREPARE:
6679 case CPU_UP_PREPARE_FROZEN:
6680 case CPU_DOWN_PREPARE:
6681 case CPU_DOWN_PREPARE_FROZEN:
6682 detach_destroy_domains(&cpu_online_map);
6685 case CPU_UP_CANCELED:
6686 case CPU_UP_CANCELED_FROZEN:
6687 case CPU_DOWN_FAILED:
6688 case CPU_DOWN_FAILED_FROZEN:
6690 case CPU_ONLINE_FROZEN:
6692 case CPU_DEAD_FROZEN:
6694 * Fall through and re-initialise the domains.
6701 /* The hotplug lock is already held by cpu_up/cpu_down */
6702 arch_init_sched_domains(&cpu_online_map);
6707 void __init sched_init_smp(void)
6709 cpumask_t non_isolated_cpus;
6711 mutex_lock(&sched_hotcpu_mutex);
6712 arch_init_sched_domains(&cpu_online_map);
6713 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6714 if (cpus_empty(non_isolated_cpus))
6715 cpu_set(smp_processor_id(), non_isolated_cpus);
6716 mutex_unlock(&sched_hotcpu_mutex);
6717 /* XXX: Theoretical race here - CPU may be hotplugged now */
6718 hotcpu_notifier(update_sched_domains, 0);
6720 /* Move init over to a non-isolated CPU */
6721 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6723 sched_init_granularity();
6726 void __init sched_init_smp(void)
6728 sched_init_granularity();
6730 #endif /* CONFIG_SMP */
6732 int in_sched_functions(unsigned long addr)
6734 return in_lock_functions(addr) ||
6735 (addr >= (unsigned long)__sched_text_start
6736 && addr < (unsigned long)__sched_text_end);
6739 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6741 cfs_rq->tasks_timeline = RB_ROOT;
6742 #ifdef CONFIG_FAIR_GROUP_SCHED
6745 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6748 void __init sched_init(void)
6750 int highest_cpu = 0;
6753 for_each_possible_cpu(i) {
6754 struct rt_prio_array *array;
6758 spin_lock_init(&rq->lock);
6759 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6762 init_cfs_rq(&rq->cfs, rq);
6763 #ifdef CONFIG_FAIR_GROUP_SCHED
6764 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6766 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6767 struct sched_entity *se =
6768 &per_cpu(init_sched_entity, i);
6770 init_cfs_rq_p[i] = cfs_rq;
6771 init_cfs_rq(cfs_rq, rq);
6772 cfs_rq->tg = &init_task_group;
6773 list_add(&cfs_rq->leaf_cfs_rq_list,
6774 &rq->leaf_cfs_rq_list);
6776 init_sched_entity_p[i] = se;
6777 se->cfs_rq = &rq->cfs;
6779 se->load.weight = init_task_group_load;
6780 se->load.inv_weight =
6781 div64_64(1ULL<<32, init_task_group_load);
6784 init_task_group.shares = init_task_group_load;
6785 spin_lock_init(&init_task_group.lock);
6788 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6789 rq->cpu_load[j] = 0;
6792 rq->active_balance = 0;
6793 rq->next_balance = jiffies;
6796 rq->migration_thread = NULL;
6797 INIT_LIST_HEAD(&rq->migration_queue);
6799 atomic_set(&rq->nr_iowait, 0);
6801 array = &rq->rt.active;
6802 for (j = 0; j < MAX_RT_PRIO; j++) {
6803 INIT_LIST_HEAD(array->queue + j);
6804 __clear_bit(j, array->bitmap);
6807 /* delimiter for bitsearch: */
6808 __set_bit(MAX_RT_PRIO, array->bitmap);
6811 set_load_weight(&init_task);
6813 #ifdef CONFIG_PREEMPT_NOTIFIERS
6814 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6818 nr_cpu_ids = highest_cpu + 1;
6819 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6822 #ifdef CONFIG_RT_MUTEXES
6823 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6827 * The boot idle thread does lazy MMU switching as well:
6829 atomic_inc(&init_mm.mm_count);
6830 enter_lazy_tlb(&init_mm, current);
6833 * Make us the idle thread. Technically, schedule() should not be
6834 * called from this thread, however somewhere below it might be,
6835 * but because we are the idle thread, we just pick up running again
6836 * when this runqueue becomes "idle".
6838 init_idle(current, smp_processor_id());
6840 * During early bootup we pretend to be a normal task:
6842 current->sched_class = &fair_sched_class;
6845 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6846 void __might_sleep(char *file, int line)
6849 static unsigned long prev_jiffy; /* ratelimiting */
6851 if ((in_atomic() || irqs_disabled()) &&
6852 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6853 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6855 prev_jiffy = jiffies;
6856 printk(KERN_ERR "BUG: sleeping function called from invalid"
6857 " context at %s:%d\n", file, line);
6858 printk("in_atomic():%d, irqs_disabled():%d\n",
6859 in_atomic(), irqs_disabled());
6860 debug_show_held_locks(current);
6861 if (irqs_disabled())
6862 print_irqtrace_events(current);
6867 EXPORT_SYMBOL(__might_sleep);
6870 #ifdef CONFIG_MAGIC_SYSRQ
6871 static void normalize_task(struct rq *rq, struct task_struct *p)
6874 update_rq_clock(rq);
6875 on_rq = p->se.on_rq;
6877 deactivate_task(rq, p, 0);
6878 __setscheduler(rq, p, SCHED_NORMAL, 0);
6880 activate_task(rq, p, 0);
6881 resched_task(rq->curr);
6885 void normalize_rt_tasks(void)
6887 struct task_struct *g, *p;
6888 unsigned long flags;
6891 read_lock_irq(&tasklist_lock);
6892 do_each_thread(g, p) {
6894 * Only normalize user tasks:
6899 p->se.exec_start = 0;
6900 #ifdef CONFIG_SCHEDSTATS
6901 p->se.wait_start = 0;
6902 p->se.sleep_start = 0;
6903 p->se.block_start = 0;
6905 task_rq(p)->clock = 0;
6909 * Renice negative nice level userspace
6912 if (TASK_NICE(p) < 0 && p->mm)
6913 set_user_nice(p, 0);
6917 spin_lock_irqsave(&p->pi_lock, flags);
6918 rq = __task_rq_lock(p);
6920 normalize_task(rq, p);
6922 __task_rq_unlock(rq);
6923 spin_unlock_irqrestore(&p->pi_lock, flags);
6924 } while_each_thread(g, p);
6926 read_unlock_irq(&tasklist_lock);
6929 #endif /* CONFIG_MAGIC_SYSRQ */
6933 * These functions are only useful for the IA64 MCA handling.
6935 * They can only be called when the whole system has been
6936 * stopped - every CPU needs to be quiescent, and no scheduling
6937 * activity can take place. Using them for anything else would
6938 * be a serious bug, and as a result, they aren't even visible
6939 * under any other configuration.
6943 * curr_task - return the current task for a given cpu.
6944 * @cpu: the processor in question.
6946 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6948 struct task_struct *curr_task(int cpu)
6950 return cpu_curr(cpu);
6954 * set_curr_task - set the current task for a given cpu.
6955 * @cpu: the processor in question.
6956 * @p: the task pointer to set.
6958 * Description: This function must only be used when non-maskable interrupts
6959 * are serviced on a separate stack. It allows the architecture to switch the
6960 * notion of the current task on a cpu in a non-blocking manner. This function
6961 * must be called with all CPU's synchronized, and interrupts disabled, the
6962 * and caller must save the original value of the current task (see
6963 * curr_task() above) and restore that value before reenabling interrupts and
6964 * re-starting the system.
6966 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6968 void set_curr_task(int cpu, struct task_struct *p)
6975 #ifdef CONFIG_FAIR_GROUP_SCHED
6977 /* allocate runqueue etc for a new task group */
6978 struct task_group *sched_create_group(void)
6980 struct task_group *tg;
6981 struct cfs_rq *cfs_rq;
6982 struct sched_entity *se;
6986 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6988 return ERR_PTR(-ENOMEM);
6990 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6993 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6997 for_each_possible_cpu(i) {
7000 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
7005 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
7010 memset(cfs_rq, 0, sizeof(struct cfs_rq));
7011 memset(se, 0, sizeof(struct sched_entity));
7013 tg->cfs_rq[i] = cfs_rq;
7014 init_cfs_rq(cfs_rq, rq);
7018 se->cfs_rq = &rq->cfs;
7020 se->load.weight = NICE_0_LOAD;
7021 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7025 for_each_possible_cpu(i) {
7027 cfs_rq = tg->cfs_rq[i];
7028 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7031 tg->shares = NICE_0_LOAD;
7032 spin_lock_init(&tg->lock);
7037 for_each_possible_cpu(i) {
7039 kfree(tg->cfs_rq[i]);
7047 return ERR_PTR(-ENOMEM);
7050 /* rcu callback to free various structures associated with a task group */
7051 static void free_sched_group(struct rcu_head *rhp)
7053 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7054 struct cfs_rq *cfs_rq;
7055 struct sched_entity *se;
7058 /* now it should be safe to free those cfs_rqs */
7059 for_each_possible_cpu(i) {
7060 cfs_rq = tg->cfs_rq[i];
7072 /* Destroy runqueue etc associated with a task group */
7073 void sched_destroy_group(struct task_group *tg)
7075 struct cfs_rq *cfs_rq = NULL;
7078 for_each_possible_cpu(i) {
7079 cfs_rq = tg->cfs_rq[i];
7080 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7085 /* wait for possible concurrent references to cfs_rqs complete */
7086 call_rcu(&tg->rcu, free_sched_group);
7089 /* change task's runqueue when it moves between groups.
7090 * The caller of this function should have put the task in its new group
7091 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7092 * reflect its new group.
7094 void sched_move_task(struct task_struct *tsk)
7097 unsigned long flags;
7100 rq = task_rq_lock(tsk, &flags);
7102 if (tsk->sched_class != &fair_sched_class) {
7103 set_task_cfs_rq(tsk, task_cpu(tsk));
7107 update_rq_clock(rq);
7109 running = task_current(rq, tsk);
7110 on_rq = tsk->se.on_rq;
7113 dequeue_task(rq, tsk, 0);
7114 if (unlikely(running))
7115 tsk->sched_class->put_prev_task(rq, tsk);
7118 set_task_cfs_rq(tsk, task_cpu(tsk));
7121 if (unlikely(running))
7122 tsk->sched_class->set_curr_task(rq);
7123 enqueue_task(rq, tsk, 0);
7127 task_rq_unlock(rq, &flags);
7130 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7132 struct cfs_rq *cfs_rq = se->cfs_rq;
7133 struct rq *rq = cfs_rq->rq;
7136 spin_lock_irq(&rq->lock);
7140 dequeue_entity(cfs_rq, se, 0);
7142 se->load.weight = shares;
7143 se->load.inv_weight = div64_64((1ULL<<32), shares);
7146 enqueue_entity(cfs_rq, se, 0);
7148 spin_unlock_irq(&rq->lock);
7151 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7155 spin_lock(&tg->lock);
7156 if (tg->shares == shares)
7159 tg->shares = shares;
7160 for_each_possible_cpu(i)
7161 set_se_shares(tg->se[i], shares);
7164 spin_unlock(&tg->lock);
7168 unsigned long sched_group_shares(struct task_group *tg)
7173 #endif /* CONFIG_FAIR_GROUP_SCHED */
7175 #ifdef CONFIG_FAIR_CGROUP_SCHED
7177 /* return corresponding task_group object of a cgroup */
7178 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7180 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7181 struct task_group, css);
7184 static struct cgroup_subsys_state *
7185 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7187 struct task_group *tg;
7189 if (!cgrp->parent) {
7190 /* This is early initialization for the top cgroup */
7191 init_task_group.css.cgroup = cgrp;
7192 return &init_task_group.css;
7195 /* we support only 1-level deep hierarchical scheduler atm */
7196 if (cgrp->parent->parent)
7197 return ERR_PTR(-EINVAL);
7199 tg = sched_create_group();
7201 return ERR_PTR(-ENOMEM);
7203 /* Bind the cgroup to task_group object we just created */
7204 tg->css.cgroup = cgrp;
7210 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7212 struct task_group *tg = cgroup_tg(cgrp);
7214 sched_destroy_group(tg);
7218 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7219 struct task_struct *tsk)
7221 /* We don't support RT-tasks being in separate groups */
7222 if (tsk->sched_class != &fair_sched_class)
7229 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7230 struct cgroup *old_cont, struct task_struct *tsk)
7232 sched_move_task(tsk);
7235 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7238 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7241 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7243 struct task_group *tg = cgroup_tg(cgrp);
7245 return (u64) tg->shares;
7248 static struct cftype cpu_files[] = {
7251 .read_uint = cpu_shares_read_uint,
7252 .write_uint = cpu_shares_write_uint,
7256 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7258 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7261 struct cgroup_subsys cpu_cgroup_subsys = {
7263 .create = cpu_cgroup_create,
7264 .destroy = cpu_cgroup_destroy,
7265 .can_attach = cpu_cgroup_can_attach,
7266 .attach = cpu_cgroup_attach,
7267 .populate = cpu_cgroup_populate,
7268 .subsys_id = cpu_cgroup_subsys_id,
7272 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7274 #ifdef CONFIG_CGROUP_CPUACCT
7277 * CPU accounting code for task groups.
7279 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7280 * (balbir@in.ibm.com).
7283 /* track cpu usage of a group of tasks */
7285 struct cgroup_subsys_state css;
7286 /* cpuusage holds pointer to a u64-type object on every cpu */
7290 struct cgroup_subsys cpuacct_subsys;
7292 /* return cpu accounting group corresponding to this container */
7293 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7295 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7296 struct cpuacct, css);
7299 /* return cpu accounting group to which this task belongs */
7300 static inline struct cpuacct *task_ca(struct task_struct *tsk)
7302 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7303 struct cpuacct, css);
7306 /* create a new cpu accounting group */
7307 static struct cgroup_subsys_state *cpuacct_create(
7308 struct cgroup_subsys *ss, struct cgroup *cont)
7310 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7313 return ERR_PTR(-ENOMEM);
7315 ca->cpuusage = alloc_percpu(u64);
7316 if (!ca->cpuusage) {
7318 return ERR_PTR(-ENOMEM);
7324 /* destroy an existing cpu accounting group */
7326 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
7328 struct cpuacct *ca = cgroup_ca(cont);
7330 free_percpu(ca->cpuusage);
7334 /* return total cpu usage (in nanoseconds) of a group */
7335 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
7337 struct cpuacct *ca = cgroup_ca(cont);
7338 u64 totalcpuusage = 0;
7341 for_each_possible_cpu(i) {
7342 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
7345 * Take rq->lock to make 64-bit addition safe on 32-bit
7348 spin_lock_irq(&cpu_rq(i)->lock);
7349 totalcpuusage += *cpuusage;
7350 spin_unlock_irq(&cpu_rq(i)->lock);
7353 return totalcpuusage;
7356 static struct cftype files[] = {
7359 .read_uint = cpuusage_read,
7363 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7365 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
7369 * charge this task's execution time to its accounting group.
7371 * called with rq->lock held.
7373 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
7377 if (!cpuacct_subsys.active)
7382 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
7384 *cpuusage += cputime;
7388 struct cgroup_subsys cpuacct_subsys = {
7390 .create = cpuacct_create,
7391 .destroy = cpuacct_destroy,
7392 .populate = cpuacct_populate,
7393 .subsys_id = cpuacct_subsys_id,
7395 #endif /* CONFIG_CGROUP_CPUACCT */