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/smp.h>
48 #include <linux/threads.h>
49 #include <linux/timer.h>
50 #include <linux/rcupdate.h>
51 #include <linux/cpu.h>
52 #include <linux/cpuset.h>
53 #include <linux/percpu.h>
54 #include <linux/kthread.h>
55 #include <linux/seq_file.h>
56 #include <linux/sysctl.h>
57 #include <linux/syscalls.h>
58 #include <linux/times.h>
59 #include <linux/tsacct_kern.h>
60 #include <linux/kprobes.h>
61 #include <linux/delayacct.h>
62 #include <linux/reciprocal_div.h>
63 #include <linux/unistd.h>
64 #include <linux/pagemap.h>
69 * Scheduler clock - returns current time in nanosec units.
70 * This is default implementation.
71 * Architectures and sub-architectures can override this.
73 unsigned long long __attribute__((weak)) sched_clock(void)
75 return (unsigned long long)jiffies * (1000000000 / HZ);
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Some helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
100 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
109 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
110 * Timeslices get refilled after they expire.
112 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
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 #define SCALE_PRIO(x, prio) \
137 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
140 * static_prio_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
141 * to time slice values: [800ms ... 100ms ... 5ms]
143 static unsigned int static_prio_timeslice(int static_prio)
145 if (static_prio == NICE_TO_PRIO(19))
148 if (static_prio < NICE_TO_PRIO(0))
149 return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
151 return SCALE_PRIO(DEF_TIMESLICE, static_prio);
154 static inline int rt_policy(int policy)
156 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
161 static inline int task_has_rt_policy(struct task_struct *p)
163 return rt_policy(p->policy);
167 * This is the priority-queue data structure of the RT scheduling class:
169 struct rt_prio_array {
170 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
171 struct list_head queue[MAX_RT_PRIO];
174 #ifdef CONFIG_FAIR_GROUP_SCHED
178 /* task group related information */
180 /* schedulable entities of this group on each cpu */
181 struct sched_entity **se;
182 /* runqueue "owned" by this group on each cpu */
183 struct cfs_rq **cfs_rq;
184 unsigned long shares;
187 /* Default task group's sched entity on each cpu */
188 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
189 /* Default task group's cfs_rq on each cpu */
190 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
192 static struct sched_entity *init_sched_entity_p[NR_CPUS];
193 static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
195 /* Default task group.
196 * Every task in system belong to this group at bootup.
198 struct task_grp init_task_grp = {
199 .se = init_sched_entity_p,
200 .cfs_rq = init_cfs_rq_p,
203 #ifdef CONFIG_FAIR_USER_SCHED
204 #define INIT_TASK_GRP_LOAD 2*NICE_0_LOAD
206 #define INIT_TASK_GRP_LOAD NICE_0_LOAD
209 static int init_task_grp_load = INIT_TASK_GRP_LOAD;
211 /* return group to which a task belongs */
212 static inline struct task_grp *task_grp(struct task_struct *p)
216 #ifdef CONFIG_FAIR_USER_SCHED
225 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
226 static inline void set_task_cfs_rq(struct task_struct *p)
228 p->se.cfs_rq = task_grp(p)->cfs_rq[task_cpu(p)];
229 p->se.parent = task_grp(p)->se[task_cpu(p)];
234 static inline void set_task_cfs_rq(struct task_struct *p) { }
236 #endif /* CONFIG_FAIR_GROUP_SCHED */
238 /* CFS-related fields in a runqueue */
240 struct load_weight load;
241 unsigned long nr_running;
246 struct rb_root tasks_timeline;
247 struct rb_node *rb_leftmost;
248 struct rb_node *rb_load_balance_curr;
249 /* 'curr' points to currently running entity on this cfs_rq.
250 * It is set to NULL otherwise (i.e when none are currently running).
252 struct sched_entity *curr;
254 unsigned long nr_spread_over;
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
259 /* leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
260 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
261 * (like users, containers etc.)
263 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
264 * list is used during load balance.
266 struct list_head leaf_cfs_rq_list; /* Better name : task_cfs_rq_list? */
267 struct task_grp *tg; /* group that "owns" this runqueue */
272 /* Real-Time classes' related field in a runqueue: */
274 struct rt_prio_array active;
275 int rt_load_balance_idx;
276 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
280 * This is the main, per-CPU runqueue data structure.
282 * Locking rule: those places that want to lock multiple runqueues
283 * (such as the load balancing or the thread migration code), lock
284 * acquire operations must be ordered by ascending &runqueue.
287 spinlock_t lock; /* runqueue lock */
290 * nr_running and cpu_load should be in the same cacheline because
291 * remote CPUs use both these fields when doing load calculation.
293 unsigned long nr_running;
294 #define CPU_LOAD_IDX_MAX 5
295 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
296 unsigned char idle_at_tick;
298 unsigned char in_nohz_recently;
300 struct load_weight load; /* capture load from *all* tasks on this cpu */
301 unsigned long nr_load_updates;
305 #ifdef CONFIG_FAIR_GROUP_SCHED
306 struct list_head leaf_cfs_rq_list; /* list of leaf cfs_rq on this cpu */
311 * This is part of a global counter where only the total sum
312 * over all CPUs matters. A task can increase this counter on
313 * one CPU and if it got migrated afterwards it may decrease
314 * it on another CPU. Always updated under the runqueue lock:
316 unsigned long nr_uninterruptible;
318 struct task_struct *curr, *idle;
319 unsigned long next_balance;
320 struct mm_struct *prev_mm;
322 u64 clock, prev_clock_raw;
325 unsigned int clock_warps, clock_overflows;
327 unsigned int clock_deep_idle_events;
333 struct sched_domain *sd;
335 /* For active balancing */
338 int cpu; /* cpu of this runqueue */
340 struct task_struct *migration_thread;
341 struct list_head migration_queue;
344 #ifdef CONFIG_SCHEDSTATS
346 struct sched_info rq_sched_info;
348 /* sys_sched_yield() stats */
349 unsigned long yld_exp_empty;
350 unsigned long yld_act_empty;
351 unsigned long yld_both_empty;
352 unsigned long yld_cnt;
354 /* schedule() stats */
355 unsigned long sched_switch;
356 unsigned long sched_cnt;
357 unsigned long sched_goidle;
359 /* try_to_wake_up() stats */
360 unsigned long ttwu_cnt;
361 unsigned long ttwu_local;
364 unsigned long bkl_cnt;
366 struct lock_class_key rq_lock_key;
369 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
370 static DEFINE_MUTEX(sched_hotcpu_mutex);
372 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
374 rq->curr->sched_class->check_preempt_curr(rq, p);
377 static inline int cpu_of(struct rq *rq)
387 * Update the per-runqueue clock, as finegrained as the platform can give
388 * us, but without assuming monotonicity, etc.:
390 static void __update_rq_clock(struct rq *rq)
392 u64 prev_raw = rq->prev_clock_raw;
393 u64 now = sched_clock();
394 s64 delta = now - prev_raw;
395 u64 clock = rq->clock;
397 #ifdef CONFIG_SCHED_DEBUG
398 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
401 * Protect against sched_clock() occasionally going backwards:
403 if (unlikely(delta < 0)) {
408 * Catch too large forward jumps too:
410 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
411 if (clock < rq->tick_timestamp + TICK_NSEC)
412 clock = rq->tick_timestamp + TICK_NSEC;
415 rq->clock_overflows++;
417 if (unlikely(delta > rq->clock_max_delta))
418 rq->clock_max_delta = delta;
423 rq->prev_clock_raw = now;
427 static void update_rq_clock(struct rq *rq)
429 if (likely(smp_processor_id() == cpu_of(rq)))
430 __update_rq_clock(rq);
434 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
435 * See detach_destroy_domains: synchronize_sched for details.
437 * The domain tree of any CPU may only be accessed from within
438 * preempt-disabled sections.
440 #define for_each_domain(cpu, __sd) \
441 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
443 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
444 #define this_rq() (&__get_cpu_var(runqueues))
445 #define task_rq(p) cpu_rq(task_cpu(p))
446 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
449 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
451 #ifdef CONFIG_SCHED_DEBUG
452 # define const_debug __read_mostly
454 # define const_debug static const
458 * Debugging: various feature bits
461 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
462 SCHED_FEAT_START_DEBIT = 2,
463 SCHED_FEAT_USE_TREE_AVG = 4,
464 SCHED_FEAT_APPROX_AVG = 8,
467 const_debug unsigned int sysctl_sched_features =
468 SCHED_FEAT_NEW_FAIR_SLEEPERS *1 |
469 SCHED_FEAT_START_DEBIT *1 |
470 SCHED_FEAT_USE_TREE_AVG *0 |
471 SCHED_FEAT_APPROX_AVG *0;
473 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
476 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
477 * clock constructed from sched_clock():
479 unsigned long long cpu_clock(int cpu)
481 unsigned long long now;
485 local_irq_save(flags);
489 local_irq_restore(flags);
494 #ifndef prepare_arch_switch
495 # define prepare_arch_switch(next) do { } while (0)
497 #ifndef finish_arch_switch
498 # define finish_arch_switch(prev) do { } while (0)
501 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
502 static inline int task_running(struct rq *rq, struct task_struct *p)
504 return rq->curr == p;
507 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
511 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
513 #ifdef CONFIG_DEBUG_SPINLOCK
514 /* this is a valid case when another task releases the spinlock */
515 rq->lock.owner = current;
518 * If we are tracking spinlock dependencies then we have to
519 * fix up the runqueue lock - which gets 'carried over' from
522 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
524 spin_unlock_irq(&rq->lock);
527 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
528 static inline int task_running(struct rq *rq, struct task_struct *p)
533 return rq->curr == p;
537 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
541 * We can optimise this out completely for !SMP, because the
542 * SMP rebalancing from interrupt is the only thing that cares
547 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
548 spin_unlock_irq(&rq->lock);
550 spin_unlock(&rq->lock);
554 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
558 * After ->oncpu is cleared, the task can be moved to a different CPU.
559 * We must ensure this doesn't happen until the switch is completely
565 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
569 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
572 * __task_rq_lock - lock the runqueue a given task resides on.
573 * Must be called interrupts disabled.
575 static inline struct rq *__task_rq_lock(struct task_struct *p)
582 spin_lock(&rq->lock);
583 if (unlikely(rq != task_rq(p))) {
584 spin_unlock(&rq->lock);
585 goto repeat_lock_task;
591 * task_rq_lock - lock the runqueue a given task resides on and disable
592 * interrupts. Note the ordering: we can safely lookup the task_rq without
593 * explicitly disabling preemption.
595 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
601 local_irq_save(*flags);
603 spin_lock(&rq->lock);
604 if (unlikely(rq != task_rq(p))) {
605 spin_unlock_irqrestore(&rq->lock, *flags);
606 goto repeat_lock_task;
611 static inline void __task_rq_unlock(struct rq *rq)
614 spin_unlock(&rq->lock);
617 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
620 spin_unlock_irqrestore(&rq->lock, *flags);
624 * this_rq_lock - lock this runqueue and disable interrupts.
626 static inline struct rq *this_rq_lock(void)
633 spin_lock(&rq->lock);
639 * We are going deep-idle (irqs are disabled):
641 void sched_clock_idle_sleep_event(void)
643 struct rq *rq = cpu_rq(smp_processor_id());
645 spin_lock(&rq->lock);
646 __update_rq_clock(rq);
647 spin_unlock(&rq->lock);
648 rq->clock_deep_idle_events++;
650 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
653 * We just idled delta nanoseconds (called with irqs disabled):
655 void sched_clock_idle_wakeup_event(u64 delta_ns)
657 struct rq *rq = cpu_rq(smp_processor_id());
658 u64 now = sched_clock();
660 rq->idle_clock += delta_ns;
662 * Override the previous timestamp and ignore all
663 * sched_clock() deltas that occured while we idled,
664 * and use the PM-provided delta_ns to advance the
667 spin_lock(&rq->lock);
668 rq->prev_clock_raw = now;
669 rq->clock += delta_ns;
670 spin_unlock(&rq->lock);
672 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
675 * resched_task - mark a task 'to be rescheduled now'.
677 * On UP this means the setting of the need_resched flag, on SMP it
678 * might also involve a cross-CPU call to trigger the scheduler on
683 #ifndef tsk_is_polling
684 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
687 static void resched_task(struct task_struct *p)
691 assert_spin_locked(&task_rq(p)->lock);
693 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
696 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
699 if (cpu == smp_processor_id())
702 /* NEED_RESCHED must be visible before we test polling */
704 if (!tsk_is_polling(p))
705 smp_send_reschedule(cpu);
708 static void resched_cpu(int cpu)
710 struct rq *rq = cpu_rq(cpu);
713 if (!spin_trylock_irqsave(&rq->lock, flags))
715 resched_task(cpu_curr(cpu));
716 spin_unlock_irqrestore(&rq->lock, flags);
719 static inline void resched_task(struct task_struct *p)
721 assert_spin_locked(&task_rq(p)->lock);
722 set_tsk_need_resched(p);
726 #if BITS_PER_LONG == 32
727 # define WMULT_CONST (~0UL)
729 # define WMULT_CONST (1UL << 32)
732 #define WMULT_SHIFT 32
735 * Shift right and round:
737 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
740 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
741 struct load_weight *lw)
745 if (unlikely(!lw->inv_weight))
746 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
748 tmp = (u64)delta_exec * weight;
750 * Check whether we'd overflow the 64-bit multiplication:
752 if (unlikely(tmp > WMULT_CONST))
753 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
756 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
758 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
761 static inline unsigned long
762 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
764 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
767 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
772 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
778 * To aid in avoiding the subversion of "niceness" due to uneven distribution
779 * of tasks with abnormal "nice" values across CPUs the contribution that
780 * each task makes to its run queue's load is weighted according to its
781 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
782 * scaled version of the new time slice allocation that they receive on time
786 #define WEIGHT_IDLEPRIO 2
787 #define WMULT_IDLEPRIO (1 << 31)
790 * Nice levels are multiplicative, with a gentle 10% change for every
791 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
792 * nice 1, it will get ~10% less CPU time than another CPU-bound task
793 * that remained on nice 0.
795 * The "10% effect" is relative and cumulative: from _any_ nice level,
796 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
797 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
798 * If a task goes up by ~10% and another task goes down by ~10% then
799 * the relative distance between them is ~25%.)
801 static const int prio_to_weight[40] = {
802 /* -20 */ 88761, 71755, 56483, 46273, 36291,
803 /* -15 */ 29154, 23254, 18705, 14949, 11916,
804 /* -10 */ 9548, 7620, 6100, 4904, 3906,
805 /* -5 */ 3121, 2501, 1991, 1586, 1277,
806 /* 0 */ 1024, 820, 655, 526, 423,
807 /* 5 */ 335, 272, 215, 172, 137,
808 /* 10 */ 110, 87, 70, 56, 45,
809 /* 15 */ 36, 29, 23, 18, 15,
813 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
815 * In cases where the weight does not change often, we can use the
816 * precalculated inverse to speed up arithmetics by turning divisions
817 * into multiplications:
819 static const u32 prio_to_wmult[40] = {
820 /* -20 */ 48388, 59856, 76040, 92818, 118348,
821 /* -15 */ 147320, 184698, 229616, 287308, 360437,
822 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
823 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
824 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
825 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
826 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
827 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
830 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
833 * runqueue iterator, to support SMP load-balancing between different
834 * scheduling classes, without having to expose their internal data
835 * structures to the load-balancing proper:
839 struct task_struct *(*start)(void *);
840 struct task_struct *(*next)(void *);
843 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
844 unsigned long max_nr_move, unsigned long max_load_move,
845 struct sched_domain *sd, enum cpu_idle_type idle,
846 int *all_pinned, unsigned long *load_moved,
847 int *this_best_prio, struct rq_iterator *iterator);
849 #include "sched_stats.h"
850 #include "sched_rt.c"
851 #include "sched_fair.c"
852 #include "sched_idletask.c"
853 #ifdef CONFIG_SCHED_DEBUG
854 # include "sched_debug.c"
857 #define sched_class_highest (&rt_sched_class)
860 * Update delta_exec, delta_fair fields for rq.
862 * delta_fair clock advances at a rate inversely proportional to
863 * total load (rq->load.weight) on the runqueue, while
864 * delta_exec advances at the same rate as wall-clock (provided
867 * delta_exec / delta_fair is a measure of the (smoothened) load on this
868 * runqueue over any given interval. This (smoothened) load is used
869 * during load balance.
871 * This function is called /before/ updating rq->load
872 * and when switching tasks.
874 static inline void inc_load(struct rq *rq, const struct task_struct *p)
876 update_load_add(&rq->load, p->se.load.weight);
879 static inline void dec_load(struct rq *rq, const struct task_struct *p)
881 update_load_sub(&rq->load, p->se.load.weight);
884 static void inc_nr_running(struct task_struct *p, struct rq *rq)
890 static void dec_nr_running(struct task_struct *p, struct rq *rq)
896 static void set_load_weight(struct task_struct *p)
898 if (task_has_rt_policy(p)) {
899 p->se.load.weight = prio_to_weight[0] * 2;
900 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
905 * SCHED_IDLE tasks get minimal weight:
907 if (p->policy == SCHED_IDLE) {
908 p->se.load.weight = WEIGHT_IDLEPRIO;
909 p->se.load.inv_weight = WMULT_IDLEPRIO;
913 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
914 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
917 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
919 sched_info_queued(p);
920 p->sched_class->enqueue_task(rq, p, wakeup);
924 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
926 p->sched_class->dequeue_task(rq, p, sleep);
931 * __normal_prio - return the priority that is based on the static prio
933 static inline int __normal_prio(struct task_struct *p)
935 return p->static_prio;
939 * Calculate the expected normal priority: i.e. priority
940 * without taking RT-inheritance into account. Might be
941 * boosted by interactivity modifiers. Changes upon fork,
942 * setprio syscalls, and whenever the interactivity
943 * estimator recalculates.
945 static inline int normal_prio(struct task_struct *p)
949 if (task_has_rt_policy(p))
950 prio = MAX_RT_PRIO-1 - p->rt_priority;
952 prio = __normal_prio(p);
957 * Calculate the current priority, i.e. the priority
958 * taken into account by the scheduler. This value might
959 * be boosted by RT tasks, or might be boosted by
960 * interactivity modifiers. Will be RT if the task got
961 * RT-boosted. If not then it returns p->normal_prio.
963 static int effective_prio(struct task_struct *p)
965 p->normal_prio = normal_prio(p);
967 * If we are RT tasks or we were boosted to RT priority,
968 * keep the priority unchanged. Otherwise, update priority
969 * to the normal priority:
971 if (!rt_prio(p->prio))
972 return p->normal_prio;
977 * activate_task - move a task to the runqueue.
979 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
981 if (p->state == TASK_UNINTERRUPTIBLE)
982 rq->nr_uninterruptible--;
984 enqueue_task(rq, p, wakeup);
985 inc_nr_running(p, rq);
989 * activate_idle_task - move idle task to the _front_ of runqueue.
991 static inline void activate_idle_task(struct task_struct *p, struct rq *rq)
995 if (p->state == TASK_UNINTERRUPTIBLE)
996 rq->nr_uninterruptible--;
998 enqueue_task(rq, p, 0);
999 inc_nr_running(p, rq);
1003 * deactivate_task - remove a task from the runqueue.
1005 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1007 if (p->state == TASK_UNINTERRUPTIBLE)
1008 rq->nr_uninterruptible++;
1010 dequeue_task(rq, p, sleep);
1011 dec_nr_running(p, rq);
1015 * task_curr - is this task currently executing on a CPU?
1016 * @p: the task in question.
1018 inline int task_curr(const struct task_struct *p)
1020 return cpu_curr(task_cpu(p)) == p;
1023 /* Used instead of source_load when we know the type == 0 */
1024 unsigned long weighted_cpuload(const int cpu)
1026 return cpu_rq(cpu)->load.weight;
1029 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1032 task_thread_info(p)->cpu = cpu;
1039 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1041 int old_cpu = task_cpu(p);
1042 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1045 clock_offset = old_rq->clock - new_rq->clock;
1047 #ifdef CONFIG_SCHEDSTATS
1048 if (p->se.wait_start)
1049 p->se.wait_start -= clock_offset;
1050 if (p->se.sleep_start)
1051 p->se.sleep_start -= clock_offset;
1052 if (p->se.block_start)
1053 p->se.block_start -= clock_offset;
1055 p->se.vruntime -= old_rq->cfs.min_vruntime - new_rq->cfs.min_vruntime;
1057 __set_task_cpu(p, new_cpu);
1060 struct migration_req {
1061 struct list_head list;
1063 struct task_struct *task;
1066 struct completion done;
1070 * The task's runqueue lock must be held.
1071 * Returns true if you have to wait for migration thread.
1074 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1076 struct rq *rq = task_rq(p);
1079 * If the task is not on a runqueue (and not running), then
1080 * it is sufficient to simply update the task's cpu field.
1082 if (!p->se.on_rq && !task_running(rq, p)) {
1083 set_task_cpu(p, dest_cpu);
1087 init_completion(&req->done);
1089 req->dest_cpu = dest_cpu;
1090 list_add(&req->list, &rq->migration_queue);
1096 * wait_task_inactive - wait for a thread to unschedule.
1098 * The caller must ensure that the task *will* unschedule sometime soon,
1099 * else this function might spin for a *long* time. This function can't
1100 * be called with interrupts off, or it may introduce deadlock with
1101 * smp_call_function() if an IPI is sent by the same process we are
1102 * waiting to become inactive.
1104 void wait_task_inactive(struct task_struct *p)
1106 unsigned long flags;
1112 * We do the initial early heuristics without holding
1113 * any task-queue locks at all. We'll only try to get
1114 * the runqueue lock when things look like they will
1120 * If the task is actively running on another CPU
1121 * still, just relax and busy-wait without holding
1124 * NOTE! Since we don't hold any locks, it's not
1125 * even sure that "rq" stays as the right runqueue!
1126 * But we don't care, since "task_running()" will
1127 * return false if the runqueue has changed and p
1128 * is actually now running somewhere else!
1130 while (task_running(rq, p))
1134 * Ok, time to look more closely! We need the rq
1135 * lock now, to be *sure*. If we're wrong, we'll
1136 * just go back and repeat.
1138 rq = task_rq_lock(p, &flags);
1139 running = task_running(rq, p);
1140 on_rq = p->se.on_rq;
1141 task_rq_unlock(rq, &flags);
1144 * Was it really running after all now that we
1145 * checked with the proper locks actually held?
1147 * Oops. Go back and try again..
1149 if (unlikely(running)) {
1155 * It's not enough that it's not actively running,
1156 * it must be off the runqueue _entirely_, and not
1159 * So if it wa still runnable (but just not actively
1160 * running right now), it's preempted, and we should
1161 * yield - it could be a while.
1163 if (unlikely(on_rq)) {
1169 * Ahh, all good. It wasn't running, and it wasn't
1170 * runnable, which means that it will never become
1171 * running in the future either. We're all done!
1176 * kick_process - kick a running thread to enter/exit the kernel
1177 * @p: the to-be-kicked thread
1179 * Cause a process which is running on another CPU to enter
1180 * kernel-mode, without any delay. (to get signals handled.)
1182 * NOTE: this function doesnt have to take the runqueue lock,
1183 * because all it wants to ensure is that the remote task enters
1184 * the kernel. If the IPI races and the task has been migrated
1185 * to another CPU then no harm is done and the purpose has been
1188 void kick_process(struct task_struct *p)
1194 if ((cpu != smp_processor_id()) && task_curr(p))
1195 smp_send_reschedule(cpu);
1200 * Return a low guess at the load of a migration-source cpu weighted
1201 * according to the scheduling class and "nice" value.
1203 * We want to under-estimate the load of migration sources, to
1204 * balance conservatively.
1206 static inline unsigned long source_load(int cpu, int type)
1208 struct rq *rq = cpu_rq(cpu);
1209 unsigned long total = weighted_cpuload(cpu);
1214 return min(rq->cpu_load[type-1], total);
1218 * Return a high guess at the load of a migration-target cpu weighted
1219 * according to the scheduling class and "nice" value.
1221 static inline unsigned long target_load(int cpu, int type)
1223 struct rq *rq = cpu_rq(cpu);
1224 unsigned long total = weighted_cpuload(cpu);
1229 return max(rq->cpu_load[type-1], total);
1233 * Return the average load per task on the cpu's run queue
1235 static inline unsigned long cpu_avg_load_per_task(int cpu)
1237 struct rq *rq = cpu_rq(cpu);
1238 unsigned long total = weighted_cpuload(cpu);
1239 unsigned long n = rq->nr_running;
1241 return n ? total / n : SCHED_LOAD_SCALE;
1245 * find_idlest_group finds and returns the least busy CPU group within the
1248 static struct sched_group *
1249 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1251 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1252 unsigned long min_load = ULONG_MAX, this_load = 0;
1253 int load_idx = sd->forkexec_idx;
1254 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1257 unsigned long load, avg_load;
1261 /* Skip over this group if it has no CPUs allowed */
1262 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1265 local_group = cpu_isset(this_cpu, group->cpumask);
1267 /* Tally up the load of all CPUs in the group */
1270 for_each_cpu_mask(i, group->cpumask) {
1271 /* Bias balancing toward cpus of our domain */
1273 load = source_load(i, load_idx);
1275 load = target_load(i, load_idx);
1280 /* Adjust by relative CPU power of the group */
1281 avg_load = sg_div_cpu_power(group,
1282 avg_load * SCHED_LOAD_SCALE);
1285 this_load = avg_load;
1287 } else if (avg_load < min_load) {
1288 min_load = avg_load;
1292 group = group->next;
1293 } while (group != sd->groups);
1295 if (!idlest || 100*this_load < imbalance*min_load)
1301 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1304 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1307 unsigned long load, min_load = ULONG_MAX;
1311 /* Traverse only the allowed CPUs */
1312 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1314 for_each_cpu_mask(i, tmp) {
1315 load = weighted_cpuload(i);
1317 if (load < min_load || (load == min_load && i == this_cpu)) {
1327 * sched_balance_self: balance the current task (running on cpu) in domains
1328 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1331 * Balance, ie. select the least loaded group.
1333 * Returns the target CPU number, or the same CPU if no balancing is needed.
1335 * preempt must be disabled.
1337 static int sched_balance_self(int cpu, int flag)
1339 struct task_struct *t = current;
1340 struct sched_domain *tmp, *sd = NULL;
1342 for_each_domain(cpu, tmp) {
1344 * If power savings logic is enabled for a domain, stop there.
1346 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1348 if (tmp->flags & flag)
1354 struct sched_group *group;
1355 int new_cpu, weight;
1357 if (!(sd->flags & flag)) {
1363 group = find_idlest_group(sd, t, cpu);
1369 new_cpu = find_idlest_cpu(group, t, cpu);
1370 if (new_cpu == -1 || new_cpu == cpu) {
1371 /* Now try balancing at a lower domain level of cpu */
1376 /* Now try balancing at a lower domain level of new_cpu */
1379 weight = cpus_weight(span);
1380 for_each_domain(cpu, tmp) {
1381 if (weight <= cpus_weight(tmp->span))
1383 if (tmp->flags & flag)
1386 /* while loop will break here if sd == NULL */
1392 #endif /* CONFIG_SMP */
1395 * wake_idle() will wake a task on an idle cpu if task->cpu is
1396 * not idle and an idle cpu is available. The span of cpus to
1397 * search starts with cpus closest then further out as needed,
1398 * so we always favor a closer, idle cpu.
1400 * Returns the CPU we should wake onto.
1402 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1403 static int wake_idle(int cpu, struct task_struct *p)
1406 struct sched_domain *sd;
1410 * If it is idle, then it is the best cpu to run this task.
1412 * This cpu is also the best, if it has more than one task already.
1413 * Siblings must be also busy(in most cases) as they didn't already
1414 * pickup the extra load from this cpu and hence we need not check
1415 * sibling runqueue info. This will avoid the checks and cache miss
1416 * penalities associated with that.
1418 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1421 for_each_domain(cpu, sd) {
1422 if (sd->flags & SD_WAKE_IDLE) {
1423 cpus_and(tmp, sd->span, p->cpus_allowed);
1424 for_each_cpu_mask(i, tmp) {
1435 static inline int wake_idle(int cpu, struct task_struct *p)
1442 * try_to_wake_up - wake up a thread
1443 * @p: the to-be-woken-up thread
1444 * @state: the mask of task states that can be woken
1445 * @sync: do a synchronous wakeup?
1447 * Put it on the run-queue if it's not already there. The "current"
1448 * thread is always on the run-queue (except when the actual
1449 * re-schedule is in progress), and as such you're allowed to do
1450 * the simpler "current->state = TASK_RUNNING" to mark yourself
1451 * runnable without the overhead of this.
1453 * returns failure only if the task is already active.
1455 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1457 int cpu, this_cpu, success = 0;
1458 unsigned long flags;
1462 struct sched_domain *sd, *this_sd = NULL;
1463 unsigned long load, this_load;
1467 rq = task_rq_lock(p, &flags);
1468 old_state = p->state;
1469 if (!(old_state & state))
1476 this_cpu = smp_processor_id();
1479 if (unlikely(task_running(rq, p)))
1484 schedstat_inc(rq, ttwu_cnt);
1485 if (cpu == this_cpu) {
1486 schedstat_inc(rq, ttwu_local);
1490 for_each_domain(this_cpu, sd) {
1491 if (cpu_isset(cpu, sd->span)) {
1492 schedstat_inc(sd, ttwu_wake_remote);
1498 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1502 * Check for affine wakeup and passive balancing possibilities.
1505 int idx = this_sd->wake_idx;
1506 unsigned int imbalance;
1508 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1510 load = source_load(cpu, idx);
1511 this_load = target_load(this_cpu, idx);
1513 new_cpu = this_cpu; /* Wake to this CPU if we can */
1515 if (this_sd->flags & SD_WAKE_AFFINE) {
1516 unsigned long tl = this_load;
1517 unsigned long tl_per_task;
1519 tl_per_task = cpu_avg_load_per_task(this_cpu);
1522 * If sync wakeup then subtract the (maximum possible)
1523 * effect of the currently running task from the load
1524 * of the current CPU:
1527 tl -= current->se.load.weight;
1530 tl + target_load(cpu, idx) <= tl_per_task) ||
1531 100*(tl + p->se.load.weight) <= imbalance*load) {
1533 * This domain has SD_WAKE_AFFINE and
1534 * p is cache cold in this domain, and
1535 * there is no bad imbalance.
1537 schedstat_inc(this_sd, ttwu_move_affine);
1543 * Start passive balancing when half the imbalance_pct
1546 if (this_sd->flags & SD_WAKE_BALANCE) {
1547 if (imbalance*this_load <= 100*load) {
1548 schedstat_inc(this_sd, ttwu_move_balance);
1554 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1556 new_cpu = wake_idle(new_cpu, p);
1557 if (new_cpu != cpu) {
1558 set_task_cpu(p, new_cpu);
1559 task_rq_unlock(rq, &flags);
1560 /* might preempt at this point */
1561 rq = task_rq_lock(p, &flags);
1562 old_state = p->state;
1563 if (!(old_state & state))
1568 this_cpu = smp_processor_id();
1573 #endif /* CONFIG_SMP */
1574 update_rq_clock(rq);
1575 activate_task(rq, p, 1);
1577 * Sync wakeups (i.e. those types of wakeups where the waker
1578 * has indicated that it will leave the CPU in short order)
1579 * don't trigger a preemption, if the woken up task will run on
1580 * this cpu. (in this case the 'I will reschedule' promise of
1581 * the waker guarantees that the freshly woken up task is going
1582 * to be considered on this CPU.)
1584 if (!sync || cpu != this_cpu)
1585 check_preempt_curr(rq, p);
1589 p->state = TASK_RUNNING;
1591 task_rq_unlock(rq, &flags);
1596 int fastcall wake_up_process(struct task_struct *p)
1598 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1599 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1601 EXPORT_SYMBOL(wake_up_process);
1603 int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1605 return try_to_wake_up(p, state, 0);
1609 * Perform scheduler related setup for a newly forked process p.
1610 * p is forked by current.
1612 * __sched_fork() is basic setup used by init_idle() too:
1614 static void __sched_fork(struct task_struct *p)
1616 p->se.exec_start = 0;
1617 p->se.sum_exec_runtime = 0;
1618 p->se.prev_sum_exec_runtime = 0;
1619 p->se.last_min_vruntime = 0;
1621 #ifdef CONFIG_SCHEDSTATS
1622 p->se.wait_start = 0;
1623 p->se.sum_sleep_runtime = 0;
1624 p->se.sleep_start = 0;
1625 p->se.block_start = 0;
1626 p->se.sleep_max = 0;
1627 p->se.block_max = 0;
1629 p->se.slice_max = 0;
1633 INIT_LIST_HEAD(&p->run_list);
1636 #ifdef CONFIG_PREEMPT_NOTIFIERS
1637 INIT_HLIST_HEAD(&p->preempt_notifiers);
1641 * We mark the process as running here, but have not actually
1642 * inserted it onto the runqueue yet. This guarantees that
1643 * nobody will actually run it, and a signal or other external
1644 * event cannot wake it up and insert it on the runqueue either.
1646 p->state = TASK_RUNNING;
1650 * fork()/clone()-time setup:
1652 void sched_fork(struct task_struct *p, int clone_flags)
1654 int cpu = get_cpu();
1659 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1661 __set_task_cpu(p, cpu);
1664 * Make sure we do not leak PI boosting priority to the child:
1666 p->prio = current->normal_prio;
1668 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1669 if (likely(sched_info_on()))
1670 memset(&p->sched_info, 0, sizeof(p->sched_info));
1672 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1675 #ifdef CONFIG_PREEMPT
1676 /* Want to start with kernel preemption disabled. */
1677 task_thread_info(p)->preempt_count = 1;
1683 * wake_up_new_task - wake up a newly created task for the first time.
1685 * This function will do some initial scheduler statistics housekeeping
1686 * that must be done for every newly created context, then puts the task
1687 * on the runqueue and wakes it.
1689 void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1691 unsigned long flags;
1695 rq = task_rq_lock(p, &flags);
1696 BUG_ON(p->state != TASK_RUNNING);
1697 this_cpu = smp_processor_id(); /* parent's CPU */
1698 update_rq_clock(rq);
1700 p->prio = effective_prio(p);
1702 if (rt_prio(p->prio))
1703 p->sched_class = &rt_sched_class;
1705 p->sched_class = &fair_sched_class;
1707 if (task_cpu(p) != this_cpu || !p->sched_class->task_new ||
1708 !current->se.on_rq) {
1709 activate_task(rq, p, 0);
1712 * Let the scheduling class do new task startup
1713 * management (if any):
1715 p->sched_class->task_new(rq, p);
1716 inc_nr_running(p, rq);
1718 check_preempt_curr(rq, p);
1719 task_rq_unlock(rq, &flags);
1722 #ifdef CONFIG_PREEMPT_NOTIFIERS
1725 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1726 * @notifier: notifier struct to register
1728 void preempt_notifier_register(struct preempt_notifier *notifier)
1730 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1732 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1735 * preempt_notifier_unregister - no longer interested in preemption notifications
1736 * @notifier: notifier struct to unregister
1738 * This is safe to call from within a preemption notifier.
1740 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1742 hlist_del(¬ifier->link);
1744 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1746 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1748 struct preempt_notifier *notifier;
1749 struct hlist_node *node;
1751 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1752 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1756 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1757 struct task_struct *next)
1759 struct preempt_notifier *notifier;
1760 struct hlist_node *node;
1762 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1763 notifier->ops->sched_out(notifier, next);
1768 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1773 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1774 struct task_struct *next)
1781 * prepare_task_switch - prepare to switch tasks
1782 * @rq: the runqueue preparing to switch
1783 * @prev: the current task that is being switched out
1784 * @next: the task we are going to switch to.
1786 * This is called with the rq lock held and interrupts off. It must
1787 * be paired with a subsequent finish_task_switch after the context
1790 * prepare_task_switch sets up locking and calls architecture specific
1794 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1795 struct task_struct *next)
1797 fire_sched_out_preempt_notifiers(prev, next);
1798 prepare_lock_switch(rq, next);
1799 prepare_arch_switch(next);
1803 * finish_task_switch - clean up after a task-switch
1804 * @rq: runqueue associated with task-switch
1805 * @prev: the thread we just switched away from.
1807 * finish_task_switch must be called after the context switch, paired
1808 * with a prepare_task_switch call before the context switch.
1809 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1810 * and do any other architecture-specific cleanup actions.
1812 * Note that we may have delayed dropping an mm in context_switch(). If
1813 * so, we finish that here outside of the runqueue lock. (Doing it
1814 * with the lock held can cause deadlocks; see schedule() for
1817 static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1818 __releases(rq->lock)
1820 struct mm_struct *mm = rq->prev_mm;
1826 * A task struct has one reference for the use as "current".
1827 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1828 * schedule one last time. The schedule call will never return, and
1829 * the scheduled task must drop that reference.
1830 * The test for TASK_DEAD must occur while the runqueue locks are
1831 * still held, otherwise prev could be scheduled on another cpu, die
1832 * there before we look at prev->state, and then the reference would
1834 * Manfred Spraul <manfred@colorfullife.com>
1836 prev_state = prev->state;
1837 finish_arch_switch(prev);
1838 finish_lock_switch(rq, prev);
1839 fire_sched_in_preempt_notifiers(current);
1842 if (unlikely(prev_state == TASK_DEAD)) {
1844 * Remove function-return probe instances associated with this
1845 * task and put them back on the free list.
1847 kprobe_flush_task(prev);
1848 put_task_struct(prev);
1853 * schedule_tail - first thing a freshly forked thread must call.
1854 * @prev: the thread we just switched away from.
1856 asmlinkage void schedule_tail(struct task_struct *prev)
1857 __releases(rq->lock)
1859 struct rq *rq = this_rq();
1861 finish_task_switch(rq, prev);
1862 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1863 /* In this case, finish_task_switch does not reenable preemption */
1866 if (current->set_child_tid)
1867 put_user(current->pid, current->set_child_tid);
1871 * context_switch - switch to the new MM and the new
1872 * thread's register state.
1875 context_switch(struct rq *rq, struct task_struct *prev,
1876 struct task_struct *next)
1878 struct mm_struct *mm, *oldmm;
1880 prepare_task_switch(rq, prev, next);
1882 oldmm = prev->active_mm;
1884 * For paravirt, this is coupled with an exit in switch_to to
1885 * combine the page table reload and the switch backend into
1888 arch_enter_lazy_cpu_mode();
1890 if (unlikely(!mm)) {
1891 next->active_mm = oldmm;
1892 atomic_inc(&oldmm->mm_count);
1893 enter_lazy_tlb(oldmm, next);
1895 switch_mm(oldmm, mm, next);
1897 if (unlikely(!prev->mm)) {
1898 prev->active_mm = NULL;
1899 rq->prev_mm = oldmm;
1902 * Since the runqueue lock will be released by the next
1903 * task (which is an invalid locking op but in the case
1904 * of the scheduler it's an obvious special-case), so we
1905 * do an early lockdep release here:
1907 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1908 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1911 /* Here we just switch the register state and the stack. */
1912 switch_to(prev, next, prev);
1916 * this_rq must be evaluated again because prev may have moved
1917 * CPUs since it called schedule(), thus the 'rq' on its stack
1918 * frame will be invalid.
1920 finish_task_switch(this_rq(), prev);
1924 * nr_running, nr_uninterruptible and nr_context_switches:
1926 * externally visible scheduler statistics: current number of runnable
1927 * threads, current number of uninterruptible-sleeping threads, total
1928 * number of context switches performed since bootup.
1930 unsigned long nr_running(void)
1932 unsigned long i, sum = 0;
1934 for_each_online_cpu(i)
1935 sum += cpu_rq(i)->nr_running;
1940 unsigned long nr_uninterruptible(void)
1942 unsigned long i, sum = 0;
1944 for_each_possible_cpu(i)
1945 sum += cpu_rq(i)->nr_uninterruptible;
1948 * Since we read the counters lockless, it might be slightly
1949 * inaccurate. Do not allow it to go below zero though:
1951 if (unlikely((long)sum < 0))
1957 unsigned long long nr_context_switches(void)
1960 unsigned long long sum = 0;
1962 for_each_possible_cpu(i)
1963 sum += cpu_rq(i)->nr_switches;
1968 unsigned long nr_iowait(void)
1970 unsigned long i, sum = 0;
1972 for_each_possible_cpu(i)
1973 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1978 unsigned long nr_active(void)
1980 unsigned long i, running = 0, uninterruptible = 0;
1982 for_each_online_cpu(i) {
1983 running += cpu_rq(i)->nr_running;
1984 uninterruptible += cpu_rq(i)->nr_uninterruptible;
1987 if (unlikely((long)uninterruptible < 0))
1988 uninterruptible = 0;
1990 return running + uninterruptible;
1994 * Update rq->cpu_load[] statistics. This function is usually called every
1995 * scheduler tick (TICK_NSEC).
1997 static void update_cpu_load(struct rq *this_rq)
1999 unsigned long this_load = this_rq->load.weight;
2002 this_rq->nr_load_updates++;
2004 /* Update our load: */
2005 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2006 unsigned long old_load, new_load;
2008 /* scale is effectively 1 << i now, and >> i divides by scale */
2010 old_load = this_rq->cpu_load[i];
2011 new_load = this_load;
2013 * Round up the averaging division if load is increasing. This
2014 * prevents us from getting stuck on 9 if the load is 10, for
2017 if (new_load > old_load)
2018 new_load += scale-1;
2019 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2026 * double_rq_lock - safely lock two runqueues
2028 * Note this does not disable interrupts like task_rq_lock,
2029 * you need to do so manually before calling.
2031 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2032 __acquires(rq1->lock)
2033 __acquires(rq2->lock)
2035 BUG_ON(!irqs_disabled());
2037 spin_lock(&rq1->lock);
2038 __acquire(rq2->lock); /* Fake it out ;) */
2041 spin_lock(&rq1->lock);
2042 spin_lock(&rq2->lock);
2044 spin_lock(&rq2->lock);
2045 spin_lock(&rq1->lock);
2048 update_rq_clock(rq1);
2049 update_rq_clock(rq2);
2053 * double_rq_unlock - safely unlock two runqueues
2055 * Note this does not restore interrupts like task_rq_unlock,
2056 * you need to do so manually after calling.
2058 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2059 __releases(rq1->lock)
2060 __releases(rq2->lock)
2062 spin_unlock(&rq1->lock);
2064 spin_unlock(&rq2->lock);
2066 __release(rq2->lock);
2070 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2072 static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2073 __releases(this_rq->lock)
2074 __acquires(busiest->lock)
2075 __acquires(this_rq->lock)
2077 if (unlikely(!irqs_disabled())) {
2078 /* printk() doesn't work good under rq->lock */
2079 spin_unlock(&this_rq->lock);
2082 if (unlikely(!spin_trylock(&busiest->lock))) {
2083 if (busiest < this_rq) {
2084 spin_unlock(&this_rq->lock);
2085 spin_lock(&busiest->lock);
2086 spin_lock(&this_rq->lock);
2088 spin_lock(&busiest->lock);
2093 * If dest_cpu is allowed for this process, migrate the task to it.
2094 * This is accomplished by forcing the cpu_allowed mask to only
2095 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2096 * the cpu_allowed mask is restored.
2098 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2100 struct migration_req req;
2101 unsigned long flags;
2104 rq = task_rq_lock(p, &flags);
2105 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2106 || unlikely(cpu_is_offline(dest_cpu)))
2109 /* force the process onto the specified CPU */
2110 if (migrate_task(p, dest_cpu, &req)) {
2111 /* Need to wait for migration thread (might exit: take ref). */
2112 struct task_struct *mt = rq->migration_thread;
2114 get_task_struct(mt);
2115 task_rq_unlock(rq, &flags);
2116 wake_up_process(mt);
2117 put_task_struct(mt);
2118 wait_for_completion(&req.done);
2123 task_rq_unlock(rq, &flags);
2127 * sched_exec - execve() is a valuable balancing opportunity, because at
2128 * this point the task has the smallest effective memory and cache footprint.
2130 void sched_exec(void)
2132 int new_cpu, this_cpu = get_cpu();
2133 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2135 if (new_cpu != this_cpu)
2136 sched_migrate_task(current, new_cpu);
2140 * pull_task - move a task from a remote runqueue to the local runqueue.
2141 * Both runqueues must be locked.
2143 static void pull_task(struct rq *src_rq, struct task_struct *p,
2144 struct rq *this_rq, int this_cpu)
2146 deactivate_task(src_rq, p, 0);
2147 set_task_cpu(p, this_cpu);
2148 activate_task(this_rq, p, 0);
2150 * Note that idle threads have a prio of MAX_PRIO, for this test
2151 * to be always true for them.
2153 check_preempt_curr(this_rq, p);
2157 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2160 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2161 struct sched_domain *sd, enum cpu_idle_type idle,
2165 * We do not migrate tasks that are:
2166 * 1) running (obviously), or
2167 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2168 * 3) are cache-hot on their current CPU.
2170 if (!cpu_isset(this_cpu, p->cpus_allowed))
2174 if (task_running(rq, p))
2180 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2181 unsigned long max_nr_move, unsigned long max_load_move,
2182 struct sched_domain *sd, enum cpu_idle_type idle,
2183 int *all_pinned, unsigned long *load_moved,
2184 int *this_best_prio, struct rq_iterator *iterator)
2186 int pulled = 0, pinned = 0, skip_for_load;
2187 struct task_struct *p;
2188 long rem_load_move = max_load_move;
2190 if (max_nr_move == 0 || max_load_move == 0)
2196 * Start the load-balancing iterator:
2198 p = iterator->start(iterator->arg);
2203 * To help distribute high priority tasks accross CPUs we don't
2204 * skip a task if it will be the highest priority task (i.e. smallest
2205 * prio value) on its new queue regardless of its load weight
2207 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2208 SCHED_LOAD_SCALE_FUZZ;
2209 if ((skip_for_load && p->prio >= *this_best_prio) ||
2210 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2211 p = iterator->next(iterator->arg);
2215 pull_task(busiest, p, this_rq, this_cpu);
2217 rem_load_move -= p->se.load.weight;
2220 * We only want to steal up to the prescribed number of tasks
2221 * and the prescribed amount of weighted load.
2223 if (pulled < max_nr_move && rem_load_move > 0) {
2224 if (p->prio < *this_best_prio)
2225 *this_best_prio = p->prio;
2226 p = iterator->next(iterator->arg);
2231 * Right now, this is the only place pull_task() is called,
2232 * so we can safely collect pull_task() stats here rather than
2233 * inside pull_task().
2235 schedstat_add(sd, lb_gained[idle], pulled);
2238 *all_pinned = pinned;
2239 *load_moved = max_load_move - rem_load_move;
2244 * move_tasks tries to move up to max_load_move weighted load from busiest to
2245 * this_rq, as part of a balancing operation within domain "sd".
2246 * Returns 1 if successful and 0 otherwise.
2248 * Called with both runqueues locked.
2250 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2251 unsigned long max_load_move,
2252 struct sched_domain *sd, enum cpu_idle_type idle,
2255 struct sched_class *class = sched_class_highest;
2256 unsigned long total_load_moved = 0;
2257 int this_best_prio = this_rq->curr->prio;
2261 class->load_balance(this_rq, this_cpu, busiest,
2262 ULONG_MAX, max_load_move - total_load_moved,
2263 sd, idle, all_pinned, &this_best_prio);
2264 class = class->next;
2265 } while (class && max_load_move > total_load_moved);
2267 return total_load_moved > 0;
2271 * move_one_task tries to move exactly one task from busiest to this_rq, as
2272 * part of active balancing operations within "domain".
2273 * Returns 1 if successful and 0 otherwise.
2275 * Called with both runqueues locked.
2277 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2278 struct sched_domain *sd, enum cpu_idle_type idle)
2280 struct sched_class *class;
2281 int this_best_prio = MAX_PRIO;
2283 for (class = sched_class_highest; class; class = class->next)
2284 if (class->load_balance(this_rq, this_cpu, busiest,
2285 1, ULONG_MAX, sd, idle, NULL,
2293 * find_busiest_group finds and returns the busiest CPU group within the
2294 * domain. It calculates and returns the amount of weighted load which
2295 * should be moved to restore balance via the imbalance parameter.
2297 static struct sched_group *
2298 find_busiest_group(struct sched_domain *sd, int this_cpu,
2299 unsigned long *imbalance, enum cpu_idle_type idle,
2300 int *sd_idle, cpumask_t *cpus, int *balance)
2302 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2303 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2304 unsigned long max_pull;
2305 unsigned long busiest_load_per_task, busiest_nr_running;
2306 unsigned long this_load_per_task, this_nr_running;
2308 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2309 int power_savings_balance = 1;
2310 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2311 unsigned long min_nr_running = ULONG_MAX;
2312 struct sched_group *group_min = NULL, *group_leader = NULL;
2315 max_load = this_load = total_load = total_pwr = 0;
2316 busiest_load_per_task = busiest_nr_running = 0;
2317 this_load_per_task = this_nr_running = 0;
2318 if (idle == CPU_NOT_IDLE)
2319 load_idx = sd->busy_idx;
2320 else if (idle == CPU_NEWLY_IDLE)
2321 load_idx = sd->newidle_idx;
2323 load_idx = sd->idle_idx;
2326 unsigned long load, group_capacity;
2329 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2330 unsigned long sum_nr_running, sum_weighted_load;
2332 local_group = cpu_isset(this_cpu, group->cpumask);
2335 balance_cpu = first_cpu(group->cpumask);
2337 /* Tally up the load of all CPUs in the group */
2338 sum_weighted_load = sum_nr_running = avg_load = 0;
2340 for_each_cpu_mask(i, group->cpumask) {
2343 if (!cpu_isset(i, *cpus))
2348 if (*sd_idle && rq->nr_running)
2351 /* Bias balancing toward cpus of our domain */
2353 if (idle_cpu(i) && !first_idle_cpu) {
2358 load = target_load(i, load_idx);
2360 load = source_load(i, load_idx);
2363 sum_nr_running += rq->nr_running;
2364 sum_weighted_load += weighted_cpuload(i);
2368 * First idle cpu or the first cpu(busiest) in this sched group
2369 * is eligible for doing load balancing at this and above
2370 * domains. In the newly idle case, we will allow all the cpu's
2371 * to do the newly idle load balance.
2373 if (idle != CPU_NEWLY_IDLE && local_group &&
2374 balance_cpu != this_cpu && balance) {
2379 total_load += avg_load;
2380 total_pwr += group->__cpu_power;
2382 /* Adjust by relative CPU power of the group */
2383 avg_load = sg_div_cpu_power(group,
2384 avg_load * SCHED_LOAD_SCALE);
2386 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2389 this_load = avg_load;
2391 this_nr_running = sum_nr_running;
2392 this_load_per_task = sum_weighted_load;
2393 } else if (avg_load > max_load &&
2394 sum_nr_running > group_capacity) {
2395 max_load = avg_load;
2397 busiest_nr_running = sum_nr_running;
2398 busiest_load_per_task = sum_weighted_load;
2401 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2403 * Busy processors will not participate in power savings
2406 if (idle == CPU_NOT_IDLE ||
2407 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2411 * If the local group is idle or completely loaded
2412 * no need to do power savings balance at this domain
2414 if (local_group && (this_nr_running >= group_capacity ||
2416 power_savings_balance = 0;
2419 * If a group is already running at full capacity or idle,
2420 * don't include that group in power savings calculations
2422 if (!power_savings_balance || sum_nr_running >= group_capacity
2427 * Calculate the group which has the least non-idle load.
2428 * This is the group from where we need to pick up the load
2431 if ((sum_nr_running < min_nr_running) ||
2432 (sum_nr_running == min_nr_running &&
2433 first_cpu(group->cpumask) <
2434 first_cpu(group_min->cpumask))) {
2436 min_nr_running = sum_nr_running;
2437 min_load_per_task = sum_weighted_load /
2442 * Calculate the group which is almost near its
2443 * capacity but still has some space to pick up some load
2444 * from other group and save more power
2446 if (sum_nr_running <= group_capacity - 1) {
2447 if (sum_nr_running > leader_nr_running ||
2448 (sum_nr_running == leader_nr_running &&
2449 first_cpu(group->cpumask) >
2450 first_cpu(group_leader->cpumask))) {
2451 group_leader = group;
2452 leader_nr_running = sum_nr_running;
2457 group = group->next;
2458 } while (group != sd->groups);
2460 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2463 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2465 if (this_load >= avg_load ||
2466 100*max_load <= sd->imbalance_pct*this_load)
2469 busiest_load_per_task /= busiest_nr_running;
2471 * We're trying to get all the cpus to the average_load, so we don't
2472 * want to push ourselves above the average load, nor do we wish to
2473 * reduce the max loaded cpu below the average load, as either of these
2474 * actions would just result in more rebalancing later, and ping-pong
2475 * tasks around. Thus we look for the minimum possible imbalance.
2476 * Negative imbalances (*we* are more loaded than anyone else) will
2477 * be counted as no imbalance for these purposes -- we can't fix that
2478 * by pulling tasks to us. Be careful of negative numbers as they'll
2479 * appear as very large values with unsigned longs.
2481 if (max_load <= busiest_load_per_task)
2485 * In the presence of smp nice balancing, certain scenarios can have
2486 * max load less than avg load(as we skip the groups at or below
2487 * its cpu_power, while calculating max_load..)
2489 if (max_load < avg_load) {
2491 goto small_imbalance;
2494 /* Don't want to pull so many tasks that a group would go idle */
2495 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2497 /* How much load to actually move to equalise the imbalance */
2498 *imbalance = min(max_pull * busiest->__cpu_power,
2499 (avg_load - this_load) * this->__cpu_power)
2503 * if *imbalance is less than the average load per runnable task
2504 * there is no gaurantee that any tasks will be moved so we'll have
2505 * a think about bumping its value to force at least one task to be
2508 if (*imbalance < busiest_load_per_task) {
2509 unsigned long tmp, pwr_now, pwr_move;
2513 pwr_move = pwr_now = 0;
2515 if (this_nr_running) {
2516 this_load_per_task /= this_nr_running;
2517 if (busiest_load_per_task > this_load_per_task)
2520 this_load_per_task = SCHED_LOAD_SCALE;
2522 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2523 busiest_load_per_task * imbn) {
2524 *imbalance = busiest_load_per_task;
2529 * OK, we don't have enough imbalance to justify moving tasks,
2530 * however we may be able to increase total CPU power used by
2534 pwr_now += busiest->__cpu_power *
2535 min(busiest_load_per_task, max_load);
2536 pwr_now += this->__cpu_power *
2537 min(this_load_per_task, this_load);
2538 pwr_now /= SCHED_LOAD_SCALE;
2540 /* Amount of load we'd subtract */
2541 tmp = sg_div_cpu_power(busiest,
2542 busiest_load_per_task * SCHED_LOAD_SCALE);
2544 pwr_move += busiest->__cpu_power *
2545 min(busiest_load_per_task, max_load - tmp);
2547 /* Amount of load we'd add */
2548 if (max_load * busiest->__cpu_power <
2549 busiest_load_per_task * SCHED_LOAD_SCALE)
2550 tmp = sg_div_cpu_power(this,
2551 max_load * busiest->__cpu_power);
2553 tmp = sg_div_cpu_power(this,
2554 busiest_load_per_task * SCHED_LOAD_SCALE);
2555 pwr_move += this->__cpu_power *
2556 min(this_load_per_task, this_load + tmp);
2557 pwr_move /= SCHED_LOAD_SCALE;
2559 /* Move if we gain throughput */
2560 if (pwr_move > pwr_now)
2561 *imbalance = busiest_load_per_task;
2567 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2568 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2571 if (this == group_leader && group_leader != group_min) {
2572 *imbalance = min_load_per_task;
2582 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2585 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2586 unsigned long imbalance, cpumask_t *cpus)
2588 struct rq *busiest = NULL, *rq;
2589 unsigned long max_load = 0;
2592 for_each_cpu_mask(i, group->cpumask) {
2595 if (!cpu_isset(i, *cpus))
2599 wl = weighted_cpuload(i);
2601 if (rq->nr_running == 1 && wl > imbalance)
2604 if (wl > max_load) {
2614 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2615 * so long as it is large enough.
2617 #define MAX_PINNED_INTERVAL 512
2620 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2621 * tasks if there is an imbalance.
2623 static int load_balance(int this_cpu, struct rq *this_rq,
2624 struct sched_domain *sd, enum cpu_idle_type idle,
2627 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2628 struct sched_group *group;
2629 unsigned long imbalance;
2631 cpumask_t cpus = CPU_MASK_ALL;
2632 unsigned long flags;
2635 * When power savings policy is enabled for the parent domain, idle
2636 * sibling can pick up load irrespective of busy siblings. In this case,
2637 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2638 * portraying it as CPU_NOT_IDLE.
2640 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2641 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2644 schedstat_inc(sd, lb_cnt[idle]);
2647 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2654 schedstat_inc(sd, lb_nobusyg[idle]);
2658 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2660 schedstat_inc(sd, lb_nobusyq[idle]);
2664 BUG_ON(busiest == this_rq);
2666 schedstat_add(sd, lb_imbalance[idle], imbalance);
2669 if (busiest->nr_running > 1) {
2671 * Attempt to move tasks. If find_busiest_group has found
2672 * an imbalance but busiest->nr_running <= 1, the group is
2673 * still unbalanced. ld_moved simply stays zero, so it is
2674 * correctly treated as an imbalance.
2676 local_irq_save(flags);
2677 double_rq_lock(this_rq, busiest);
2678 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2679 imbalance, sd, idle, &all_pinned);
2680 double_rq_unlock(this_rq, busiest);
2681 local_irq_restore(flags);
2684 * some other cpu did the load balance for us.
2686 if (ld_moved && this_cpu != smp_processor_id())
2687 resched_cpu(this_cpu);
2689 /* All tasks on this runqueue were pinned by CPU affinity */
2690 if (unlikely(all_pinned)) {
2691 cpu_clear(cpu_of(busiest), cpus);
2692 if (!cpus_empty(cpus))
2699 schedstat_inc(sd, lb_failed[idle]);
2700 sd->nr_balance_failed++;
2702 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2704 spin_lock_irqsave(&busiest->lock, flags);
2706 /* don't kick the migration_thread, if the curr
2707 * task on busiest cpu can't be moved to this_cpu
2709 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2710 spin_unlock_irqrestore(&busiest->lock, flags);
2712 goto out_one_pinned;
2715 if (!busiest->active_balance) {
2716 busiest->active_balance = 1;
2717 busiest->push_cpu = this_cpu;
2720 spin_unlock_irqrestore(&busiest->lock, flags);
2722 wake_up_process(busiest->migration_thread);
2725 * We've kicked active balancing, reset the failure
2728 sd->nr_balance_failed = sd->cache_nice_tries+1;
2731 sd->nr_balance_failed = 0;
2733 if (likely(!active_balance)) {
2734 /* We were unbalanced, so reset the balancing interval */
2735 sd->balance_interval = sd->min_interval;
2738 * If we've begun active balancing, start to back off. This
2739 * case may not be covered by the all_pinned logic if there
2740 * is only 1 task on the busy runqueue (because we don't call
2743 if (sd->balance_interval < sd->max_interval)
2744 sd->balance_interval *= 2;
2747 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2748 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2753 schedstat_inc(sd, lb_balanced[idle]);
2755 sd->nr_balance_failed = 0;
2758 /* tune up the balancing interval */
2759 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2760 (sd->balance_interval < sd->max_interval))
2761 sd->balance_interval *= 2;
2763 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2764 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2770 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2771 * tasks if there is an imbalance.
2773 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2774 * this_rq is locked.
2777 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2779 struct sched_group *group;
2780 struct rq *busiest = NULL;
2781 unsigned long imbalance;
2785 cpumask_t cpus = CPU_MASK_ALL;
2788 * When power savings policy is enabled for the parent domain, idle
2789 * sibling can pick up load irrespective of busy siblings. In this case,
2790 * let the state of idle sibling percolate up as IDLE, instead of
2791 * portraying it as CPU_NOT_IDLE.
2793 if (sd->flags & SD_SHARE_CPUPOWER &&
2794 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2797 schedstat_inc(sd, lb_cnt[CPU_NEWLY_IDLE]);
2799 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2800 &sd_idle, &cpus, NULL);
2802 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2806 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2809 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2813 BUG_ON(busiest == this_rq);
2815 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2818 if (busiest->nr_running > 1) {
2819 /* Attempt to move tasks */
2820 double_lock_balance(this_rq, busiest);
2821 /* this_rq->clock is already updated */
2822 update_rq_clock(busiest);
2823 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2824 imbalance, sd, CPU_NEWLY_IDLE,
2826 spin_unlock(&busiest->lock);
2828 if (unlikely(all_pinned)) {
2829 cpu_clear(cpu_of(busiest), cpus);
2830 if (!cpus_empty(cpus))
2836 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2837 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2838 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2841 sd->nr_balance_failed = 0;
2846 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2847 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2848 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2850 sd->nr_balance_failed = 0;
2856 * idle_balance is called by schedule() if this_cpu is about to become
2857 * idle. Attempts to pull tasks from other CPUs.
2859 static void idle_balance(int this_cpu, struct rq *this_rq)
2861 struct sched_domain *sd;
2862 int pulled_task = -1;
2863 unsigned long next_balance = jiffies + HZ;
2865 for_each_domain(this_cpu, sd) {
2866 unsigned long interval;
2868 if (!(sd->flags & SD_LOAD_BALANCE))
2871 if (sd->flags & SD_BALANCE_NEWIDLE)
2872 /* If we've pulled tasks over stop searching: */
2873 pulled_task = load_balance_newidle(this_cpu,
2876 interval = msecs_to_jiffies(sd->balance_interval);
2877 if (time_after(next_balance, sd->last_balance + interval))
2878 next_balance = sd->last_balance + interval;
2882 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
2884 * We are going idle. next_balance may be set based on
2885 * a busy processor. So reset next_balance.
2887 this_rq->next_balance = next_balance;
2892 * active_load_balance is run by migration threads. It pushes running tasks
2893 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2894 * running on each physical CPU where possible, and avoids physical /
2895 * logical imbalances.
2897 * Called with busiest_rq locked.
2899 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2901 int target_cpu = busiest_rq->push_cpu;
2902 struct sched_domain *sd;
2903 struct rq *target_rq;
2905 /* Is there any task to move? */
2906 if (busiest_rq->nr_running <= 1)
2909 target_rq = cpu_rq(target_cpu);
2912 * This condition is "impossible", if it occurs
2913 * we need to fix it. Originally reported by
2914 * Bjorn Helgaas on a 128-cpu setup.
2916 BUG_ON(busiest_rq == target_rq);
2918 /* move a task from busiest_rq to target_rq */
2919 double_lock_balance(busiest_rq, target_rq);
2920 update_rq_clock(busiest_rq);
2921 update_rq_clock(target_rq);
2923 /* Search for an sd spanning us and the target CPU. */
2924 for_each_domain(target_cpu, sd) {
2925 if ((sd->flags & SD_LOAD_BALANCE) &&
2926 cpu_isset(busiest_cpu, sd->span))
2931 schedstat_inc(sd, alb_cnt);
2933 if (move_one_task(target_rq, target_cpu, busiest_rq,
2935 schedstat_inc(sd, alb_pushed);
2937 schedstat_inc(sd, alb_failed);
2939 spin_unlock(&target_rq->lock);
2944 atomic_t load_balancer;
2946 } nohz ____cacheline_aligned = {
2947 .load_balancer = ATOMIC_INIT(-1),
2948 .cpu_mask = CPU_MASK_NONE,
2952 * This routine will try to nominate the ilb (idle load balancing)
2953 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
2954 * load balancing on behalf of all those cpus. If all the cpus in the system
2955 * go into this tickless mode, then there will be no ilb owner (as there is
2956 * no need for one) and all the cpus will sleep till the next wakeup event
2959 * For the ilb owner, tick is not stopped. And this tick will be used
2960 * for idle load balancing. ilb owner will still be part of
2963 * While stopping the tick, this cpu will become the ilb owner if there
2964 * is no other owner. And will be the owner till that cpu becomes busy
2965 * or if all cpus in the system stop their ticks at which point
2966 * there is no need for ilb owner.
2968 * When the ilb owner becomes busy, it nominates another owner, during the
2969 * next busy scheduler_tick()
2971 int select_nohz_load_balancer(int stop_tick)
2973 int cpu = smp_processor_id();
2976 cpu_set(cpu, nohz.cpu_mask);
2977 cpu_rq(cpu)->in_nohz_recently = 1;
2980 * If we are going offline and still the leader, give up!
2982 if (cpu_is_offline(cpu) &&
2983 atomic_read(&nohz.load_balancer) == cpu) {
2984 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
2989 /* time for ilb owner also to sleep */
2990 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
2991 if (atomic_read(&nohz.load_balancer) == cpu)
2992 atomic_set(&nohz.load_balancer, -1);
2996 if (atomic_read(&nohz.load_balancer) == -1) {
2997 /* make me the ilb owner */
2998 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3000 } else if (atomic_read(&nohz.load_balancer) == cpu)
3003 if (!cpu_isset(cpu, nohz.cpu_mask))
3006 cpu_clear(cpu, nohz.cpu_mask);
3008 if (atomic_read(&nohz.load_balancer) == cpu)
3009 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3016 static DEFINE_SPINLOCK(balancing);
3019 * It checks each scheduling domain to see if it is due to be balanced,
3020 * and initiates a balancing operation if so.
3022 * Balancing parameters are set up in arch_init_sched_domains.
3024 static inline void rebalance_domains(int cpu, enum cpu_idle_type idle)
3027 struct rq *rq = cpu_rq(cpu);
3028 unsigned long interval;
3029 struct sched_domain *sd;
3030 /* Earliest time when we have to do rebalance again */
3031 unsigned long next_balance = jiffies + 60*HZ;
3032 int update_next_balance = 0;
3034 for_each_domain(cpu, sd) {
3035 if (!(sd->flags & SD_LOAD_BALANCE))
3038 interval = sd->balance_interval;
3039 if (idle != CPU_IDLE)
3040 interval *= sd->busy_factor;
3042 /* scale ms to jiffies */
3043 interval = msecs_to_jiffies(interval);
3044 if (unlikely(!interval))
3046 if (interval > HZ*NR_CPUS/10)
3047 interval = HZ*NR_CPUS/10;
3050 if (sd->flags & SD_SERIALIZE) {
3051 if (!spin_trylock(&balancing))
3055 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3056 if (load_balance(cpu, rq, sd, idle, &balance)) {
3058 * We've pulled tasks over so either we're no
3059 * longer idle, or one of our SMT siblings is
3062 idle = CPU_NOT_IDLE;
3064 sd->last_balance = jiffies;
3066 if (sd->flags & SD_SERIALIZE)
3067 spin_unlock(&balancing);
3069 if (time_after(next_balance, sd->last_balance + interval)) {
3070 next_balance = sd->last_balance + interval;
3071 update_next_balance = 1;
3075 * Stop the load balance at this level. There is another
3076 * CPU in our sched group which is doing load balancing more
3084 * next_balance will be updated only when there is a need.
3085 * When the cpu is attached to null domain for ex, it will not be
3088 if (likely(update_next_balance))
3089 rq->next_balance = next_balance;
3093 * run_rebalance_domains is triggered when needed from the scheduler tick.
3094 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3095 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3097 static void run_rebalance_domains(struct softirq_action *h)
3099 int this_cpu = smp_processor_id();
3100 struct rq *this_rq = cpu_rq(this_cpu);
3101 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3102 CPU_IDLE : CPU_NOT_IDLE;
3104 rebalance_domains(this_cpu, idle);
3108 * If this cpu is the owner for idle load balancing, then do the
3109 * balancing on behalf of the other idle cpus whose ticks are
3112 if (this_rq->idle_at_tick &&
3113 atomic_read(&nohz.load_balancer) == this_cpu) {
3114 cpumask_t cpus = nohz.cpu_mask;
3118 cpu_clear(this_cpu, cpus);
3119 for_each_cpu_mask(balance_cpu, cpus) {
3121 * If this cpu gets work to do, stop the load balancing
3122 * work being done for other cpus. Next load
3123 * balancing owner will pick it up.
3128 rebalance_domains(balance_cpu, CPU_IDLE);
3130 rq = cpu_rq(balance_cpu);
3131 if (time_after(this_rq->next_balance, rq->next_balance))
3132 this_rq->next_balance = rq->next_balance;
3139 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3141 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3142 * idle load balancing owner or decide to stop the periodic load balancing,
3143 * if the whole system is idle.
3145 static inline void trigger_load_balance(struct rq *rq, int cpu)
3149 * If we were in the nohz mode recently and busy at the current
3150 * scheduler tick, then check if we need to nominate new idle
3153 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3154 rq->in_nohz_recently = 0;
3156 if (atomic_read(&nohz.load_balancer) == cpu) {
3157 cpu_clear(cpu, nohz.cpu_mask);
3158 atomic_set(&nohz.load_balancer, -1);
3161 if (atomic_read(&nohz.load_balancer) == -1) {
3163 * simple selection for now: Nominate the
3164 * first cpu in the nohz list to be the next
3167 * TBD: Traverse the sched domains and nominate
3168 * the nearest cpu in the nohz.cpu_mask.
3170 int ilb = first_cpu(nohz.cpu_mask);
3178 * If this cpu is idle and doing idle load balancing for all the
3179 * cpus with ticks stopped, is it time for that to stop?
3181 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3182 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3188 * If this cpu is idle and the idle load balancing is done by
3189 * someone else, then no need raise the SCHED_SOFTIRQ
3191 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3192 cpu_isset(cpu, nohz.cpu_mask))
3195 if (time_after_eq(jiffies, rq->next_balance))
3196 raise_softirq(SCHED_SOFTIRQ);
3199 #else /* CONFIG_SMP */
3202 * on UP we do not need to balance between CPUs:
3204 static inline void idle_balance(int cpu, struct rq *rq)
3208 /* Avoid "used but not defined" warning on UP */
3209 static int balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3210 unsigned long max_nr_move, unsigned long max_load_move,
3211 struct sched_domain *sd, enum cpu_idle_type idle,
3212 int *all_pinned, unsigned long *load_moved,
3213 int *this_best_prio, struct rq_iterator *iterator)
3222 DEFINE_PER_CPU(struct kernel_stat, kstat);
3224 EXPORT_PER_CPU_SYMBOL(kstat);
3227 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3228 * that have not yet been banked in case the task is currently running.
3230 unsigned long long task_sched_runtime(struct task_struct *p)
3232 unsigned long flags;
3236 rq = task_rq_lock(p, &flags);
3237 ns = p->se.sum_exec_runtime;
3238 if (rq->curr == p) {
3239 update_rq_clock(rq);
3240 delta_exec = rq->clock - p->se.exec_start;
3241 if ((s64)delta_exec > 0)
3244 task_rq_unlock(rq, &flags);
3250 * Account user cpu time to a process.
3251 * @p: the process that the cpu time gets accounted to
3252 * @hardirq_offset: the offset to subtract from hardirq_count()
3253 * @cputime: the cpu time spent in user space since the last update
3255 void account_user_time(struct task_struct *p, cputime_t cputime)
3257 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3260 p->utime = cputime_add(p->utime, cputime);
3262 /* Add user time to cpustat. */
3263 tmp = cputime_to_cputime64(cputime);
3264 if (TASK_NICE(p) > 0)
3265 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3267 cpustat->user = cputime64_add(cpustat->user, tmp);
3271 * Account system cpu time to a process.
3272 * @p: the process that the cpu time gets accounted to
3273 * @hardirq_offset: the offset to subtract from hardirq_count()
3274 * @cputime: the cpu time spent in kernel space since the last update
3276 void account_system_time(struct task_struct *p, int hardirq_offset,
3279 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3280 struct rq *rq = this_rq();
3283 p->stime = cputime_add(p->stime, cputime);
3285 /* Add system time to cpustat. */
3286 tmp = cputime_to_cputime64(cputime);
3287 if (hardirq_count() - hardirq_offset)
3288 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3289 else if (softirq_count())
3290 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3291 else if (p != rq->idle)
3292 cpustat->system = cputime64_add(cpustat->system, tmp);
3293 else if (atomic_read(&rq->nr_iowait) > 0)
3294 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3296 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3297 /* Account for system time used */
3298 acct_update_integrals(p);
3302 * Account for involuntary wait time.
3303 * @p: the process from which the cpu time has been stolen
3304 * @steal: the cpu time spent in involuntary wait
3306 void account_steal_time(struct task_struct *p, cputime_t steal)
3308 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3309 cputime64_t tmp = cputime_to_cputime64(steal);
3310 struct rq *rq = this_rq();
3312 if (p == rq->idle) {
3313 p->stime = cputime_add(p->stime, steal);
3314 if (atomic_read(&rq->nr_iowait) > 0)
3315 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3317 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3319 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3323 * This function gets called by the timer code, with HZ frequency.
3324 * We call it with interrupts disabled.
3326 * It also gets called by the fork code, when changing the parent's
3329 void scheduler_tick(void)
3331 int cpu = smp_processor_id();
3332 struct rq *rq = cpu_rq(cpu);
3333 struct task_struct *curr = rq->curr;
3334 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3336 spin_lock(&rq->lock);
3337 __update_rq_clock(rq);
3339 * Let rq->clock advance by at least TICK_NSEC:
3341 if (unlikely(rq->clock < next_tick))
3342 rq->clock = next_tick;
3343 rq->tick_timestamp = rq->clock;
3344 update_cpu_load(rq);
3345 if (curr != rq->idle) /* FIXME: needed? */
3346 curr->sched_class->task_tick(rq, curr);
3347 spin_unlock(&rq->lock);
3350 rq->idle_at_tick = idle_cpu(cpu);
3351 trigger_load_balance(rq, cpu);
3355 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3357 void fastcall add_preempt_count(int val)
3362 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3364 preempt_count() += val;
3366 * Spinlock count overflowing soon?
3368 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3371 EXPORT_SYMBOL(add_preempt_count);
3373 void fastcall sub_preempt_count(int val)
3378 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3381 * Is the spinlock portion underflowing?
3383 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3384 !(preempt_count() & PREEMPT_MASK)))
3387 preempt_count() -= val;
3389 EXPORT_SYMBOL(sub_preempt_count);
3394 * Print scheduling while atomic bug:
3396 static noinline void __schedule_bug(struct task_struct *prev)
3398 printk(KERN_ERR "BUG: scheduling while atomic: %s/0x%08x/%d\n",
3399 prev->comm, preempt_count(), prev->pid);
3400 debug_show_held_locks(prev);
3401 if (irqs_disabled())
3402 print_irqtrace_events(prev);
3407 * Various schedule()-time debugging checks and statistics:
3409 static inline void schedule_debug(struct task_struct *prev)
3412 * Test if we are atomic. Since do_exit() needs to call into
3413 * schedule() atomically, we ignore that path for now.
3414 * Otherwise, whine if we are scheduling when we should not be.
3416 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3417 __schedule_bug(prev);
3419 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3421 schedstat_inc(this_rq(), sched_cnt);
3422 #ifdef CONFIG_SCHEDSTATS
3423 if (unlikely(prev->lock_depth >= 0)) {
3424 schedstat_inc(this_rq(), bkl_cnt);
3425 schedstat_inc(prev, sched_info.bkl_cnt);
3431 * Pick up the highest-prio task:
3433 static inline struct task_struct *
3434 pick_next_task(struct rq *rq, struct task_struct *prev)
3436 struct sched_class *class;
3437 struct task_struct *p;
3440 * Optimization: we know that if all tasks are in
3441 * the fair class we can call that function directly:
3443 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3444 p = fair_sched_class.pick_next_task(rq);
3449 class = sched_class_highest;
3451 p = class->pick_next_task(rq);
3455 * Will never be NULL as the idle class always
3456 * returns a non-NULL p:
3458 class = class->next;
3463 * schedule() is the main scheduler function.
3465 asmlinkage void __sched schedule(void)
3467 struct task_struct *prev, *next;
3474 cpu = smp_processor_id();
3478 switch_count = &prev->nivcsw;
3480 release_kernel_lock(prev);
3481 need_resched_nonpreemptible:
3483 schedule_debug(prev);
3485 spin_lock_irq(&rq->lock);
3486 clear_tsk_need_resched(prev);
3487 __update_rq_clock(rq);
3489 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3490 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3491 unlikely(signal_pending(prev)))) {
3492 prev->state = TASK_RUNNING;
3494 deactivate_task(rq, prev, 1);
3496 switch_count = &prev->nvcsw;
3499 if (unlikely(!rq->nr_running))
3500 idle_balance(cpu, rq);
3502 prev->sched_class->put_prev_task(rq, prev);
3503 next = pick_next_task(rq, prev);
3505 sched_info_switch(prev, next);
3507 if (likely(prev != next)) {
3512 context_switch(rq, prev, next); /* unlocks the rq */
3514 spin_unlock_irq(&rq->lock);
3516 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3517 cpu = smp_processor_id();
3519 goto need_resched_nonpreemptible;
3521 preempt_enable_no_resched();
3522 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3525 EXPORT_SYMBOL(schedule);
3527 #ifdef CONFIG_PREEMPT
3529 * this is the entry point to schedule() from in-kernel preemption
3530 * off of preempt_enable. Kernel preemptions off return from interrupt
3531 * occur there and call schedule directly.
3533 asmlinkage void __sched preempt_schedule(void)
3535 struct thread_info *ti = current_thread_info();
3536 #ifdef CONFIG_PREEMPT_BKL
3537 struct task_struct *task = current;
3538 int saved_lock_depth;
3541 * If there is a non-zero preempt_count or interrupts are disabled,
3542 * we do not want to preempt the current task. Just return..
3544 if (likely(ti->preempt_count || irqs_disabled()))
3548 add_preempt_count(PREEMPT_ACTIVE);
3550 * We keep the big kernel semaphore locked, but we
3551 * clear ->lock_depth so that schedule() doesnt
3552 * auto-release the semaphore:
3554 #ifdef CONFIG_PREEMPT_BKL
3555 saved_lock_depth = task->lock_depth;
3556 task->lock_depth = -1;
3559 #ifdef CONFIG_PREEMPT_BKL
3560 task->lock_depth = saved_lock_depth;
3562 sub_preempt_count(PREEMPT_ACTIVE);
3564 /* we could miss a preemption opportunity between schedule and now */
3566 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3569 EXPORT_SYMBOL(preempt_schedule);
3572 * this is the entry point to schedule() from kernel preemption
3573 * off of irq context.
3574 * Note, that this is called and return with irqs disabled. This will
3575 * protect us against recursive calling from irq.
3577 asmlinkage void __sched preempt_schedule_irq(void)
3579 struct thread_info *ti = current_thread_info();
3580 #ifdef CONFIG_PREEMPT_BKL
3581 struct task_struct *task = current;
3582 int saved_lock_depth;
3584 /* Catch callers which need to be fixed */
3585 BUG_ON(ti->preempt_count || !irqs_disabled());
3588 add_preempt_count(PREEMPT_ACTIVE);
3590 * We keep the big kernel semaphore locked, but we
3591 * clear ->lock_depth so that schedule() doesnt
3592 * auto-release the semaphore:
3594 #ifdef CONFIG_PREEMPT_BKL
3595 saved_lock_depth = task->lock_depth;
3596 task->lock_depth = -1;
3600 local_irq_disable();
3601 #ifdef CONFIG_PREEMPT_BKL
3602 task->lock_depth = saved_lock_depth;
3604 sub_preempt_count(PREEMPT_ACTIVE);
3606 /* we could miss a preemption opportunity between schedule and now */
3608 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3612 #endif /* CONFIG_PREEMPT */
3614 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3617 return try_to_wake_up(curr->private, mode, sync);
3619 EXPORT_SYMBOL(default_wake_function);
3622 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3623 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3624 * number) then we wake all the non-exclusive tasks and one exclusive task.
3626 * There are circumstances in which we can try to wake a task which has already
3627 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3628 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3630 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3631 int nr_exclusive, int sync, void *key)
3633 wait_queue_t *curr, *next;
3635 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3636 unsigned flags = curr->flags;
3638 if (curr->func(curr, mode, sync, key) &&
3639 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3645 * __wake_up - wake up threads blocked on a waitqueue.
3647 * @mode: which threads
3648 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3649 * @key: is directly passed to the wakeup function
3651 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3652 int nr_exclusive, void *key)
3654 unsigned long flags;
3656 spin_lock_irqsave(&q->lock, flags);
3657 __wake_up_common(q, mode, nr_exclusive, 0, key);
3658 spin_unlock_irqrestore(&q->lock, flags);
3660 EXPORT_SYMBOL(__wake_up);
3663 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3665 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3667 __wake_up_common(q, mode, 1, 0, NULL);
3671 * __wake_up_sync - wake up threads blocked on a waitqueue.
3673 * @mode: which threads
3674 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3676 * The sync wakeup differs that the waker knows that it will schedule
3677 * away soon, so while the target thread will be woken up, it will not
3678 * be migrated to another CPU - ie. the two threads are 'synchronized'
3679 * with each other. This can prevent needless bouncing between CPUs.
3681 * On UP it can prevent extra preemption.
3684 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3686 unsigned long flags;
3692 if (unlikely(!nr_exclusive))
3695 spin_lock_irqsave(&q->lock, flags);
3696 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3697 spin_unlock_irqrestore(&q->lock, flags);
3699 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3701 void fastcall complete(struct completion *x)
3703 unsigned long flags;
3705 spin_lock_irqsave(&x->wait.lock, flags);
3707 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3709 spin_unlock_irqrestore(&x->wait.lock, flags);
3711 EXPORT_SYMBOL(complete);
3713 void fastcall complete_all(struct completion *x)
3715 unsigned long flags;
3717 spin_lock_irqsave(&x->wait.lock, flags);
3718 x->done += UINT_MAX/2;
3719 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3721 spin_unlock_irqrestore(&x->wait.lock, flags);
3723 EXPORT_SYMBOL(complete_all);
3725 void fastcall __sched wait_for_completion(struct completion *x)
3729 spin_lock_irq(&x->wait.lock);
3731 DECLARE_WAITQUEUE(wait, current);
3733 wait.flags |= WQ_FLAG_EXCLUSIVE;
3734 __add_wait_queue_tail(&x->wait, &wait);
3736 __set_current_state(TASK_UNINTERRUPTIBLE);
3737 spin_unlock_irq(&x->wait.lock);
3739 spin_lock_irq(&x->wait.lock);
3741 __remove_wait_queue(&x->wait, &wait);
3744 spin_unlock_irq(&x->wait.lock);
3746 EXPORT_SYMBOL(wait_for_completion);
3748 unsigned long fastcall __sched
3749 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3753 spin_lock_irq(&x->wait.lock);
3755 DECLARE_WAITQUEUE(wait, current);
3757 wait.flags |= WQ_FLAG_EXCLUSIVE;
3758 __add_wait_queue_tail(&x->wait, &wait);
3760 __set_current_state(TASK_UNINTERRUPTIBLE);
3761 spin_unlock_irq(&x->wait.lock);
3762 timeout = schedule_timeout(timeout);
3763 spin_lock_irq(&x->wait.lock);
3765 __remove_wait_queue(&x->wait, &wait);
3769 __remove_wait_queue(&x->wait, &wait);
3773 spin_unlock_irq(&x->wait.lock);
3776 EXPORT_SYMBOL(wait_for_completion_timeout);
3778 int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3784 spin_lock_irq(&x->wait.lock);
3786 DECLARE_WAITQUEUE(wait, current);
3788 wait.flags |= WQ_FLAG_EXCLUSIVE;
3789 __add_wait_queue_tail(&x->wait, &wait);
3791 if (signal_pending(current)) {
3793 __remove_wait_queue(&x->wait, &wait);
3796 __set_current_state(TASK_INTERRUPTIBLE);
3797 spin_unlock_irq(&x->wait.lock);
3799 spin_lock_irq(&x->wait.lock);
3801 __remove_wait_queue(&x->wait, &wait);
3805 spin_unlock_irq(&x->wait.lock);
3809 EXPORT_SYMBOL(wait_for_completion_interruptible);
3811 unsigned long fastcall __sched
3812 wait_for_completion_interruptible_timeout(struct completion *x,
3813 unsigned long timeout)
3817 spin_lock_irq(&x->wait.lock);
3819 DECLARE_WAITQUEUE(wait, current);
3821 wait.flags |= WQ_FLAG_EXCLUSIVE;
3822 __add_wait_queue_tail(&x->wait, &wait);
3824 if (signal_pending(current)) {
3825 timeout = -ERESTARTSYS;
3826 __remove_wait_queue(&x->wait, &wait);
3829 __set_current_state(TASK_INTERRUPTIBLE);
3830 spin_unlock_irq(&x->wait.lock);
3831 timeout = schedule_timeout(timeout);
3832 spin_lock_irq(&x->wait.lock);
3834 __remove_wait_queue(&x->wait, &wait);
3838 __remove_wait_queue(&x->wait, &wait);
3842 spin_unlock_irq(&x->wait.lock);
3845 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3848 sleep_on_head(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3850 spin_lock_irqsave(&q->lock, *flags);
3851 __add_wait_queue(q, wait);
3852 spin_unlock(&q->lock);
3856 sleep_on_tail(wait_queue_head_t *q, wait_queue_t *wait, unsigned long *flags)
3858 spin_lock_irq(&q->lock);
3859 __remove_wait_queue(q, wait);
3860 spin_unlock_irqrestore(&q->lock, *flags);
3863 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3865 unsigned long flags;
3868 init_waitqueue_entry(&wait, current);
3870 current->state = TASK_INTERRUPTIBLE;
3872 sleep_on_head(q, &wait, &flags);
3874 sleep_on_tail(q, &wait, &flags);
3876 EXPORT_SYMBOL(interruptible_sleep_on);
3879 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3881 unsigned long flags;
3884 init_waitqueue_entry(&wait, current);
3886 current->state = TASK_INTERRUPTIBLE;
3888 sleep_on_head(q, &wait, &flags);
3889 timeout = schedule_timeout(timeout);
3890 sleep_on_tail(q, &wait, &flags);
3894 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3896 void __sched sleep_on(wait_queue_head_t *q)
3898 unsigned long flags;
3901 init_waitqueue_entry(&wait, current);
3903 current->state = TASK_UNINTERRUPTIBLE;
3905 sleep_on_head(q, &wait, &flags);
3907 sleep_on_tail(q, &wait, &flags);
3909 EXPORT_SYMBOL(sleep_on);
3911 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3913 unsigned long flags;
3916 init_waitqueue_entry(&wait, current);
3918 current->state = TASK_UNINTERRUPTIBLE;
3920 sleep_on_head(q, &wait, &flags);
3921 timeout = schedule_timeout(timeout);
3922 sleep_on_tail(q, &wait, &flags);
3926 EXPORT_SYMBOL(sleep_on_timeout);
3928 #ifdef CONFIG_RT_MUTEXES
3931 * rt_mutex_setprio - set the current priority of a task
3933 * @prio: prio value (kernel-internal form)
3935 * This function changes the 'effective' priority of a task. It does
3936 * not touch ->normal_prio like __setscheduler().
3938 * Used by the rt_mutex code to implement priority inheritance logic.
3940 void rt_mutex_setprio(struct task_struct *p, int prio)
3942 unsigned long flags;
3943 int oldprio, on_rq, running;
3946 BUG_ON(prio < 0 || prio > MAX_PRIO);
3948 rq = task_rq_lock(p, &flags);
3949 update_rq_clock(rq);
3952 on_rq = p->se.on_rq;
3953 running = task_running(rq, p);
3955 dequeue_task(rq, p, 0);
3957 p->sched_class->put_prev_task(rq, p);
3961 p->sched_class = &rt_sched_class;
3963 p->sched_class = &fair_sched_class;
3969 p->sched_class->set_curr_task(rq);
3970 enqueue_task(rq, p, 0);
3972 * Reschedule if we are currently running on this runqueue and
3973 * our priority decreased, or if we are not currently running on
3974 * this runqueue and our priority is higher than the current's
3977 if (p->prio > oldprio)
3978 resched_task(rq->curr);
3980 check_preempt_curr(rq, p);
3983 task_rq_unlock(rq, &flags);
3988 void set_user_nice(struct task_struct *p, long nice)
3990 int old_prio, delta, on_rq;
3991 unsigned long flags;
3994 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3997 * We have to be careful, if called from sys_setpriority(),
3998 * the task might be in the middle of scheduling on another CPU.
4000 rq = task_rq_lock(p, &flags);
4001 update_rq_clock(rq);
4003 * The RT priorities are set via sched_setscheduler(), but we still
4004 * allow the 'normal' nice value to be set - but as expected
4005 * it wont have any effect on scheduling until the task is
4006 * SCHED_FIFO/SCHED_RR:
4008 if (task_has_rt_policy(p)) {
4009 p->static_prio = NICE_TO_PRIO(nice);
4012 on_rq = p->se.on_rq;
4014 dequeue_task(rq, p, 0);
4018 p->static_prio = NICE_TO_PRIO(nice);
4021 p->prio = effective_prio(p);
4022 delta = p->prio - old_prio;
4025 enqueue_task(rq, p, 0);
4028 * If the task increased its priority or is running and
4029 * lowered its priority, then reschedule its CPU:
4031 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4032 resched_task(rq->curr);
4035 task_rq_unlock(rq, &flags);
4037 EXPORT_SYMBOL(set_user_nice);
4040 * can_nice - check if a task can reduce its nice value
4044 int can_nice(const struct task_struct *p, const int nice)
4046 /* convert nice value [19,-20] to rlimit style value [1,40] */
4047 int nice_rlim = 20 - nice;
4049 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4050 capable(CAP_SYS_NICE));
4053 #ifdef __ARCH_WANT_SYS_NICE
4056 * sys_nice - change the priority of the current process.
4057 * @increment: priority increment
4059 * sys_setpriority is a more generic, but much slower function that
4060 * does similar things.
4062 asmlinkage long sys_nice(int increment)
4067 * Setpriority might change our priority at the same moment.
4068 * We don't have to worry. Conceptually one call occurs first
4069 * and we have a single winner.
4071 if (increment < -40)
4076 nice = PRIO_TO_NICE(current->static_prio) + increment;
4082 if (increment < 0 && !can_nice(current, nice))
4085 retval = security_task_setnice(current, nice);
4089 set_user_nice(current, nice);
4096 * task_prio - return the priority value of a given task.
4097 * @p: the task in question.
4099 * This is the priority value as seen by users in /proc.
4100 * RT tasks are offset by -200. Normal tasks are centered
4101 * around 0, value goes from -16 to +15.
4103 int task_prio(const struct task_struct *p)
4105 return p->prio - MAX_RT_PRIO;
4109 * task_nice - return the nice value of a given task.
4110 * @p: the task in question.
4112 int task_nice(const struct task_struct *p)
4114 return TASK_NICE(p);
4116 EXPORT_SYMBOL_GPL(task_nice);
4119 * idle_cpu - is a given cpu idle currently?
4120 * @cpu: the processor in question.
4122 int idle_cpu(int cpu)
4124 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4128 * idle_task - return the idle task for a given cpu.
4129 * @cpu: the processor in question.
4131 struct task_struct *idle_task(int cpu)
4133 return cpu_rq(cpu)->idle;
4137 * find_process_by_pid - find a process with a matching PID value.
4138 * @pid: the pid in question.
4140 static inline struct task_struct *find_process_by_pid(pid_t pid)
4142 return pid ? find_task_by_pid(pid) : current;
4145 /* Actually do priority change: must hold rq lock. */
4147 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4149 BUG_ON(p->se.on_rq);
4152 switch (p->policy) {
4156 p->sched_class = &fair_sched_class;
4160 p->sched_class = &rt_sched_class;
4164 p->rt_priority = prio;
4165 p->normal_prio = normal_prio(p);
4166 /* we are holding p->pi_lock already */
4167 p->prio = rt_mutex_getprio(p);
4172 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4173 * @p: the task in question.
4174 * @policy: new policy.
4175 * @param: structure containing the new RT priority.
4177 * NOTE that the task may be already dead.
4179 int sched_setscheduler(struct task_struct *p, int policy,
4180 struct sched_param *param)
4182 int retval, oldprio, oldpolicy = -1, on_rq, running;
4183 unsigned long flags;
4186 /* may grab non-irq protected spin_locks */
4187 BUG_ON(in_interrupt());
4189 /* double check policy once rq lock held */
4191 policy = oldpolicy = p->policy;
4192 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4193 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4194 policy != SCHED_IDLE)
4197 * Valid priorities for SCHED_FIFO and SCHED_RR are
4198 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4199 * SCHED_BATCH and SCHED_IDLE is 0.
4201 if (param->sched_priority < 0 ||
4202 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4203 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4205 if (rt_policy(policy) != (param->sched_priority != 0))
4209 * Allow unprivileged RT tasks to decrease priority:
4211 if (!capable(CAP_SYS_NICE)) {
4212 if (rt_policy(policy)) {
4213 unsigned long rlim_rtprio;
4215 if (!lock_task_sighand(p, &flags))
4217 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4218 unlock_task_sighand(p, &flags);
4220 /* can't set/change the rt policy */
4221 if (policy != p->policy && !rlim_rtprio)
4224 /* can't increase priority */
4225 if (param->sched_priority > p->rt_priority &&
4226 param->sched_priority > rlim_rtprio)
4230 * Like positive nice levels, dont allow tasks to
4231 * move out of SCHED_IDLE either:
4233 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4236 /* can't change other user's priorities */
4237 if ((current->euid != p->euid) &&
4238 (current->euid != p->uid))
4242 retval = security_task_setscheduler(p, policy, param);
4246 * make sure no PI-waiters arrive (or leave) while we are
4247 * changing the priority of the task:
4249 spin_lock_irqsave(&p->pi_lock, flags);
4251 * To be able to change p->policy safely, the apropriate
4252 * runqueue lock must be held.
4254 rq = __task_rq_lock(p);
4255 /* recheck policy now with rq lock held */
4256 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4257 policy = oldpolicy = -1;
4258 __task_rq_unlock(rq);
4259 spin_unlock_irqrestore(&p->pi_lock, flags);
4262 update_rq_clock(rq);
4263 on_rq = p->se.on_rq;
4264 running = task_running(rq, p);
4266 deactivate_task(rq, p, 0);
4268 p->sched_class->put_prev_task(rq, p);
4272 __setscheduler(rq, p, policy, param->sched_priority);
4276 p->sched_class->set_curr_task(rq);
4277 activate_task(rq, p, 0);
4279 * Reschedule if we are currently running on this runqueue and
4280 * our priority decreased, or if we are not currently running on
4281 * this runqueue and our priority is higher than the current's
4284 if (p->prio > oldprio)
4285 resched_task(rq->curr);
4287 check_preempt_curr(rq, p);
4290 __task_rq_unlock(rq);
4291 spin_unlock_irqrestore(&p->pi_lock, flags);
4293 rt_mutex_adjust_pi(p);
4297 EXPORT_SYMBOL_GPL(sched_setscheduler);
4300 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4302 struct sched_param lparam;
4303 struct task_struct *p;
4306 if (!param || pid < 0)
4308 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4313 p = find_process_by_pid(pid);
4315 retval = sched_setscheduler(p, policy, &lparam);
4322 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4323 * @pid: the pid in question.
4324 * @policy: new policy.
4325 * @param: structure containing the new RT priority.
4327 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4328 struct sched_param __user *param)
4330 /* negative values for policy are not valid */
4334 return do_sched_setscheduler(pid, policy, param);
4338 * sys_sched_setparam - set/change the RT priority of a thread
4339 * @pid: the pid in question.
4340 * @param: structure containing the new RT priority.
4342 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4344 return do_sched_setscheduler(pid, -1, param);
4348 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4349 * @pid: the pid in question.
4351 asmlinkage long sys_sched_getscheduler(pid_t pid)
4353 struct task_struct *p;
4354 int retval = -EINVAL;
4360 read_lock(&tasklist_lock);
4361 p = find_process_by_pid(pid);
4363 retval = security_task_getscheduler(p);
4367 read_unlock(&tasklist_lock);
4374 * sys_sched_getscheduler - get the RT priority of a thread
4375 * @pid: the pid in question.
4376 * @param: structure containing the RT priority.
4378 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4380 struct sched_param lp;
4381 struct task_struct *p;
4382 int retval = -EINVAL;
4384 if (!param || pid < 0)
4387 read_lock(&tasklist_lock);
4388 p = find_process_by_pid(pid);
4393 retval = security_task_getscheduler(p);
4397 lp.sched_priority = p->rt_priority;
4398 read_unlock(&tasklist_lock);
4401 * This one might sleep, we cannot do it with a spinlock held ...
4403 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4409 read_unlock(&tasklist_lock);
4413 long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4415 cpumask_t cpus_allowed;
4416 struct task_struct *p;
4419 mutex_lock(&sched_hotcpu_mutex);
4420 read_lock(&tasklist_lock);
4422 p = find_process_by_pid(pid);
4424 read_unlock(&tasklist_lock);
4425 mutex_unlock(&sched_hotcpu_mutex);
4430 * It is not safe to call set_cpus_allowed with the
4431 * tasklist_lock held. We will bump the task_struct's
4432 * usage count and then drop tasklist_lock.
4435 read_unlock(&tasklist_lock);
4438 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4439 !capable(CAP_SYS_NICE))
4442 retval = security_task_setscheduler(p, 0, NULL);
4446 cpus_allowed = cpuset_cpus_allowed(p);
4447 cpus_and(new_mask, new_mask, cpus_allowed);
4448 retval = set_cpus_allowed(p, new_mask);
4452 mutex_unlock(&sched_hotcpu_mutex);
4456 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4457 cpumask_t *new_mask)
4459 if (len < sizeof(cpumask_t)) {
4460 memset(new_mask, 0, sizeof(cpumask_t));
4461 } else if (len > sizeof(cpumask_t)) {
4462 len = sizeof(cpumask_t);
4464 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4468 * sys_sched_setaffinity - set the cpu affinity of a process
4469 * @pid: pid of the process
4470 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4471 * @user_mask_ptr: user-space pointer to the new cpu mask
4473 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4474 unsigned long __user *user_mask_ptr)
4479 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4483 return sched_setaffinity(pid, new_mask);
4487 * Represents all cpu's present in the system
4488 * In systems capable of hotplug, this map could dynamically grow
4489 * as new cpu's are detected in the system via any platform specific
4490 * method, such as ACPI for e.g.
4493 cpumask_t cpu_present_map __read_mostly;
4494 EXPORT_SYMBOL(cpu_present_map);
4497 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4498 EXPORT_SYMBOL(cpu_online_map);
4500 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4501 EXPORT_SYMBOL(cpu_possible_map);
4504 long sched_getaffinity(pid_t pid, cpumask_t *mask)
4506 struct task_struct *p;
4509 mutex_lock(&sched_hotcpu_mutex);
4510 read_lock(&tasklist_lock);
4513 p = find_process_by_pid(pid);
4517 retval = security_task_getscheduler(p);
4521 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4524 read_unlock(&tasklist_lock);
4525 mutex_unlock(&sched_hotcpu_mutex);
4531 * sys_sched_getaffinity - get the cpu affinity of a process
4532 * @pid: pid of the process
4533 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4534 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4536 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4537 unsigned long __user *user_mask_ptr)
4542 if (len < sizeof(cpumask_t))
4545 ret = sched_getaffinity(pid, &mask);
4549 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4552 return sizeof(cpumask_t);
4556 * sys_sched_yield - yield the current processor to other threads.
4558 * This function yields the current CPU to other tasks. If there are no
4559 * other threads running on this CPU then this function will return.
4561 asmlinkage long sys_sched_yield(void)
4563 struct rq *rq = this_rq_lock();
4565 schedstat_inc(rq, yld_cnt);
4566 current->sched_class->yield_task(rq);
4569 * Since we are going to call schedule() anyway, there's
4570 * no need to preempt or enable interrupts:
4572 __release(rq->lock);
4573 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4574 _raw_spin_unlock(&rq->lock);
4575 preempt_enable_no_resched();
4582 static void __cond_resched(void)
4584 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4585 __might_sleep(__FILE__, __LINE__);
4588 * The BKS might be reacquired before we have dropped
4589 * PREEMPT_ACTIVE, which could trigger a second
4590 * cond_resched() call.
4593 add_preempt_count(PREEMPT_ACTIVE);
4595 sub_preempt_count(PREEMPT_ACTIVE);
4596 } while (need_resched());
4599 int __sched cond_resched(void)
4601 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4602 system_state == SYSTEM_RUNNING) {
4608 EXPORT_SYMBOL(cond_resched);
4611 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4612 * call schedule, and on return reacquire the lock.
4614 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4615 * operations here to prevent schedule() from being called twice (once via
4616 * spin_unlock(), once by hand).
4618 int cond_resched_lock(spinlock_t *lock)
4622 if (need_lockbreak(lock)) {
4628 if (need_resched() && system_state == SYSTEM_RUNNING) {
4629 spin_release(&lock->dep_map, 1, _THIS_IP_);
4630 _raw_spin_unlock(lock);
4631 preempt_enable_no_resched();
4638 EXPORT_SYMBOL(cond_resched_lock);
4640 int __sched cond_resched_softirq(void)
4642 BUG_ON(!in_softirq());
4644 if (need_resched() && system_state == SYSTEM_RUNNING) {
4652 EXPORT_SYMBOL(cond_resched_softirq);
4655 * yield - yield the current processor to other threads.
4657 * This is a shortcut for kernel-space yielding - it marks the
4658 * thread runnable and calls sys_sched_yield().
4660 void __sched yield(void)
4662 set_current_state(TASK_RUNNING);
4665 EXPORT_SYMBOL(yield);
4668 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4669 * that process accounting knows that this is a task in IO wait state.
4671 * But don't do that if it is a deliberate, throttling IO wait (this task
4672 * has set its backing_dev_info: the queue against which it should throttle)
4674 void __sched io_schedule(void)
4676 struct rq *rq = &__raw_get_cpu_var(runqueues);
4678 delayacct_blkio_start();
4679 atomic_inc(&rq->nr_iowait);
4681 atomic_dec(&rq->nr_iowait);
4682 delayacct_blkio_end();
4684 EXPORT_SYMBOL(io_schedule);
4686 long __sched io_schedule_timeout(long timeout)
4688 struct rq *rq = &__raw_get_cpu_var(runqueues);
4691 delayacct_blkio_start();
4692 atomic_inc(&rq->nr_iowait);
4693 ret = schedule_timeout(timeout);
4694 atomic_dec(&rq->nr_iowait);
4695 delayacct_blkio_end();
4700 * sys_sched_get_priority_max - return maximum RT priority.
4701 * @policy: scheduling class.
4703 * this syscall returns the maximum rt_priority that can be used
4704 * by a given scheduling class.
4706 asmlinkage long sys_sched_get_priority_max(int policy)
4713 ret = MAX_USER_RT_PRIO-1;
4725 * sys_sched_get_priority_min - return minimum RT priority.
4726 * @policy: scheduling class.
4728 * this syscall returns the minimum rt_priority that can be used
4729 * by a given scheduling class.
4731 asmlinkage long sys_sched_get_priority_min(int policy)
4749 * sys_sched_rr_get_interval - return the default timeslice of a process.
4750 * @pid: pid of the process.
4751 * @interval: userspace pointer to the timeslice value.
4753 * this syscall writes the default timeslice value of a given process
4754 * into the user-space timespec buffer. A value of '0' means infinity.
4757 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4759 struct task_struct *p;
4760 int retval = -EINVAL;
4767 read_lock(&tasklist_lock);
4768 p = find_process_by_pid(pid);
4772 retval = security_task_getscheduler(p);
4776 jiffies_to_timespec(p->policy == SCHED_FIFO ?
4777 0 : static_prio_timeslice(p->static_prio), &t);
4778 read_unlock(&tasklist_lock);
4779 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4783 read_unlock(&tasklist_lock);
4787 static const char stat_nam[] = "RSDTtZX";
4789 static void show_task(struct task_struct *p)
4791 unsigned long free = 0;
4794 state = p->state ? __ffs(p->state) + 1 : 0;
4795 printk("%-13.13s %c", p->comm,
4796 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4797 #if BITS_PER_LONG == 32
4798 if (state == TASK_RUNNING)
4799 printk(" running ");
4801 printk(" %08lx ", thread_saved_pc(p));
4803 if (state == TASK_RUNNING)
4804 printk(" running task ");
4806 printk(" %016lx ", thread_saved_pc(p));
4808 #ifdef CONFIG_DEBUG_STACK_USAGE
4810 unsigned long *n = end_of_stack(p);
4813 free = (unsigned long)n - (unsigned long)end_of_stack(p);
4816 printk("%5lu %5d %6d\n", free, p->pid, p->parent->pid);
4818 if (state != TASK_RUNNING)
4819 show_stack(p, NULL);
4822 void show_state_filter(unsigned long state_filter)
4824 struct task_struct *g, *p;
4826 #if BITS_PER_LONG == 32
4828 " task PC stack pid father\n");
4831 " task PC stack pid father\n");
4833 read_lock(&tasklist_lock);
4834 do_each_thread(g, p) {
4836 * reset the NMI-timeout, listing all files on a slow
4837 * console might take alot of time:
4839 touch_nmi_watchdog();
4840 if (!state_filter || (p->state & state_filter))
4842 } while_each_thread(g, p);
4844 touch_all_softlockup_watchdogs();
4846 #ifdef CONFIG_SCHED_DEBUG
4847 sysrq_sched_debug_show();
4849 read_unlock(&tasklist_lock);
4851 * Only show locks if all tasks are dumped:
4853 if (state_filter == -1)
4854 debug_show_all_locks();
4857 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4859 idle->sched_class = &idle_sched_class;
4863 * init_idle - set up an idle thread for a given CPU
4864 * @idle: task in question
4865 * @cpu: cpu the idle task belongs to
4867 * NOTE: this function does not set the idle thread's NEED_RESCHED
4868 * flag, to make booting more robust.
4870 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4872 struct rq *rq = cpu_rq(cpu);
4873 unsigned long flags;
4876 idle->se.exec_start = sched_clock();
4878 idle->prio = idle->normal_prio = MAX_PRIO;
4879 idle->cpus_allowed = cpumask_of_cpu(cpu);
4880 __set_task_cpu(idle, cpu);
4882 spin_lock_irqsave(&rq->lock, flags);
4883 rq->curr = rq->idle = idle;
4884 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4887 spin_unlock_irqrestore(&rq->lock, flags);
4889 /* Set the preempt count _outside_ the spinlocks! */
4890 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4891 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4893 task_thread_info(idle)->preempt_count = 0;
4896 * The idle tasks have their own, simple scheduling class:
4898 idle->sched_class = &idle_sched_class;
4902 * In a system that switches off the HZ timer nohz_cpu_mask
4903 * indicates which cpus entered this state. This is used
4904 * in the rcu update to wait only for active cpus. For system
4905 * which do not switch off the HZ timer nohz_cpu_mask should
4906 * always be CPU_MASK_NONE.
4908 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4912 * This is how migration works:
4914 * 1) we queue a struct migration_req structure in the source CPU's
4915 * runqueue and wake up that CPU's migration thread.
4916 * 2) we down() the locked semaphore => thread blocks.
4917 * 3) migration thread wakes up (implicitly it forces the migrated
4918 * thread off the CPU)
4919 * 4) it gets the migration request and checks whether the migrated
4920 * task is still in the wrong runqueue.
4921 * 5) if it's in the wrong runqueue then the migration thread removes
4922 * it and puts it into the right queue.
4923 * 6) migration thread up()s the semaphore.
4924 * 7) we wake up and the migration is done.
4928 * Change a given task's CPU affinity. Migrate the thread to a
4929 * proper CPU and schedule it away if the CPU it's executing on
4930 * is removed from the allowed bitmask.
4932 * NOTE: the caller must have a valid reference to the task, the
4933 * task must not exit() & deallocate itself prematurely. The
4934 * call is not atomic; no spinlocks may be held.
4936 int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4938 struct migration_req req;
4939 unsigned long flags;
4943 rq = task_rq_lock(p, &flags);
4944 if (!cpus_intersects(new_mask, cpu_online_map)) {
4949 p->cpus_allowed = new_mask;
4950 /* Can the task run on the task's current CPU? If so, we're done */
4951 if (cpu_isset(task_cpu(p), new_mask))
4954 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4955 /* Need help from migration thread: drop lock and wait. */
4956 task_rq_unlock(rq, &flags);
4957 wake_up_process(rq->migration_thread);
4958 wait_for_completion(&req.done);
4959 tlb_migrate_finish(p->mm);
4963 task_rq_unlock(rq, &flags);
4967 EXPORT_SYMBOL_GPL(set_cpus_allowed);
4970 * Move (not current) task off this cpu, onto dest cpu. We're doing
4971 * this because either it can't run here any more (set_cpus_allowed()
4972 * away from this CPU, or CPU going down), or because we're
4973 * attempting to rebalance this task on exec (sched_exec).
4975 * So we race with normal scheduler movements, but that's OK, as long
4976 * as the task is no longer on this CPU.
4978 * Returns non-zero if task was successfully migrated.
4980 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4982 struct rq *rq_dest, *rq_src;
4985 if (unlikely(cpu_is_offline(dest_cpu)))
4988 rq_src = cpu_rq(src_cpu);
4989 rq_dest = cpu_rq(dest_cpu);
4991 double_rq_lock(rq_src, rq_dest);
4992 /* Already moved. */
4993 if (task_cpu(p) != src_cpu)
4995 /* Affinity changed (again). */
4996 if (!cpu_isset(dest_cpu, p->cpus_allowed))
4999 on_rq = p->se.on_rq;
5001 deactivate_task(rq_src, p, 0);
5003 set_task_cpu(p, dest_cpu);
5005 activate_task(rq_dest, p, 0);
5006 check_preempt_curr(rq_dest, p);
5010 double_rq_unlock(rq_src, rq_dest);
5015 * migration_thread - this is a highprio system thread that performs
5016 * thread migration by bumping thread off CPU then 'pushing' onto
5019 static int migration_thread(void *data)
5021 int cpu = (long)data;
5025 BUG_ON(rq->migration_thread != current);
5027 set_current_state(TASK_INTERRUPTIBLE);
5028 while (!kthread_should_stop()) {
5029 struct migration_req *req;
5030 struct list_head *head;
5032 spin_lock_irq(&rq->lock);
5034 if (cpu_is_offline(cpu)) {
5035 spin_unlock_irq(&rq->lock);
5039 if (rq->active_balance) {
5040 active_load_balance(rq, cpu);
5041 rq->active_balance = 0;
5044 head = &rq->migration_queue;
5046 if (list_empty(head)) {
5047 spin_unlock_irq(&rq->lock);
5049 set_current_state(TASK_INTERRUPTIBLE);
5052 req = list_entry(head->next, struct migration_req, list);
5053 list_del_init(head->next);
5055 spin_unlock(&rq->lock);
5056 __migrate_task(req->task, cpu, req->dest_cpu);
5059 complete(&req->done);
5061 __set_current_state(TASK_RUNNING);
5065 /* Wait for kthread_stop */
5066 set_current_state(TASK_INTERRUPTIBLE);
5067 while (!kthread_should_stop()) {
5069 set_current_state(TASK_INTERRUPTIBLE);
5071 __set_current_state(TASK_RUNNING);
5075 #ifdef CONFIG_HOTPLUG_CPU
5077 * Figure out where task on dead CPU should go, use force if neccessary.
5078 * NOTE: interrupts should be disabled by the caller
5080 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5082 unsigned long flags;
5089 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5090 cpus_and(mask, mask, p->cpus_allowed);
5091 dest_cpu = any_online_cpu(mask);
5093 /* On any allowed CPU? */
5094 if (dest_cpu == NR_CPUS)
5095 dest_cpu = any_online_cpu(p->cpus_allowed);
5097 /* No more Mr. Nice Guy. */
5098 if (dest_cpu == NR_CPUS) {
5099 rq = task_rq_lock(p, &flags);
5100 cpus_setall(p->cpus_allowed);
5101 dest_cpu = any_online_cpu(p->cpus_allowed);
5102 task_rq_unlock(rq, &flags);
5105 * Don't tell them about moving exiting tasks or
5106 * kernel threads (both mm NULL), since they never
5109 if (p->mm && printk_ratelimit())
5110 printk(KERN_INFO "process %d (%s) no "
5111 "longer affine to cpu%d\n",
5112 p->pid, p->comm, dead_cpu);
5114 if (!__migrate_task(p, dead_cpu, dest_cpu))
5119 * While a dead CPU has no uninterruptible tasks queued at this point,
5120 * it might still have a nonzero ->nr_uninterruptible counter, because
5121 * for performance reasons the counter is not stricly tracking tasks to
5122 * their home CPUs. So we just add the counter to another CPU's counter,
5123 * to keep the global sum constant after CPU-down:
5125 static void migrate_nr_uninterruptible(struct rq *rq_src)
5127 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5128 unsigned long flags;
5130 local_irq_save(flags);
5131 double_rq_lock(rq_src, rq_dest);
5132 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5133 rq_src->nr_uninterruptible = 0;
5134 double_rq_unlock(rq_src, rq_dest);
5135 local_irq_restore(flags);
5138 /* Run through task list and migrate tasks from the dead cpu. */
5139 static void migrate_live_tasks(int src_cpu)
5141 struct task_struct *p, *t;
5143 write_lock_irq(&tasklist_lock);
5145 do_each_thread(t, p) {
5149 if (task_cpu(p) == src_cpu)
5150 move_task_off_dead_cpu(src_cpu, p);
5151 } while_each_thread(t, p);
5153 write_unlock_irq(&tasklist_lock);
5157 * Schedules idle task to be the next runnable task on current CPU.
5158 * It does so by boosting its priority to highest possible and adding it to
5159 * the _front_ of the runqueue. Used by CPU offline code.
5161 void sched_idle_next(void)
5163 int this_cpu = smp_processor_id();
5164 struct rq *rq = cpu_rq(this_cpu);
5165 struct task_struct *p = rq->idle;
5166 unsigned long flags;
5168 /* cpu has to be offline */
5169 BUG_ON(cpu_online(this_cpu));
5172 * Strictly not necessary since rest of the CPUs are stopped by now
5173 * and interrupts disabled on the current cpu.
5175 spin_lock_irqsave(&rq->lock, flags);
5177 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5179 /* Add idle task to the _front_ of its priority queue: */
5180 activate_idle_task(p, rq);
5182 spin_unlock_irqrestore(&rq->lock, flags);
5186 * Ensures that the idle task is using init_mm right before its cpu goes
5189 void idle_task_exit(void)
5191 struct mm_struct *mm = current->active_mm;
5193 BUG_ON(cpu_online(smp_processor_id()));
5196 switch_mm(mm, &init_mm, current);
5200 /* called under rq->lock with disabled interrupts */
5201 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5203 struct rq *rq = cpu_rq(dead_cpu);
5205 /* Must be exiting, otherwise would be on tasklist. */
5206 BUG_ON(p->exit_state != EXIT_ZOMBIE && p->exit_state != EXIT_DEAD);
5208 /* Cannot have done final schedule yet: would have vanished. */
5209 BUG_ON(p->state == TASK_DEAD);
5214 * Drop lock around migration; if someone else moves it,
5215 * that's OK. No task can be added to this CPU, so iteration is
5217 * NOTE: interrupts should be left disabled --dev@
5219 spin_unlock(&rq->lock);
5220 move_task_off_dead_cpu(dead_cpu, p);
5221 spin_lock(&rq->lock);
5226 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5227 static void migrate_dead_tasks(unsigned int dead_cpu)
5229 struct rq *rq = cpu_rq(dead_cpu);
5230 struct task_struct *next;
5233 if (!rq->nr_running)
5235 update_rq_clock(rq);
5236 next = pick_next_task(rq, rq->curr);
5239 migrate_dead(dead_cpu, next);
5243 #endif /* CONFIG_HOTPLUG_CPU */
5245 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5247 static struct ctl_table sd_ctl_dir[] = {
5249 .procname = "sched_domain",
5255 static struct ctl_table sd_ctl_root[] = {
5257 .ctl_name = CTL_KERN,
5258 .procname = "kernel",
5260 .child = sd_ctl_dir,
5265 static struct ctl_table *sd_alloc_ctl_entry(int n)
5267 struct ctl_table *entry =
5268 kmalloc(n * sizeof(struct ctl_table), GFP_KERNEL);
5271 memset(entry, 0, n * sizeof(struct ctl_table));
5277 set_table_entry(struct ctl_table *entry,
5278 const char *procname, void *data, int maxlen,
5279 mode_t mode, proc_handler *proc_handler)
5281 entry->procname = procname;
5283 entry->maxlen = maxlen;
5285 entry->proc_handler = proc_handler;
5288 static struct ctl_table *
5289 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5291 struct ctl_table *table = sd_alloc_ctl_entry(14);
5293 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5294 sizeof(long), 0644, proc_doulongvec_minmax);
5295 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5296 sizeof(long), 0644, proc_doulongvec_minmax);
5297 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5298 sizeof(int), 0644, proc_dointvec_minmax);
5299 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5300 sizeof(int), 0644, proc_dointvec_minmax);
5301 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5302 sizeof(int), 0644, proc_dointvec_minmax);
5303 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5304 sizeof(int), 0644, proc_dointvec_minmax);
5305 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5306 sizeof(int), 0644, proc_dointvec_minmax);
5307 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5308 sizeof(int), 0644, proc_dointvec_minmax);
5309 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5310 sizeof(int), 0644, proc_dointvec_minmax);
5311 set_table_entry(&table[10], "cache_nice_tries",
5312 &sd->cache_nice_tries,
5313 sizeof(int), 0644, proc_dointvec_minmax);
5314 set_table_entry(&table[12], "flags", &sd->flags,
5315 sizeof(int), 0644, proc_dointvec_minmax);
5320 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5322 struct ctl_table *entry, *table;
5323 struct sched_domain *sd;
5324 int domain_num = 0, i;
5327 for_each_domain(cpu, sd)
5329 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5332 for_each_domain(cpu, sd) {
5333 snprintf(buf, 32, "domain%d", i);
5334 entry->procname = kstrdup(buf, GFP_KERNEL);
5336 entry->child = sd_alloc_ctl_domain_table(sd);
5343 static struct ctl_table_header *sd_sysctl_header;
5344 static void init_sched_domain_sysctl(void)
5346 int i, cpu_num = num_online_cpus();
5347 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5350 sd_ctl_dir[0].child = entry;
5352 for (i = 0; i < cpu_num; i++, entry++) {
5353 snprintf(buf, 32, "cpu%d", i);
5354 entry->procname = kstrdup(buf, GFP_KERNEL);
5356 entry->child = sd_alloc_ctl_cpu_table(i);
5358 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5361 static void init_sched_domain_sysctl(void)
5367 * migration_call - callback that gets triggered when a CPU is added.
5368 * Here we can start up the necessary migration thread for the new CPU.
5370 static int __cpuinit
5371 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5373 struct task_struct *p;
5374 int cpu = (long)hcpu;
5375 unsigned long flags;
5379 case CPU_LOCK_ACQUIRE:
5380 mutex_lock(&sched_hotcpu_mutex);
5383 case CPU_UP_PREPARE:
5384 case CPU_UP_PREPARE_FROZEN:
5385 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5388 kthread_bind(p, cpu);
5389 /* Must be high prio: stop_machine expects to yield to it. */
5390 rq = task_rq_lock(p, &flags);
5391 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5392 task_rq_unlock(rq, &flags);
5393 cpu_rq(cpu)->migration_thread = p;
5397 case CPU_ONLINE_FROZEN:
5398 /* Strictly unneccessary, as first user will wake it. */
5399 wake_up_process(cpu_rq(cpu)->migration_thread);
5402 #ifdef CONFIG_HOTPLUG_CPU
5403 case CPU_UP_CANCELED:
5404 case CPU_UP_CANCELED_FROZEN:
5405 if (!cpu_rq(cpu)->migration_thread)
5407 /* Unbind it from offline cpu so it can run. Fall thru. */
5408 kthread_bind(cpu_rq(cpu)->migration_thread,
5409 any_online_cpu(cpu_online_map));
5410 kthread_stop(cpu_rq(cpu)->migration_thread);
5411 cpu_rq(cpu)->migration_thread = NULL;
5415 case CPU_DEAD_FROZEN:
5416 migrate_live_tasks(cpu);
5418 kthread_stop(rq->migration_thread);
5419 rq->migration_thread = NULL;
5420 /* Idle task back to normal (off runqueue, low prio) */
5421 rq = task_rq_lock(rq->idle, &flags);
5422 update_rq_clock(rq);
5423 deactivate_task(rq, rq->idle, 0);
5424 rq->idle->static_prio = MAX_PRIO;
5425 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5426 rq->idle->sched_class = &idle_sched_class;
5427 migrate_dead_tasks(cpu);
5428 task_rq_unlock(rq, &flags);
5429 migrate_nr_uninterruptible(rq);
5430 BUG_ON(rq->nr_running != 0);
5432 /* No need to migrate the tasks: it was best-effort if
5433 * they didn't take sched_hotcpu_mutex. Just wake up
5434 * the requestors. */
5435 spin_lock_irq(&rq->lock);
5436 while (!list_empty(&rq->migration_queue)) {
5437 struct migration_req *req;
5439 req = list_entry(rq->migration_queue.next,
5440 struct migration_req, list);
5441 list_del_init(&req->list);
5442 complete(&req->done);
5444 spin_unlock_irq(&rq->lock);
5447 case CPU_LOCK_RELEASE:
5448 mutex_unlock(&sched_hotcpu_mutex);
5454 /* Register at highest priority so that task migration (migrate_all_tasks)
5455 * happens before everything else.
5457 static struct notifier_block __cpuinitdata migration_notifier = {
5458 .notifier_call = migration_call,
5462 int __init migration_init(void)
5464 void *cpu = (void *)(long)smp_processor_id();
5467 /* Start one for the boot CPU: */
5468 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5469 BUG_ON(err == NOTIFY_BAD);
5470 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5471 register_cpu_notifier(&migration_notifier);
5479 /* Number of possible processor ids */
5480 int nr_cpu_ids __read_mostly = NR_CPUS;
5481 EXPORT_SYMBOL(nr_cpu_ids);
5483 #undef SCHED_DOMAIN_DEBUG
5484 #ifdef SCHED_DOMAIN_DEBUG
5485 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5490 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5494 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5499 struct sched_group *group = sd->groups;
5500 cpumask_t groupmask;
5502 cpumask_scnprintf(str, NR_CPUS, sd->span);
5503 cpus_clear(groupmask);
5506 for (i = 0; i < level + 1; i++)
5508 printk("domain %d: ", level);
5510 if (!(sd->flags & SD_LOAD_BALANCE)) {
5511 printk("does not load-balance\n");
5513 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5518 printk("span %s\n", str);
5520 if (!cpu_isset(cpu, sd->span))
5521 printk(KERN_ERR "ERROR: domain->span does not contain "
5523 if (!cpu_isset(cpu, group->cpumask))
5524 printk(KERN_ERR "ERROR: domain->groups does not contain"
5528 for (i = 0; i < level + 2; i++)
5534 printk(KERN_ERR "ERROR: group is NULL\n");
5538 if (!group->__cpu_power) {
5540 printk(KERN_ERR "ERROR: domain->cpu_power not "
5544 if (!cpus_weight(group->cpumask)) {
5546 printk(KERN_ERR "ERROR: empty group\n");
5549 if (cpus_intersects(groupmask, group->cpumask)) {
5551 printk(KERN_ERR "ERROR: repeated CPUs\n");
5554 cpus_or(groupmask, groupmask, group->cpumask);
5556 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5559 group = group->next;
5560 } while (group != sd->groups);
5563 if (!cpus_equal(sd->span, groupmask))
5564 printk(KERN_ERR "ERROR: groups don't span "
5572 if (!cpus_subset(groupmask, sd->span))
5573 printk(KERN_ERR "ERROR: parent span is not a superset "
5574 "of domain->span\n");
5579 # define sched_domain_debug(sd, cpu) do { } while (0)
5582 static int sd_degenerate(struct sched_domain *sd)
5584 if (cpus_weight(sd->span) == 1)
5587 /* Following flags need at least 2 groups */
5588 if (sd->flags & (SD_LOAD_BALANCE |
5589 SD_BALANCE_NEWIDLE |
5593 SD_SHARE_PKG_RESOURCES)) {
5594 if (sd->groups != sd->groups->next)
5598 /* Following flags don't use groups */
5599 if (sd->flags & (SD_WAKE_IDLE |
5608 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5610 unsigned long cflags = sd->flags, pflags = parent->flags;
5612 if (sd_degenerate(parent))
5615 if (!cpus_equal(sd->span, parent->span))
5618 /* Does parent contain flags not in child? */
5619 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5620 if (cflags & SD_WAKE_AFFINE)
5621 pflags &= ~SD_WAKE_BALANCE;
5622 /* Flags needing groups don't count if only 1 group in parent */
5623 if (parent->groups == parent->groups->next) {
5624 pflags &= ~(SD_LOAD_BALANCE |
5625 SD_BALANCE_NEWIDLE |
5629 SD_SHARE_PKG_RESOURCES);
5631 if (~cflags & pflags)
5638 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5639 * hold the hotplug lock.
5641 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
5643 struct rq *rq = cpu_rq(cpu);
5644 struct sched_domain *tmp;
5646 /* Remove the sched domains which do not contribute to scheduling. */
5647 for (tmp = sd; tmp; tmp = tmp->parent) {
5648 struct sched_domain *parent = tmp->parent;
5651 if (sd_parent_degenerate(tmp, parent)) {
5652 tmp->parent = parent->parent;
5654 parent->parent->child = tmp;
5658 if (sd && sd_degenerate(sd)) {
5664 sched_domain_debug(sd, cpu);
5666 rcu_assign_pointer(rq->sd, sd);
5669 /* cpus with isolated domains */
5670 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
5672 /* Setup the mask of cpus configured for isolated domains */
5673 static int __init isolated_cpu_setup(char *str)
5675 int ints[NR_CPUS], i;
5677 str = get_options(str, ARRAY_SIZE(ints), ints);
5678 cpus_clear(cpu_isolated_map);
5679 for (i = 1; i <= ints[0]; i++)
5680 if (ints[i] < NR_CPUS)
5681 cpu_set(ints[i], cpu_isolated_map);
5685 __setup ("isolcpus=", isolated_cpu_setup);
5688 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5689 * to a function which identifies what group(along with sched group) a CPU
5690 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5691 * (due to the fact that we keep track of groups covered with a cpumask_t).
5693 * init_sched_build_groups will build a circular linked list of the groups
5694 * covered by the given span, and will set each group's ->cpumask correctly,
5695 * and ->cpu_power to 0.
5698 init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5699 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5700 struct sched_group **sg))
5702 struct sched_group *first = NULL, *last = NULL;
5703 cpumask_t covered = CPU_MASK_NONE;
5706 for_each_cpu_mask(i, span) {
5707 struct sched_group *sg;
5708 int group = group_fn(i, cpu_map, &sg);
5711 if (cpu_isset(i, covered))
5714 sg->cpumask = CPU_MASK_NONE;
5715 sg->__cpu_power = 0;
5717 for_each_cpu_mask(j, span) {
5718 if (group_fn(j, cpu_map, NULL) != group)
5721 cpu_set(j, covered);
5722 cpu_set(j, sg->cpumask);
5733 #define SD_NODES_PER_DOMAIN 16
5738 * find_next_best_node - find the next node to include in a sched_domain
5739 * @node: node whose sched_domain we're building
5740 * @used_nodes: nodes already in the sched_domain
5742 * Find the next node to include in a given scheduling domain. Simply
5743 * finds the closest node not already in the @used_nodes map.
5745 * Should use nodemask_t.
5747 static int find_next_best_node(int node, unsigned long *used_nodes)
5749 int i, n, val, min_val, best_node = 0;
5753 for (i = 0; i < MAX_NUMNODES; i++) {
5754 /* Start at @node */
5755 n = (node + i) % MAX_NUMNODES;
5757 if (!nr_cpus_node(n))
5760 /* Skip already used nodes */
5761 if (test_bit(n, used_nodes))
5764 /* Simple min distance search */
5765 val = node_distance(node, n);
5767 if (val < min_val) {
5773 set_bit(best_node, used_nodes);
5778 * sched_domain_node_span - get a cpumask for a node's sched_domain
5779 * @node: node whose cpumask we're constructing
5780 * @size: number of nodes to include in this span
5782 * Given a node, construct a good cpumask for its sched_domain to span. It
5783 * should be one that prevents unnecessary balancing, but also spreads tasks
5786 static cpumask_t sched_domain_node_span(int node)
5788 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
5789 cpumask_t span, nodemask;
5793 bitmap_zero(used_nodes, MAX_NUMNODES);
5795 nodemask = node_to_cpumask(node);
5796 cpus_or(span, span, nodemask);
5797 set_bit(node, used_nodes);
5799 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5800 int next_node = find_next_best_node(node, used_nodes);
5802 nodemask = node_to_cpumask(next_node);
5803 cpus_or(span, span, nodemask);
5810 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5813 * SMT sched-domains:
5815 #ifdef CONFIG_SCHED_SMT
5816 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
5817 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
5819 static int cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map,
5820 struct sched_group **sg)
5823 *sg = &per_cpu(sched_group_cpus, cpu);
5829 * multi-core sched-domains:
5831 #ifdef CONFIG_SCHED_MC
5832 static DEFINE_PER_CPU(struct sched_domain, core_domains);
5833 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
5836 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
5837 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5838 struct sched_group **sg)
5841 cpumask_t mask = cpu_sibling_map[cpu];
5842 cpus_and(mask, mask, *cpu_map);
5843 group = first_cpu(mask);
5845 *sg = &per_cpu(sched_group_core, group);
5848 #elif defined(CONFIG_SCHED_MC)
5849 static int cpu_to_core_group(int cpu, const cpumask_t *cpu_map,
5850 struct sched_group **sg)
5853 *sg = &per_cpu(sched_group_core, cpu);
5858 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
5859 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
5861 static int cpu_to_phys_group(int cpu, const cpumask_t *cpu_map,
5862 struct sched_group **sg)
5865 #ifdef CONFIG_SCHED_MC
5866 cpumask_t mask = cpu_coregroup_map(cpu);
5867 cpus_and(mask, mask, *cpu_map);
5868 group = first_cpu(mask);
5869 #elif defined(CONFIG_SCHED_SMT)
5870 cpumask_t mask = cpu_sibling_map[cpu];
5871 cpus_and(mask, mask, *cpu_map);
5872 group = first_cpu(mask);
5877 *sg = &per_cpu(sched_group_phys, group);
5883 * The init_sched_build_groups can't handle what we want to do with node
5884 * groups, so roll our own. Now each node has its own list of groups which
5885 * gets dynamically allocated.
5887 static DEFINE_PER_CPU(struct sched_domain, node_domains);
5888 static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
5890 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
5891 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
5893 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
5894 struct sched_group **sg)
5896 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
5899 cpus_and(nodemask, nodemask, *cpu_map);
5900 group = first_cpu(nodemask);
5903 *sg = &per_cpu(sched_group_allnodes, group);
5907 static void init_numa_sched_groups_power(struct sched_group *group_head)
5909 struct sched_group *sg = group_head;
5915 for_each_cpu_mask(j, sg->cpumask) {
5916 struct sched_domain *sd;
5918 sd = &per_cpu(phys_domains, j);
5919 if (j != first_cpu(sd->groups->cpumask)) {
5921 * Only add "power" once for each
5927 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
5930 if (sg != group_head)
5936 /* Free memory allocated for various sched_group structures */
5937 static void free_sched_groups(const cpumask_t *cpu_map)
5941 for_each_cpu_mask(cpu, *cpu_map) {
5942 struct sched_group **sched_group_nodes
5943 = sched_group_nodes_bycpu[cpu];
5945 if (!sched_group_nodes)
5948 for (i = 0; i < MAX_NUMNODES; i++) {
5949 cpumask_t nodemask = node_to_cpumask(i);
5950 struct sched_group *oldsg, *sg = sched_group_nodes[i];
5952 cpus_and(nodemask, nodemask, *cpu_map);
5953 if (cpus_empty(nodemask))
5963 if (oldsg != sched_group_nodes[i])
5966 kfree(sched_group_nodes);
5967 sched_group_nodes_bycpu[cpu] = NULL;
5971 static void free_sched_groups(const cpumask_t *cpu_map)
5977 * Initialize sched groups cpu_power.
5979 * cpu_power indicates the capacity of sched group, which is used while
5980 * distributing the load between different sched groups in a sched domain.
5981 * Typically cpu_power for all the groups in a sched domain will be same unless
5982 * there are asymmetries in the topology. If there are asymmetries, group
5983 * having more cpu_power will pickup more load compared to the group having
5986 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
5987 * the maximum number of tasks a group can handle in the presence of other idle
5988 * or lightly loaded groups in the same sched domain.
5990 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5992 struct sched_domain *child;
5993 struct sched_group *group;
5995 WARN_ON(!sd || !sd->groups);
5997 if (cpu != first_cpu(sd->groups->cpumask))
6002 sd->groups->__cpu_power = 0;
6005 * For perf policy, if the groups in child domain share resources
6006 * (for example cores sharing some portions of the cache hierarchy
6007 * or SMT), then set this domain groups cpu_power such that each group
6008 * can handle only one task, when there are other idle groups in the
6009 * same sched domain.
6011 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6013 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6014 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6019 * add cpu_power of each child group to this groups cpu_power
6021 group = child->groups;
6023 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6024 group = group->next;
6025 } while (group != child->groups);
6029 * Build sched domains for a given set of cpus and attach the sched domains
6030 * to the individual cpus
6032 static int build_sched_domains(const cpumask_t *cpu_map)
6036 struct sched_group **sched_group_nodes = NULL;
6037 int sd_allnodes = 0;
6040 * Allocate the per-node list of sched groups
6042 sched_group_nodes = kzalloc(sizeof(struct sched_group *)*MAX_NUMNODES,
6044 if (!sched_group_nodes) {
6045 printk(KERN_WARNING "Can not alloc sched group node list\n");
6048 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6052 * Set up domains for cpus specified by the cpu_map.
6054 for_each_cpu_mask(i, *cpu_map) {
6055 struct sched_domain *sd = NULL, *p;
6056 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6058 cpus_and(nodemask, nodemask, *cpu_map);
6061 if (cpus_weight(*cpu_map) >
6062 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6063 sd = &per_cpu(allnodes_domains, i);
6064 *sd = SD_ALLNODES_INIT;
6065 sd->span = *cpu_map;
6066 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6072 sd = &per_cpu(node_domains, i);
6074 sd->span = sched_domain_node_span(cpu_to_node(i));
6078 cpus_and(sd->span, sd->span, *cpu_map);
6082 sd = &per_cpu(phys_domains, i);
6084 sd->span = nodemask;
6088 cpu_to_phys_group(i, cpu_map, &sd->groups);
6090 #ifdef CONFIG_SCHED_MC
6092 sd = &per_cpu(core_domains, i);
6094 sd->span = cpu_coregroup_map(i);
6095 cpus_and(sd->span, sd->span, *cpu_map);
6098 cpu_to_core_group(i, cpu_map, &sd->groups);
6101 #ifdef CONFIG_SCHED_SMT
6103 sd = &per_cpu(cpu_domains, i);
6104 *sd = SD_SIBLING_INIT;
6105 sd->span = cpu_sibling_map[i];
6106 cpus_and(sd->span, sd->span, *cpu_map);
6109 cpu_to_cpu_group(i, cpu_map, &sd->groups);
6113 #ifdef CONFIG_SCHED_SMT
6114 /* Set up CPU (sibling) groups */
6115 for_each_cpu_mask(i, *cpu_map) {
6116 cpumask_t this_sibling_map = cpu_sibling_map[i];
6117 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6118 if (i != first_cpu(this_sibling_map))
6121 init_sched_build_groups(this_sibling_map, cpu_map,
6126 #ifdef CONFIG_SCHED_MC
6127 /* Set up multi-core groups */
6128 for_each_cpu_mask(i, *cpu_map) {
6129 cpumask_t this_core_map = cpu_coregroup_map(i);
6130 cpus_and(this_core_map, this_core_map, *cpu_map);
6131 if (i != first_cpu(this_core_map))
6133 init_sched_build_groups(this_core_map, cpu_map,
6134 &cpu_to_core_group);
6138 /* Set up physical groups */
6139 for (i = 0; i < MAX_NUMNODES; i++) {
6140 cpumask_t nodemask = node_to_cpumask(i);
6142 cpus_and(nodemask, nodemask, *cpu_map);
6143 if (cpus_empty(nodemask))
6146 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6150 /* Set up node groups */
6152 init_sched_build_groups(*cpu_map, cpu_map,
6153 &cpu_to_allnodes_group);
6155 for (i = 0; i < MAX_NUMNODES; i++) {
6156 /* Set up node groups */
6157 struct sched_group *sg, *prev;
6158 cpumask_t nodemask = node_to_cpumask(i);
6159 cpumask_t domainspan;
6160 cpumask_t covered = CPU_MASK_NONE;
6163 cpus_and(nodemask, nodemask, *cpu_map);
6164 if (cpus_empty(nodemask)) {
6165 sched_group_nodes[i] = NULL;
6169 domainspan = sched_domain_node_span(i);
6170 cpus_and(domainspan, domainspan, *cpu_map);
6172 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6174 printk(KERN_WARNING "Can not alloc domain group for "
6178 sched_group_nodes[i] = sg;
6179 for_each_cpu_mask(j, nodemask) {
6180 struct sched_domain *sd;
6182 sd = &per_cpu(node_domains, j);
6185 sg->__cpu_power = 0;
6186 sg->cpumask = nodemask;
6188 cpus_or(covered, covered, nodemask);
6191 for (j = 0; j < MAX_NUMNODES; j++) {
6192 cpumask_t tmp, notcovered;
6193 int n = (i + j) % MAX_NUMNODES;
6195 cpus_complement(notcovered, covered);
6196 cpus_and(tmp, notcovered, *cpu_map);
6197 cpus_and(tmp, tmp, domainspan);
6198 if (cpus_empty(tmp))
6201 nodemask = node_to_cpumask(n);
6202 cpus_and(tmp, tmp, nodemask);
6203 if (cpus_empty(tmp))
6206 sg = kmalloc_node(sizeof(struct sched_group),
6210 "Can not alloc domain group for node %d\n", j);
6213 sg->__cpu_power = 0;
6215 sg->next = prev->next;
6216 cpus_or(covered, covered, tmp);
6223 /* Calculate CPU power for physical packages and nodes */
6224 #ifdef CONFIG_SCHED_SMT
6225 for_each_cpu_mask(i, *cpu_map) {
6226 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6228 init_sched_groups_power(i, sd);
6231 #ifdef CONFIG_SCHED_MC
6232 for_each_cpu_mask(i, *cpu_map) {
6233 struct sched_domain *sd = &per_cpu(core_domains, i);
6235 init_sched_groups_power(i, sd);
6239 for_each_cpu_mask(i, *cpu_map) {
6240 struct sched_domain *sd = &per_cpu(phys_domains, i);
6242 init_sched_groups_power(i, sd);
6246 for (i = 0; i < MAX_NUMNODES; i++)
6247 init_numa_sched_groups_power(sched_group_nodes[i]);
6250 struct sched_group *sg;
6252 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6253 init_numa_sched_groups_power(sg);
6257 /* Attach the domains */
6258 for_each_cpu_mask(i, *cpu_map) {
6259 struct sched_domain *sd;
6260 #ifdef CONFIG_SCHED_SMT
6261 sd = &per_cpu(cpu_domains, i);
6262 #elif defined(CONFIG_SCHED_MC)
6263 sd = &per_cpu(core_domains, i);
6265 sd = &per_cpu(phys_domains, i);
6267 cpu_attach_domain(sd, i);
6274 free_sched_groups(cpu_map);
6279 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6281 static int arch_init_sched_domains(const cpumask_t *cpu_map)
6283 cpumask_t cpu_default_map;
6287 * Setup mask for cpus without special case scheduling requirements.
6288 * For now this just excludes isolated cpus, but could be used to
6289 * exclude other special cases in the future.
6291 cpus_andnot(cpu_default_map, *cpu_map, cpu_isolated_map);
6293 err = build_sched_domains(&cpu_default_map);
6298 static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6300 free_sched_groups(cpu_map);
6304 * Detach sched domains from a group of cpus specified in cpu_map
6305 * These cpus will now be attached to the NULL domain
6307 static void detach_destroy_domains(const cpumask_t *cpu_map)
6311 for_each_cpu_mask(i, *cpu_map)
6312 cpu_attach_domain(NULL, i);
6313 synchronize_sched();
6314 arch_destroy_sched_domains(cpu_map);
6318 * Partition sched domains as specified by the cpumasks below.
6319 * This attaches all cpus from the cpumasks to the NULL domain,
6320 * waits for a RCU quiescent period, recalculates sched
6321 * domain information and then attaches them back to the
6322 * correct sched domains
6323 * Call with hotplug lock held
6325 int partition_sched_domains(cpumask_t *partition1, cpumask_t *partition2)
6327 cpumask_t change_map;
6330 cpus_and(*partition1, *partition1, cpu_online_map);
6331 cpus_and(*partition2, *partition2, cpu_online_map);
6332 cpus_or(change_map, *partition1, *partition2);
6334 /* Detach sched domains from all of the affected cpus */
6335 detach_destroy_domains(&change_map);
6336 if (!cpus_empty(*partition1))
6337 err = build_sched_domains(partition1);
6338 if (!err && !cpus_empty(*partition2))
6339 err = build_sched_domains(partition2);
6344 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6345 static int arch_reinit_sched_domains(void)
6349 mutex_lock(&sched_hotcpu_mutex);
6350 detach_destroy_domains(&cpu_online_map);
6351 err = arch_init_sched_domains(&cpu_online_map);
6352 mutex_unlock(&sched_hotcpu_mutex);
6357 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6361 if (buf[0] != '0' && buf[0] != '1')
6365 sched_smt_power_savings = (buf[0] == '1');
6367 sched_mc_power_savings = (buf[0] == '1');
6369 ret = arch_reinit_sched_domains();
6371 return ret ? ret : count;
6374 #ifdef CONFIG_SCHED_MC
6375 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6377 return sprintf(page, "%u\n", sched_mc_power_savings);
6379 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6380 const char *buf, size_t count)
6382 return sched_power_savings_store(buf, count, 0);
6384 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6385 sched_mc_power_savings_store);
6388 #ifdef CONFIG_SCHED_SMT
6389 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6391 return sprintf(page, "%u\n", sched_smt_power_savings);
6393 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6394 const char *buf, size_t count)
6396 return sched_power_savings_store(buf, count, 1);
6398 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6399 sched_smt_power_savings_store);
6402 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6406 #ifdef CONFIG_SCHED_SMT
6408 err = sysfs_create_file(&cls->kset.kobj,
6409 &attr_sched_smt_power_savings.attr);
6411 #ifdef CONFIG_SCHED_MC
6412 if (!err && mc_capable())
6413 err = sysfs_create_file(&cls->kset.kobj,
6414 &attr_sched_mc_power_savings.attr);
6421 * Force a reinitialization of the sched domains hierarchy. The domains
6422 * and groups cannot be updated in place without racing with the balancing
6423 * code, so we temporarily attach all running cpus to the NULL domain
6424 * which will prevent rebalancing while the sched domains are recalculated.
6426 static int update_sched_domains(struct notifier_block *nfb,
6427 unsigned long action, void *hcpu)
6430 case CPU_UP_PREPARE:
6431 case CPU_UP_PREPARE_FROZEN:
6432 case CPU_DOWN_PREPARE:
6433 case CPU_DOWN_PREPARE_FROZEN:
6434 detach_destroy_domains(&cpu_online_map);
6437 case CPU_UP_CANCELED:
6438 case CPU_UP_CANCELED_FROZEN:
6439 case CPU_DOWN_FAILED:
6440 case CPU_DOWN_FAILED_FROZEN:
6442 case CPU_ONLINE_FROZEN:
6444 case CPU_DEAD_FROZEN:
6446 * Fall through and re-initialise the domains.
6453 /* The hotplug lock is already held by cpu_up/cpu_down */
6454 arch_init_sched_domains(&cpu_online_map);
6459 void __init sched_init_smp(void)
6461 cpumask_t non_isolated_cpus;
6463 mutex_lock(&sched_hotcpu_mutex);
6464 arch_init_sched_domains(&cpu_online_map);
6465 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
6466 if (cpus_empty(non_isolated_cpus))
6467 cpu_set(smp_processor_id(), non_isolated_cpus);
6468 mutex_unlock(&sched_hotcpu_mutex);
6469 /* XXX: Theoretical race here - CPU may be hotplugged now */
6470 hotcpu_notifier(update_sched_domains, 0);
6472 init_sched_domain_sysctl();
6474 /* Move init over to a non-isolated CPU */
6475 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6479 void __init sched_init_smp(void)
6482 #endif /* CONFIG_SMP */
6484 int in_sched_functions(unsigned long addr)
6486 /* Linker adds these: start and end of __sched functions */
6487 extern char __sched_text_start[], __sched_text_end[];
6489 return in_lock_functions(addr) ||
6490 (addr >= (unsigned long)__sched_text_start
6491 && addr < (unsigned long)__sched_text_end);
6494 static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
6496 cfs_rq->tasks_timeline = RB_ROOT;
6497 #ifdef CONFIG_FAIR_GROUP_SCHED
6500 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6503 void __init sched_init(void)
6505 int highest_cpu = 0;
6509 * Link up the scheduling class hierarchy:
6511 rt_sched_class.next = &fair_sched_class;
6512 fair_sched_class.next = &idle_sched_class;
6513 idle_sched_class.next = NULL;
6515 for_each_possible_cpu(i) {
6516 struct rt_prio_array *array;
6520 spin_lock_init(&rq->lock);
6521 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
6524 init_cfs_rq(&rq->cfs, rq);
6525 #ifdef CONFIG_FAIR_GROUP_SCHED
6526 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6528 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6529 struct sched_entity *se =
6530 &per_cpu(init_sched_entity, i);
6532 init_cfs_rq_p[i] = cfs_rq;
6533 init_cfs_rq(cfs_rq, rq);
6534 cfs_rq->tg = &init_task_grp;
6535 list_add(&cfs_rq->leaf_cfs_rq_list,
6536 &rq->leaf_cfs_rq_list);
6538 init_sched_entity_p[i] = se;
6539 se->cfs_rq = &rq->cfs;
6541 se->load.weight = init_task_grp_load;
6542 se->load.inv_weight =
6543 div64_64(1ULL<<32, init_task_grp_load);
6546 init_task_grp.shares = init_task_grp_load;
6549 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6550 rq->cpu_load[j] = 0;
6553 rq->active_balance = 0;
6554 rq->next_balance = jiffies;
6557 rq->migration_thread = NULL;
6558 INIT_LIST_HEAD(&rq->migration_queue);
6560 atomic_set(&rq->nr_iowait, 0);
6562 array = &rq->rt.active;
6563 for (j = 0; j < MAX_RT_PRIO; j++) {
6564 INIT_LIST_HEAD(array->queue + j);
6565 __clear_bit(j, array->bitmap);
6568 /* delimiter for bitsearch: */
6569 __set_bit(MAX_RT_PRIO, array->bitmap);
6572 set_load_weight(&init_task);
6574 #ifdef CONFIG_PREEMPT_NOTIFIERS
6575 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6579 nr_cpu_ids = highest_cpu + 1;
6580 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6583 #ifdef CONFIG_RT_MUTEXES
6584 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6588 * The boot idle thread does lazy MMU switching as well:
6590 atomic_inc(&init_mm.mm_count);
6591 enter_lazy_tlb(&init_mm, current);
6594 * Make us the idle thread. Technically, schedule() should not be
6595 * called from this thread, however somewhere below it might be,
6596 * but because we are the idle thread, we just pick up running again
6597 * when this runqueue becomes "idle".
6599 init_idle(current, smp_processor_id());
6601 * During early bootup we pretend to be a normal task:
6603 current->sched_class = &fair_sched_class;
6606 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6607 void __might_sleep(char *file, int line)
6610 static unsigned long prev_jiffy; /* ratelimiting */
6612 if ((in_atomic() || irqs_disabled()) &&
6613 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6614 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6616 prev_jiffy = jiffies;
6617 printk(KERN_ERR "BUG: sleeping function called from invalid"
6618 " context at %s:%d\n", file, line);
6619 printk("in_atomic():%d, irqs_disabled():%d\n",
6620 in_atomic(), irqs_disabled());
6621 debug_show_held_locks(current);
6622 if (irqs_disabled())
6623 print_irqtrace_events(current);
6628 EXPORT_SYMBOL(__might_sleep);
6631 #ifdef CONFIG_MAGIC_SYSRQ
6632 void normalize_rt_tasks(void)
6634 struct task_struct *g, *p;
6635 unsigned long flags;
6639 read_lock_irq(&tasklist_lock);
6640 do_each_thread(g, p) {
6641 p->se.exec_start = 0;
6642 #ifdef CONFIG_SCHEDSTATS
6643 p->se.wait_start = 0;
6644 p->se.sleep_start = 0;
6645 p->se.block_start = 0;
6647 task_rq(p)->clock = 0;
6651 * Renice negative nice level userspace
6654 if (TASK_NICE(p) < 0 && p->mm)
6655 set_user_nice(p, 0);
6659 spin_lock_irqsave(&p->pi_lock, flags);
6660 rq = __task_rq_lock(p);
6663 * Do not touch the migration thread:
6665 if (p == rq->migration_thread)
6669 update_rq_clock(rq);
6670 on_rq = p->se.on_rq;
6672 deactivate_task(rq, p, 0);
6673 __setscheduler(rq, p, SCHED_NORMAL, 0);
6675 activate_task(rq, p, 0);
6676 resched_task(rq->curr);
6681 __task_rq_unlock(rq);
6682 spin_unlock_irqrestore(&p->pi_lock, flags);
6683 } while_each_thread(g, p);
6685 read_unlock_irq(&tasklist_lock);
6688 #endif /* CONFIG_MAGIC_SYSRQ */
6692 * These functions are only useful for the IA64 MCA handling.
6694 * They can only be called when the whole system has been
6695 * stopped - every CPU needs to be quiescent, and no scheduling
6696 * activity can take place. Using them for anything else would
6697 * be a serious bug, and as a result, they aren't even visible
6698 * under any other configuration.
6702 * curr_task - return the current task for a given cpu.
6703 * @cpu: the processor in question.
6705 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6707 struct task_struct *curr_task(int cpu)
6709 return cpu_curr(cpu);
6713 * set_curr_task - set the current task for a given cpu.
6714 * @cpu: the processor in question.
6715 * @p: the task pointer to set.
6717 * Description: This function must only be used when non-maskable interrupts
6718 * are serviced on a separate stack. It allows the architecture to switch the
6719 * notion of the current task on a cpu in a non-blocking manner. This function
6720 * must be called with all CPU's synchronized, and interrupts disabled, the
6721 * and caller must save the original value of the current task (see
6722 * curr_task() above) and restore that value before reenabling interrupts and
6723 * re-starting the system.
6725 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6727 void set_curr_task(int cpu, struct task_struct *p)
6734 #ifdef CONFIG_FAIR_GROUP_SCHED
6736 /* allocate runqueue etc for a new task group */
6737 struct task_grp *sched_create_group(void)
6739 struct task_grp *tg;
6740 struct cfs_rq *cfs_rq;
6741 struct sched_entity *se;
6745 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6747 return ERR_PTR(-ENOMEM);
6749 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
6752 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
6756 for_each_possible_cpu(i) {
6759 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
6764 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
6769 memset(cfs_rq, 0, sizeof(struct cfs_rq));
6770 memset(se, 0, sizeof(struct sched_entity));
6772 tg->cfs_rq[i] = cfs_rq;
6773 init_cfs_rq(cfs_rq, rq);
6777 se->cfs_rq = &rq->cfs;
6779 se->load.weight = NICE_0_LOAD;
6780 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
6784 for_each_possible_cpu(i) {
6786 cfs_rq = tg->cfs_rq[i];
6787 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
6790 tg->shares = NICE_0_LOAD;
6795 for_each_possible_cpu(i) {
6796 if (tg->cfs_rq && tg->cfs_rq[i])
6797 kfree(tg->cfs_rq[i]);
6798 if (tg->se && tg->se[i])
6808 return ERR_PTR(-ENOMEM);
6811 /* rcu callback to free various structures associated with a task group */
6812 static void free_sched_group(struct rcu_head *rhp)
6814 struct cfs_rq *cfs_rq = container_of(rhp, struct cfs_rq, rcu);
6815 struct task_grp *tg = cfs_rq->tg;
6816 struct sched_entity *se;
6819 /* now it should be safe to free those cfs_rqs */
6820 for_each_possible_cpu(i) {
6821 cfs_rq = tg->cfs_rq[i];
6833 /* Destroy runqueue etc associated with a task group */
6834 void sched_destroy_group(struct task_grp *tg)
6836 struct cfs_rq *cfs_rq;
6839 for_each_possible_cpu(i) {
6840 cfs_rq = tg->cfs_rq[i];
6841 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
6844 cfs_rq = tg->cfs_rq[0];
6846 /* wait for possible concurrent references to cfs_rqs complete */
6847 call_rcu(&cfs_rq->rcu, free_sched_group);
6850 /* change task's runqueue when it moves between groups.
6851 * The caller of this function should have put the task in its new group
6852 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6853 * reflect its new group.
6855 void sched_move_task(struct task_struct *tsk)
6858 unsigned long flags;
6861 rq = task_rq_lock(tsk, &flags);
6863 if (tsk->sched_class != &fair_sched_class)
6866 update_rq_clock(rq);
6868 running = task_running(rq, tsk);
6869 on_rq = tsk->se.on_rq;
6872 dequeue_task(rq, tsk, 0);
6873 if (unlikely(running))
6874 tsk->sched_class->put_prev_task(rq, tsk);
6877 set_task_cfs_rq(tsk);
6880 if (unlikely(running))
6881 tsk->sched_class->set_curr_task(rq);
6882 enqueue_task(rq, tsk, 0);
6886 task_rq_unlock(rq, &flags);
6889 static void set_se_shares(struct sched_entity *se, unsigned long shares)
6891 struct cfs_rq *cfs_rq = se->cfs_rq;
6892 struct rq *rq = cfs_rq->rq;
6895 spin_lock_irq(&rq->lock);
6899 dequeue_entity(cfs_rq, se, 0);
6901 se->load.weight = shares;
6902 se->load.inv_weight = div64_64((1ULL<<32), shares);
6905 enqueue_entity(cfs_rq, se, 0);
6907 spin_unlock_irq(&rq->lock);
6910 int sched_group_set_shares(struct task_grp *tg, unsigned long shares)
6914 if (tg->shares == shares)
6917 /* return -EINVAL if the new value is not sane */
6919 tg->shares = shares;
6920 for_each_possible_cpu(i)
6921 set_se_shares(tg->se[i], shares);
6926 #endif /* CONFIG_FAIR_GROUP_SCHED */