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;
173 * shares assigned to a task group governs how much of cpu bandwidth
174 * is allocated to the group. The more shares a group has, the more is
175 * the cpu bandwidth allocated to it.
177 * For ex, lets say that there are three task groups, A, B and C which
178 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
179 * cpu bandwidth allocated by the scheduler to task groups A, B and C
182 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
183 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
184 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
186 * The weight assigned to a task group's schedulable entities on every
187 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
188 * group's shares. For ex: lets say that task group A has been
189 * assigned shares of 1000 and there are two CPUs in a system. Then,
191 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
193 * Note: It's not necessary that each of a task's group schedulable
194 * entity have the same weight on all CPUs. If the group
195 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
196 * better distribution of weight could be:
198 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
199 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
201 * rebalance_shares() is responsible for distributing the shares of a
202 * task groups like this among the group's schedulable entities across
206 unsigned long shares;
211 /* Default task group's sched entity on each cpu */
212 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
213 /* Default task group's cfs_rq on each cpu */
214 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
216 static struct sched_entity *init_sched_entity_p[NR_CPUS];
217 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
219 /* task_group_mutex serializes add/remove of task groups and also changes to
220 * a task group's cpu shares.
222 static DEFINE_MUTEX(task_group_mutex);
224 /* doms_cur_mutex serializes access to doms_cur[] array */
225 static DEFINE_MUTEX(doms_cur_mutex);
228 /* kernel thread that runs rebalance_shares() periodically */
229 static struct task_struct *lb_monitor_task;
230 static int load_balance_monitor(void *unused);
233 static void set_se_shares(struct sched_entity *se, unsigned long shares);
235 /* Default task group.
236 * Every task in system belong to this group at bootup.
238 struct task_group init_task_group = {
239 .se = init_sched_entity_p,
240 .cfs_rq = init_cfs_rq_p,
243 #ifdef CONFIG_FAIR_USER_SCHED
244 # define INIT_TASK_GROUP_LOAD 2*NICE_0_LOAD
246 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
249 #define MIN_GROUP_SHARES 2
251 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
253 /* return group to which a task belongs */
254 static inline struct task_group *task_group(struct task_struct *p)
256 struct task_group *tg;
258 #ifdef CONFIG_FAIR_USER_SCHED
260 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
261 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
262 struct task_group, css);
264 tg = &init_task_group;
269 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
270 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu)
272 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
273 p->se.parent = task_group(p)->se[cpu];
276 static inline void lock_task_group_list(void)
278 mutex_lock(&task_group_mutex);
281 static inline void unlock_task_group_list(void)
283 mutex_unlock(&task_group_mutex);
286 static inline void lock_doms_cur(void)
288 mutex_lock(&doms_cur_mutex);
291 static inline void unlock_doms_cur(void)
293 mutex_unlock(&doms_cur_mutex);
298 static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu) { }
299 static inline void lock_task_group_list(void) { }
300 static inline void unlock_task_group_list(void) { }
301 static inline void lock_doms_cur(void) { }
302 static inline void unlock_doms_cur(void) { }
304 #endif /* CONFIG_FAIR_GROUP_SCHED */
306 /* CFS-related fields in a runqueue */
308 struct load_weight load;
309 unsigned long nr_running;
314 struct rb_root tasks_timeline;
315 struct rb_node *rb_leftmost;
316 struct rb_node *rb_load_balance_curr;
317 /* 'curr' points to currently running entity on this cfs_rq.
318 * It is set to NULL otherwise (i.e when none are currently running).
320 struct sched_entity *curr;
322 unsigned long nr_spread_over;
324 #ifdef CONFIG_FAIR_GROUP_SCHED
325 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
328 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
329 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
330 * (like users, containers etc.)
332 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
333 * list is used during load balance.
335 struct list_head leaf_cfs_rq_list;
336 struct task_group *tg; /* group that "owns" this runqueue */
340 /* Real-Time classes' related field in a runqueue: */
342 struct rt_prio_array active;
343 int rt_load_balance_idx;
344 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
345 unsigned long rt_nr_running;
346 unsigned long rt_nr_migratory;
347 /* highest queued rt task prio */
355 * We add the notion of a root-domain which will be used to define per-domain
356 * variables. Each exclusive cpuset essentially defines an island domain by
357 * fully partitioning the member cpus from any other cpuset. Whenever a new
358 * exclusive cpuset is created, we also create and attach a new root-domain
361 * By default the system creates a single root-domain with all cpus as
362 * members (mimicking the global state we have today).
370 static struct root_domain def_root_domain;
375 * This is the main, per-CPU runqueue data structure.
377 * Locking rule: those places that want to lock multiple runqueues
378 * (such as the load balancing or the thread migration code), lock
379 * acquire operations must be ordered by ascending &runqueue.
386 * nr_running and cpu_load should be in the same cacheline because
387 * remote CPUs use both these fields when doing load calculation.
389 unsigned long nr_running;
390 #define CPU_LOAD_IDX_MAX 5
391 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
392 unsigned char idle_at_tick;
394 unsigned char in_nohz_recently;
396 /* capture load from *all* tasks on this cpu: */
397 struct load_weight load;
398 unsigned long nr_load_updates;
402 #ifdef CONFIG_FAIR_GROUP_SCHED
403 /* list of leaf cfs_rq on this cpu: */
404 struct list_head leaf_cfs_rq_list;
409 * This is part of a global counter where only the total sum
410 * over all CPUs matters. A task can increase this counter on
411 * one CPU and if it got migrated afterwards it may decrease
412 * it on another CPU. Always updated under the runqueue lock:
414 unsigned long nr_uninterruptible;
416 struct task_struct *curr, *idle;
417 unsigned long next_balance;
418 struct mm_struct *prev_mm;
420 u64 clock, prev_clock_raw;
423 unsigned int clock_warps, clock_overflows;
425 unsigned int clock_deep_idle_events;
431 struct root_domain *rd;
432 struct sched_domain *sd;
434 /* For active balancing */
437 /* cpu of this runqueue: */
440 struct task_struct *migration_thread;
441 struct list_head migration_queue;
444 #ifdef CONFIG_SCHEDSTATS
446 struct sched_info rq_sched_info;
448 /* sys_sched_yield() stats */
449 unsigned int yld_exp_empty;
450 unsigned int yld_act_empty;
451 unsigned int yld_both_empty;
452 unsigned int yld_count;
454 /* schedule() stats */
455 unsigned int sched_switch;
456 unsigned int sched_count;
457 unsigned int sched_goidle;
459 /* try_to_wake_up() stats */
460 unsigned int ttwu_count;
461 unsigned int ttwu_local;
464 unsigned int bkl_count;
466 struct lock_class_key rq_lock_key;
469 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
471 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
473 rq->curr->sched_class->check_preempt_curr(rq, p);
476 static inline int cpu_of(struct rq *rq)
486 * Update the per-runqueue clock, as finegrained as the platform can give
487 * us, but without assuming monotonicity, etc.:
489 static void __update_rq_clock(struct rq *rq)
491 u64 prev_raw = rq->prev_clock_raw;
492 u64 now = sched_clock();
493 s64 delta = now - prev_raw;
494 u64 clock = rq->clock;
496 #ifdef CONFIG_SCHED_DEBUG
497 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
500 * Protect against sched_clock() occasionally going backwards:
502 if (unlikely(delta < 0)) {
507 * Catch too large forward jumps too:
509 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
510 if (clock < rq->tick_timestamp + TICK_NSEC)
511 clock = rq->tick_timestamp + TICK_NSEC;
514 rq->clock_overflows++;
516 if (unlikely(delta > rq->clock_max_delta))
517 rq->clock_max_delta = delta;
522 rq->prev_clock_raw = now;
526 static void update_rq_clock(struct rq *rq)
528 if (likely(smp_processor_id() == cpu_of(rq)))
529 __update_rq_clock(rq);
533 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
534 * See detach_destroy_domains: synchronize_sched for details.
536 * The domain tree of any CPU may only be accessed from within
537 * preempt-disabled sections.
539 #define for_each_domain(cpu, __sd) \
540 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
542 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
543 #define this_rq() (&__get_cpu_var(runqueues))
544 #define task_rq(p) cpu_rq(task_cpu(p))
545 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
548 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
550 #ifdef CONFIG_SCHED_DEBUG
551 # define const_debug __read_mostly
553 # define const_debug static const
557 * Debugging: various feature bits
560 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
561 SCHED_FEAT_WAKEUP_PREEMPT = 2,
562 SCHED_FEAT_START_DEBIT = 4,
563 SCHED_FEAT_TREE_AVG = 8,
564 SCHED_FEAT_APPROX_AVG = 16,
567 const_debug unsigned int sysctl_sched_features =
568 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
569 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
570 SCHED_FEAT_START_DEBIT * 1 |
571 SCHED_FEAT_TREE_AVG * 0 |
572 SCHED_FEAT_APPROX_AVG * 0;
574 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
577 * Number of tasks to iterate in a single balance run.
578 * Limited because this is done with IRQs disabled.
580 const_debug unsigned int sysctl_sched_nr_migrate = 32;
583 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
584 * clock constructed from sched_clock():
586 unsigned long long cpu_clock(int cpu)
588 unsigned long long now;
592 local_irq_save(flags);
595 * Only call sched_clock() if the scheduler has already been
596 * initialized (some code might call cpu_clock() very early):
601 local_irq_restore(flags);
605 EXPORT_SYMBOL_GPL(cpu_clock);
607 #ifndef prepare_arch_switch
608 # define prepare_arch_switch(next) do { } while (0)
610 #ifndef finish_arch_switch
611 # define finish_arch_switch(prev) do { } while (0)
614 static inline int task_current(struct rq *rq, struct task_struct *p)
616 return rq->curr == p;
619 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
620 static inline int task_running(struct rq *rq, struct task_struct *p)
622 return task_current(rq, p);
625 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
629 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
631 #ifdef CONFIG_DEBUG_SPINLOCK
632 /* this is a valid case when another task releases the spinlock */
633 rq->lock.owner = current;
636 * If we are tracking spinlock dependencies then we have to
637 * fix up the runqueue lock - which gets 'carried over' from
640 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
642 spin_unlock_irq(&rq->lock);
645 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
646 static inline int task_running(struct rq *rq, struct task_struct *p)
651 return task_current(rq, p);
655 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
659 * We can optimise this out completely for !SMP, because the
660 * SMP rebalancing from interrupt is the only thing that cares
665 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
666 spin_unlock_irq(&rq->lock);
668 spin_unlock(&rq->lock);
672 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
676 * After ->oncpu is cleared, the task can be moved to a different CPU.
677 * We must ensure this doesn't happen until the switch is completely
683 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
687 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
690 * __task_rq_lock - lock the runqueue a given task resides on.
691 * Must be called interrupts disabled.
693 static inline struct rq *__task_rq_lock(struct task_struct *p)
697 struct rq *rq = task_rq(p);
698 spin_lock(&rq->lock);
699 if (likely(rq == task_rq(p)))
701 spin_unlock(&rq->lock);
706 * task_rq_lock - lock the runqueue a given task resides on and disable
707 * interrupts. Note the ordering: we can safely lookup the task_rq without
708 * explicitly disabling preemption.
710 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
716 local_irq_save(*flags);
718 spin_lock(&rq->lock);
719 if (likely(rq == task_rq(p)))
721 spin_unlock_irqrestore(&rq->lock, *flags);
725 static void __task_rq_unlock(struct rq *rq)
728 spin_unlock(&rq->lock);
731 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
734 spin_unlock_irqrestore(&rq->lock, *flags);
738 * this_rq_lock - lock this runqueue and disable interrupts.
740 static struct rq *this_rq_lock(void)
747 spin_lock(&rq->lock);
753 * We are going deep-idle (irqs are disabled):
755 void sched_clock_idle_sleep_event(void)
757 struct rq *rq = cpu_rq(smp_processor_id());
759 spin_lock(&rq->lock);
760 __update_rq_clock(rq);
761 spin_unlock(&rq->lock);
762 rq->clock_deep_idle_events++;
764 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
767 * We just idled delta nanoseconds (called with irqs disabled):
769 void sched_clock_idle_wakeup_event(u64 delta_ns)
771 struct rq *rq = cpu_rq(smp_processor_id());
772 u64 now = sched_clock();
774 touch_softlockup_watchdog();
775 rq->idle_clock += delta_ns;
777 * Override the previous timestamp and ignore all
778 * sched_clock() deltas that occured while we idled,
779 * and use the PM-provided delta_ns to advance the
782 spin_lock(&rq->lock);
783 rq->prev_clock_raw = now;
784 rq->clock += delta_ns;
785 spin_unlock(&rq->lock);
787 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
790 * resched_task - mark a task 'to be rescheduled now'.
792 * On UP this means the setting of the need_resched flag, on SMP it
793 * might also involve a cross-CPU call to trigger the scheduler on
798 #ifndef tsk_is_polling
799 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
802 static void resched_task(struct task_struct *p)
806 assert_spin_locked(&task_rq(p)->lock);
808 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
811 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
814 if (cpu == smp_processor_id())
817 /* NEED_RESCHED must be visible before we test polling */
819 if (!tsk_is_polling(p))
820 smp_send_reschedule(cpu);
823 static void resched_cpu(int cpu)
825 struct rq *rq = cpu_rq(cpu);
828 if (!spin_trylock_irqsave(&rq->lock, flags))
830 resched_task(cpu_curr(cpu));
831 spin_unlock_irqrestore(&rq->lock, flags);
834 static inline void resched_task(struct task_struct *p)
836 assert_spin_locked(&task_rq(p)->lock);
837 set_tsk_need_resched(p);
841 #if BITS_PER_LONG == 32
842 # define WMULT_CONST (~0UL)
844 # define WMULT_CONST (1UL << 32)
847 #define WMULT_SHIFT 32
850 * Shift right and round:
852 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
855 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
856 struct load_weight *lw)
860 if (unlikely(!lw->inv_weight))
861 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
863 tmp = (u64)delta_exec * weight;
865 * Check whether we'd overflow the 64-bit multiplication:
867 if (unlikely(tmp > WMULT_CONST))
868 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
871 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
873 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
876 static inline unsigned long
877 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
879 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
882 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
887 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
893 * To aid in avoiding the subversion of "niceness" due to uneven distribution
894 * of tasks with abnormal "nice" values across CPUs the contribution that
895 * each task makes to its run queue's load is weighted according to its
896 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
897 * scaled version of the new time slice allocation that they receive on time
901 #define WEIGHT_IDLEPRIO 2
902 #define WMULT_IDLEPRIO (1 << 31)
905 * Nice levels are multiplicative, with a gentle 10% change for every
906 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
907 * nice 1, it will get ~10% less CPU time than another CPU-bound task
908 * that remained on nice 0.
910 * The "10% effect" is relative and cumulative: from _any_ nice level,
911 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
912 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
913 * If a task goes up by ~10% and another task goes down by ~10% then
914 * the relative distance between them is ~25%.)
916 static const int prio_to_weight[40] = {
917 /* -20 */ 88761, 71755, 56483, 46273, 36291,
918 /* -15 */ 29154, 23254, 18705, 14949, 11916,
919 /* -10 */ 9548, 7620, 6100, 4904, 3906,
920 /* -5 */ 3121, 2501, 1991, 1586, 1277,
921 /* 0 */ 1024, 820, 655, 526, 423,
922 /* 5 */ 335, 272, 215, 172, 137,
923 /* 10 */ 110, 87, 70, 56, 45,
924 /* 15 */ 36, 29, 23, 18, 15,
928 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
930 * In cases where the weight does not change often, we can use the
931 * precalculated inverse to speed up arithmetics by turning divisions
932 * into multiplications:
934 static const u32 prio_to_wmult[40] = {
935 /* -20 */ 48388, 59856, 76040, 92818, 118348,
936 /* -15 */ 147320, 184698, 229616, 287308, 360437,
937 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
938 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
939 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
940 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
941 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
942 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
945 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
948 * runqueue iterator, to support SMP load-balancing between different
949 * scheduling classes, without having to expose their internal data
950 * structures to the load-balancing proper:
954 struct task_struct *(*start)(void *);
955 struct task_struct *(*next)(void *);
960 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
961 unsigned long max_load_move, struct sched_domain *sd,
962 enum cpu_idle_type idle, int *all_pinned,
963 int *this_best_prio, struct rq_iterator *iterator);
966 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
967 struct sched_domain *sd, enum cpu_idle_type idle,
968 struct rq_iterator *iterator);
971 #ifdef CONFIG_CGROUP_CPUACCT
972 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
974 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
977 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
979 update_load_add(&rq->load, load);
982 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
984 update_load_sub(&rq->load, load);
988 static unsigned long source_load(int cpu, int type);
989 static unsigned long target_load(int cpu, int type);
990 static unsigned long cpu_avg_load_per_task(int cpu);
991 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
992 #endif /* CONFIG_SMP */
994 #include "sched_stats.h"
995 #include "sched_idletask.c"
996 #include "sched_fair.c"
997 #include "sched_rt.c"
998 #ifdef CONFIG_SCHED_DEBUG
999 # include "sched_debug.c"
1002 #define sched_class_highest (&rt_sched_class)
1004 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1009 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1014 static void set_load_weight(struct task_struct *p)
1016 if (task_has_rt_policy(p)) {
1017 p->se.load.weight = prio_to_weight[0] * 2;
1018 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1023 * SCHED_IDLE tasks get minimal weight:
1025 if (p->policy == SCHED_IDLE) {
1026 p->se.load.weight = WEIGHT_IDLEPRIO;
1027 p->se.load.inv_weight = WMULT_IDLEPRIO;
1031 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1032 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1035 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1037 sched_info_queued(p);
1038 p->sched_class->enqueue_task(rq, p, wakeup);
1042 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1044 p->sched_class->dequeue_task(rq, p, sleep);
1049 * __normal_prio - return the priority that is based on the static prio
1051 static inline int __normal_prio(struct task_struct *p)
1053 return p->static_prio;
1057 * Calculate the expected normal priority: i.e. priority
1058 * without taking RT-inheritance into account. Might be
1059 * boosted by interactivity modifiers. Changes upon fork,
1060 * setprio syscalls, and whenever the interactivity
1061 * estimator recalculates.
1063 static inline int normal_prio(struct task_struct *p)
1067 if (task_has_rt_policy(p))
1068 prio = MAX_RT_PRIO-1 - p->rt_priority;
1070 prio = __normal_prio(p);
1075 * Calculate the current priority, i.e. the priority
1076 * taken into account by the scheduler. This value might
1077 * be boosted by RT tasks, or might be boosted by
1078 * interactivity modifiers. Will be RT if the task got
1079 * RT-boosted. If not then it returns p->normal_prio.
1081 static int effective_prio(struct task_struct *p)
1083 p->normal_prio = normal_prio(p);
1085 * If we are RT tasks or we were boosted to RT priority,
1086 * keep the priority unchanged. Otherwise, update priority
1087 * to the normal priority:
1089 if (!rt_prio(p->prio))
1090 return p->normal_prio;
1095 * activate_task - move a task to the runqueue.
1097 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1099 if (p->state == TASK_UNINTERRUPTIBLE)
1100 rq->nr_uninterruptible--;
1102 enqueue_task(rq, p, wakeup);
1103 inc_nr_running(p, rq);
1107 * deactivate_task - remove a task from the runqueue.
1109 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1111 if (p->state == TASK_UNINTERRUPTIBLE)
1112 rq->nr_uninterruptible++;
1114 dequeue_task(rq, p, sleep);
1115 dec_nr_running(p, rq);
1119 * task_curr - is this task currently executing on a CPU?
1120 * @p: the task in question.
1122 inline int task_curr(const struct task_struct *p)
1124 return cpu_curr(task_cpu(p)) == p;
1127 /* Used instead of source_load when we know the type == 0 */
1128 unsigned long weighted_cpuload(const int cpu)
1130 return cpu_rq(cpu)->load.weight;
1133 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1135 set_task_cfs_rq(p, cpu);
1138 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1139 * successfuly executed on another CPU. We must ensure that updates of
1140 * per-task data have been completed by this moment.
1143 task_thread_info(p)->cpu = cpu;
1150 * Is this task likely cache-hot:
1153 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1157 if (p->sched_class != &fair_sched_class)
1160 if (sysctl_sched_migration_cost == -1)
1162 if (sysctl_sched_migration_cost == 0)
1165 delta = now - p->se.exec_start;
1167 return delta < (s64)sysctl_sched_migration_cost;
1171 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1173 int old_cpu = task_cpu(p);
1174 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1175 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1176 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1179 clock_offset = old_rq->clock - new_rq->clock;
1181 #ifdef CONFIG_SCHEDSTATS
1182 if (p->se.wait_start)
1183 p->se.wait_start -= clock_offset;
1184 if (p->se.sleep_start)
1185 p->se.sleep_start -= clock_offset;
1186 if (p->se.block_start)
1187 p->se.block_start -= clock_offset;
1188 if (old_cpu != new_cpu) {
1189 schedstat_inc(p, se.nr_migrations);
1190 if (task_hot(p, old_rq->clock, NULL))
1191 schedstat_inc(p, se.nr_forced2_migrations);
1194 p->se.vruntime -= old_cfsrq->min_vruntime -
1195 new_cfsrq->min_vruntime;
1197 __set_task_cpu(p, new_cpu);
1200 struct migration_req {
1201 struct list_head list;
1203 struct task_struct *task;
1206 struct completion done;
1210 * The task's runqueue lock must be held.
1211 * Returns true if you have to wait for migration thread.
1214 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1216 struct rq *rq = task_rq(p);
1219 * If the task is not on a runqueue (and not running), then
1220 * it is sufficient to simply update the task's cpu field.
1222 if (!p->se.on_rq && !task_running(rq, p)) {
1223 set_task_cpu(p, dest_cpu);
1227 init_completion(&req->done);
1229 req->dest_cpu = dest_cpu;
1230 list_add(&req->list, &rq->migration_queue);
1236 * wait_task_inactive - wait for a thread to unschedule.
1238 * The caller must ensure that the task *will* unschedule sometime soon,
1239 * else this function might spin for a *long* time. This function can't
1240 * be called with interrupts off, or it may introduce deadlock with
1241 * smp_call_function() if an IPI is sent by the same process we are
1242 * waiting to become inactive.
1244 void wait_task_inactive(struct task_struct *p)
1246 unsigned long flags;
1252 * We do the initial early heuristics without holding
1253 * any task-queue locks at all. We'll only try to get
1254 * the runqueue lock when things look like they will
1260 * If the task is actively running on another CPU
1261 * still, just relax and busy-wait without holding
1264 * NOTE! Since we don't hold any locks, it's not
1265 * even sure that "rq" stays as the right runqueue!
1266 * But we don't care, since "task_running()" will
1267 * return false if the runqueue has changed and p
1268 * is actually now running somewhere else!
1270 while (task_running(rq, p))
1274 * Ok, time to look more closely! We need the rq
1275 * lock now, to be *sure*. If we're wrong, we'll
1276 * just go back and repeat.
1278 rq = task_rq_lock(p, &flags);
1279 running = task_running(rq, p);
1280 on_rq = p->se.on_rq;
1281 task_rq_unlock(rq, &flags);
1284 * Was it really running after all now that we
1285 * checked with the proper locks actually held?
1287 * Oops. Go back and try again..
1289 if (unlikely(running)) {
1295 * It's not enough that it's not actively running,
1296 * it must be off the runqueue _entirely_, and not
1299 * So if it wa still runnable (but just not actively
1300 * running right now), it's preempted, and we should
1301 * yield - it could be a while.
1303 if (unlikely(on_rq)) {
1304 schedule_timeout_uninterruptible(1);
1309 * Ahh, all good. It wasn't running, and it wasn't
1310 * runnable, which means that it will never become
1311 * running in the future either. We're all done!
1318 * kick_process - kick a running thread to enter/exit the kernel
1319 * @p: the to-be-kicked thread
1321 * Cause a process which is running on another CPU to enter
1322 * kernel-mode, without any delay. (to get signals handled.)
1324 * NOTE: this function doesnt have to take the runqueue lock,
1325 * because all it wants to ensure is that the remote task enters
1326 * the kernel. If the IPI races and the task has been migrated
1327 * to another CPU then no harm is done and the purpose has been
1330 void kick_process(struct task_struct *p)
1336 if ((cpu != smp_processor_id()) && task_curr(p))
1337 smp_send_reschedule(cpu);
1342 * Return a low guess at the load of a migration-source cpu weighted
1343 * according to the scheduling class and "nice" value.
1345 * We want to under-estimate the load of migration sources, to
1346 * balance conservatively.
1348 static unsigned long source_load(int cpu, int type)
1350 struct rq *rq = cpu_rq(cpu);
1351 unsigned long total = weighted_cpuload(cpu);
1356 return min(rq->cpu_load[type-1], total);
1360 * Return a high guess at the load of a migration-target cpu weighted
1361 * according to the scheduling class and "nice" value.
1363 static unsigned long target_load(int cpu, int type)
1365 struct rq *rq = cpu_rq(cpu);
1366 unsigned long total = weighted_cpuload(cpu);
1371 return max(rq->cpu_load[type-1], total);
1375 * Return the average load per task on the cpu's run queue
1377 static unsigned long cpu_avg_load_per_task(int cpu)
1379 struct rq *rq = cpu_rq(cpu);
1380 unsigned long total = weighted_cpuload(cpu);
1381 unsigned long n = rq->nr_running;
1383 return n ? total / n : SCHED_LOAD_SCALE;
1387 * find_idlest_group finds and returns the least busy CPU group within the
1390 static struct sched_group *
1391 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1393 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1394 unsigned long min_load = ULONG_MAX, this_load = 0;
1395 int load_idx = sd->forkexec_idx;
1396 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1399 unsigned long load, avg_load;
1403 /* Skip over this group if it has no CPUs allowed */
1404 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1407 local_group = cpu_isset(this_cpu, group->cpumask);
1409 /* Tally up the load of all CPUs in the group */
1412 for_each_cpu_mask(i, group->cpumask) {
1413 /* Bias balancing toward cpus of our domain */
1415 load = source_load(i, load_idx);
1417 load = target_load(i, load_idx);
1422 /* Adjust by relative CPU power of the group */
1423 avg_load = sg_div_cpu_power(group,
1424 avg_load * SCHED_LOAD_SCALE);
1427 this_load = avg_load;
1429 } else if (avg_load < min_load) {
1430 min_load = avg_load;
1433 } while (group = group->next, group != sd->groups);
1435 if (!idlest || 100*this_load < imbalance*min_load)
1441 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1444 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1447 unsigned long load, min_load = ULONG_MAX;
1451 /* Traverse only the allowed CPUs */
1452 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1454 for_each_cpu_mask(i, tmp) {
1455 load = weighted_cpuload(i);
1457 if (load < min_load || (load == min_load && i == this_cpu)) {
1467 * sched_balance_self: balance the current task (running on cpu) in domains
1468 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1471 * Balance, ie. select the least loaded group.
1473 * Returns the target CPU number, or the same CPU if no balancing is needed.
1475 * preempt must be disabled.
1477 static int sched_balance_self(int cpu, int flag)
1479 struct task_struct *t = current;
1480 struct sched_domain *tmp, *sd = NULL;
1482 for_each_domain(cpu, tmp) {
1484 * If power savings logic is enabled for a domain, stop there.
1486 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1488 if (tmp->flags & flag)
1494 struct sched_group *group;
1495 int new_cpu, weight;
1497 if (!(sd->flags & flag)) {
1503 group = find_idlest_group(sd, t, cpu);
1509 new_cpu = find_idlest_cpu(group, t, cpu);
1510 if (new_cpu == -1 || new_cpu == cpu) {
1511 /* Now try balancing at a lower domain level of cpu */
1516 /* Now try balancing at a lower domain level of new_cpu */
1519 weight = cpus_weight(span);
1520 for_each_domain(cpu, tmp) {
1521 if (weight <= cpus_weight(tmp->span))
1523 if (tmp->flags & flag)
1526 /* while loop will break here if sd == NULL */
1532 #endif /* CONFIG_SMP */
1535 * try_to_wake_up - wake up a thread
1536 * @p: the to-be-woken-up thread
1537 * @state: the mask of task states that can be woken
1538 * @sync: do a synchronous wakeup?
1540 * Put it on the run-queue if it's not already there. The "current"
1541 * thread is always on the run-queue (except when the actual
1542 * re-schedule is in progress), and as such you're allowed to do
1543 * the simpler "current->state = TASK_RUNNING" to mark yourself
1544 * runnable without the overhead of this.
1546 * returns failure only if the task is already active.
1548 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1550 int cpu, orig_cpu, this_cpu, success = 0;
1551 unsigned long flags;
1558 rq = task_rq_lock(p, &flags);
1559 old_state = p->state;
1560 if (!(old_state & state))
1568 this_cpu = smp_processor_id();
1571 if (unlikely(task_running(rq, p)))
1574 new_cpu = p->sched_class->select_task_rq(p, sync);
1575 if (new_cpu != cpu) {
1576 set_task_cpu(p, new_cpu);
1577 task_rq_unlock(rq, &flags);
1578 /* might preempt at this point */
1579 rq = task_rq_lock(p, &flags);
1580 old_state = p->state;
1581 if (!(old_state & state))
1586 this_cpu = smp_processor_id();
1590 #ifdef CONFIG_SCHEDSTATS
1591 schedstat_inc(rq, ttwu_count);
1592 if (cpu == this_cpu)
1593 schedstat_inc(rq, ttwu_local);
1595 struct sched_domain *sd;
1596 for_each_domain(this_cpu, sd) {
1597 if (cpu_isset(cpu, sd->span)) {
1598 schedstat_inc(sd, ttwu_wake_remote);
1608 #endif /* CONFIG_SMP */
1609 schedstat_inc(p, se.nr_wakeups);
1611 schedstat_inc(p, se.nr_wakeups_sync);
1612 if (orig_cpu != cpu)
1613 schedstat_inc(p, se.nr_wakeups_migrate);
1614 if (cpu == this_cpu)
1615 schedstat_inc(p, se.nr_wakeups_local);
1617 schedstat_inc(p, se.nr_wakeups_remote);
1618 update_rq_clock(rq);
1619 activate_task(rq, p, 1);
1620 check_preempt_curr(rq, p);
1624 p->state = TASK_RUNNING;
1625 wakeup_balance_rt(rq, p);
1627 task_rq_unlock(rq, &flags);
1632 int fastcall wake_up_process(struct task_struct *p)
1634 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1635 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1637 EXPORT_SYMBOL(wake_up_process);
1639 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1641 return try_to_wake_up(p, state, 0);
1645 * Perform scheduler related setup for a newly forked process p.
1646 * p is forked by current.
1648 * __sched_fork() is basic setup used by init_idle() too:
1650 static void __sched_fork(struct task_struct *p)
1652 p->se.exec_start = 0;
1653 p->se.sum_exec_runtime = 0;
1654 p->se.prev_sum_exec_runtime = 0;
1656 #ifdef CONFIG_SCHEDSTATS
1657 p->se.wait_start = 0;
1658 p->se.sum_sleep_runtime = 0;
1659 p->se.sleep_start = 0;
1660 p->se.block_start = 0;
1661 p->se.sleep_max = 0;
1662 p->se.block_max = 0;
1664 p->se.slice_max = 0;
1668 INIT_LIST_HEAD(&p->run_list);
1671 #ifdef CONFIG_PREEMPT_NOTIFIERS
1672 INIT_HLIST_HEAD(&p->preempt_notifiers);
1676 * We mark the process as running here, but have not actually
1677 * inserted it onto the runqueue yet. This guarantees that
1678 * nobody will actually run it, and a signal or other external
1679 * event cannot wake it up and insert it on the runqueue either.
1681 p->state = TASK_RUNNING;
1685 * fork()/clone()-time setup:
1687 void sched_fork(struct task_struct *p, int clone_flags)
1689 int cpu = get_cpu();
1694 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1696 set_task_cpu(p, cpu);
1699 * Make sure we do not leak PI boosting priority to the child:
1701 p->prio = current->normal_prio;
1702 if (!rt_prio(p->prio))
1703 p->sched_class = &fair_sched_class;
1705 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1706 if (likely(sched_info_on()))
1707 memset(&p->sched_info, 0, sizeof(p->sched_info));
1709 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1712 #ifdef CONFIG_PREEMPT
1713 /* Want to start with kernel preemption disabled. */
1714 task_thread_info(p)->preempt_count = 1;
1720 * wake_up_new_task - wake up a newly created task for the first time.
1722 * This function will do some initial scheduler statistics housekeeping
1723 * that must be done for every newly created context, then puts the task
1724 * on the runqueue and wakes it.
1726 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1728 unsigned long flags;
1731 rq = task_rq_lock(p, &flags);
1732 BUG_ON(p->state != TASK_RUNNING);
1733 update_rq_clock(rq);
1735 p->prio = effective_prio(p);
1737 if (!p->sched_class->task_new || !current->se.on_rq) {
1738 activate_task(rq, p, 0);
1741 * Let the scheduling class do new task startup
1742 * management (if any):
1744 p->sched_class->task_new(rq, p);
1745 inc_nr_running(p, rq);
1747 check_preempt_curr(rq, p);
1748 wakeup_balance_rt(rq, p);
1749 task_rq_unlock(rq, &flags);
1752 #ifdef CONFIG_PREEMPT_NOTIFIERS
1755 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1756 * @notifier: notifier struct to register
1758 void preempt_notifier_register(struct preempt_notifier *notifier)
1760 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1762 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1765 * preempt_notifier_unregister - no longer interested in preemption notifications
1766 * @notifier: notifier struct to unregister
1768 * This is safe to call from within a preemption notifier.
1770 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1772 hlist_del(¬ifier->link);
1774 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1776 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1778 struct preempt_notifier *notifier;
1779 struct hlist_node *node;
1781 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1782 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1786 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1787 struct task_struct *next)
1789 struct preempt_notifier *notifier;
1790 struct hlist_node *node;
1792 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1793 notifier->ops->sched_out(notifier, next);
1798 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1803 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1804 struct task_struct *next)
1811 * prepare_task_switch - prepare to switch tasks
1812 * @rq: the runqueue preparing to switch
1813 * @prev: the current task that is being switched out
1814 * @next: the task we are going to switch to.
1816 * This is called with the rq lock held and interrupts off. It must
1817 * be paired with a subsequent finish_task_switch after the context
1820 * prepare_task_switch sets up locking and calls architecture specific
1824 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1825 struct task_struct *next)
1827 fire_sched_out_preempt_notifiers(prev, next);
1828 prepare_lock_switch(rq, next);
1829 prepare_arch_switch(next);
1833 * finish_task_switch - clean up after a task-switch
1834 * @rq: runqueue associated with task-switch
1835 * @prev: the thread we just switched away from.
1837 * finish_task_switch must be called after the context switch, paired
1838 * with a prepare_task_switch call before the context switch.
1839 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1840 * and do any other architecture-specific cleanup actions.
1842 * Note that we may have delayed dropping an mm in context_switch(). If
1843 * so, we finish that here outside of the runqueue lock. (Doing it
1844 * with the lock held can cause deadlocks; see schedule() for
1847 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1848 __releases(rq->lock)
1850 struct mm_struct *mm = rq->prev_mm;
1856 * A task struct has one reference for the use as "current".
1857 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1858 * schedule one last time. The schedule call will never return, and
1859 * the scheduled task must drop that reference.
1860 * The test for TASK_DEAD must occur while the runqueue locks are
1861 * still held, otherwise prev could be scheduled on another cpu, die
1862 * there before we look at prev->state, and then the reference would
1864 * Manfred Spraul <manfred@colorfullife.com>
1866 prev_state = prev->state;
1867 finish_arch_switch(prev);
1868 finish_lock_switch(rq, prev);
1869 schedule_tail_balance_rt(rq);
1871 fire_sched_in_preempt_notifiers(current);
1874 if (unlikely(prev_state == TASK_DEAD)) {
1876 * Remove function-return probe instances associated with this
1877 * task and put them back on the free list.
1879 kprobe_flush_task(prev);
1880 put_task_struct(prev);
1885 * schedule_tail - first thing a freshly forked thread must call.
1886 * @prev: the thread we just switched away from.
1888 asmlinkage void schedule_tail(struct task_struct *prev)
1889 __releases(rq->lock)
1891 struct rq *rq = this_rq();
1893 finish_task_switch(rq, prev);
1894 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1895 /* In this case, finish_task_switch does not reenable preemption */
1898 if (current->set_child_tid)
1899 put_user(task_pid_vnr(current), current->set_child_tid);
1903 * context_switch - switch to the new MM and the new
1904 * thread's register state.
1907 context_switch(struct rq *rq, struct task_struct *prev,
1908 struct task_struct *next)
1910 struct mm_struct *mm, *oldmm;
1912 prepare_task_switch(rq, prev, next);
1914 oldmm = prev->active_mm;
1916 * For paravirt, this is coupled with an exit in switch_to to
1917 * combine the page table reload and the switch backend into
1920 arch_enter_lazy_cpu_mode();
1922 if (unlikely(!mm)) {
1923 next->active_mm = oldmm;
1924 atomic_inc(&oldmm->mm_count);
1925 enter_lazy_tlb(oldmm, next);
1927 switch_mm(oldmm, mm, next);
1929 if (unlikely(!prev->mm)) {
1930 prev->active_mm = NULL;
1931 rq->prev_mm = oldmm;
1934 * Since the runqueue lock will be released by the next
1935 * task (which is an invalid locking op but in the case
1936 * of the scheduler it's an obvious special-case), so we
1937 * do an early lockdep release here:
1939 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1940 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1943 /* Here we just switch the register state and the stack. */
1944 switch_to(prev, next, prev);
1948 * this_rq must be evaluated again because prev may have moved
1949 * CPUs since it called schedule(), thus the 'rq' on its stack
1950 * frame will be invalid.
1952 finish_task_switch(this_rq(), prev);
1956 * nr_running, nr_uninterruptible and nr_context_switches:
1958 * externally visible scheduler statistics: current number of runnable
1959 * threads, current number of uninterruptible-sleeping threads, total
1960 * number of context switches performed since bootup.
1962 unsigned long nr_running(void)
1964 unsigned long i, sum = 0;
1966 for_each_online_cpu(i)
1967 sum += cpu_rq(i)->nr_running;
1972 unsigned long nr_uninterruptible(void)
1974 unsigned long i, sum = 0;
1976 for_each_possible_cpu(i)
1977 sum += cpu_rq(i)->nr_uninterruptible;
1980 * Since we read the counters lockless, it might be slightly
1981 * inaccurate. Do not allow it to go below zero though:
1983 if (unlikely((long)sum < 0))
1989 unsigned long long nr_context_switches(void)
1992 unsigned long long sum = 0;
1994 for_each_possible_cpu(i)
1995 sum += cpu_rq(i)->nr_switches;
2000 unsigned long nr_iowait(void)
2002 unsigned long i, sum = 0;
2004 for_each_possible_cpu(i)
2005 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2010 unsigned long nr_active(void)
2012 unsigned long i, running = 0, uninterruptible = 0;
2014 for_each_online_cpu(i) {
2015 running += cpu_rq(i)->nr_running;
2016 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2019 if (unlikely((long)uninterruptible < 0))
2020 uninterruptible = 0;
2022 return running + uninterruptible;
2026 * Update rq->cpu_load[] statistics. This function is usually called every
2027 * scheduler tick (TICK_NSEC).
2029 static void update_cpu_load(struct rq *this_rq)
2031 unsigned long this_load = this_rq->load.weight;
2034 this_rq->nr_load_updates++;
2036 /* Update our load: */
2037 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2038 unsigned long old_load, new_load;
2040 /* scale is effectively 1 << i now, and >> i divides by scale */
2042 old_load = this_rq->cpu_load[i];
2043 new_load = this_load;
2045 * Round up the averaging division if load is increasing. This
2046 * prevents us from getting stuck on 9 if the load is 10, for
2049 if (new_load > old_load)
2050 new_load += scale-1;
2051 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2058 * double_rq_lock - safely lock two runqueues
2060 * Note this does not disable interrupts like task_rq_lock,
2061 * you need to do so manually before calling.
2063 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2064 __acquires(rq1->lock)
2065 __acquires(rq2->lock)
2067 BUG_ON(!irqs_disabled());
2069 spin_lock(&rq1->lock);
2070 __acquire(rq2->lock); /* Fake it out ;) */
2073 spin_lock(&rq1->lock);
2074 spin_lock(&rq2->lock);
2076 spin_lock(&rq2->lock);
2077 spin_lock(&rq1->lock);
2080 update_rq_clock(rq1);
2081 update_rq_clock(rq2);
2085 * double_rq_unlock - safely unlock two runqueues
2087 * Note this does not restore interrupts like task_rq_unlock,
2088 * you need to do so manually after calling.
2090 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2091 __releases(rq1->lock)
2092 __releases(rq2->lock)
2094 spin_unlock(&rq1->lock);
2096 spin_unlock(&rq2->lock);
2098 __release(rq2->lock);
2102 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2104 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2105 __releases(this_rq->lock)
2106 __acquires(busiest->lock)
2107 __acquires(this_rq->lock)
2111 if (unlikely(!irqs_disabled())) {
2112 /* printk() doesn't work good under rq->lock */
2113 spin_unlock(&this_rq->lock);
2116 if (unlikely(!spin_trylock(&busiest->lock))) {
2117 if (busiest < this_rq) {
2118 spin_unlock(&this_rq->lock);
2119 spin_lock(&busiest->lock);
2120 spin_lock(&this_rq->lock);
2123 spin_lock(&busiest->lock);
2129 * If dest_cpu is allowed for this process, migrate the task to it.
2130 * This is accomplished by forcing the cpu_allowed mask to only
2131 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2132 * the cpu_allowed mask is restored.
2134 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2136 struct migration_req req;
2137 unsigned long flags;
2140 rq = task_rq_lock(p, &flags);
2141 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2142 || unlikely(cpu_is_offline(dest_cpu)))
2145 /* force the process onto the specified CPU */
2146 if (migrate_task(p, dest_cpu, &req)) {
2147 /* Need to wait for migration thread (might exit: take ref). */
2148 struct task_struct *mt = rq->migration_thread;
2150 get_task_struct(mt);
2151 task_rq_unlock(rq, &flags);
2152 wake_up_process(mt);
2153 put_task_struct(mt);
2154 wait_for_completion(&req.done);
2159 task_rq_unlock(rq, &flags);
2163 * sched_exec - execve() is a valuable balancing opportunity, because at
2164 * this point the task has the smallest effective memory and cache footprint.
2166 void sched_exec(void)
2168 int new_cpu, this_cpu = get_cpu();
2169 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2171 if (new_cpu != this_cpu)
2172 sched_migrate_task(current, new_cpu);
2176 * pull_task - move a task from a remote runqueue to the local runqueue.
2177 * Both runqueues must be locked.
2179 static void pull_task(struct rq *src_rq, struct task_struct *p,
2180 struct rq *this_rq, int this_cpu)
2182 deactivate_task(src_rq, p, 0);
2183 set_task_cpu(p, this_cpu);
2184 activate_task(this_rq, p, 0);
2186 * Note that idle threads have a prio of MAX_PRIO, for this test
2187 * to be always true for them.
2189 check_preempt_curr(this_rq, p);
2193 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2196 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2197 struct sched_domain *sd, enum cpu_idle_type idle,
2201 * We do not migrate tasks that are:
2202 * 1) running (obviously), or
2203 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2204 * 3) are cache-hot on their current CPU.
2206 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2207 schedstat_inc(p, se.nr_failed_migrations_affine);
2212 if (task_running(rq, p)) {
2213 schedstat_inc(p, se.nr_failed_migrations_running);
2218 * Aggressive migration if:
2219 * 1) task is cache cold, or
2220 * 2) too many balance attempts have failed.
2223 if (!task_hot(p, rq->clock, sd) ||
2224 sd->nr_balance_failed > sd->cache_nice_tries) {
2225 #ifdef CONFIG_SCHEDSTATS
2226 if (task_hot(p, rq->clock, sd)) {
2227 schedstat_inc(sd, lb_hot_gained[idle]);
2228 schedstat_inc(p, se.nr_forced_migrations);
2234 if (task_hot(p, rq->clock, sd)) {
2235 schedstat_inc(p, se.nr_failed_migrations_hot);
2241 static unsigned long
2242 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2243 unsigned long max_load_move, struct sched_domain *sd,
2244 enum cpu_idle_type idle, int *all_pinned,
2245 int *this_best_prio, struct rq_iterator *iterator)
2247 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2248 struct task_struct *p;
2249 long rem_load_move = max_load_move;
2251 if (max_load_move == 0)
2257 * Start the load-balancing iterator:
2259 p = iterator->start(iterator->arg);
2261 if (!p || loops++ > sysctl_sched_nr_migrate)
2264 * To help distribute high priority tasks across CPUs we don't
2265 * skip a task if it will be the highest priority task (i.e. smallest
2266 * prio value) on its new queue regardless of its load weight
2268 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2269 SCHED_LOAD_SCALE_FUZZ;
2270 if ((skip_for_load && p->prio >= *this_best_prio) ||
2271 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2272 p = iterator->next(iterator->arg);
2276 pull_task(busiest, p, this_rq, this_cpu);
2278 rem_load_move -= p->se.load.weight;
2281 * We only want to steal up to the prescribed amount of weighted load.
2283 if (rem_load_move > 0) {
2284 if (p->prio < *this_best_prio)
2285 *this_best_prio = p->prio;
2286 p = iterator->next(iterator->arg);
2291 * Right now, this is one of only two places pull_task() is called,
2292 * so we can safely collect pull_task() stats here rather than
2293 * inside pull_task().
2295 schedstat_add(sd, lb_gained[idle], pulled);
2298 *all_pinned = pinned;
2300 return max_load_move - rem_load_move;
2304 * move_tasks tries to move up to max_load_move weighted load from busiest to
2305 * this_rq, as part of a balancing operation within domain "sd".
2306 * Returns 1 if successful and 0 otherwise.
2308 * Called with both runqueues locked.
2310 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2311 unsigned long max_load_move,
2312 struct sched_domain *sd, enum cpu_idle_type idle,
2315 const struct sched_class *class = sched_class_highest;
2316 unsigned long total_load_moved = 0;
2317 int this_best_prio = this_rq->curr->prio;
2321 class->load_balance(this_rq, this_cpu, busiest,
2322 max_load_move - total_load_moved,
2323 sd, idle, all_pinned, &this_best_prio);
2324 class = class->next;
2325 } while (class && max_load_move > total_load_moved);
2327 return total_load_moved > 0;
2331 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2332 struct sched_domain *sd, enum cpu_idle_type idle,
2333 struct rq_iterator *iterator)
2335 struct task_struct *p = iterator->start(iterator->arg);
2339 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2340 pull_task(busiest, p, this_rq, this_cpu);
2342 * Right now, this is only the second place pull_task()
2343 * is called, so we can safely collect pull_task()
2344 * stats here rather than inside pull_task().
2346 schedstat_inc(sd, lb_gained[idle]);
2350 p = iterator->next(iterator->arg);
2357 * move_one_task tries to move exactly one task from busiest to this_rq, as
2358 * part of active balancing operations within "domain".
2359 * Returns 1 if successful and 0 otherwise.
2361 * Called with both runqueues locked.
2363 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2364 struct sched_domain *sd, enum cpu_idle_type idle)
2366 const struct sched_class *class;
2368 for (class = sched_class_highest; class; class = class->next)
2369 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2376 * find_busiest_group finds and returns the busiest CPU group within the
2377 * domain. It calculates and returns the amount of weighted load which
2378 * should be moved to restore balance via the imbalance parameter.
2380 static struct sched_group *
2381 find_busiest_group(struct sched_domain *sd, int this_cpu,
2382 unsigned long *imbalance, enum cpu_idle_type idle,
2383 int *sd_idle, cpumask_t *cpus, int *balance)
2385 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2386 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2387 unsigned long max_pull;
2388 unsigned long busiest_load_per_task, busiest_nr_running;
2389 unsigned long this_load_per_task, this_nr_running;
2390 int load_idx, group_imb = 0;
2391 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2392 int power_savings_balance = 1;
2393 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2394 unsigned long min_nr_running = ULONG_MAX;
2395 struct sched_group *group_min = NULL, *group_leader = NULL;
2398 max_load = this_load = total_load = total_pwr = 0;
2399 busiest_load_per_task = busiest_nr_running = 0;
2400 this_load_per_task = this_nr_running = 0;
2401 if (idle == CPU_NOT_IDLE)
2402 load_idx = sd->busy_idx;
2403 else if (idle == CPU_NEWLY_IDLE)
2404 load_idx = sd->newidle_idx;
2406 load_idx = sd->idle_idx;
2409 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2412 int __group_imb = 0;
2413 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2414 unsigned long sum_nr_running, sum_weighted_load;
2416 local_group = cpu_isset(this_cpu, group->cpumask);
2419 balance_cpu = first_cpu(group->cpumask);
2421 /* Tally up the load of all CPUs in the group */
2422 sum_weighted_load = sum_nr_running = avg_load = 0;
2424 min_cpu_load = ~0UL;
2426 for_each_cpu_mask(i, group->cpumask) {
2429 if (!cpu_isset(i, *cpus))
2434 if (*sd_idle && rq->nr_running)
2437 /* Bias balancing toward cpus of our domain */
2439 if (idle_cpu(i) && !first_idle_cpu) {
2444 load = target_load(i, load_idx);
2446 load = source_load(i, load_idx);
2447 if (load > max_cpu_load)
2448 max_cpu_load = load;
2449 if (min_cpu_load > load)
2450 min_cpu_load = load;
2454 sum_nr_running += rq->nr_running;
2455 sum_weighted_load += weighted_cpuload(i);
2459 * First idle cpu or the first cpu(busiest) in this sched group
2460 * is eligible for doing load balancing at this and above
2461 * domains. In the newly idle case, we will allow all the cpu's
2462 * to do the newly idle load balance.
2464 if (idle != CPU_NEWLY_IDLE && local_group &&
2465 balance_cpu != this_cpu && balance) {
2470 total_load += avg_load;
2471 total_pwr += group->__cpu_power;
2473 /* Adjust by relative CPU power of the group */
2474 avg_load = sg_div_cpu_power(group,
2475 avg_load * SCHED_LOAD_SCALE);
2477 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2480 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2483 this_load = avg_load;
2485 this_nr_running = sum_nr_running;
2486 this_load_per_task = sum_weighted_load;
2487 } else if (avg_load > max_load &&
2488 (sum_nr_running > group_capacity || __group_imb)) {
2489 max_load = avg_load;
2491 busiest_nr_running = sum_nr_running;
2492 busiest_load_per_task = sum_weighted_load;
2493 group_imb = __group_imb;
2496 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2498 * Busy processors will not participate in power savings
2501 if (idle == CPU_NOT_IDLE ||
2502 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2506 * If the local group is idle or completely loaded
2507 * no need to do power savings balance at this domain
2509 if (local_group && (this_nr_running >= group_capacity ||
2511 power_savings_balance = 0;
2514 * If a group is already running at full capacity or idle,
2515 * don't include that group in power savings calculations
2517 if (!power_savings_balance || sum_nr_running >= group_capacity
2522 * Calculate the group which has the least non-idle load.
2523 * This is the group from where we need to pick up the load
2526 if ((sum_nr_running < min_nr_running) ||
2527 (sum_nr_running == min_nr_running &&
2528 first_cpu(group->cpumask) <
2529 first_cpu(group_min->cpumask))) {
2531 min_nr_running = sum_nr_running;
2532 min_load_per_task = sum_weighted_load /
2537 * Calculate the group which is almost near its
2538 * capacity but still has some space to pick up some load
2539 * from other group and save more power
2541 if (sum_nr_running <= group_capacity - 1) {
2542 if (sum_nr_running > leader_nr_running ||
2543 (sum_nr_running == leader_nr_running &&
2544 first_cpu(group->cpumask) >
2545 first_cpu(group_leader->cpumask))) {
2546 group_leader = group;
2547 leader_nr_running = sum_nr_running;
2552 group = group->next;
2553 } while (group != sd->groups);
2555 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2558 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2560 if (this_load >= avg_load ||
2561 100*max_load <= sd->imbalance_pct*this_load)
2564 busiest_load_per_task /= busiest_nr_running;
2566 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2569 * We're trying to get all the cpus to the average_load, so we don't
2570 * want to push ourselves above the average load, nor do we wish to
2571 * reduce the max loaded cpu below the average load, as either of these
2572 * actions would just result in more rebalancing later, and ping-pong
2573 * tasks around. Thus we look for the minimum possible imbalance.
2574 * Negative imbalances (*we* are more loaded than anyone else) will
2575 * be counted as no imbalance for these purposes -- we can't fix that
2576 * by pulling tasks to us. Be careful of negative numbers as they'll
2577 * appear as very large values with unsigned longs.
2579 if (max_load <= busiest_load_per_task)
2583 * In the presence of smp nice balancing, certain scenarios can have
2584 * max load less than avg load(as we skip the groups at or below
2585 * its cpu_power, while calculating max_load..)
2587 if (max_load < avg_load) {
2589 goto small_imbalance;
2592 /* Don't want to pull so many tasks that a group would go idle */
2593 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2595 /* How much load to actually move to equalise the imbalance */
2596 *imbalance = min(max_pull * busiest->__cpu_power,
2597 (avg_load - this_load) * this->__cpu_power)
2601 * if *imbalance is less than the average load per runnable task
2602 * there is no gaurantee that any tasks will be moved so we'll have
2603 * a think about bumping its value to force at least one task to be
2606 if (*imbalance < busiest_load_per_task) {
2607 unsigned long tmp, pwr_now, pwr_move;
2611 pwr_move = pwr_now = 0;
2613 if (this_nr_running) {
2614 this_load_per_task /= this_nr_running;
2615 if (busiest_load_per_task > this_load_per_task)
2618 this_load_per_task = SCHED_LOAD_SCALE;
2620 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2621 busiest_load_per_task * imbn) {
2622 *imbalance = busiest_load_per_task;
2627 * OK, we don't have enough imbalance to justify moving tasks,
2628 * however we may be able to increase total CPU power used by
2632 pwr_now += busiest->__cpu_power *
2633 min(busiest_load_per_task, max_load);
2634 pwr_now += this->__cpu_power *
2635 min(this_load_per_task, this_load);
2636 pwr_now /= SCHED_LOAD_SCALE;
2638 /* Amount of load we'd subtract */
2639 tmp = sg_div_cpu_power(busiest,
2640 busiest_load_per_task * SCHED_LOAD_SCALE);
2642 pwr_move += busiest->__cpu_power *
2643 min(busiest_load_per_task, max_load - tmp);
2645 /* Amount of load we'd add */
2646 if (max_load * busiest->__cpu_power <
2647 busiest_load_per_task * SCHED_LOAD_SCALE)
2648 tmp = sg_div_cpu_power(this,
2649 max_load * busiest->__cpu_power);
2651 tmp = sg_div_cpu_power(this,
2652 busiest_load_per_task * SCHED_LOAD_SCALE);
2653 pwr_move += this->__cpu_power *
2654 min(this_load_per_task, this_load + tmp);
2655 pwr_move /= SCHED_LOAD_SCALE;
2657 /* Move if we gain throughput */
2658 if (pwr_move > pwr_now)
2659 *imbalance = busiest_load_per_task;
2665 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2666 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2669 if (this == group_leader && group_leader != group_min) {
2670 *imbalance = min_load_per_task;
2680 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2683 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2684 unsigned long imbalance, cpumask_t *cpus)
2686 struct rq *busiest = NULL, *rq;
2687 unsigned long max_load = 0;
2690 for_each_cpu_mask(i, group->cpumask) {
2693 if (!cpu_isset(i, *cpus))
2697 wl = weighted_cpuload(i);
2699 if (rq->nr_running == 1 && wl > imbalance)
2702 if (wl > max_load) {
2712 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2713 * so long as it is large enough.
2715 #define MAX_PINNED_INTERVAL 512
2718 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2719 * tasks if there is an imbalance.
2721 static int load_balance(int this_cpu, struct rq *this_rq,
2722 struct sched_domain *sd, enum cpu_idle_type idle,
2725 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2726 struct sched_group *group;
2727 unsigned long imbalance;
2729 cpumask_t cpus = CPU_MASK_ALL;
2730 unsigned long flags;
2733 * When power savings policy is enabled for the parent domain, idle
2734 * sibling can pick up load irrespective of busy siblings. In this case,
2735 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2736 * portraying it as CPU_NOT_IDLE.
2738 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2739 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2742 schedstat_inc(sd, lb_count[idle]);
2745 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2752 schedstat_inc(sd, lb_nobusyg[idle]);
2756 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2758 schedstat_inc(sd, lb_nobusyq[idle]);
2762 BUG_ON(busiest == this_rq);
2764 schedstat_add(sd, lb_imbalance[idle], imbalance);
2767 if (busiest->nr_running > 1) {
2769 * Attempt to move tasks. If find_busiest_group has found
2770 * an imbalance but busiest->nr_running <= 1, the group is
2771 * still unbalanced. ld_moved simply stays zero, so it is
2772 * correctly treated as an imbalance.
2774 local_irq_save(flags);
2775 double_rq_lock(this_rq, busiest);
2776 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2777 imbalance, sd, idle, &all_pinned);
2778 double_rq_unlock(this_rq, busiest);
2779 local_irq_restore(flags);
2782 * some other cpu did the load balance for us.
2784 if (ld_moved && this_cpu != smp_processor_id())
2785 resched_cpu(this_cpu);
2787 /* All tasks on this runqueue were pinned by CPU affinity */
2788 if (unlikely(all_pinned)) {
2789 cpu_clear(cpu_of(busiest), cpus);
2790 if (!cpus_empty(cpus))
2797 schedstat_inc(sd, lb_failed[idle]);
2798 sd->nr_balance_failed++;
2800 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2802 spin_lock_irqsave(&busiest->lock, flags);
2804 /* don't kick the migration_thread, if the curr
2805 * task on busiest cpu can't be moved to this_cpu
2807 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2808 spin_unlock_irqrestore(&busiest->lock, flags);
2810 goto out_one_pinned;
2813 if (!busiest->active_balance) {
2814 busiest->active_balance = 1;
2815 busiest->push_cpu = this_cpu;
2818 spin_unlock_irqrestore(&busiest->lock, flags);
2820 wake_up_process(busiest->migration_thread);
2823 * We've kicked active balancing, reset the failure
2826 sd->nr_balance_failed = sd->cache_nice_tries+1;
2829 sd->nr_balance_failed = 0;
2831 if (likely(!active_balance)) {
2832 /* We were unbalanced, so reset the balancing interval */
2833 sd->balance_interval = sd->min_interval;
2836 * If we've begun active balancing, start to back off. This
2837 * case may not be covered by the all_pinned logic if there
2838 * is only 1 task on the busy runqueue (because we don't call
2841 if (sd->balance_interval < sd->max_interval)
2842 sd->balance_interval *= 2;
2845 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2846 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2851 schedstat_inc(sd, lb_balanced[idle]);
2853 sd->nr_balance_failed = 0;
2856 /* tune up the balancing interval */
2857 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2858 (sd->balance_interval < sd->max_interval))
2859 sd->balance_interval *= 2;
2861 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2862 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2868 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2869 * tasks if there is an imbalance.
2871 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2872 * this_rq is locked.
2875 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2877 struct sched_group *group;
2878 struct rq *busiest = NULL;
2879 unsigned long imbalance;
2883 cpumask_t cpus = CPU_MASK_ALL;
2886 * When power savings policy is enabled for the parent domain, idle
2887 * sibling can pick up load irrespective of busy siblings. In this case,
2888 * let the state of idle sibling percolate up as IDLE, instead of
2889 * portraying it as CPU_NOT_IDLE.
2891 if (sd->flags & SD_SHARE_CPUPOWER &&
2892 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2895 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2897 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2898 &sd_idle, &cpus, NULL);
2900 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2904 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2907 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2911 BUG_ON(busiest == this_rq);
2913 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2916 if (busiest->nr_running > 1) {
2917 /* Attempt to move tasks */
2918 double_lock_balance(this_rq, busiest);
2919 /* this_rq->clock is already updated */
2920 update_rq_clock(busiest);
2921 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2922 imbalance, sd, CPU_NEWLY_IDLE,
2924 spin_unlock(&busiest->lock);
2926 if (unlikely(all_pinned)) {
2927 cpu_clear(cpu_of(busiest), cpus);
2928 if (!cpus_empty(cpus))
2934 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2935 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2936 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2939 sd->nr_balance_failed = 0;
2944 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2945 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2946 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2948 sd->nr_balance_failed = 0;
2954 * idle_balance is called by schedule() if this_cpu is about to become
2955 * idle. Attempts to pull tasks from other CPUs.
2957 static void idle_balance(int this_cpu, struct rq *this_rq)
2959 struct sched_domain *sd;
2960 int pulled_task = -1;
2961 unsigned long next_balance = jiffies + HZ;
2963 for_each_domain(this_cpu, sd) {
2964 unsigned long interval;
2966 if (!(sd->flags & SD_LOAD_BALANCE))
2969 if (sd->flags & SD_BALANCE_NEWIDLE)
2970 /* If we've pulled tasks over stop searching: */
2971 pulled_task = load_balance_newidle(this_cpu,
2974 interval = msecs_to_jiffies(sd->balance_interval);
2975 if (time_after(next_balance, sd->last_balance + interval))
2976 next_balance = sd->last_balance + interval;
2980 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2982 * We are going idle. next_balance may be set based on
2983 * a busy processor. So reset next_balance.
2985 this_rq->next_balance = next_balance;
2990 * active_load_balance is run by migration threads. It pushes running tasks
2991 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2992 * running on each physical CPU where possible, and avoids physical /
2993 * logical imbalances.
2995 * Called with busiest_rq locked.
2997 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2999 int target_cpu = busiest_rq->push_cpu;
3000 struct sched_domain *sd;
3001 struct rq *target_rq;
3003 /* Is there any task to move? */
3004 if (busiest_rq->nr_running <= 1)
3007 target_rq = cpu_rq(target_cpu);
3010 * This condition is "impossible", if it occurs
3011 * we need to fix it. Originally reported by
3012 * Bjorn Helgaas on a 128-cpu setup.
3014 BUG_ON(busiest_rq == target_rq);
3016 /* move a task from busiest_rq to target_rq */
3017 double_lock_balance(busiest_rq, target_rq);
3018 update_rq_clock(busiest_rq);
3019 update_rq_clock(target_rq);
3021 /* Search for an sd spanning us and the target CPU. */
3022 for_each_domain(target_cpu, sd) {
3023 if ((sd->flags & SD_LOAD_BALANCE) &&
3024 cpu_isset(busiest_cpu, sd->span))
3029 schedstat_inc(sd, alb_count);
3031 if (move_one_task(target_rq, target_cpu, busiest_rq,
3033 schedstat_inc(sd, alb_pushed);
3035 schedstat_inc(sd, alb_failed);
3037 spin_unlock(&target_rq->lock);
3042 atomic_t load_balancer;
3044 } nohz ____cacheline_aligned = {
3045 .load_balancer = ATOMIC_INIT(-1),
3046 .cpu_mask = CPU_MASK_NONE,
3050 * This routine will try to nominate the ilb (idle load balancing)
3051 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3052 * load balancing on behalf of all those cpus. If all the cpus in the system
3053 * go into this tickless mode, then there will be no ilb owner (as there is
3054 * no need for one) and all the cpus will sleep till the next wakeup event
3057 * For the ilb owner, tick is not stopped. And this tick will be used
3058 * for idle load balancing. ilb owner will still be part of
3061 * While stopping the tick, this cpu will become the ilb owner if there
3062 * is no other owner. And will be the owner till that cpu becomes busy
3063 * or if all cpus in the system stop their ticks at which point
3064 * there is no need for ilb owner.
3066 * When the ilb owner becomes busy, it nominates another owner, during the
3067 * next busy scheduler_tick()
3069 int select_nohz_load_balancer(int stop_tick)
3071 int cpu = smp_processor_id();
3074 cpu_set(cpu, nohz.cpu_mask);
3075 cpu_rq(cpu)->in_nohz_recently = 1;
3078 * If we are going offline and still the leader, give up!
3080 if (cpu_is_offline(cpu) &&
3081 atomic_read(&nohz.load_balancer) == cpu) {
3082 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3087 /* time for ilb owner also to sleep */
3088 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3089 if (atomic_read(&nohz.load_balancer) == cpu)
3090 atomic_set(&nohz.load_balancer, -1);
3094 if (atomic_read(&nohz.load_balancer) == -1) {
3095 /* make me the ilb owner */
3096 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3098 } else if (atomic_read(&nohz.load_balancer) == cpu)
3101 if (!cpu_isset(cpu, nohz.cpu_mask))
3104 cpu_clear(cpu, nohz.cpu_mask);
3106 if (atomic_read(&nohz.load_balancer) == cpu)
3107 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3114 static DEFINE_SPINLOCK(balancing);
3117 * It checks each scheduling domain to see if it is due to be balanced,
3118 * and initiates a balancing operation if so.
3120 * Balancing parameters are set up in arch_init_sched_domains.
3122 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3125 struct rq *rq = cpu_rq(cpu);
3126 unsigned long interval;
3127 struct sched_domain *sd;
3128 /* Earliest time when we have to do rebalance again */
3129 unsigned long next_balance = jiffies + 60*HZ;
3130 int update_next_balance = 0;
3132 for_each_domain(cpu, sd) {
3133 if (!(sd->flags & SD_LOAD_BALANCE))
3136 interval = sd->balance_interval;
3137 if (idle != CPU_IDLE)
3138 interval *= sd->busy_factor;
3140 /* scale ms to jiffies */
3141 interval = msecs_to_jiffies(interval);
3142 if (unlikely(!interval))
3144 if (interval > HZ*NR_CPUS/10)
3145 interval = HZ*NR_CPUS/10;
3148 if (sd->flags & SD_SERIALIZE) {
3149 if (!spin_trylock(&balancing))
3153 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3154 if (load_balance(cpu, rq, sd, idle, &balance)) {
3156 * We've pulled tasks over so either we're no
3157 * longer idle, or one of our SMT siblings is
3160 idle = CPU_NOT_IDLE;
3162 sd->last_balance = jiffies;
3164 if (sd->flags & SD_SERIALIZE)
3165 spin_unlock(&balancing);
3167 if (time_after(next_balance, sd->last_balance + interval)) {
3168 next_balance = sd->last_balance + interval;
3169 update_next_balance = 1;
3173 * Stop the load balance at this level. There is another
3174 * CPU in our sched group which is doing load balancing more
3182 * next_balance will be updated only when there is a need.
3183 * When the cpu is attached to null domain for ex, it will not be
3186 if (likely(update_next_balance))
3187 rq->next_balance = next_balance;
3191 * run_rebalance_domains is triggered when needed from the scheduler tick.
3192 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3193 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3195 static void run_rebalance_domains(struct softirq_action *h)
3197 int this_cpu = smp_processor_id();
3198 struct rq *this_rq = cpu_rq(this_cpu);
3199 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3200 CPU_IDLE : CPU_NOT_IDLE;
3202 rebalance_domains(this_cpu, idle);
3206 * If this cpu is the owner for idle load balancing, then do the
3207 * balancing on behalf of the other idle cpus whose ticks are
3210 if (this_rq->idle_at_tick &&
3211 atomic_read(&nohz.load_balancer) == this_cpu) {
3212 cpumask_t cpus = nohz.cpu_mask;
3216 cpu_clear(this_cpu, cpus);
3217 for_each_cpu_mask(balance_cpu, cpus) {
3219 * If this cpu gets work to do, stop the load balancing
3220 * work being done for other cpus. Next load
3221 * balancing owner will pick it up.
3226 rebalance_domains(balance_cpu, CPU_IDLE);
3228 rq = cpu_rq(balance_cpu);
3229 if (time_after(this_rq->next_balance, rq->next_balance))
3230 this_rq->next_balance = rq->next_balance;
3237 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3239 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3240 * idle load balancing owner or decide to stop the periodic load balancing,
3241 * if the whole system is idle.
3243 static inline void trigger_load_balance(struct rq *rq, int cpu)
3247 * If we were in the nohz mode recently and busy at the current
3248 * scheduler tick, then check if we need to nominate new idle
3251 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3252 rq->in_nohz_recently = 0;
3254 if (atomic_read(&nohz.load_balancer) == cpu) {
3255 cpu_clear(cpu, nohz.cpu_mask);
3256 atomic_set(&nohz.load_balancer, -1);
3259 if (atomic_read(&nohz.load_balancer) == -1) {
3261 * simple selection for now: Nominate the
3262 * first cpu in the nohz list to be the next
3265 * TBD: Traverse the sched domains and nominate
3266 * the nearest cpu in the nohz.cpu_mask.
3268 int ilb = first_cpu(nohz.cpu_mask);
3276 * If this cpu is idle and doing idle load balancing for all the
3277 * cpus with ticks stopped, is it time for that to stop?
3279 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3280 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3286 * If this cpu is idle and the idle load balancing is done by
3287 * someone else, then no need raise the SCHED_SOFTIRQ
3289 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3290 cpu_isset(cpu, nohz.cpu_mask))
3293 if (time_after_eq(jiffies, rq->next_balance))
3294 raise_softirq(SCHED_SOFTIRQ);
3297 #else /* CONFIG_SMP */
3300 * on UP we do not need to balance between CPUs:
3302 static inline void idle_balance(int cpu, struct rq *rq)
3308 DEFINE_PER_CPU(struct kernel_stat, kstat);
3310 EXPORT_PER_CPU_SYMBOL(kstat);
3313 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3314 * that have not yet been banked in case the task is currently running.
3316 unsigned long long task_sched_runtime(struct task_struct *p)
3318 unsigned long flags;
3322 rq = task_rq_lock(p, &flags);
3323 ns = p->se.sum_exec_runtime;
3324 if (task_current(rq, p)) {
3325 update_rq_clock(rq);
3326 delta_exec = rq->clock - p->se.exec_start;
3327 if ((s64)delta_exec > 0)
3330 task_rq_unlock(rq, &flags);
3336 * Account user cpu time to a process.
3337 * @p: the process that the cpu time gets accounted to
3338 * @cputime: the cpu time spent in user space since the last update
3340 void account_user_time(struct task_struct *p, cputime_t cputime)
3342 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3345 p->utime = cputime_add(p->utime, cputime);
3347 /* Add user time to cpustat. */
3348 tmp = cputime_to_cputime64(cputime);
3349 if (TASK_NICE(p) > 0)
3350 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3352 cpustat->user = cputime64_add(cpustat->user, tmp);
3356 * Account guest cpu time to a process.
3357 * @p: the process that the cpu time gets accounted to
3358 * @cputime: the cpu time spent in virtual machine since the last update
3360 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3363 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3365 tmp = cputime_to_cputime64(cputime);
3367 p->utime = cputime_add(p->utime, cputime);
3368 p->gtime = cputime_add(p->gtime, cputime);
3370 cpustat->user = cputime64_add(cpustat->user, tmp);
3371 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3375 * Account scaled user cpu time to a process.
3376 * @p: the process that the cpu time gets accounted to
3377 * @cputime: the cpu time spent in user space since the last update
3379 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3381 p->utimescaled = cputime_add(p->utimescaled, cputime);
3385 * Account system cpu time to a process.
3386 * @p: the process that the cpu time gets accounted to
3387 * @hardirq_offset: the offset to subtract from hardirq_count()
3388 * @cputime: the cpu time spent in kernel space since the last update
3390 void account_system_time(struct task_struct *p, int hardirq_offset,
3393 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3394 struct rq *rq = this_rq();
3397 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3398 return account_guest_time(p, cputime);
3400 p->stime = cputime_add(p->stime, cputime);
3402 /* Add system time to cpustat. */
3403 tmp = cputime_to_cputime64(cputime);
3404 if (hardirq_count() - hardirq_offset)
3405 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3406 else if (softirq_count())
3407 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3408 else if (p != rq->idle)
3409 cpustat->system = cputime64_add(cpustat->system, tmp);
3410 else if (atomic_read(&rq->nr_iowait) > 0)
3411 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3413 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3414 /* Account for system time used */
3415 acct_update_integrals(p);
3419 * Account scaled system cpu time to a process.
3420 * @p: the process that the cpu time gets accounted to
3421 * @hardirq_offset: the offset to subtract from hardirq_count()
3422 * @cputime: the cpu time spent in kernel space since the last update
3424 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3426 p->stimescaled = cputime_add(p->stimescaled, cputime);
3430 * Account for involuntary wait time.
3431 * @p: the process from which the cpu time has been stolen
3432 * @steal: the cpu time spent in involuntary wait
3434 void account_steal_time(struct task_struct *p, cputime_t steal)
3436 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3437 cputime64_t tmp = cputime_to_cputime64(steal);
3438 struct rq *rq = this_rq();
3440 if (p == rq->idle) {
3441 p->stime = cputime_add(p->stime, steal);
3442 if (atomic_read(&rq->nr_iowait) > 0)
3443 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3445 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3447 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3451 * This function gets called by the timer code, with HZ frequency.
3452 * We call it with interrupts disabled.
3454 * It also gets called by the fork code, when changing the parent's
3457 void scheduler_tick(void)
3459 int cpu = smp_processor_id();
3460 struct rq *rq = cpu_rq(cpu);
3461 struct task_struct *curr = rq->curr;
3462 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3464 spin_lock(&rq->lock);
3465 __update_rq_clock(rq);
3467 * Let rq->clock advance by at least TICK_NSEC:
3469 if (unlikely(rq->clock < next_tick))
3470 rq->clock = next_tick;
3471 rq->tick_timestamp = rq->clock;
3472 update_cpu_load(rq);
3473 if (curr != rq->idle) /* FIXME: needed? */
3474 curr->sched_class->task_tick(rq, curr);
3475 spin_unlock(&rq->lock);
3478 rq->idle_at_tick = idle_cpu(cpu);
3479 trigger_load_balance(rq, cpu);
3483 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3485 void fastcall add_preempt_count(int val)
3490 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3492 preempt_count() += val;
3494 * Spinlock count overflowing soon?
3496 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3499 EXPORT_SYMBOL(add_preempt_count);
3501 void fastcall sub_preempt_count(int val)
3506 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3509 * Is the spinlock portion underflowing?
3511 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3512 !(preempt_count() & PREEMPT_MASK)))
3515 preempt_count() -= val;
3517 EXPORT_SYMBOL(sub_preempt_count);
3522 * Print scheduling while atomic bug:
3524 static noinline void __schedule_bug(struct task_struct *prev)
3526 struct pt_regs *regs = get_irq_regs();
3528 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3529 prev->comm, prev->pid, preempt_count());
3531 debug_show_held_locks(prev);
3532 if (irqs_disabled())
3533 print_irqtrace_events(prev);
3542 * Various schedule()-time debugging checks and statistics:
3544 static inline void schedule_debug(struct task_struct *prev)
3547 * Test if we are atomic. Since do_exit() needs to call into
3548 * schedule() atomically, we ignore that path for now.
3549 * Otherwise, whine if we are scheduling when we should not be.
3551 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3552 __schedule_bug(prev);
3554 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3556 schedstat_inc(this_rq(), sched_count);
3557 #ifdef CONFIG_SCHEDSTATS
3558 if (unlikely(prev->lock_depth >= 0)) {
3559 schedstat_inc(this_rq(), bkl_count);
3560 schedstat_inc(prev, sched_info.bkl_count);
3566 * Pick up the highest-prio task:
3568 static inline struct task_struct *
3569 pick_next_task(struct rq *rq, struct task_struct *prev)
3571 const struct sched_class *class;
3572 struct task_struct *p;
3575 * Optimization: we know that if all tasks are in
3576 * the fair class we can call that function directly:
3578 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3579 p = fair_sched_class.pick_next_task(rq);
3584 class = sched_class_highest;
3586 p = class->pick_next_task(rq);
3590 * Will never be NULL as the idle class always
3591 * returns a non-NULL p:
3593 class = class->next;
3598 * schedule() is the main scheduler function.
3600 asmlinkage void __sched schedule(void)
3602 struct task_struct *prev, *next;
3609 cpu = smp_processor_id();
3613 switch_count = &prev->nivcsw;
3615 release_kernel_lock(prev);
3616 need_resched_nonpreemptible:
3618 schedule_debug(prev);
3621 * Do the rq-clock update outside the rq lock:
3623 local_irq_disable();
3624 __update_rq_clock(rq);
3625 spin_lock(&rq->lock);
3626 clear_tsk_need_resched(prev);
3628 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3629 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3630 unlikely(signal_pending(prev)))) {
3631 prev->state = TASK_RUNNING;
3633 deactivate_task(rq, prev, 1);
3635 switch_count = &prev->nvcsw;
3638 schedule_balance_rt(rq, prev);
3640 if (unlikely(!rq->nr_running))
3641 idle_balance(cpu, rq);
3643 prev->sched_class->put_prev_task(rq, prev);
3644 next = pick_next_task(rq, prev);
3646 sched_info_switch(prev, next);
3648 if (likely(prev != next)) {
3653 context_switch(rq, prev, next); /* unlocks the rq */
3655 spin_unlock_irq(&rq->lock);
3657 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3658 cpu = smp_processor_id();
3660 goto need_resched_nonpreemptible;
3662 preempt_enable_no_resched();
3663 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3666 EXPORT_SYMBOL(schedule);
3668 #ifdef CONFIG_PREEMPT
3670 * this is the entry point to schedule() from in-kernel preemption
3671 * off of preempt_enable. Kernel preemptions off return from interrupt
3672 * occur there and call schedule directly.
3674 asmlinkage void __sched preempt_schedule(void)
3676 struct thread_info *ti = current_thread_info();
3677 #ifdef CONFIG_PREEMPT_BKL
3678 struct task_struct *task = current;
3679 int saved_lock_depth;
3682 * If there is a non-zero preempt_count or interrupts are disabled,
3683 * we do not want to preempt the current task. Just return..
3685 if (likely(ti->preempt_count || irqs_disabled()))
3689 add_preempt_count(PREEMPT_ACTIVE);
3692 * We keep the big kernel semaphore locked, but we
3693 * clear ->lock_depth so that schedule() doesnt
3694 * auto-release the semaphore:
3696 #ifdef CONFIG_PREEMPT_BKL
3697 saved_lock_depth = task->lock_depth;
3698 task->lock_depth = -1;
3701 #ifdef CONFIG_PREEMPT_BKL
3702 task->lock_depth = saved_lock_depth;
3704 sub_preempt_count(PREEMPT_ACTIVE);
3707 * Check again in case we missed a preemption opportunity
3708 * between schedule and now.
3711 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3713 EXPORT_SYMBOL(preempt_schedule);
3716 * this is the entry point to schedule() from kernel preemption
3717 * off of irq context.
3718 * Note, that this is called and return with irqs disabled. This will
3719 * protect us against recursive calling from irq.
3721 asmlinkage void __sched preempt_schedule_irq(void)
3723 struct thread_info *ti = current_thread_info();
3724 #ifdef CONFIG_PREEMPT_BKL
3725 struct task_struct *task = current;
3726 int saved_lock_depth;
3728 /* Catch callers which need to be fixed */
3729 BUG_ON(ti->preempt_count || !irqs_disabled());
3732 add_preempt_count(PREEMPT_ACTIVE);
3735 * We keep the big kernel semaphore locked, but we
3736 * clear ->lock_depth so that schedule() doesnt
3737 * auto-release the semaphore:
3739 #ifdef CONFIG_PREEMPT_BKL
3740 saved_lock_depth = task->lock_depth;
3741 task->lock_depth = -1;
3745 local_irq_disable();
3746 #ifdef CONFIG_PREEMPT_BKL
3747 task->lock_depth = saved_lock_depth;
3749 sub_preempt_count(PREEMPT_ACTIVE);
3752 * Check again in case we missed a preemption opportunity
3753 * between schedule and now.
3756 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3759 #endif /* CONFIG_PREEMPT */
3761 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3764 return try_to_wake_up(curr->private, mode, sync);
3766 EXPORT_SYMBOL(default_wake_function);
3769 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3770 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3771 * number) then we wake all the non-exclusive tasks and one exclusive task.
3773 * There are circumstances in which we can try to wake a task which has already
3774 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3775 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3777 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3778 int nr_exclusive, int sync, void *key)
3780 wait_queue_t *curr, *next;
3782 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3783 unsigned flags = curr->flags;
3785 if (curr->func(curr, mode, sync, key) &&
3786 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3792 * __wake_up - wake up threads blocked on a waitqueue.
3794 * @mode: which threads
3795 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3796 * @key: is directly passed to the wakeup function
3798 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3799 int nr_exclusive, void *key)
3801 unsigned long flags;
3803 spin_lock_irqsave(&q->lock, flags);
3804 __wake_up_common(q, mode, nr_exclusive, 0, key);
3805 spin_unlock_irqrestore(&q->lock, flags);
3807 EXPORT_SYMBOL(__wake_up);
3810 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3812 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3814 __wake_up_common(q, mode, 1, 0, NULL);
3818 * __wake_up_sync - wake up threads blocked on a waitqueue.
3820 * @mode: which threads
3821 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3823 * The sync wakeup differs that the waker knows that it will schedule
3824 * away soon, so while the target thread will be woken up, it will not
3825 * be migrated to another CPU - ie. the two threads are 'synchronized'
3826 * with each other. This can prevent needless bouncing between CPUs.
3828 * On UP it can prevent extra preemption.
3831 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3833 unsigned long flags;
3839 if (unlikely(!nr_exclusive))
3842 spin_lock_irqsave(&q->lock, flags);
3843 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3844 spin_unlock_irqrestore(&q->lock, flags);
3846 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3848 void complete(struct completion *x)
3850 unsigned long flags;
3852 spin_lock_irqsave(&x->wait.lock, flags);
3854 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3856 spin_unlock_irqrestore(&x->wait.lock, flags);
3858 EXPORT_SYMBOL(complete);
3860 void complete_all(struct completion *x)
3862 unsigned long flags;
3864 spin_lock_irqsave(&x->wait.lock, flags);
3865 x->done += UINT_MAX/2;
3866 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3868 spin_unlock_irqrestore(&x->wait.lock, flags);
3870 EXPORT_SYMBOL(complete_all);
3872 static inline long __sched
3873 do_wait_for_common(struct completion *x, long timeout, int state)
3876 DECLARE_WAITQUEUE(wait, current);
3878 wait.flags |= WQ_FLAG_EXCLUSIVE;
3879 __add_wait_queue_tail(&x->wait, &wait);
3881 if (state == TASK_INTERRUPTIBLE &&
3882 signal_pending(current)) {
3883 __remove_wait_queue(&x->wait, &wait);
3884 return -ERESTARTSYS;
3886 __set_current_state(state);
3887 spin_unlock_irq(&x->wait.lock);
3888 timeout = schedule_timeout(timeout);
3889 spin_lock_irq(&x->wait.lock);
3891 __remove_wait_queue(&x->wait, &wait);
3895 __remove_wait_queue(&x->wait, &wait);
3902 wait_for_common(struct completion *x, long timeout, int state)
3906 spin_lock_irq(&x->wait.lock);
3907 timeout = do_wait_for_common(x, timeout, state);
3908 spin_unlock_irq(&x->wait.lock);
3912 void __sched wait_for_completion(struct completion *x)
3914 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3916 EXPORT_SYMBOL(wait_for_completion);
3918 unsigned long __sched
3919 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3921 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3923 EXPORT_SYMBOL(wait_for_completion_timeout);
3925 int __sched wait_for_completion_interruptible(struct completion *x)
3927 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3928 if (t == -ERESTARTSYS)
3932 EXPORT_SYMBOL(wait_for_completion_interruptible);
3934 unsigned long __sched
3935 wait_for_completion_interruptible_timeout(struct completion *x,
3936 unsigned long timeout)
3938 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3940 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3943 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3945 unsigned long flags;
3948 init_waitqueue_entry(&wait, current);
3950 __set_current_state(state);
3952 spin_lock_irqsave(&q->lock, flags);
3953 __add_wait_queue(q, &wait);
3954 spin_unlock(&q->lock);
3955 timeout = schedule_timeout(timeout);
3956 spin_lock_irq(&q->lock);
3957 __remove_wait_queue(q, &wait);
3958 spin_unlock_irqrestore(&q->lock, flags);
3963 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3965 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3967 EXPORT_SYMBOL(interruptible_sleep_on);
3970 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3972 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3974 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3976 void __sched sleep_on(wait_queue_head_t *q)
3978 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3980 EXPORT_SYMBOL(sleep_on);
3982 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3984 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3986 EXPORT_SYMBOL(sleep_on_timeout);
3988 #ifdef CONFIG_RT_MUTEXES
3991 * rt_mutex_setprio - set the current priority of a task
3993 * @prio: prio value (kernel-internal form)
3995 * This function changes the 'effective' priority of a task. It does
3996 * not touch ->normal_prio like __setscheduler().
3998 * Used by the rt_mutex code to implement priority inheritance logic.
4000 void rt_mutex_setprio(struct task_struct *p, int prio)
4002 unsigned long flags;
4003 int oldprio, on_rq, running;
4006 BUG_ON(prio < 0 || prio > MAX_PRIO);
4008 rq = task_rq_lock(p, &flags);
4009 update_rq_clock(rq);
4012 on_rq = p->se.on_rq;
4013 running = task_current(rq, p);
4015 dequeue_task(rq, p, 0);
4017 p->sched_class->put_prev_task(rq, p);
4021 p->sched_class = &rt_sched_class;
4023 p->sched_class = &fair_sched_class;
4029 p->sched_class->set_curr_task(rq);
4030 enqueue_task(rq, p, 0);
4032 * Reschedule if we are currently running on this runqueue and
4033 * our priority decreased, or if we are not currently running on
4034 * this runqueue and our priority is higher than the current's
4037 if (p->prio > oldprio)
4038 resched_task(rq->curr);
4040 check_preempt_curr(rq, p);
4043 task_rq_unlock(rq, &flags);
4048 void set_user_nice(struct task_struct *p, long nice)
4050 int old_prio, delta, on_rq;
4051 unsigned long flags;
4054 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4057 * We have to be careful, if called from sys_setpriority(),
4058 * the task might be in the middle of scheduling on another CPU.
4060 rq = task_rq_lock(p, &flags);
4061 update_rq_clock(rq);
4063 * The RT priorities are set via sched_setscheduler(), but we still
4064 * allow the 'normal' nice value to be set - but as expected
4065 * it wont have any effect on scheduling until the task is
4066 * SCHED_FIFO/SCHED_RR:
4068 if (task_has_rt_policy(p)) {
4069 p->static_prio = NICE_TO_PRIO(nice);
4072 on_rq = p->se.on_rq;
4074 dequeue_task(rq, p, 0);
4076 p->static_prio = NICE_TO_PRIO(nice);
4079 p->prio = effective_prio(p);
4080 delta = p->prio - old_prio;
4083 enqueue_task(rq, p, 0);
4085 * If the task increased its priority or is running and
4086 * lowered its priority, then reschedule its CPU:
4088 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4089 resched_task(rq->curr);
4092 task_rq_unlock(rq, &flags);
4094 EXPORT_SYMBOL(set_user_nice);
4097 * can_nice - check if a task can reduce its nice value
4101 int can_nice(const struct task_struct *p, const int nice)
4103 /* convert nice value [19,-20] to rlimit style value [1,40] */
4104 int nice_rlim = 20 - nice;
4106 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4107 capable(CAP_SYS_NICE));
4110 #ifdef __ARCH_WANT_SYS_NICE
4113 * sys_nice - change the priority of the current process.
4114 * @increment: priority increment
4116 * sys_setpriority is a more generic, but much slower function that
4117 * does similar things.
4119 asmlinkage long sys_nice(int increment)
4124 * Setpriority might change our priority at the same moment.
4125 * We don't have to worry. Conceptually one call occurs first
4126 * and we have a single winner.
4128 if (increment < -40)
4133 nice = PRIO_TO_NICE(current->static_prio) + increment;
4139 if (increment < 0 && !can_nice(current, nice))
4142 retval = security_task_setnice(current, nice);
4146 set_user_nice(current, nice);
4153 * task_prio - return the priority value of a given task.
4154 * @p: the task in question.
4156 * This is the priority value as seen by users in /proc.
4157 * RT tasks are offset by -200. Normal tasks are centered
4158 * around 0, value goes from -16 to +15.
4160 int task_prio(const struct task_struct *p)
4162 return p->prio - MAX_RT_PRIO;
4166 * task_nice - return the nice value of a given task.
4167 * @p: the task in question.
4169 int task_nice(const struct task_struct *p)
4171 return TASK_NICE(p);
4173 EXPORT_SYMBOL_GPL(task_nice);
4176 * idle_cpu - is a given cpu idle currently?
4177 * @cpu: the processor in question.
4179 int idle_cpu(int cpu)
4181 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4185 * idle_task - return the idle task for a given cpu.
4186 * @cpu: the processor in question.
4188 struct task_struct *idle_task(int cpu)
4190 return cpu_rq(cpu)->idle;
4194 * find_process_by_pid - find a process with a matching PID value.
4195 * @pid: the pid in question.
4197 static struct task_struct *find_process_by_pid(pid_t pid)
4199 return pid ? find_task_by_vpid(pid) : current;
4202 /* Actually do priority change: must hold rq lock. */
4204 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4206 BUG_ON(p->se.on_rq);
4209 switch (p->policy) {
4213 p->sched_class = &fair_sched_class;
4217 p->sched_class = &rt_sched_class;
4221 p->rt_priority = prio;
4222 p->normal_prio = normal_prio(p);
4223 /* we are holding p->pi_lock already */
4224 p->prio = rt_mutex_getprio(p);
4229 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4230 * @p: the task in question.
4231 * @policy: new policy.
4232 * @param: structure containing the new RT priority.
4234 * NOTE that the task may be already dead.
4236 int sched_setscheduler(struct task_struct *p, int policy,
4237 struct sched_param *param)
4239 int retval, oldprio, oldpolicy = -1, on_rq, running;
4240 unsigned long flags;
4243 /* may grab non-irq protected spin_locks */
4244 BUG_ON(in_interrupt());
4246 /* double check policy once rq lock held */
4248 policy = oldpolicy = p->policy;
4249 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4250 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4251 policy != SCHED_IDLE)
4254 * Valid priorities for SCHED_FIFO and SCHED_RR are
4255 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4256 * SCHED_BATCH and SCHED_IDLE is 0.
4258 if (param->sched_priority < 0 ||
4259 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4260 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4262 if (rt_policy(policy) != (param->sched_priority != 0))
4266 * Allow unprivileged RT tasks to decrease priority:
4268 if (!capable(CAP_SYS_NICE)) {
4269 if (rt_policy(policy)) {
4270 unsigned long rlim_rtprio;
4272 if (!lock_task_sighand(p, &flags))
4274 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4275 unlock_task_sighand(p, &flags);
4277 /* can't set/change the rt policy */
4278 if (policy != p->policy && !rlim_rtprio)
4281 /* can't increase priority */
4282 if (param->sched_priority > p->rt_priority &&
4283 param->sched_priority > rlim_rtprio)
4287 * Like positive nice levels, dont allow tasks to
4288 * move out of SCHED_IDLE either:
4290 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4293 /* can't change other user's priorities */
4294 if ((current->euid != p->euid) &&
4295 (current->euid != p->uid))
4299 retval = security_task_setscheduler(p, policy, param);
4303 * make sure no PI-waiters arrive (or leave) while we are
4304 * changing the priority of the task:
4306 spin_lock_irqsave(&p->pi_lock, flags);
4308 * To be able to change p->policy safely, the apropriate
4309 * runqueue lock must be held.
4311 rq = __task_rq_lock(p);
4312 /* recheck policy now with rq lock held */
4313 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4314 policy = oldpolicy = -1;
4315 __task_rq_unlock(rq);
4316 spin_unlock_irqrestore(&p->pi_lock, flags);
4319 update_rq_clock(rq);
4320 on_rq = p->se.on_rq;
4321 running = task_current(rq, p);
4323 deactivate_task(rq, p, 0);
4325 p->sched_class->put_prev_task(rq, p);
4329 __setscheduler(rq, p, policy, param->sched_priority);
4333 p->sched_class->set_curr_task(rq);
4334 activate_task(rq, p, 0);
4336 * Reschedule if we are currently running on this runqueue and
4337 * our priority decreased, or if we are not currently running on
4338 * this runqueue and our priority is higher than the current's
4341 if (p->prio > oldprio)
4342 resched_task(rq->curr);
4344 check_preempt_curr(rq, p);
4347 __task_rq_unlock(rq);
4348 spin_unlock_irqrestore(&p->pi_lock, flags);
4350 rt_mutex_adjust_pi(p);
4354 EXPORT_SYMBOL_GPL(sched_setscheduler);
4357 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4359 struct sched_param lparam;
4360 struct task_struct *p;
4363 if (!param || pid < 0)
4365 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4370 p = find_process_by_pid(pid);
4372 retval = sched_setscheduler(p, policy, &lparam);
4379 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4380 * @pid: the pid in question.
4381 * @policy: new policy.
4382 * @param: structure containing the new RT priority.
4385 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4387 /* negative values for policy are not valid */
4391 return do_sched_setscheduler(pid, policy, param);
4395 * sys_sched_setparam - set/change the RT priority of a thread
4396 * @pid: the pid in question.
4397 * @param: structure containing the new RT priority.
4399 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4401 return do_sched_setscheduler(pid, -1, param);
4405 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4406 * @pid: the pid in question.
4408 asmlinkage long sys_sched_getscheduler(pid_t pid)
4410 struct task_struct *p;
4417 read_lock(&tasklist_lock);
4418 p = find_process_by_pid(pid);
4420 retval = security_task_getscheduler(p);
4424 read_unlock(&tasklist_lock);
4429 * sys_sched_getscheduler - get the RT priority of a thread
4430 * @pid: the pid in question.
4431 * @param: structure containing the RT priority.
4433 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4435 struct sched_param lp;
4436 struct task_struct *p;
4439 if (!param || pid < 0)
4442 read_lock(&tasklist_lock);
4443 p = find_process_by_pid(pid);
4448 retval = security_task_getscheduler(p);
4452 lp.sched_priority = p->rt_priority;
4453 read_unlock(&tasklist_lock);
4456 * This one might sleep, we cannot do it with a spinlock held ...
4458 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4463 read_unlock(&tasklist_lock);
4467 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4469 cpumask_t cpus_allowed;
4470 struct task_struct *p;
4474 read_lock(&tasklist_lock);
4476 p = find_process_by_pid(pid);
4478 read_unlock(&tasklist_lock);
4484 * It is not safe to call set_cpus_allowed with the
4485 * tasklist_lock held. We will bump the task_struct's
4486 * usage count and then drop tasklist_lock.
4489 read_unlock(&tasklist_lock);
4492 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4493 !capable(CAP_SYS_NICE))
4496 retval = security_task_setscheduler(p, 0, NULL);
4500 cpus_allowed = cpuset_cpus_allowed(p);
4501 cpus_and(new_mask, new_mask, cpus_allowed);
4503 retval = set_cpus_allowed(p, new_mask);
4506 cpus_allowed = cpuset_cpus_allowed(p);
4507 if (!cpus_subset(new_mask, cpus_allowed)) {
4509 * We must have raced with a concurrent cpuset
4510 * update. Just reset the cpus_allowed to the
4511 * cpuset's cpus_allowed
4513 new_mask = cpus_allowed;
4523 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4524 cpumask_t *new_mask)
4526 if (len < sizeof(cpumask_t)) {
4527 memset(new_mask, 0, sizeof(cpumask_t));
4528 } else if (len > sizeof(cpumask_t)) {
4529 len = sizeof(cpumask_t);
4531 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4535 * sys_sched_setaffinity - set the cpu affinity of a process
4536 * @pid: pid of the process
4537 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4538 * @user_mask_ptr: user-space pointer to the new cpu mask
4540 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4541 unsigned long __user *user_mask_ptr)
4546 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4550 return sched_setaffinity(pid, new_mask);
4554 * Represents all cpu's present in the system
4555 * In systems capable of hotplug, this map could dynamically grow
4556 * as new cpu's are detected in the system via any platform specific
4557 * method, such as ACPI for e.g.
4560 cpumask_t cpu_present_map __read_mostly;
4561 EXPORT_SYMBOL(cpu_present_map);
4564 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4565 EXPORT_SYMBOL(cpu_online_map);
4567 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4568 EXPORT_SYMBOL(cpu_possible_map);
4571 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4573 struct task_struct *p;
4577 read_lock(&tasklist_lock);
4580 p = find_process_by_pid(pid);
4584 retval = security_task_getscheduler(p);
4588 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4591 read_unlock(&tasklist_lock);
4598 * sys_sched_getaffinity - get the cpu affinity of a process
4599 * @pid: pid of the process
4600 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4601 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4603 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4604 unsigned long __user *user_mask_ptr)
4609 if (len < sizeof(cpumask_t))
4612 ret = sched_getaffinity(pid, &mask);
4616 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4619 return sizeof(cpumask_t);
4623 * sys_sched_yield - yield the current processor to other threads.
4625 * This function yields the current CPU to other tasks. If there are no
4626 * other threads running on this CPU then this function will return.
4628 asmlinkage long sys_sched_yield(void)
4630 struct rq *rq = this_rq_lock();
4632 schedstat_inc(rq, yld_count);
4633 current->sched_class->yield_task(rq);
4636 * Since we are going to call schedule() anyway, there's
4637 * no need to preempt or enable interrupts:
4639 __release(rq->lock);
4640 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4641 _raw_spin_unlock(&rq->lock);
4642 preempt_enable_no_resched();
4649 static void __cond_resched(void)
4651 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4652 __might_sleep(__FILE__, __LINE__);
4655 * The BKS might be reacquired before we have dropped
4656 * PREEMPT_ACTIVE, which could trigger a second
4657 * cond_resched() call.
4660 add_preempt_count(PREEMPT_ACTIVE);
4662 sub_preempt_count(PREEMPT_ACTIVE);
4663 } while (need_resched());
4666 int __sched cond_resched(void)
4668 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4669 system_state == SYSTEM_RUNNING) {
4675 EXPORT_SYMBOL(cond_resched);
4678 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4679 * call schedule, and on return reacquire the lock.
4681 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4682 * operations here to prevent schedule() from being called twice (once via
4683 * spin_unlock(), once by hand).
4685 int cond_resched_lock(spinlock_t *lock)
4689 if (need_lockbreak(lock)) {
4695 if (need_resched() && system_state == SYSTEM_RUNNING) {
4696 spin_release(&lock->dep_map, 1, _THIS_IP_);
4697 _raw_spin_unlock(lock);
4698 preempt_enable_no_resched();
4705 EXPORT_SYMBOL(cond_resched_lock);
4707 int __sched cond_resched_softirq(void)
4709 BUG_ON(!in_softirq());
4711 if (need_resched() && system_state == SYSTEM_RUNNING) {
4719 EXPORT_SYMBOL(cond_resched_softirq);
4722 * yield - yield the current processor to other threads.
4724 * This is a shortcut for kernel-space yielding - it marks the
4725 * thread runnable and calls sys_sched_yield().
4727 void __sched yield(void)
4729 set_current_state(TASK_RUNNING);
4732 EXPORT_SYMBOL(yield);
4735 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4736 * that process accounting knows that this is a task in IO wait state.
4738 * But don't do that if it is a deliberate, throttling IO wait (this task
4739 * has set its backing_dev_info: the queue against which it should throttle)
4741 void __sched io_schedule(void)
4743 struct rq *rq = &__raw_get_cpu_var(runqueues);
4745 delayacct_blkio_start();
4746 atomic_inc(&rq->nr_iowait);
4748 atomic_dec(&rq->nr_iowait);
4749 delayacct_blkio_end();
4751 EXPORT_SYMBOL(io_schedule);
4753 long __sched io_schedule_timeout(long timeout)
4755 struct rq *rq = &__raw_get_cpu_var(runqueues);
4758 delayacct_blkio_start();
4759 atomic_inc(&rq->nr_iowait);
4760 ret = schedule_timeout(timeout);
4761 atomic_dec(&rq->nr_iowait);
4762 delayacct_blkio_end();
4767 * sys_sched_get_priority_max - return maximum RT priority.
4768 * @policy: scheduling class.
4770 * this syscall returns the maximum rt_priority that can be used
4771 * by a given scheduling class.
4773 asmlinkage long sys_sched_get_priority_max(int policy)
4780 ret = MAX_USER_RT_PRIO-1;
4792 * sys_sched_get_priority_min - return minimum RT priority.
4793 * @policy: scheduling class.
4795 * this syscall returns the minimum rt_priority that can be used
4796 * by a given scheduling class.
4798 asmlinkage long sys_sched_get_priority_min(int policy)
4816 * sys_sched_rr_get_interval - return the default timeslice of a process.
4817 * @pid: pid of the process.
4818 * @interval: userspace pointer to the timeslice value.
4820 * this syscall writes the default timeslice value of a given process
4821 * into the user-space timespec buffer. A value of '0' means infinity.
4824 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4826 struct task_struct *p;
4827 unsigned int time_slice;
4835 read_lock(&tasklist_lock);
4836 p = find_process_by_pid(pid);
4840 retval = security_task_getscheduler(p);
4845 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4846 * tasks that are on an otherwise idle runqueue:
4849 if (p->policy == SCHED_RR) {
4850 time_slice = DEF_TIMESLICE;
4852 struct sched_entity *se = &p->se;
4853 unsigned long flags;
4856 rq = task_rq_lock(p, &flags);
4857 if (rq->cfs.load.weight)
4858 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4859 task_rq_unlock(rq, &flags);
4861 read_unlock(&tasklist_lock);
4862 jiffies_to_timespec(time_slice, &t);
4863 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4867 read_unlock(&tasklist_lock);
4871 static const char stat_nam[] = "RSDTtZX";
4873 void sched_show_task(struct task_struct *p)
4875 unsigned long free = 0;
4878 state = p->state ? __ffs(p->state) + 1 : 0;
4879 printk(KERN_INFO "%-13.13s %c", p->comm,
4880 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4881 #if BITS_PER_LONG == 32
4882 if (state == TASK_RUNNING)
4883 printk(KERN_CONT " running ");
4885 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4887 if (state == TASK_RUNNING)
4888 printk(KERN_CONT " running task ");
4890 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4892 #ifdef CONFIG_DEBUG_STACK_USAGE
4894 unsigned long *n = end_of_stack(p);
4897 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4900 printk(KERN_CONT "%5lu %5d %6d\n", free,
4901 task_pid_nr(p), task_pid_nr(p->real_parent));
4903 if (state != TASK_RUNNING)
4904 show_stack(p, NULL);
4907 void show_state_filter(unsigned long state_filter)
4909 struct task_struct *g, *p;
4911 #if BITS_PER_LONG == 32
4913 " task PC stack pid father\n");
4916 " task PC stack pid father\n");
4918 read_lock(&tasklist_lock);
4919 do_each_thread(g, p) {
4921 * reset the NMI-timeout, listing all files on a slow
4922 * console might take alot of time:
4924 touch_nmi_watchdog();
4925 if (!state_filter || (p->state & state_filter))
4927 } while_each_thread(g, p);
4929 touch_all_softlockup_watchdogs();
4931 #ifdef CONFIG_SCHED_DEBUG
4932 sysrq_sched_debug_show();
4934 read_unlock(&tasklist_lock);
4936 * Only show locks if all tasks are dumped:
4938 if (state_filter == -1)
4939 debug_show_all_locks();
4942 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4944 idle->sched_class = &idle_sched_class;
4948 * init_idle - set up an idle thread for a given CPU
4949 * @idle: task in question
4950 * @cpu: cpu the idle task belongs to
4952 * NOTE: this function does not set the idle thread's NEED_RESCHED
4953 * flag, to make booting more robust.
4955 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4957 struct rq *rq = cpu_rq(cpu);
4958 unsigned long flags;
4961 idle->se.exec_start = sched_clock();
4963 idle->prio = idle->normal_prio = MAX_PRIO;
4964 idle->cpus_allowed = cpumask_of_cpu(cpu);
4965 __set_task_cpu(idle, cpu);
4967 spin_lock_irqsave(&rq->lock, flags);
4968 rq->curr = rq->idle = idle;
4969 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4972 spin_unlock_irqrestore(&rq->lock, flags);
4974 /* Set the preempt count _outside_ the spinlocks! */
4975 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4976 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4978 task_thread_info(idle)->preempt_count = 0;
4981 * The idle tasks have their own, simple scheduling class:
4983 idle->sched_class = &idle_sched_class;
4987 * In a system that switches off the HZ timer nohz_cpu_mask
4988 * indicates which cpus entered this state. This is used
4989 * in the rcu update to wait only for active cpus. For system
4990 * which do not switch off the HZ timer nohz_cpu_mask should
4991 * always be CPU_MASK_NONE.
4993 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4996 * Increase the granularity value when there are more CPUs,
4997 * because with more CPUs the 'effective latency' as visible
4998 * to users decreases. But the relationship is not linear,
4999 * so pick a second-best guess by going with the log2 of the
5002 * This idea comes from the SD scheduler of Con Kolivas:
5004 static inline void sched_init_granularity(void)
5006 unsigned int factor = 1 + ilog2(num_online_cpus());
5007 const unsigned long limit = 200000000;
5009 sysctl_sched_min_granularity *= factor;
5010 if (sysctl_sched_min_granularity > limit)
5011 sysctl_sched_min_granularity = limit;
5013 sysctl_sched_latency *= factor;
5014 if (sysctl_sched_latency > limit)
5015 sysctl_sched_latency = limit;
5017 sysctl_sched_wakeup_granularity *= factor;
5018 sysctl_sched_batch_wakeup_granularity *= factor;
5023 * This is how migration works:
5025 * 1) we queue a struct migration_req structure in the source CPU's
5026 * runqueue and wake up that CPU's migration thread.
5027 * 2) we down() the locked semaphore => thread blocks.
5028 * 3) migration thread wakes up (implicitly it forces the migrated
5029 * thread off the CPU)
5030 * 4) it gets the migration request and checks whether the migrated
5031 * task is still in the wrong runqueue.
5032 * 5) if it's in the wrong runqueue then the migration thread removes
5033 * it and puts it into the right queue.
5034 * 6) migration thread up()s the semaphore.
5035 * 7) we wake up and the migration is done.
5039 * Change a given task's CPU affinity. Migrate the thread to a
5040 * proper CPU and schedule it away if the CPU it's executing on
5041 * is removed from the allowed bitmask.
5043 * NOTE: the caller must have a valid reference to the task, the
5044 * task must not exit() & deallocate itself prematurely. The
5045 * call is not atomic; no spinlocks may be held.
5047 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5049 struct migration_req req;
5050 unsigned long flags;
5054 rq = task_rq_lock(p, &flags);
5055 if (!cpus_intersects(new_mask, cpu_online_map)) {
5060 if (p->sched_class->set_cpus_allowed)
5061 p->sched_class->set_cpus_allowed(p, &new_mask);
5063 p->cpus_allowed = new_mask;
5064 p->nr_cpus_allowed = cpus_weight(new_mask);
5067 /* Can the task run on the task's current CPU? If so, we're done */
5068 if (cpu_isset(task_cpu(p), new_mask))
5071 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5072 /* Need help from migration thread: drop lock and wait. */
5073 task_rq_unlock(rq, &flags);
5074 wake_up_process(rq->migration_thread);
5075 wait_for_completion(&req.done);
5076 tlb_migrate_finish(p->mm);
5080 task_rq_unlock(rq, &flags);
5084 EXPORT_SYMBOL_GPL(set_cpus_allowed);
5087 * Move (not current) task off this cpu, onto dest cpu. We're doing
5088 * this because either it can't run here any more (set_cpus_allowed()
5089 * away from this CPU, or CPU going down), or because we're
5090 * attempting to rebalance this task on exec (sched_exec).
5092 * So we race with normal scheduler movements, but that's OK, as long
5093 * as the task is no longer on this CPU.
5095 * Returns non-zero if task was successfully migrated.
5097 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5099 struct rq *rq_dest, *rq_src;
5102 if (unlikely(cpu_is_offline(dest_cpu)))
5105 rq_src = cpu_rq(src_cpu);
5106 rq_dest = cpu_rq(dest_cpu);
5108 double_rq_lock(rq_src, rq_dest);
5109 /* Already moved. */
5110 if (task_cpu(p) != src_cpu)
5112 /* Affinity changed (again). */
5113 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5116 on_rq = p->se.on_rq;
5118 deactivate_task(rq_src, p, 0);
5120 set_task_cpu(p, dest_cpu);
5122 activate_task(rq_dest, p, 0);
5123 check_preempt_curr(rq_dest, p);
5127 double_rq_unlock(rq_src, rq_dest);
5132 * migration_thread - this is a highprio system thread that performs
5133 * thread migration by bumping thread off CPU then 'pushing' onto
5136 static int migration_thread(void *data)
5138 int cpu = (long)data;
5142 BUG_ON(rq->migration_thread != current);
5144 set_current_state(TASK_INTERRUPTIBLE);
5145 while (!kthread_should_stop()) {
5146 struct migration_req *req;
5147 struct list_head *head;
5149 spin_lock_irq(&rq->lock);
5151 if (cpu_is_offline(cpu)) {
5152 spin_unlock_irq(&rq->lock);
5156 if (rq->active_balance) {
5157 active_load_balance(rq, cpu);
5158 rq->active_balance = 0;
5161 head = &rq->migration_queue;
5163 if (list_empty(head)) {
5164 spin_unlock_irq(&rq->lock);
5166 set_current_state(TASK_INTERRUPTIBLE);
5169 req = list_entry(head->next, struct migration_req, list);
5170 list_del_init(head->next);
5172 spin_unlock(&rq->lock);
5173 __migrate_task(req->task, cpu, req->dest_cpu);
5176 complete(&req->done);
5178 __set_current_state(TASK_RUNNING);
5182 /* Wait for kthread_stop */
5183 set_current_state(TASK_INTERRUPTIBLE);
5184 while (!kthread_should_stop()) {
5186 set_current_state(TASK_INTERRUPTIBLE);
5188 __set_current_state(TASK_RUNNING);
5192 #ifdef CONFIG_HOTPLUG_CPU
5194 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5198 local_irq_disable();
5199 ret = __migrate_task(p, src_cpu, dest_cpu);
5205 * Figure out where task on dead CPU should go, use force if necessary.
5206 * NOTE: interrupts should be disabled by the caller
5208 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5210 unsigned long flags;
5217 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5218 cpus_and(mask, mask, p->cpus_allowed);
5219 dest_cpu = any_online_cpu(mask);
5221 /* On any allowed CPU? */
5222 if (dest_cpu == NR_CPUS)
5223 dest_cpu = any_online_cpu(p->cpus_allowed);
5225 /* No more Mr. Nice Guy. */
5226 if (dest_cpu == NR_CPUS) {
5227 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5229 * Try to stay on the same cpuset, where the
5230 * current cpuset may be a subset of all cpus.
5231 * The cpuset_cpus_allowed_locked() variant of
5232 * cpuset_cpus_allowed() will not block. It must be
5233 * called within calls to cpuset_lock/cpuset_unlock.
5235 rq = task_rq_lock(p, &flags);
5236 p->cpus_allowed = cpus_allowed;
5237 dest_cpu = any_online_cpu(p->cpus_allowed);
5238 task_rq_unlock(rq, &flags);
5241 * Don't tell them about moving exiting tasks or
5242 * kernel threads (both mm NULL), since they never
5245 if (p->mm && printk_ratelimit()) {
5246 printk(KERN_INFO "process %d (%s) no "
5247 "longer affine to cpu%d\n",
5248 task_pid_nr(p), p->comm, dead_cpu);
5251 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5255 * While a dead CPU has no uninterruptible tasks queued at this point,
5256 * it might still have a nonzero ->nr_uninterruptible counter, because
5257 * for performance reasons the counter is not stricly tracking tasks to
5258 * their home CPUs. So we just add the counter to another CPU's counter,
5259 * to keep the global sum constant after CPU-down:
5261 static void migrate_nr_uninterruptible(struct rq *rq_src)
5263 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5264 unsigned long flags;
5266 local_irq_save(flags);
5267 double_rq_lock(rq_src, rq_dest);
5268 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5269 rq_src->nr_uninterruptible = 0;
5270 double_rq_unlock(rq_src, rq_dest);
5271 local_irq_restore(flags);
5274 /* Run through task list and migrate tasks from the dead cpu. */
5275 static void migrate_live_tasks(int src_cpu)
5277 struct task_struct *p, *t;
5279 read_lock(&tasklist_lock);
5281 do_each_thread(t, p) {
5285 if (task_cpu(p) == src_cpu)
5286 move_task_off_dead_cpu(src_cpu, p);
5287 } while_each_thread(t, p);
5289 read_unlock(&tasklist_lock);
5293 * Schedules idle task to be the next runnable task on current CPU.
5294 * It does so by boosting its priority to highest possible.
5295 * Used by CPU offline code.
5297 void sched_idle_next(void)
5299 int this_cpu = smp_processor_id();
5300 struct rq *rq = cpu_rq(this_cpu);
5301 struct task_struct *p = rq->idle;
5302 unsigned long flags;
5304 /* cpu has to be offline */
5305 BUG_ON(cpu_online(this_cpu));
5308 * Strictly not necessary since rest of the CPUs are stopped by now
5309 * and interrupts disabled on the current cpu.
5311 spin_lock_irqsave(&rq->lock, flags);
5313 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5315 update_rq_clock(rq);
5316 activate_task(rq, p, 0);
5318 spin_unlock_irqrestore(&rq->lock, flags);
5322 * Ensures that the idle task is using init_mm right before its cpu goes
5325 void idle_task_exit(void)
5327 struct mm_struct *mm = current->active_mm;
5329 BUG_ON(cpu_online(smp_processor_id()));
5332 switch_mm(mm, &init_mm, current);
5336 /* called under rq->lock with disabled interrupts */
5337 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5339 struct rq *rq = cpu_rq(dead_cpu);
5341 /* Must be exiting, otherwise would be on tasklist. */
5342 BUG_ON(!p->exit_state);
5344 /* Cannot have done final schedule yet: would have vanished. */
5345 BUG_ON(p->state == TASK_DEAD);
5350 * Drop lock around migration; if someone else moves it,
5351 * that's OK. No task can be added to this CPU, so iteration is
5354 spin_unlock_irq(&rq->lock);
5355 move_task_off_dead_cpu(dead_cpu, p);
5356 spin_lock_irq(&rq->lock);
5361 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5362 static void migrate_dead_tasks(unsigned int dead_cpu)
5364 struct rq *rq = cpu_rq(dead_cpu);
5365 struct task_struct *next;
5368 if (!rq->nr_running)
5370 update_rq_clock(rq);
5371 next = pick_next_task(rq, rq->curr);
5374 migrate_dead(dead_cpu, next);
5378 #endif /* CONFIG_HOTPLUG_CPU */
5380 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5382 static struct ctl_table sd_ctl_dir[] = {
5384 .procname = "sched_domain",
5390 static struct ctl_table sd_ctl_root[] = {
5392 .ctl_name = CTL_KERN,
5393 .procname = "kernel",
5395 .child = sd_ctl_dir,
5400 static struct ctl_table *sd_alloc_ctl_entry(int n)
5402 struct ctl_table *entry =
5403 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5408 static void sd_free_ctl_entry(struct ctl_table **tablep)
5410 struct ctl_table *entry;
5413 * In the intermediate directories, both the child directory and
5414 * procname are dynamically allocated and could fail but the mode
5415 * will always be set. In the lowest directory the names are
5416 * static strings and all have proc handlers.
5418 for (entry = *tablep; entry->mode; entry++) {
5420 sd_free_ctl_entry(&entry->child);
5421 if (entry->proc_handler == NULL)
5422 kfree(entry->procname);
5430 set_table_entry(struct ctl_table *entry,
5431 const char *procname, void *data, int maxlen,
5432 mode_t mode, proc_handler *proc_handler)
5434 entry->procname = procname;
5436 entry->maxlen = maxlen;
5438 entry->proc_handler = proc_handler;
5441 static struct ctl_table *
5442 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5444 struct ctl_table *table = sd_alloc_ctl_entry(12);
5449 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5450 sizeof(long), 0644, proc_doulongvec_minmax);
5451 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5452 sizeof(long), 0644, proc_doulongvec_minmax);
5453 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5454 sizeof(int), 0644, proc_dointvec_minmax);
5455 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5456 sizeof(int), 0644, proc_dointvec_minmax);
5457 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5458 sizeof(int), 0644, proc_dointvec_minmax);
5459 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5460 sizeof(int), 0644, proc_dointvec_minmax);
5461 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5462 sizeof(int), 0644, proc_dointvec_minmax);
5463 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5464 sizeof(int), 0644, proc_dointvec_minmax);
5465 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5466 sizeof(int), 0644, proc_dointvec_minmax);
5467 set_table_entry(&table[9], "cache_nice_tries",
5468 &sd->cache_nice_tries,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 set_table_entry(&table[10], "flags", &sd->flags,
5471 sizeof(int), 0644, proc_dointvec_minmax);
5472 /* &table[11] is terminator */
5477 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5479 struct ctl_table *entry, *table;
5480 struct sched_domain *sd;
5481 int domain_num = 0, i;
5484 for_each_domain(cpu, sd)
5486 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5491 for_each_domain(cpu, sd) {
5492 snprintf(buf, 32, "domain%d", i);
5493 entry->procname = kstrdup(buf, GFP_KERNEL);
5495 entry->child = sd_alloc_ctl_domain_table(sd);
5502 static struct ctl_table_header *sd_sysctl_header;
5503 static void register_sched_domain_sysctl(void)
5505 int i, cpu_num = num_online_cpus();
5506 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5509 WARN_ON(sd_ctl_dir[0].child);
5510 sd_ctl_dir[0].child = entry;
5515 for_each_online_cpu(i) {
5516 snprintf(buf, 32, "cpu%d", i);
5517 entry->procname = kstrdup(buf, GFP_KERNEL);
5519 entry->child = sd_alloc_ctl_cpu_table(i);
5523 WARN_ON(sd_sysctl_header);
5524 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5527 /* may be called multiple times per register */
5528 static void unregister_sched_domain_sysctl(void)
5530 if (sd_sysctl_header)
5531 unregister_sysctl_table(sd_sysctl_header);
5532 sd_sysctl_header = NULL;
5533 if (sd_ctl_dir[0].child)
5534 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5537 static void register_sched_domain_sysctl(void)
5540 static void unregister_sched_domain_sysctl(void)
5546 * migration_call - callback that gets triggered when a CPU is added.
5547 * Here we can start up the necessary migration thread for the new CPU.
5549 static int __cpuinit
5550 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5552 struct task_struct *p;
5553 int cpu = (long)hcpu;
5554 unsigned long flags;
5559 case CPU_UP_PREPARE:
5560 case CPU_UP_PREPARE_FROZEN:
5561 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5564 kthread_bind(p, cpu);
5565 /* Must be high prio: stop_machine expects to yield to it. */
5566 rq = task_rq_lock(p, &flags);
5567 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5568 task_rq_unlock(rq, &flags);
5569 cpu_rq(cpu)->migration_thread = p;
5573 case CPU_ONLINE_FROZEN:
5574 /* Strictly unnecessary, as first user will wake it. */
5575 wake_up_process(cpu_rq(cpu)->migration_thread);
5577 /* Update our root-domain */
5579 spin_lock_irqsave(&rq->lock, flags);
5581 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5582 cpu_set(cpu, rq->rd->online);
5584 spin_unlock_irqrestore(&rq->lock, flags);
5587 #ifdef CONFIG_HOTPLUG_CPU
5588 case CPU_UP_CANCELED:
5589 case CPU_UP_CANCELED_FROZEN:
5590 if (!cpu_rq(cpu)->migration_thread)
5592 /* Unbind it from offline cpu so it can run. Fall thru. */
5593 kthread_bind(cpu_rq(cpu)->migration_thread,
5594 any_online_cpu(cpu_online_map));
5595 kthread_stop(cpu_rq(cpu)->migration_thread);
5596 cpu_rq(cpu)->migration_thread = NULL;
5600 case CPU_DEAD_FROZEN:
5601 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5602 migrate_live_tasks(cpu);
5604 kthread_stop(rq->migration_thread);
5605 rq->migration_thread = NULL;
5606 /* Idle task back to normal (off runqueue, low prio) */
5607 spin_lock_irq(&rq->lock);
5608 update_rq_clock(rq);
5609 deactivate_task(rq, rq->idle, 0);
5610 rq->idle->static_prio = MAX_PRIO;
5611 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5612 rq->idle->sched_class = &idle_sched_class;
5613 migrate_dead_tasks(cpu);
5614 spin_unlock_irq(&rq->lock);
5616 migrate_nr_uninterruptible(rq);
5617 BUG_ON(rq->nr_running != 0);
5620 * No need to migrate the tasks: it was best-effort if
5621 * they didn't take sched_hotcpu_mutex. Just wake up
5624 spin_lock_irq(&rq->lock);
5625 while (!list_empty(&rq->migration_queue)) {
5626 struct migration_req *req;
5628 req = list_entry(rq->migration_queue.next,
5629 struct migration_req, list);
5630 list_del_init(&req->list);
5631 complete(&req->done);
5633 spin_unlock_irq(&rq->lock);
5636 case CPU_DOWN_PREPARE:
5637 /* Update our root-domain */
5639 spin_lock_irqsave(&rq->lock, flags);
5641 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5642 cpu_clear(cpu, rq->rd->online);
5644 spin_unlock_irqrestore(&rq->lock, flags);
5651 /* Register at highest priority so that task migration (migrate_all_tasks)
5652 * happens before everything else.
5654 static struct notifier_block __cpuinitdata migration_notifier = {
5655 .notifier_call = migration_call,
5659 void __init migration_init(void)
5661 void *cpu = (void *)(long)smp_processor_id();
5664 /* Start one for the boot CPU: */
5665 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5666 BUG_ON(err == NOTIFY_BAD);
5667 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5668 register_cpu_notifier(&migration_notifier);
5674 /* Number of possible processor ids */
5675 int nr_cpu_ids __read_mostly = NR_CPUS;
5676 EXPORT_SYMBOL(nr_cpu_ids);
5678 #ifdef CONFIG_SCHED_DEBUG
5680 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
5682 struct sched_group *group = sd->groups;
5683 cpumask_t groupmask;
5686 cpumask_scnprintf(str, NR_CPUS, sd->span);
5687 cpus_clear(groupmask);
5689 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5691 if (!(sd->flags & SD_LOAD_BALANCE)) {
5692 printk("does not load-balance\n");
5694 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5699 printk(KERN_CONT "span %s\n", str);
5701 if (!cpu_isset(cpu, sd->span)) {
5702 printk(KERN_ERR "ERROR: domain->span does not contain "
5705 if (!cpu_isset(cpu, group->cpumask)) {
5706 printk(KERN_ERR "ERROR: domain->groups does not contain"
5710 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5714 printk(KERN_ERR "ERROR: group is NULL\n");
5718 if (!group->__cpu_power) {
5719 printk(KERN_CONT "\n");
5720 printk(KERN_ERR "ERROR: domain->cpu_power not "
5725 if (!cpus_weight(group->cpumask)) {
5726 printk(KERN_CONT "\n");
5727 printk(KERN_ERR "ERROR: empty group\n");
5731 if (cpus_intersects(groupmask, group->cpumask)) {
5732 printk(KERN_CONT "\n");
5733 printk(KERN_ERR "ERROR: repeated CPUs\n");
5737 cpus_or(groupmask, groupmask, group->cpumask);
5739 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5740 printk(KERN_CONT " %s", str);
5742 group = group->next;
5743 } while (group != sd->groups);
5744 printk(KERN_CONT "\n");
5746 if (!cpus_equal(sd->span, groupmask))
5747 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5749 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5750 printk(KERN_ERR "ERROR: parent span is not a superset "
5751 "of domain->span\n");
5755 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5760 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5764 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5767 if (sched_domain_debug_one(sd, cpu, level))
5776 # define sched_domain_debug(sd, cpu) do { } while (0)
5779 static int sd_degenerate(struct sched_domain *sd)
5781 if (cpus_weight(sd->span) == 1)
5784 /* Following flags need at least 2 groups */
5785 if (sd->flags & (SD_LOAD_BALANCE |
5786 SD_BALANCE_NEWIDLE |
5790 SD_SHARE_PKG_RESOURCES)) {
5791 if (sd->groups != sd->groups->next)
5795 /* Following flags don't use groups */
5796 if (sd->flags & (SD_WAKE_IDLE |
5805 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5807 unsigned long cflags = sd->flags, pflags = parent->flags;
5809 if (sd_degenerate(parent))
5812 if (!cpus_equal(sd->span, parent->span))
5815 /* Does parent contain flags not in child? */
5816 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5817 if (cflags & SD_WAKE_AFFINE)
5818 pflags &= ~SD_WAKE_BALANCE;
5819 /* Flags needing groups don't count if only 1 group in parent */
5820 if (parent->groups == parent->groups->next) {
5821 pflags &= ~(SD_LOAD_BALANCE |
5822 SD_BALANCE_NEWIDLE |
5826 SD_SHARE_PKG_RESOURCES);
5828 if (~cflags & pflags)
5834 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5836 unsigned long flags;
5837 const struct sched_class *class;
5839 spin_lock_irqsave(&rq->lock, flags);
5842 struct root_domain *old_rd = rq->rd;
5844 for (class = sched_class_highest; class; class = class->next)
5845 if (class->leave_domain)
5846 class->leave_domain(rq);
5848 if (atomic_dec_and_test(&old_rd->refcount))
5852 atomic_inc(&rd->refcount);
5855 for (class = sched_class_highest; class; class = class->next)
5856 if (class->join_domain)
5857 class->join_domain(rq);
5859 spin_unlock_irqrestore(&rq->lock, flags);
5862 static void init_rootdomain(struct root_domain *rd, const cpumask_t *map)
5864 memset(rd, 0, sizeof(*rd));
5867 cpus_and(rd->online, rd->span, cpu_online_map);
5870 static void init_defrootdomain(void)
5872 cpumask_t cpus = CPU_MASK_ALL;
5874 init_rootdomain(&def_root_domain, &cpus);
5875 atomic_set(&def_root_domain.refcount, 1);
5878 static struct root_domain *alloc_rootdomain(const cpumask_t *map)
5880 struct root_domain *rd;
5882 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5886 init_rootdomain(rd, map);
5892 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5893 * hold the hotplug lock.
5895 static void cpu_attach_domain(struct sched_domain *sd,
5896 struct root_domain *rd, int cpu)
5898 struct rq *rq = cpu_rq(cpu);
5899 struct sched_domain *tmp;
5901 /* Remove the sched domains which do not contribute to scheduling. */
5902 for (tmp = sd; tmp; tmp = tmp->parent) {
5903 struct sched_domain *parent = tmp->parent;
5906 if (sd_parent_degenerate(tmp, parent)) {
5907 tmp->parent = parent->parent;
5909 parent->parent->child = tmp;
5913 if (sd && sd_degenerate(sd)) {
5919 sched_domain_debug(sd, cpu);
5921 rq_attach_root(rq, rd);
5922 rcu_assign_pointer(rq->sd, sd);
5925 /* cpus with isolated domains */
5926 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5928 /* Setup the mask of cpus configured for isolated domains */
5929 static int __init isolated_cpu_setup(char *str)
5931 int ints[NR_CPUS], i;
5933 str = get_options(str, ARRAY_SIZE(ints), ints);
5934 cpus_clear(cpu_isolated_map);
5935 for (i = 1; i <= ints[0]; i++)
5936 if (ints[i] < NR_CPUS)
5937 cpu_set(ints[i], cpu_isolated_map);
5941 __setup("isolcpus=", isolated_cpu_setup);
5944 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5945 * to a function which identifies what group(along with sched group) a CPU
5946 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5947 * (due to the fact that we keep track of groups covered with a cpumask_t).
5949 * init_sched_build_groups will build a circular linked list of the groups
5950 * covered by the given span, and will set each group's ->cpumask correctly,
5951 * and ->cpu_power to 0.
5954 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5955 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5956 struct sched_group **sg))
5958 struct sched_group *first = NULL, *last = NULL;
5959 cpumask_t covered = CPU_MASK_NONE;
5962 for_each_cpu_mask(i, span) {
5963 struct sched_group *sg;
5964 int group = group_fn(i, cpu_map, &sg);
5967 if (cpu_isset(i, covered))
5970 sg->cpumask = CPU_MASK_NONE;
5971 sg->__cpu_power = 0;
5973 for_each_cpu_mask(j, span) {
5974 if (group_fn(j, cpu_map, NULL) != group)
5977 cpu_set(j, covered);
5978 cpu_set(j, sg->cpumask);
5989 #define SD_NODES_PER_DOMAIN 16
5994 * find_next_best_node - find the next node to include in a sched_domain
5995 * @node: node whose sched_domain we're building
5996 * @used_nodes: nodes already in the sched_domain
5998 * Find the next node to include in a given scheduling domain. Simply
5999 * finds the closest node not already in the @used_nodes map.
6001 * Should use nodemask_t.
6003 static int find_next_best_node(int node, unsigned long *used_nodes)
6005 int i, n, val, min_val, best_node = 0;
6009 for (i = 0; i < MAX_NUMNODES; i++) {
6010 /* Start at @node */
6011 n = (node + i) % MAX_NUMNODES;
6013 if (!nr_cpus_node(n))
6016 /* Skip already used nodes */
6017 if (test_bit(n, used_nodes))
6020 /* Simple min distance search */
6021 val = node_distance(node, n);
6023 if (val < min_val) {
6029 set_bit(best_node, used_nodes);
6034 * sched_domain_node_span - get a cpumask for a node's sched_domain
6035 * @node: node whose cpumask we're constructing
6036 * @size: number of nodes to include in this span
6038 * Given a node, construct a good cpumask for its sched_domain to span. It
6039 * should be one that prevents unnecessary balancing, but also spreads tasks
6042 static cpumask_t sched_domain_node_span(int node)
6044 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6045 cpumask_t span, nodemask;
6049 bitmap_zero(used_nodes, MAX_NUMNODES);
6051 nodemask = node_to_cpumask(node);
6052 cpus_or(span, span, nodemask);
6053 set_bit(node, used_nodes);
6055 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6056 int next_node = find_next_best_node(node, used_nodes);
6058 nodemask = node_to_cpumask(next_node);
6059 cpus_or(span, span, nodemask);
6066 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6069 * SMT sched-domains:
6071 #ifdef CONFIG_SCHED_SMT
6072 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6073 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6076 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6079 *sg = &per_cpu(sched_group_cpus, cpu);
6085 * multi-core sched-domains:
6087 #ifdef CONFIG_SCHED_MC
6088 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6089 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6092 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6094 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6097 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6098 cpus_and(mask, mask, *cpu_map);
6099 group = first_cpu(mask);
6101 *sg = &per_cpu(sched_group_core, group);
6104 #elif defined(CONFIG_SCHED_MC)
6106 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6109 *sg = &per_cpu(sched_group_core, cpu);
6114 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6115 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6118 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6121 #ifdef CONFIG_SCHED_MC
6122 cpumask_t mask = cpu_coregroup_map(cpu);
6123 cpus_and(mask, mask, *cpu_map);
6124 group = first_cpu(mask);
6125 #elif defined(CONFIG_SCHED_SMT)
6126 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6127 cpus_and(mask, mask, *cpu_map);
6128 group = first_cpu(mask);
6133 *sg = &per_cpu(sched_group_phys, group);
6139 * The init_sched_build_groups can't handle what we want to do with node
6140 * groups, so roll our own. Now each node has its own list of groups which
6141 * gets dynamically allocated.
6143 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6144 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6146 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6147 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6149 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6150 struct sched_group **sg)
6152 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6155 cpus_and(nodemask, nodemask, *cpu_map);
6156 group = first_cpu(nodemask);
6159 *sg = &per_cpu(sched_group_allnodes, group);
6163 static void init_numa_sched_groups_power(struct sched_group *group_head)
6165 struct sched_group *sg = group_head;
6171 for_each_cpu_mask(j, sg->cpumask) {
6172 struct sched_domain *sd;
6174 sd = &per_cpu(phys_domains, j);
6175 if (j != first_cpu(sd->groups->cpumask)) {
6177 * Only add "power" once for each
6183 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6186 } while (sg != group_head);
6191 /* Free memory allocated for various sched_group structures */
6192 static void free_sched_groups(const cpumask_t *cpu_map)
6196 for_each_cpu_mask(cpu, *cpu_map) {
6197 struct sched_group **sched_group_nodes
6198 = sched_group_nodes_bycpu[cpu];
6200 if (!sched_group_nodes)
6203 for (i = 0; i < MAX_NUMNODES; i++) {
6204 cpumask_t nodemask = node_to_cpumask(i);
6205 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6207 cpus_and(nodemask, nodemask, *cpu_map);
6208 if (cpus_empty(nodemask))
6218 if (oldsg != sched_group_nodes[i])
6221 kfree(sched_group_nodes);
6222 sched_group_nodes_bycpu[cpu] = NULL;
6226 static void free_sched_groups(const cpumask_t *cpu_map)
6232 * Initialize sched groups cpu_power.
6234 * cpu_power indicates the capacity of sched group, which is used while
6235 * distributing the load between different sched groups in a sched domain.
6236 * Typically cpu_power for all the groups in a sched domain will be same unless
6237 * there are asymmetries in the topology. If there are asymmetries, group
6238 * having more cpu_power will pickup more load compared to the group having
6241 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6242 * the maximum number of tasks a group can handle in the presence of other idle
6243 * or lightly loaded groups in the same sched domain.
6245 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6247 struct sched_domain *child;
6248 struct sched_group *group;
6250 WARN_ON(!sd || !sd->groups);
6252 if (cpu != first_cpu(sd->groups->cpumask))
6257 sd->groups->__cpu_power = 0;
6260 * For perf policy, if the groups in child domain share resources
6261 * (for example cores sharing some portions of the cache hierarchy
6262 * or SMT), then set this domain groups cpu_power such that each group
6263 * can handle only one task, when there are other idle groups in the
6264 * same sched domain.
6266 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6268 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6269 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6274 * add cpu_power of each child group to this groups cpu_power
6276 group = child->groups;
6278 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6279 group = group->next;
6280 } while (group != child->groups);
6284 * Build sched domains for a given set of cpus and attach the sched domains
6285 * to the individual cpus
6287 static int build_sched_domains(const cpumask_t *cpu_map)
6290 struct root_domain *rd;
6292 struct sched_group **sched_group_nodes = NULL;
6293 int sd_allnodes = 0;
6296 * Allocate the per-node list of sched groups
6298 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6300 if (!sched_group_nodes) {
6301 printk(KERN_WARNING "Can not alloc sched group node list\n");
6304 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6307 rd = alloc_rootdomain(cpu_map);
6309 printk(KERN_WARNING "Cannot alloc root domain\n");
6314 * Set up domains for cpus specified by the cpu_map.
6316 for_each_cpu_mask(i, *cpu_map) {
6317 struct sched_domain *sd = NULL, *p;
6318 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6320 cpus_and(nodemask, nodemask, *cpu_map);
6323 if (cpus_weight(*cpu_map) >
6324 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6325 sd = &per_cpu(allnodes_domains, i);
6326 *sd = SD_ALLNODES_INIT;
6327 sd->span = *cpu_map;
6328 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6334 sd = &per_cpu(node_domains, i);
6336 sd->span = sched_domain_node_span(cpu_to_node(i));
6340 cpus_and(sd->span, sd->span, *cpu_map);
6344 sd = &per_cpu(phys_domains, i);
6346 sd->span = nodemask;
6350 cpu_to_phys_group(i, cpu_map, &sd->groups);
6352 #ifdef CONFIG_SCHED_MC
6354 sd = &per_cpu(core_domains, i);
6356 sd->span = cpu_coregroup_map(i);
6357 cpus_and(sd->span, sd->span, *cpu_map);
6360 cpu_to_core_group(i, cpu_map, &sd->groups);
6363 #ifdef CONFIG_SCHED_SMT
6365 sd = &per_cpu(cpu_domains, i);
6366 *sd = SD_SIBLING_INIT;
6367 sd->span = per_cpu(cpu_sibling_map, i);
6368 cpus_and(sd->span, sd->span, *cpu_map);
6371 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6375 #ifdef CONFIG_SCHED_SMT
6376 /* Set up CPU (sibling) groups */
6377 for_each_cpu_mask(i, *cpu_map) {
6378 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6379 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6380 if (i != first_cpu(this_sibling_map))
6383 init_sched_build_groups(this_sibling_map, cpu_map,
6388 #ifdef CONFIG_SCHED_MC
6389 /* Set up multi-core groups */
6390 for_each_cpu_mask(i, *cpu_map) {
6391 cpumask_t this_core_map = cpu_coregroup_map(i);
6392 cpus_and(this_core_map, this_core_map, *cpu_map);
6393 if (i != first_cpu(this_core_map))
6395 init_sched_build_groups(this_core_map, cpu_map,
6396 &cpu_to_core_group);
6400 /* Set up physical groups */
6401 for (i = 0; i < MAX_NUMNODES; i++) {
6402 cpumask_t nodemask = node_to_cpumask(i);
6404 cpus_and(nodemask, nodemask, *cpu_map);
6405 if (cpus_empty(nodemask))
6408 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6412 /* Set up node groups */
6414 init_sched_build_groups(*cpu_map, cpu_map,
6415 &cpu_to_allnodes_group);
6417 for (i = 0; i < MAX_NUMNODES; i++) {
6418 /* Set up node groups */
6419 struct sched_group *sg, *prev;
6420 cpumask_t nodemask = node_to_cpumask(i);
6421 cpumask_t domainspan;
6422 cpumask_t covered = CPU_MASK_NONE;
6425 cpus_and(nodemask, nodemask, *cpu_map);
6426 if (cpus_empty(nodemask)) {
6427 sched_group_nodes[i] = NULL;
6431 domainspan = sched_domain_node_span(i);
6432 cpus_and(domainspan, domainspan, *cpu_map);
6434 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6436 printk(KERN_WARNING "Can not alloc domain group for "
6440 sched_group_nodes[i] = sg;
6441 for_each_cpu_mask(j, nodemask) {
6442 struct sched_domain *sd;
6444 sd = &per_cpu(node_domains, j);
6447 sg->__cpu_power = 0;
6448 sg->cpumask = nodemask;
6450 cpus_or(covered, covered, nodemask);
6453 for (j = 0; j < MAX_NUMNODES; j++) {
6454 cpumask_t tmp, notcovered;
6455 int n = (i + j) % MAX_NUMNODES;
6457 cpus_complement(notcovered, covered);
6458 cpus_and(tmp, notcovered, *cpu_map);
6459 cpus_and(tmp, tmp, domainspan);
6460 if (cpus_empty(tmp))
6463 nodemask = node_to_cpumask(n);
6464 cpus_and(tmp, tmp, nodemask);
6465 if (cpus_empty(tmp))
6468 sg = kmalloc_node(sizeof(struct sched_group),
6472 "Can not alloc domain group for node %d\n", j);
6475 sg->__cpu_power = 0;
6477 sg->next = prev->next;
6478 cpus_or(covered, covered, tmp);
6485 /* Calculate CPU power for physical packages and nodes */
6486 #ifdef CONFIG_SCHED_SMT
6487 for_each_cpu_mask(i, *cpu_map) {
6488 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6490 init_sched_groups_power(i, sd);
6493 #ifdef CONFIG_SCHED_MC
6494 for_each_cpu_mask(i, *cpu_map) {
6495 struct sched_domain *sd = &per_cpu(core_domains, i);
6497 init_sched_groups_power(i, sd);
6501 for_each_cpu_mask(i, *cpu_map) {
6502 struct sched_domain *sd = &per_cpu(phys_domains, i);
6504 init_sched_groups_power(i, sd);
6508 for (i = 0; i < MAX_NUMNODES; i++)
6509 init_numa_sched_groups_power(sched_group_nodes[i]);
6512 struct sched_group *sg;
6514 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6515 init_numa_sched_groups_power(sg);
6519 /* Attach the domains */
6520 for_each_cpu_mask(i, *cpu_map) {
6521 struct sched_domain *sd;
6522 #ifdef CONFIG_SCHED_SMT
6523 sd = &per_cpu(cpu_domains, i);
6524 #elif defined(CONFIG_SCHED_MC)
6525 sd = &per_cpu(core_domains, i);
6527 sd = &per_cpu(phys_domains, i);
6529 cpu_attach_domain(sd, rd, i);
6536 free_sched_groups(cpu_map);
6541 static cpumask_t *doms_cur; /* current sched domains */
6542 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6545 * Special case: If a kmalloc of a doms_cur partition (array of
6546 * cpumask_t) fails, then fallback to a single sched domain,
6547 * as determined by the single cpumask_t fallback_doms.
6549 static cpumask_t fallback_doms;
6552 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6553 * For now this just excludes isolated cpus, but could be used to
6554 * exclude other special cases in the future.
6556 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6561 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6563 doms_cur = &fallback_doms;
6564 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6565 err = build_sched_domains(doms_cur);
6566 register_sched_domain_sysctl();
6571 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6573 free_sched_groups(cpu_map);
6577 * Detach sched domains from a group of cpus specified in cpu_map
6578 * These cpus will now be attached to the NULL domain
6580 static void detach_destroy_domains(const cpumask_t *cpu_map)
6584 unregister_sched_domain_sysctl();
6586 for_each_cpu_mask(i, *cpu_map)
6587 cpu_attach_domain(NULL, &def_root_domain, i);
6588 synchronize_sched();
6589 arch_destroy_sched_domains(cpu_map);
6593 * Partition sched domains as specified by the 'ndoms_new'
6594 * cpumasks in the array doms_new[] of cpumasks. This compares
6595 * doms_new[] to the current sched domain partitioning, doms_cur[].
6596 * It destroys each deleted domain and builds each new domain.
6598 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6599 * The masks don't intersect (don't overlap.) We should setup one
6600 * sched domain for each mask. CPUs not in any of the cpumasks will
6601 * not be load balanced. If the same cpumask appears both in the
6602 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6605 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6606 * ownership of it and will kfree it when done with it. If the caller
6607 * failed the kmalloc call, then it can pass in doms_new == NULL,
6608 * and partition_sched_domains() will fallback to the single partition
6611 * Call with hotplug lock held
6613 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6619 /* always unregister in case we don't destroy any domains */
6620 unregister_sched_domain_sysctl();
6622 if (doms_new == NULL) {
6624 doms_new = &fallback_doms;
6625 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6628 /* Destroy deleted domains */
6629 for (i = 0; i < ndoms_cur; i++) {
6630 for (j = 0; j < ndoms_new; j++) {
6631 if (cpus_equal(doms_cur[i], doms_new[j]))
6634 /* no match - a current sched domain not in new doms_new[] */
6635 detach_destroy_domains(doms_cur + i);
6640 /* Build new domains */
6641 for (i = 0; i < ndoms_new; i++) {
6642 for (j = 0; j < ndoms_cur; j++) {
6643 if (cpus_equal(doms_new[i], doms_cur[j]))
6646 /* no match - add a new doms_new */
6647 build_sched_domains(doms_new + i);
6652 /* Remember the new sched domains */
6653 if (doms_cur != &fallback_doms)
6655 doms_cur = doms_new;
6656 ndoms_cur = ndoms_new;
6658 register_sched_domain_sysctl();
6663 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6664 static int arch_reinit_sched_domains(void)
6669 detach_destroy_domains(&cpu_online_map);
6670 err = arch_init_sched_domains(&cpu_online_map);
6676 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6680 if (buf[0] != '0' && buf[0] != '1')
6684 sched_smt_power_savings = (buf[0] == '1');
6686 sched_mc_power_savings = (buf[0] == '1');
6688 ret = arch_reinit_sched_domains();
6690 return ret ? ret : count;
6693 #ifdef CONFIG_SCHED_MC
6694 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6696 return sprintf(page, "%u\n", sched_mc_power_savings);
6698 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6699 const char *buf, size_t count)
6701 return sched_power_savings_store(buf, count, 0);
6703 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6704 sched_mc_power_savings_store);
6707 #ifdef CONFIG_SCHED_SMT
6708 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6710 return sprintf(page, "%u\n", sched_smt_power_savings);
6712 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6713 const char *buf, size_t count)
6715 return sched_power_savings_store(buf, count, 1);
6717 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6718 sched_smt_power_savings_store);
6721 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6725 #ifdef CONFIG_SCHED_SMT
6727 err = sysfs_create_file(&cls->kset.kobj,
6728 &attr_sched_smt_power_savings.attr);
6730 #ifdef CONFIG_SCHED_MC
6731 if (!err && mc_capable())
6732 err = sysfs_create_file(&cls->kset.kobj,
6733 &attr_sched_mc_power_savings.attr);
6740 * Force a reinitialization of the sched domains hierarchy. The domains
6741 * and groups cannot be updated in place without racing with the balancing
6742 * code, so we temporarily attach all running cpus to the NULL domain
6743 * which will prevent rebalancing while the sched domains are recalculated.
6745 static int update_sched_domains(struct notifier_block *nfb,
6746 unsigned long action, void *hcpu)
6749 case CPU_UP_PREPARE:
6750 case CPU_UP_PREPARE_FROZEN:
6751 case CPU_DOWN_PREPARE:
6752 case CPU_DOWN_PREPARE_FROZEN:
6753 detach_destroy_domains(&cpu_online_map);
6756 case CPU_UP_CANCELED:
6757 case CPU_UP_CANCELED_FROZEN:
6758 case CPU_DOWN_FAILED:
6759 case CPU_DOWN_FAILED_FROZEN:
6761 case CPU_ONLINE_FROZEN:
6763 case CPU_DEAD_FROZEN:
6765 * Fall through and re-initialise the domains.
6772 /* The hotplug lock is already held by cpu_up/cpu_down */
6773 arch_init_sched_domains(&cpu_online_map);
6778 void __init sched_init_smp(void)
6780 cpumask_t non_isolated_cpus;
6783 arch_init_sched_domains(&cpu_online_map);
6784 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6785 if (cpus_empty(non_isolated_cpus))
6786 cpu_set(smp_processor_id(), non_isolated_cpus);
6788 /* XXX: Theoretical race here - CPU may be hotplugged now */
6789 hotcpu_notifier(update_sched_domains, 0);
6791 /* Move init over to a non-isolated CPU */
6792 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6794 sched_init_granularity();
6796 #ifdef CONFIG_FAIR_GROUP_SCHED
6797 if (nr_cpu_ids == 1)
6800 lb_monitor_task = kthread_create(load_balance_monitor, NULL,
6802 if (!IS_ERR(lb_monitor_task)) {
6803 lb_monitor_task->flags |= PF_NOFREEZE;
6804 wake_up_process(lb_monitor_task);
6806 printk(KERN_ERR "Could not create load balance monitor thread"
6807 "(error = %ld) \n", PTR_ERR(lb_monitor_task));
6812 void __init sched_init_smp(void)
6814 sched_init_granularity();
6816 #endif /* CONFIG_SMP */
6818 int in_sched_functions(unsigned long addr)
6820 return in_lock_functions(addr) ||
6821 (addr >= (unsigned long)__sched_text_start
6822 && addr < (unsigned long)__sched_text_end);
6825 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6827 cfs_rq->tasks_timeline = RB_ROOT;
6828 #ifdef CONFIG_FAIR_GROUP_SCHED
6831 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6834 void __init sched_init(void)
6836 int highest_cpu = 0;
6840 init_defrootdomain();
6843 for_each_possible_cpu(i) {
6844 struct rt_prio_array *array;
6848 spin_lock_init(&rq->lock);
6849 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6852 init_cfs_rq(&rq->cfs, rq);
6853 #ifdef CONFIG_FAIR_GROUP_SCHED
6854 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6856 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6857 struct sched_entity *se =
6858 &per_cpu(init_sched_entity, i);
6860 init_cfs_rq_p[i] = cfs_rq;
6861 init_cfs_rq(cfs_rq, rq);
6862 cfs_rq->tg = &init_task_group;
6863 list_add(&cfs_rq->leaf_cfs_rq_list,
6864 &rq->leaf_cfs_rq_list);
6866 init_sched_entity_p[i] = se;
6867 se->cfs_rq = &rq->cfs;
6869 se->load.weight = init_task_group_load;
6870 se->load.inv_weight =
6871 div64_64(1ULL<<32, init_task_group_load);
6874 init_task_group.shares = init_task_group_load;
6877 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6878 rq->cpu_load[j] = 0;
6882 rq_attach_root(rq, &def_root_domain);
6883 rq->active_balance = 0;
6884 rq->next_balance = jiffies;
6887 rq->migration_thread = NULL;
6888 INIT_LIST_HEAD(&rq->migration_queue);
6889 rq->rt.highest_prio = MAX_RT_PRIO;
6890 rq->rt.overloaded = 0;
6892 atomic_set(&rq->nr_iowait, 0);
6894 array = &rq->rt.active;
6895 for (j = 0; j < MAX_RT_PRIO; j++) {
6896 INIT_LIST_HEAD(array->queue + j);
6897 __clear_bit(j, array->bitmap);
6900 /* delimiter for bitsearch: */
6901 __set_bit(MAX_RT_PRIO, array->bitmap);
6904 set_load_weight(&init_task);
6906 #ifdef CONFIG_PREEMPT_NOTIFIERS
6907 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6911 nr_cpu_ids = highest_cpu + 1;
6912 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6915 #ifdef CONFIG_RT_MUTEXES
6916 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6920 * The boot idle thread does lazy MMU switching as well:
6922 atomic_inc(&init_mm.mm_count);
6923 enter_lazy_tlb(&init_mm, current);
6926 * Make us the idle thread. Technically, schedule() should not be
6927 * called from this thread, however somewhere below it might be,
6928 * but because we are the idle thread, we just pick up running again
6929 * when this runqueue becomes "idle".
6931 init_idle(current, smp_processor_id());
6933 * During early bootup we pretend to be a normal task:
6935 current->sched_class = &fair_sched_class;
6938 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6939 void __might_sleep(char *file, int line)
6942 static unsigned long prev_jiffy; /* ratelimiting */
6944 if ((in_atomic() || irqs_disabled()) &&
6945 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6946 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6948 prev_jiffy = jiffies;
6949 printk(KERN_ERR "BUG: sleeping function called from invalid"
6950 " context at %s:%d\n", file, line);
6951 printk("in_atomic():%d, irqs_disabled():%d\n",
6952 in_atomic(), irqs_disabled());
6953 debug_show_held_locks(current);
6954 if (irqs_disabled())
6955 print_irqtrace_events(current);
6960 EXPORT_SYMBOL(__might_sleep);
6963 #ifdef CONFIG_MAGIC_SYSRQ
6964 static void normalize_task(struct rq *rq, struct task_struct *p)
6967 update_rq_clock(rq);
6968 on_rq = p->se.on_rq;
6970 deactivate_task(rq, p, 0);
6971 __setscheduler(rq, p, SCHED_NORMAL, 0);
6973 activate_task(rq, p, 0);
6974 resched_task(rq->curr);
6978 void normalize_rt_tasks(void)
6980 struct task_struct *g, *p;
6981 unsigned long flags;
6984 read_lock_irq(&tasklist_lock);
6985 do_each_thread(g, p) {
6987 * Only normalize user tasks:
6992 p->se.exec_start = 0;
6993 #ifdef CONFIG_SCHEDSTATS
6994 p->se.wait_start = 0;
6995 p->se.sleep_start = 0;
6996 p->se.block_start = 0;
6998 task_rq(p)->clock = 0;
7002 * Renice negative nice level userspace
7005 if (TASK_NICE(p) < 0 && p->mm)
7006 set_user_nice(p, 0);
7010 spin_lock_irqsave(&p->pi_lock, flags);
7011 rq = __task_rq_lock(p);
7013 normalize_task(rq, p);
7015 __task_rq_unlock(rq);
7016 spin_unlock_irqrestore(&p->pi_lock, flags);
7017 } while_each_thread(g, p);
7019 read_unlock_irq(&tasklist_lock);
7022 #endif /* CONFIG_MAGIC_SYSRQ */
7026 * These functions are only useful for the IA64 MCA handling.
7028 * They can only be called when the whole system has been
7029 * stopped - every CPU needs to be quiescent, and no scheduling
7030 * activity can take place. Using them for anything else would
7031 * be a serious bug, and as a result, they aren't even visible
7032 * under any other configuration.
7036 * curr_task - return the current task for a given cpu.
7037 * @cpu: the processor in question.
7039 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7041 struct task_struct *curr_task(int cpu)
7043 return cpu_curr(cpu);
7047 * set_curr_task - set the current task for a given cpu.
7048 * @cpu: the processor in question.
7049 * @p: the task pointer to set.
7051 * Description: This function must only be used when non-maskable interrupts
7052 * are serviced on a separate stack. It allows the architecture to switch the
7053 * notion of the current task on a cpu in a non-blocking manner. This function
7054 * must be called with all CPU's synchronized, and interrupts disabled, the
7055 * and caller must save the original value of the current task (see
7056 * curr_task() above) and restore that value before reenabling interrupts and
7057 * re-starting the system.
7059 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7061 void set_curr_task(int cpu, struct task_struct *p)
7068 #ifdef CONFIG_FAIR_GROUP_SCHED
7072 * distribute shares of all task groups among their schedulable entities,
7073 * to reflect load distrbution across cpus.
7075 static int rebalance_shares(struct sched_domain *sd, int this_cpu)
7077 struct cfs_rq *cfs_rq;
7078 struct rq *rq = cpu_rq(this_cpu);
7079 cpumask_t sdspan = sd->span;
7082 /* Walk thr' all the task groups that we have */
7083 for_each_leaf_cfs_rq(rq, cfs_rq) {
7085 unsigned long total_load = 0, total_shares;
7086 struct task_group *tg = cfs_rq->tg;
7088 /* Gather total task load of this group across cpus */
7089 for_each_cpu_mask(i, sdspan)
7090 total_load += tg->cfs_rq[i]->load.weight;
7092 /* Nothing to do if this group has no load */
7097 * tg->shares represents the number of cpu shares the task group
7098 * is eligible to hold on a single cpu. On N cpus, it is
7099 * eligible to hold (N * tg->shares) number of cpu shares.
7101 total_shares = tg->shares * cpus_weight(sdspan);
7104 * redistribute total_shares across cpus as per the task load
7107 for_each_cpu_mask(i, sdspan) {
7108 unsigned long local_load, local_shares;
7110 local_load = tg->cfs_rq[i]->load.weight;
7111 local_shares = (local_load * total_shares) / total_load;
7113 local_shares = MIN_GROUP_SHARES;
7114 if (local_shares == tg->se[i]->load.weight)
7117 spin_lock_irq(&cpu_rq(i)->lock);
7118 set_se_shares(tg->se[i], local_shares);
7119 spin_unlock_irq(&cpu_rq(i)->lock);
7128 * How frequently should we rebalance_shares() across cpus?
7130 * The more frequently we rebalance shares, the more accurate is the fairness
7131 * of cpu bandwidth distribution between task groups. However higher frequency
7132 * also implies increased scheduling overhead.
7134 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7135 * consecutive calls to rebalance_shares() in the same sched domain.
7137 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7138 * consecutive calls to rebalance_shares() in the same sched domain.
7140 * These settings allows for the appropriate tradeoff between accuracy of
7141 * fairness and the associated overhead.
7145 /* default: 8ms, units: milliseconds */
7146 const_debug unsigned int sysctl_sched_min_bal_int_shares = 8;
7148 /* default: 128ms, units: milliseconds */
7149 const_debug unsigned int sysctl_sched_max_bal_int_shares = 128;
7151 /* kernel thread that runs rebalance_shares() periodically */
7152 static int load_balance_monitor(void *unused)
7154 unsigned int timeout = sysctl_sched_min_bal_int_shares;
7155 struct sched_param schedparm;
7159 * We don't want this thread's execution to be limited by the shares
7160 * assigned to default group (init_task_group). Hence make it run
7161 * as a SCHED_RR RT task at the lowest priority.
7163 schedparm.sched_priority = 1;
7164 ret = sched_setscheduler(current, SCHED_RR, &schedparm);
7166 printk(KERN_ERR "Couldn't set SCHED_RR policy for load balance"
7167 " monitor thread (error = %d) \n", ret);
7169 while (!kthread_should_stop()) {
7170 int i, cpu, balanced = 1;
7172 /* Prevent cpus going down or coming up */
7174 /* lockout changes to doms_cur[] array */
7177 * Enter a rcu read-side critical section to safely walk rq->sd
7178 * chain on various cpus and to walk task group list
7179 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7183 for (i = 0; i < ndoms_cur; i++) {
7184 cpumask_t cpumap = doms_cur[i];
7185 struct sched_domain *sd = NULL, *sd_prev = NULL;
7187 cpu = first_cpu(cpumap);
7189 /* Find the highest domain at which to balance shares */
7190 for_each_domain(cpu, sd) {
7191 if (!(sd->flags & SD_LOAD_BALANCE))
7197 /* sd == NULL? No load balance reqd in this domain */
7201 balanced &= rebalance_shares(sd, cpu);
7210 timeout = sysctl_sched_min_bal_int_shares;
7211 else if (timeout < sysctl_sched_max_bal_int_shares)
7214 msleep_interruptible(timeout);
7219 #endif /* CONFIG_SMP */
7221 /* allocate runqueue etc for a new task group */
7222 struct task_group *sched_create_group(void)
7224 struct task_group *tg;
7225 struct cfs_rq *cfs_rq;
7226 struct sched_entity *se;
7230 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7232 return ERR_PTR(-ENOMEM);
7234 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7237 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7241 for_each_possible_cpu(i) {
7244 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
7249 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
7254 memset(cfs_rq, 0, sizeof(struct cfs_rq));
7255 memset(se, 0, sizeof(struct sched_entity));
7257 tg->cfs_rq[i] = cfs_rq;
7258 init_cfs_rq(cfs_rq, rq);
7262 se->cfs_rq = &rq->cfs;
7264 se->load.weight = NICE_0_LOAD;
7265 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7269 tg->shares = NICE_0_LOAD;
7271 lock_task_group_list();
7272 for_each_possible_cpu(i) {
7274 cfs_rq = tg->cfs_rq[i];
7275 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7277 unlock_task_group_list();
7282 for_each_possible_cpu(i) {
7284 kfree(tg->cfs_rq[i]);
7292 return ERR_PTR(-ENOMEM);
7295 /* rcu callback to free various structures associated with a task group */
7296 static void free_sched_group(struct rcu_head *rhp)
7298 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7299 struct cfs_rq *cfs_rq;
7300 struct sched_entity *se;
7303 /* now it should be safe to free those cfs_rqs */
7304 for_each_possible_cpu(i) {
7305 cfs_rq = tg->cfs_rq[i];
7317 /* Destroy runqueue etc associated with a task group */
7318 void sched_destroy_group(struct task_group *tg)
7320 struct cfs_rq *cfs_rq = NULL;
7323 lock_task_group_list();
7324 for_each_possible_cpu(i) {
7325 cfs_rq = tg->cfs_rq[i];
7326 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7328 unlock_task_group_list();
7332 /* wait for possible concurrent references to cfs_rqs complete */
7333 call_rcu(&tg->rcu, free_sched_group);
7336 /* change task's runqueue when it moves between groups.
7337 * The caller of this function should have put the task in its new group
7338 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7339 * reflect its new group.
7341 void sched_move_task(struct task_struct *tsk)
7344 unsigned long flags;
7347 rq = task_rq_lock(tsk, &flags);
7349 if (tsk->sched_class != &fair_sched_class) {
7350 set_task_cfs_rq(tsk, task_cpu(tsk));
7354 update_rq_clock(rq);
7356 running = task_current(rq, tsk);
7357 on_rq = tsk->se.on_rq;
7360 dequeue_task(rq, tsk, 0);
7361 if (unlikely(running))
7362 tsk->sched_class->put_prev_task(rq, tsk);
7365 set_task_cfs_rq(tsk, task_cpu(tsk));
7368 if (unlikely(running))
7369 tsk->sched_class->set_curr_task(rq);
7370 enqueue_task(rq, tsk, 0);
7374 task_rq_unlock(rq, &flags);
7377 /* rq->lock to be locked by caller */
7378 static void set_se_shares(struct sched_entity *se, unsigned long shares)
7380 struct cfs_rq *cfs_rq = se->cfs_rq;
7381 struct rq *rq = cfs_rq->rq;
7385 shares = MIN_GROUP_SHARES;
7389 dequeue_entity(cfs_rq, se, 0);
7390 dec_cpu_load(rq, se->load.weight);
7393 se->load.weight = shares;
7394 se->load.inv_weight = div64_64((1ULL<<32), shares);
7397 enqueue_entity(cfs_rq, se, 0);
7398 inc_cpu_load(rq, se->load.weight);
7402 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7405 struct cfs_rq *cfs_rq;
7408 lock_task_group_list();
7409 if (tg->shares == shares)
7412 if (shares < MIN_GROUP_SHARES)
7413 shares = MIN_GROUP_SHARES;
7416 * Prevent any load balance activity (rebalance_shares,
7417 * load_balance_fair) from referring to this group first,
7418 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7420 for_each_possible_cpu(i) {
7421 cfs_rq = tg->cfs_rq[i];
7422 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7425 /* wait for any ongoing reference to this group to finish */
7426 synchronize_sched();
7429 * Now we are free to modify the group's share on each cpu
7430 * w/o tripping rebalance_share or load_balance_fair.
7432 tg->shares = shares;
7433 for_each_possible_cpu(i) {
7434 spin_lock_irq(&cpu_rq(i)->lock);
7435 set_se_shares(tg->se[i], shares);
7436 spin_unlock_irq(&cpu_rq(i)->lock);
7440 * Enable load balance activity on this group, by inserting it back on
7441 * each cpu's rq->leaf_cfs_rq_list.
7443 for_each_possible_cpu(i) {
7445 cfs_rq = tg->cfs_rq[i];
7446 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7449 unlock_task_group_list();
7453 unsigned long sched_group_shares(struct task_group *tg)
7458 #endif /* CONFIG_FAIR_GROUP_SCHED */
7460 #ifdef CONFIG_FAIR_CGROUP_SCHED
7462 /* return corresponding task_group object of a cgroup */
7463 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7465 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7466 struct task_group, css);
7469 static struct cgroup_subsys_state *
7470 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7472 struct task_group *tg;
7474 if (!cgrp->parent) {
7475 /* This is early initialization for the top cgroup */
7476 init_task_group.css.cgroup = cgrp;
7477 return &init_task_group.css;
7480 /* we support only 1-level deep hierarchical scheduler atm */
7481 if (cgrp->parent->parent)
7482 return ERR_PTR(-EINVAL);
7484 tg = sched_create_group();
7486 return ERR_PTR(-ENOMEM);
7488 /* Bind the cgroup to task_group object we just created */
7489 tg->css.cgroup = cgrp;
7495 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7497 struct task_group *tg = cgroup_tg(cgrp);
7499 sched_destroy_group(tg);
7503 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7504 struct task_struct *tsk)
7506 /* We don't support RT-tasks being in separate groups */
7507 if (tsk->sched_class != &fair_sched_class)
7514 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7515 struct cgroup *old_cont, struct task_struct *tsk)
7517 sched_move_task(tsk);
7520 static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7523 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7526 static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7528 struct task_group *tg = cgroup_tg(cgrp);
7530 return (u64) tg->shares;
7533 static struct cftype cpu_files[] = {
7536 .read_uint = cpu_shares_read_uint,
7537 .write_uint = cpu_shares_write_uint,
7541 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7543 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7546 struct cgroup_subsys cpu_cgroup_subsys = {
7548 .create = cpu_cgroup_create,
7549 .destroy = cpu_cgroup_destroy,
7550 .can_attach = cpu_cgroup_can_attach,
7551 .attach = cpu_cgroup_attach,
7552 .populate = cpu_cgroup_populate,
7553 .subsys_id = cpu_cgroup_subsys_id,
7557 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7559 #ifdef CONFIG_CGROUP_CPUACCT
7562 * CPU accounting code for task groups.
7564 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7565 * (balbir@in.ibm.com).
7568 /* track cpu usage of a group of tasks */
7570 struct cgroup_subsys_state css;
7571 /* cpuusage holds pointer to a u64-type object on every cpu */
7575 struct cgroup_subsys cpuacct_subsys;
7577 /* return cpu accounting group corresponding to this container */
7578 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7580 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7581 struct cpuacct, css);
7584 /* return cpu accounting group to which this task belongs */
7585 static inline struct cpuacct *task_ca(struct task_struct *tsk)
7587 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7588 struct cpuacct, css);
7591 /* create a new cpu accounting group */
7592 static struct cgroup_subsys_state *cpuacct_create(
7593 struct cgroup_subsys *ss, struct cgroup *cont)
7595 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7598 return ERR_PTR(-ENOMEM);
7600 ca->cpuusage = alloc_percpu(u64);
7601 if (!ca->cpuusage) {
7603 return ERR_PTR(-ENOMEM);
7609 /* destroy an existing cpu accounting group */
7611 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
7613 struct cpuacct *ca = cgroup_ca(cont);
7615 free_percpu(ca->cpuusage);
7619 /* return total cpu usage (in nanoseconds) of a group */
7620 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
7622 struct cpuacct *ca = cgroup_ca(cont);
7623 u64 totalcpuusage = 0;
7626 for_each_possible_cpu(i) {
7627 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
7630 * Take rq->lock to make 64-bit addition safe on 32-bit
7633 spin_lock_irq(&cpu_rq(i)->lock);
7634 totalcpuusage += *cpuusage;
7635 spin_unlock_irq(&cpu_rq(i)->lock);
7638 return totalcpuusage;
7641 static struct cftype files[] = {
7644 .read_uint = cpuusage_read,
7648 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7650 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
7654 * charge this task's execution time to its accounting group.
7656 * called with rq->lock held.
7658 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
7662 if (!cpuacct_subsys.active)
7667 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
7669 *cpuusage += cputime;
7673 struct cgroup_subsys cpuacct_subsys = {
7675 .create = cpuacct_create,
7676 .destroy = cpuacct_destroy,
7677 .populate = cpuacct_populate,
7678 .subsys_id = cpuacct_subsys_id,
7680 #endif /* CONFIG_CGROUP_CPUACCT */