4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
28 #include <linux/module.h>
29 #include <linux/nmi.h>
30 #include <linux/init.h>
31 #include <linux/uaccess.h>
32 #include <linux/highmem.h>
33 #include <linux/smp_lock.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/security.h>
41 #include <linux/notifier.h>
42 #include <linux/profile.h>
43 #include <linux/freezer.h>
44 #include <linux/vmalloc.h>
45 #include <linux/blkdev.h>
46 #include <linux/delay.h>
47 #include <linux/pid_namespace.h>
48 #include <linux/smp.h>
49 #include <linux/threads.h>
50 #include <linux/timer.h>
51 #include <linux/rcupdate.h>
52 #include <linux/cpu.h>
53 #include <linux/cpuset.h>
54 #include <linux/percpu.h>
55 #include <linux/kthread.h>
56 #include <linux/seq_file.h>
57 #include <linux/sysctl.h>
58 #include <linux/syscalls.h>
59 #include <linux/times.h>
60 #include <linux/tsacct_kern.h>
61 #include <linux/kprobes.h>
62 #include <linux/delayacct.h>
63 #include <linux/reciprocal_div.h>
64 #include <linux/unistd.h>
65 #include <linux/pagemap.h>
68 #include <asm/irq_regs.h>
71 * Scheduler clock - returns current time in nanosec units.
72 * This is default implementation.
73 * Architectures and sub-architectures can override this.
75 unsigned long long __attribute__((weak
)) sched_clock(void)
77 return (unsigned long long)jiffies
* (NSEC_PER_SEC
/ HZ
);
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Some helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define JIFFIES_TO_NS(TIME) ((TIME) * (NSEC_PER_SEC / HZ))
104 #define NICE_0_LOAD SCHED_LOAD_SCALE
105 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
108 * These are the 'tuning knobs' of the scheduler:
110 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
111 * Timeslices get refilled after they expire.
113 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
122 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
126 * Each time a sched group cpu_power is changed,
127 * we must compute its reciprocal value
129 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
131 sg
->__cpu_power
+= val
;
132 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
136 static inline int rt_policy(int policy
)
138 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
143 static inline int task_has_rt_policy(struct task_struct
*p
)
145 return rt_policy(p
->policy
);
149 * This is the priority-queue data structure of the RT scheduling class:
151 struct rt_prio_array
{
152 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
153 struct list_head queue
[MAX_RT_PRIO
];
156 #ifdef CONFIG_FAIR_GROUP_SCHED
158 #include <linux/cgroup.h>
162 /* task group related information */
164 #ifdef CONFIG_FAIR_CGROUP_SCHED
165 struct cgroup_subsys_state css
;
167 /* schedulable entities of this group on each cpu */
168 struct sched_entity
**se
;
169 /* runqueue "owned" by this group on each cpu */
170 struct cfs_rq
**cfs_rq
;
173 * shares assigned to a task group governs how much of cpu bandwidth
174 * is allocated to the group. The more shares a group has, the more is
175 * the cpu bandwidth allocated to it.
177 * For ex, lets say that there are three task groups, A, B and C which
178 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
179 * cpu bandwidth allocated by the scheduler to task groups A, B and C
182 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
183 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
184 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
186 * The weight assigned to a task group's schedulable entities on every
187 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
188 * group's shares. For ex: lets say that task group A has been
189 * assigned shares of 1000 and there are two CPUs in a system. Then,
191 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
193 * Note: It's not necessary that each of a task's group schedulable
194 * entity have the same weight on all CPUs. If the group
195 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
196 * better distribution of weight could be:
198 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
199 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
201 * rebalance_shares() is responsible for distributing the shares of a
202 * task groups like this among the group's schedulable entities across
206 unsigned long shares
;
211 /* Default task group's sched entity on each cpu */
212 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
213 /* Default task group's cfs_rq on each cpu */
214 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
216 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
217 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
219 /* task_group_mutex serializes add/remove of task groups and also changes to
220 * a task group's cpu shares.
222 static DEFINE_MUTEX(task_group_mutex
);
224 /* doms_cur_mutex serializes access to doms_cur[] array */
225 static DEFINE_MUTEX(doms_cur_mutex
);
228 /* kernel thread that runs rebalance_shares() periodically */
229 static struct task_struct
*lb_monitor_task
;
230 static int load_balance_monitor(void *unused
);
233 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
);
235 /* Default task group.
236 * Every task in system belong to this group at bootup.
238 struct task_group init_task_group
= {
239 .se
= init_sched_entity_p
,
240 .cfs_rq
= init_cfs_rq_p
,
243 #ifdef CONFIG_FAIR_USER_SCHED
244 # define INIT_TASK_GROUP_LOAD 2*NICE_0_LOAD
246 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
249 #define MIN_GROUP_SHARES 2
251 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
253 /* return group to which a task belongs */
254 static inline struct task_group
*task_group(struct task_struct
*p
)
256 struct task_group
*tg
;
258 #ifdef CONFIG_FAIR_USER_SCHED
260 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
261 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
262 struct task_group
, css
);
264 tg
= &init_task_group
;
269 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
270 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
)
272 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
273 p
->se
.parent
= task_group(p
)->se
[cpu
];
276 static inline void lock_task_group_list(void)
278 mutex_lock(&task_group_mutex
);
281 static inline void unlock_task_group_list(void)
283 mutex_unlock(&task_group_mutex
);
286 static inline void lock_doms_cur(void)
288 mutex_lock(&doms_cur_mutex
);
291 static inline void unlock_doms_cur(void)
293 mutex_unlock(&doms_cur_mutex
);
298 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
) { }
299 static inline void lock_task_group_list(void) { }
300 static inline void unlock_task_group_list(void) { }
301 static inline void lock_doms_cur(void) { }
302 static inline void unlock_doms_cur(void) { }
304 #endif /* CONFIG_FAIR_GROUP_SCHED */
306 /* CFS-related fields in a runqueue */
308 struct load_weight load
;
309 unsigned long nr_running
;
314 struct rb_root tasks_timeline
;
315 struct rb_node
*rb_leftmost
;
316 struct rb_node
*rb_load_balance_curr
;
317 /* 'curr' points to currently running entity on this cfs_rq.
318 * It is set to NULL otherwise (i.e when none are currently running).
320 struct sched_entity
*curr
;
322 unsigned long nr_spread_over
;
324 #ifdef CONFIG_FAIR_GROUP_SCHED
325 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
328 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
329 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
330 * (like users, containers etc.)
332 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
333 * list is used during load balance.
335 struct list_head leaf_cfs_rq_list
;
336 struct task_group
*tg
; /* group that "owns" this runqueue */
340 /* Real-Time classes' related field in a runqueue: */
342 struct rt_prio_array active
;
343 int rt_load_balance_idx
;
344 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
345 unsigned long rt_nr_running
;
346 /* highest queued rt task prio */
351 * This is the main, per-CPU runqueue data structure.
353 * Locking rule: those places that want to lock multiple runqueues
354 * (such as the load balancing or the thread migration code), lock
355 * acquire operations must be ordered by ascending &runqueue.
362 * nr_running and cpu_load should be in the same cacheline because
363 * remote CPUs use both these fields when doing load calculation.
365 unsigned long nr_running
;
366 #define CPU_LOAD_IDX_MAX 5
367 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
368 unsigned char idle_at_tick
;
370 unsigned char in_nohz_recently
;
372 /* capture load from *all* tasks on this cpu: */
373 struct load_weight load
;
374 unsigned long nr_load_updates
;
378 #ifdef CONFIG_FAIR_GROUP_SCHED
379 /* list of leaf cfs_rq on this cpu: */
380 struct list_head leaf_cfs_rq_list
;
385 * This is part of a global counter where only the total sum
386 * over all CPUs matters. A task can increase this counter on
387 * one CPU and if it got migrated afterwards it may decrease
388 * it on another CPU. Always updated under the runqueue lock:
390 unsigned long nr_uninterruptible
;
392 struct task_struct
*curr
, *idle
;
393 unsigned long next_balance
;
394 struct mm_struct
*prev_mm
;
396 u64 clock
, prev_clock_raw
;
399 unsigned int clock_warps
, clock_overflows
;
401 unsigned int clock_deep_idle_events
;
407 struct sched_domain
*sd
;
409 /* For active balancing */
412 /* cpu of this runqueue: */
415 struct task_struct
*migration_thread
;
416 struct list_head migration_queue
;
419 #ifdef CONFIG_SCHEDSTATS
421 struct sched_info rq_sched_info
;
423 /* sys_sched_yield() stats */
424 unsigned int yld_exp_empty
;
425 unsigned int yld_act_empty
;
426 unsigned int yld_both_empty
;
427 unsigned int yld_count
;
429 /* schedule() stats */
430 unsigned int sched_switch
;
431 unsigned int sched_count
;
432 unsigned int sched_goidle
;
434 /* try_to_wake_up() stats */
435 unsigned int ttwu_count
;
436 unsigned int ttwu_local
;
439 unsigned int bkl_count
;
441 struct lock_class_key rq_lock_key
;
444 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
446 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
448 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
451 static inline int cpu_of(struct rq
*rq
)
461 * Update the per-runqueue clock, as finegrained as the platform can give
462 * us, but without assuming monotonicity, etc.:
464 static void __update_rq_clock(struct rq
*rq
)
466 u64 prev_raw
= rq
->prev_clock_raw
;
467 u64 now
= sched_clock();
468 s64 delta
= now
- prev_raw
;
469 u64 clock
= rq
->clock
;
471 #ifdef CONFIG_SCHED_DEBUG
472 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
475 * Protect against sched_clock() occasionally going backwards:
477 if (unlikely(delta
< 0)) {
482 * Catch too large forward jumps too:
484 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
485 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
486 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
489 rq
->clock_overflows
++;
491 if (unlikely(delta
> rq
->clock_max_delta
))
492 rq
->clock_max_delta
= delta
;
497 rq
->prev_clock_raw
= now
;
501 static void update_rq_clock(struct rq
*rq
)
503 if (likely(smp_processor_id() == cpu_of(rq
)))
504 __update_rq_clock(rq
);
508 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
509 * See detach_destroy_domains: synchronize_sched for details.
511 * The domain tree of any CPU may only be accessed from within
512 * preempt-disabled sections.
514 #define for_each_domain(cpu, __sd) \
515 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
517 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
518 #define this_rq() (&__get_cpu_var(runqueues))
519 #define task_rq(p) cpu_rq(task_cpu(p))
520 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
523 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
525 #ifdef CONFIG_SCHED_DEBUG
526 # define const_debug __read_mostly
528 # define const_debug static const
532 * Debugging: various feature bits
535 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
536 SCHED_FEAT_WAKEUP_PREEMPT
= 2,
537 SCHED_FEAT_START_DEBIT
= 4,
538 SCHED_FEAT_TREE_AVG
= 8,
539 SCHED_FEAT_APPROX_AVG
= 16,
542 const_debug
unsigned int sysctl_sched_features
=
543 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
544 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
545 SCHED_FEAT_START_DEBIT
* 1 |
546 SCHED_FEAT_TREE_AVG
* 0 |
547 SCHED_FEAT_APPROX_AVG
* 0;
549 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
552 * Number of tasks to iterate in a single balance run.
553 * Limited because this is done with IRQs disabled.
555 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
558 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
559 * clock constructed from sched_clock():
561 unsigned long long cpu_clock(int cpu
)
563 unsigned long long now
;
567 local_irq_save(flags
);
570 * Only call sched_clock() if the scheduler has already been
571 * initialized (some code might call cpu_clock() very early):
576 local_irq_restore(flags
);
580 EXPORT_SYMBOL_GPL(cpu_clock
);
582 #ifndef prepare_arch_switch
583 # define prepare_arch_switch(next) do { } while (0)
585 #ifndef finish_arch_switch
586 # define finish_arch_switch(prev) do { } while (0)
589 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
591 return rq
->curr
== p
;
594 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
595 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
597 return task_current(rq
, p
);
600 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
604 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
606 #ifdef CONFIG_DEBUG_SPINLOCK
607 /* this is a valid case when another task releases the spinlock */
608 rq
->lock
.owner
= current
;
611 * If we are tracking spinlock dependencies then we have to
612 * fix up the runqueue lock - which gets 'carried over' from
615 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
617 spin_unlock_irq(&rq
->lock
);
620 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
621 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
626 return task_current(rq
, p
);
630 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
634 * We can optimise this out completely for !SMP, because the
635 * SMP rebalancing from interrupt is the only thing that cares
640 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
641 spin_unlock_irq(&rq
->lock
);
643 spin_unlock(&rq
->lock
);
647 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
651 * After ->oncpu is cleared, the task can be moved to a different CPU.
652 * We must ensure this doesn't happen until the switch is completely
658 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
662 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
665 * __task_rq_lock - lock the runqueue a given task resides on.
666 * Must be called interrupts disabled.
668 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
672 struct rq
*rq
= task_rq(p
);
673 spin_lock(&rq
->lock
);
674 if (likely(rq
== task_rq(p
)))
676 spin_unlock(&rq
->lock
);
681 * task_rq_lock - lock the runqueue a given task resides on and disable
682 * interrupts. Note the ordering: we can safely lookup the task_rq without
683 * explicitly disabling preemption.
685 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
691 local_irq_save(*flags
);
693 spin_lock(&rq
->lock
);
694 if (likely(rq
== task_rq(p
)))
696 spin_unlock_irqrestore(&rq
->lock
, *flags
);
700 static void __task_rq_unlock(struct rq
*rq
)
703 spin_unlock(&rq
->lock
);
706 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
709 spin_unlock_irqrestore(&rq
->lock
, *flags
);
713 * this_rq_lock - lock this runqueue and disable interrupts.
715 static struct rq
*this_rq_lock(void)
722 spin_lock(&rq
->lock
);
728 * We are going deep-idle (irqs are disabled):
730 void sched_clock_idle_sleep_event(void)
732 struct rq
*rq
= cpu_rq(smp_processor_id());
734 spin_lock(&rq
->lock
);
735 __update_rq_clock(rq
);
736 spin_unlock(&rq
->lock
);
737 rq
->clock_deep_idle_events
++;
739 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
742 * We just idled delta nanoseconds (called with irqs disabled):
744 void sched_clock_idle_wakeup_event(u64 delta_ns
)
746 struct rq
*rq
= cpu_rq(smp_processor_id());
747 u64 now
= sched_clock();
749 touch_softlockup_watchdog();
750 rq
->idle_clock
+= delta_ns
;
752 * Override the previous timestamp and ignore all
753 * sched_clock() deltas that occured while we idled,
754 * and use the PM-provided delta_ns to advance the
757 spin_lock(&rq
->lock
);
758 rq
->prev_clock_raw
= now
;
759 rq
->clock
+= delta_ns
;
760 spin_unlock(&rq
->lock
);
762 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
765 * resched_task - mark a task 'to be rescheduled now'.
767 * On UP this means the setting of the need_resched flag, on SMP it
768 * might also involve a cross-CPU call to trigger the scheduler on
773 #ifndef tsk_is_polling
774 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
777 static void resched_task(struct task_struct
*p
)
781 assert_spin_locked(&task_rq(p
)->lock
);
783 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
786 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
789 if (cpu
== smp_processor_id())
792 /* NEED_RESCHED must be visible before we test polling */
794 if (!tsk_is_polling(p
))
795 smp_send_reschedule(cpu
);
798 static void resched_cpu(int cpu
)
800 struct rq
*rq
= cpu_rq(cpu
);
803 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
805 resched_task(cpu_curr(cpu
));
806 spin_unlock_irqrestore(&rq
->lock
, flags
);
809 static inline void resched_task(struct task_struct
*p
)
811 assert_spin_locked(&task_rq(p
)->lock
);
812 set_tsk_need_resched(p
);
816 #if BITS_PER_LONG == 32
817 # define WMULT_CONST (~0UL)
819 # define WMULT_CONST (1UL << 32)
822 #define WMULT_SHIFT 32
825 * Shift right and round:
827 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
830 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
831 struct load_weight
*lw
)
835 if (unlikely(!lw
->inv_weight
))
836 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
838 tmp
= (u64
)delta_exec
* weight
;
840 * Check whether we'd overflow the 64-bit multiplication:
842 if (unlikely(tmp
> WMULT_CONST
))
843 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
846 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
848 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
851 static inline unsigned long
852 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
854 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
857 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
862 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
868 * To aid in avoiding the subversion of "niceness" due to uneven distribution
869 * of tasks with abnormal "nice" values across CPUs the contribution that
870 * each task makes to its run queue's load is weighted according to its
871 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
872 * scaled version of the new time slice allocation that they receive on time
876 #define WEIGHT_IDLEPRIO 2
877 #define WMULT_IDLEPRIO (1 << 31)
880 * Nice levels are multiplicative, with a gentle 10% change for every
881 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
882 * nice 1, it will get ~10% less CPU time than another CPU-bound task
883 * that remained on nice 0.
885 * The "10% effect" is relative and cumulative: from _any_ nice level,
886 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
887 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
888 * If a task goes up by ~10% and another task goes down by ~10% then
889 * the relative distance between them is ~25%.)
891 static const int prio_to_weight
[40] = {
892 /* -20 */ 88761, 71755, 56483, 46273, 36291,
893 /* -15 */ 29154, 23254, 18705, 14949, 11916,
894 /* -10 */ 9548, 7620, 6100, 4904, 3906,
895 /* -5 */ 3121, 2501, 1991, 1586, 1277,
896 /* 0 */ 1024, 820, 655, 526, 423,
897 /* 5 */ 335, 272, 215, 172, 137,
898 /* 10 */ 110, 87, 70, 56, 45,
899 /* 15 */ 36, 29, 23, 18, 15,
903 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
905 * In cases where the weight does not change often, we can use the
906 * precalculated inverse to speed up arithmetics by turning divisions
907 * into multiplications:
909 static const u32 prio_to_wmult
[40] = {
910 /* -20 */ 48388, 59856, 76040, 92818, 118348,
911 /* -15 */ 147320, 184698, 229616, 287308, 360437,
912 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
913 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
914 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
915 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
916 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
917 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
920 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
923 * runqueue iterator, to support SMP load-balancing between different
924 * scheduling classes, without having to expose their internal data
925 * structures to the load-balancing proper:
929 struct task_struct
*(*start
)(void *);
930 struct task_struct
*(*next
)(void *);
935 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
936 unsigned long max_load_move
, struct sched_domain
*sd
,
937 enum cpu_idle_type idle
, int *all_pinned
,
938 int *this_best_prio
, struct rq_iterator
*iterator
);
941 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
942 struct sched_domain
*sd
, enum cpu_idle_type idle
,
943 struct rq_iterator
*iterator
);
946 #ifdef CONFIG_CGROUP_CPUACCT
947 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
949 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
952 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
954 update_load_add(&rq
->load
, load
);
957 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
959 update_load_sub(&rq
->load
, load
);
962 #include "sched_stats.h"
963 #include "sched_idletask.c"
964 #include "sched_fair.c"
965 #include "sched_rt.c"
966 #ifdef CONFIG_SCHED_DEBUG
967 # include "sched_debug.c"
970 #define sched_class_highest (&rt_sched_class)
972 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
977 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
982 static void set_load_weight(struct task_struct
*p
)
984 if (task_has_rt_policy(p
)) {
985 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
986 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
991 * SCHED_IDLE tasks get minimal weight:
993 if (p
->policy
== SCHED_IDLE
) {
994 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
995 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
999 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1000 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1003 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1005 sched_info_queued(p
);
1006 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1010 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1012 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1017 * __normal_prio - return the priority that is based on the static prio
1019 static inline int __normal_prio(struct task_struct
*p
)
1021 return p
->static_prio
;
1025 * Calculate the expected normal priority: i.e. priority
1026 * without taking RT-inheritance into account. Might be
1027 * boosted by interactivity modifiers. Changes upon fork,
1028 * setprio syscalls, and whenever the interactivity
1029 * estimator recalculates.
1031 static inline int normal_prio(struct task_struct
*p
)
1035 if (task_has_rt_policy(p
))
1036 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1038 prio
= __normal_prio(p
);
1043 * Calculate the current priority, i.e. the priority
1044 * taken into account by the scheduler. This value might
1045 * be boosted by RT tasks, or might be boosted by
1046 * interactivity modifiers. Will be RT if the task got
1047 * RT-boosted. If not then it returns p->normal_prio.
1049 static int effective_prio(struct task_struct
*p
)
1051 p
->normal_prio
= normal_prio(p
);
1053 * If we are RT tasks or we were boosted to RT priority,
1054 * keep the priority unchanged. Otherwise, update priority
1055 * to the normal priority:
1057 if (!rt_prio(p
->prio
))
1058 return p
->normal_prio
;
1063 * activate_task - move a task to the runqueue.
1065 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1067 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1068 rq
->nr_uninterruptible
--;
1070 enqueue_task(rq
, p
, wakeup
);
1071 inc_nr_running(p
, rq
);
1075 * deactivate_task - remove a task from the runqueue.
1077 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1079 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1080 rq
->nr_uninterruptible
++;
1082 dequeue_task(rq
, p
, sleep
);
1083 dec_nr_running(p
, rq
);
1087 * task_curr - is this task currently executing on a CPU?
1088 * @p: the task in question.
1090 inline int task_curr(const struct task_struct
*p
)
1092 return cpu_curr(task_cpu(p
)) == p
;
1095 /* Used instead of source_load when we know the type == 0 */
1096 unsigned long weighted_cpuload(const int cpu
)
1098 return cpu_rq(cpu
)->load
.weight
;
1101 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1103 set_task_cfs_rq(p
, cpu
);
1106 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1107 * successfuly executed on another CPU. We must ensure that updates of
1108 * per-task data have been completed by this moment.
1111 task_thread_info(p
)->cpu
= cpu
;
1118 * Is this task likely cache-hot:
1121 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1125 if (p
->sched_class
!= &fair_sched_class
)
1128 if (sysctl_sched_migration_cost
== -1)
1130 if (sysctl_sched_migration_cost
== 0)
1133 delta
= now
- p
->se
.exec_start
;
1135 return delta
< (s64
)sysctl_sched_migration_cost
;
1139 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1141 int old_cpu
= task_cpu(p
);
1142 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1143 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1144 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1147 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1149 #ifdef CONFIG_SCHEDSTATS
1150 if (p
->se
.wait_start
)
1151 p
->se
.wait_start
-= clock_offset
;
1152 if (p
->se
.sleep_start
)
1153 p
->se
.sleep_start
-= clock_offset
;
1154 if (p
->se
.block_start
)
1155 p
->se
.block_start
-= clock_offset
;
1156 if (old_cpu
!= new_cpu
) {
1157 schedstat_inc(p
, se
.nr_migrations
);
1158 if (task_hot(p
, old_rq
->clock
, NULL
))
1159 schedstat_inc(p
, se
.nr_forced2_migrations
);
1162 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1163 new_cfsrq
->min_vruntime
;
1165 __set_task_cpu(p
, new_cpu
);
1168 struct migration_req
{
1169 struct list_head list
;
1171 struct task_struct
*task
;
1174 struct completion done
;
1178 * The task's runqueue lock must be held.
1179 * Returns true if you have to wait for migration thread.
1182 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1184 struct rq
*rq
= task_rq(p
);
1187 * If the task is not on a runqueue (and not running), then
1188 * it is sufficient to simply update the task's cpu field.
1190 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1191 set_task_cpu(p
, dest_cpu
);
1195 init_completion(&req
->done
);
1197 req
->dest_cpu
= dest_cpu
;
1198 list_add(&req
->list
, &rq
->migration_queue
);
1204 * wait_task_inactive - wait for a thread to unschedule.
1206 * The caller must ensure that the task *will* unschedule sometime soon,
1207 * else this function might spin for a *long* time. This function can't
1208 * be called with interrupts off, or it may introduce deadlock with
1209 * smp_call_function() if an IPI is sent by the same process we are
1210 * waiting to become inactive.
1212 void wait_task_inactive(struct task_struct
*p
)
1214 unsigned long flags
;
1220 * We do the initial early heuristics without holding
1221 * any task-queue locks at all. We'll only try to get
1222 * the runqueue lock when things look like they will
1228 * If the task is actively running on another CPU
1229 * still, just relax and busy-wait without holding
1232 * NOTE! Since we don't hold any locks, it's not
1233 * even sure that "rq" stays as the right runqueue!
1234 * But we don't care, since "task_running()" will
1235 * return false if the runqueue has changed and p
1236 * is actually now running somewhere else!
1238 while (task_running(rq
, p
))
1242 * Ok, time to look more closely! We need the rq
1243 * lock now, to be *sure*. If we're wrong, we'll
1244 * just go back and repeat.
1246 rq
= task_rq_lock(p
, &flags
);
1247 running
= task_running(rq
, p
);
1248 on_rq
= p
->se
.on_rq
;
1249 task_rq_unlock(rq
, &flags
);
1252 * Was it really running after all now that we
1253 * checked with the proper locks actually held?
1255 * Oops. Go back and try again..
1257 if (unlikely(running
)) {
1263 * It's not enough that it's not actively running,
1264 * it must be off the runqueue _entirely_, and not
1267 * So if it wa still runnable (but just not actively
1268 * running right now), it's preempted, and we should
1269 * yield - it could be a while.
1271 if (unlikely(on_rq
)) {
1272 schedule_timeout_uninterruptible(1);
1277 * Ahh, all good. It wasn't running, and it wasn't
1278 * runnable, which means that it will never become
1279 * running in the future either. We're all done!
1286 * kick_process - kick a running thread to enter/exit the kernel
1287 * @p: the to-be-kicked thread
1289 * Cause a process which is running on another CPU to enter
1290 * kernel-mode, without any delay. (to get signals handled.)
1292 * NOTE: this function doesnt have to take the runqueue lock,
1293 * because all it wants to ensure is that the remote task enters
1294 * the kernel. If the IPI races and the task has been migrated
1295 * to another CPU then no harm is done and the purpose has been
1298 void kick_process(struct task_struct
*p
)
1304 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1305 smp_send_reschedule(cpu
);
1310 * Return a low guess at the load of a migration-source cpu weighted
1311 * according to the scheduling class and "nice" value.
1313 * We want to under-estimate the load of migration sources, to
1314 * balance conservatively.
1316 static unsigned long source_load(int cpu
, int type
)
1318 struct rq
*rq
= cpu_rq(cpu
);
1319 unsigned long total
= weighted_cpuload(cpu
);
1324 return min(rq
->cpu_load
[type
-1], total
);
1328 * Return a high guess at the load of a migration-target cpu weighted
1329 * according to the scheduling class and "nice" value.
1331 static unsigned long target_load(int cpu
, int type
)
1333 struct rq
*rq
= cpu_rq(cpu
);
1334 unsigned long total
= weighted_cpuload(cpu
);
1339 return max(rq
->cpu_load
[type
-1], total
);
1343 * Return the average load per task on the cpu's run queue
1345 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1347 struct rq
*rq
= cpu_rq(cpu
);
1348 unsigned long total
= weighted_cpuload(cpu
);
1349 unsigned long n
= rq
->nr_running
;
1351 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1355 * find_idlest_group finds and returns the least busy CPU group within the
1358 static struct sched_group
*
1359 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1361 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1362 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1363 int load_idx
= sd
->forkexec_idx
;
1364 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1367 unsigned long load
, avg_load
;
1371 /* Skip over this group if it has no CPUs allowed */
1372 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1375 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1377 /* Tally up the load of all CPUs in the group */
1380 for_each_cpu_mask(i
, group
->cpumask
) {
1381 /* Bias balancing toward cpus of our domain */
1383 load
= source_load(i
, load_idx
);
1385 load
= target_load(i
, load_idx
);
1390 /* Adjust by relative CPU power of the group */
1391 avg_load
= sg_div_cpu_power(group
,
1392 avg_load
* SCHED_LOAD_SCALE
);
1395 this_load
= avg_load
;
1397 } else if (avg_load
< min_load
) {
1398 min_load
= avg_load
;
1401 } while (group
= group
->next
, group
!= sd
->groups
);
1403 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1409 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1412 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1415 unsigned long load
, min_load
= ULONG_MAX
;
1419 /* Traverse only the allowed CPUs */
1420 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1422 for_each_cpu_mask(i
, tmp
) {
1423 load
= weighted_cpuload(i
);
1425 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1435 * sched_balance_self: balance the current task (running on cpu) in domains
1436 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1439 * Balance, ie. select the least loaded group.
1441 * Returns the target CPU number, or the same CPU if no balancing is needed.
1443 * preempt must be disabled.
1445 static int sched_balance_self(int cpu
, int flag
)
1447 struct task_struct
*t
= current
;
1448 struct sched_domain
*tmp
, *sd
= NULL
;
1450 for_each_domain(cpu
, tmp
) {
1452 * If power savings logic is enabled for a domain, stop there.
1454 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1456 if (tmp
->flags
& flag
)
1462 struct sched_group
*group
;
1463 int new_cpu
, weight
;
1465 if (!(sd
->flags
& flag
)) {
1471 group
= find_idlest_group(sd
, t
, cpu
);
1477 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1478 if (new_cpu
== -1 || new_cpu
== cpu
) {
1479 /* Now try balancing at a lower domain level of cpu */
1484 /* Now try balancing at a lower domain level of new_cpu */
1487 weight
= cpus_weight(span
);
1488 for_each_domain(cpu
, tmp
) {
1489 if (weight
<= cpus_weight(tmp
->span
))
1491 if (tmp
->flags
& flag
)
1494 /* while loop will break here if sd == NULL */
1500 #endif /* CONFIG_SMP */
1503 * wake_idle() will wake a task on an idle cpu if task->cpu is
1504 * not idle and an idle cpu is available. The span of cpus to
1505 * search starts with cpus closest then further out as needed,
1506 * so we always favor a closer, idle cpu.
1508 * Returns the CPU we should wake onto.
1510 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1511 static int wake_idle(int cpu
, struct task_struct
*p
)
1514 struct sched_domain
*sd
;
1518 * If it is idle, then it is the best cpu to run this task.
1520 * This cpu is also the best, if it has more than one task already.
1521 * Siblings must be also busy(in most cases) as they didn't already
1522 * pickup the extra load from this cpu and hence we need not check
1523 * sibling runqueue info. This will avoid the checks and cache miss
1524 * penalities associated with that.
1526 if (idle_cpu(cpu
) || cpu_rq(cpu
)->nr_running
> 1)
1529 for_each_domain(cpu
, sd
) {
1530 if (sd
->flags
& SD_WAKE_IDLE
) {
1531 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1532 for_each_cpu_mask(i
, tmp
) {
1534 if (i
!= task_cpu(p
)) {
1536 se
.nr_wakeups_idle
);
1548 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1555 * try_to_wake_up - wake up a thread
1556 * @p: the to-be-woken-up thread
1557 * @state: the mask of task states that can be woken
1558 * @sync: do a synchronous wakeup?
1560 * Put it on the run-queue if it's not already there. The "current"
1561 * thread is always on the run-queue (except when the actual
1562 * re-schedule is in progress), and as such you're allowed to do
1563 * the simpler "current->state = TASK_RUNNING" to mark yourself
1564 * runnable without the overhead of this.
1566 * returns failure only if the task is already active.
1568 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1570 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1571 unsigned long flags
;
1575 struct sched_domain
*sd
, *this_sd
= NULL
;
1576 unsigned long load
, this_load
;
1580 rq
= task_rq_lock(p
, &flags
);
1581 old_state
= p
->state
;
1582 if (!(old_state
& state
))
1590 this_cpu
= smp_processor_id();
1593 if (unlikely(task_running(rq
, p
)))
1598 schedstat_inc(rq
, ttwu_count
);
1599 if (cpu
== this_cpu
) {
1600 schedstat_inc(rq
, ttwu_local
);
1604 for_each_domain(this_cpu
, sd
) {
1605 if (cpu_isset(cpu
, sd
->span
)) {
1606 schedstat_inc(sd
, ttwu_wake_remote
);
1612 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1616 * Check for affine wakeup and passive balancing possibilities.
1619 int idx
= this_sd
->wake_idx
;
1620 unsigned int imbalance
;
1622 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1624 load
= source_load(cpu
, idx
);
1625 this_load
= target_load(this_cpu
, idx
);
1627 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1629 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1630 unsigned long tl
= this_load
;
1631 unsigned long tl_per_task
;
1634 * Attract cache-cold tasks on sync wakeups:
1636 if (sync
&& !task_hot(p
, rq
->clock
, this_sd
))
1639 schedstat_inc(p
, se
.nr_wakeups_affine_attempts
);
1640 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1643 * If sync wakeup then subtract the (maximum possible)
1644 * effect of the currently running task from the load
1645 * of the current CPU:
1648 tl
-= current
->se
.load
.weight
;
1651 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1652 100*(tl
+ p
->se
.load
.weight
) <= imbalance
*load
) {
1654 * This domain has SD_WAKE_AFFINE and
1655 * p is cache cold in this domain, and
1656 * there is no bad imbalance.
1658 schedstat_inc(this_sd
, ttwu_move_affine
);
1659 schedstat_inc(p
, se
.nr_wakeups_affine
);
1665 * Start passive balancing when half the imbalance_pct
1668 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1669 if (imbalance
*this_load
<= 100*load
) {
1670 schedstat_inc(this_sd
, ttwu_move_balance
);
1671 schedstat_inc(p
, se
.nr_wakeups_passive
);
1677 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1679 new_cpu
= wake_idle(new_cpu
, p
);
1680 if (new_cpu
!= cpu
) {
1681 set_task_cpu(p
, new_cpu
);
1682 task_rq_unlock(rq
, &flags
);
1683 /* might preempt at this point */
1684 rq
= task_rq_lock(p
, &flags
);
1685 old_state
= p
->state
;
1686 if (!(old_state
& state
))
1691 this_cpu
= smp_processor_id();
1696 #endif /* CONFIG_SMP */
1697 schedstat_inc(p
, se
.nr_wakeups
);
1699 schedstat_inc(p
, se
.nr_wakeups_sync
);
1700 if (orig_cpu
!= cpu
)
1701 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1702 if (cpu
== this_cpu
)
1703 schedstat_inc(p
, se
.nr_wakeups_local
);
1705 schedstat_inc(p
, se
.nr_wakeups_remote
);
1706 update_rq_clock(rq
);
1707 activate_task(rq
, p
, 1);
1708 check_preempt_curr(rq
, p
);
1712 p
->state
= TASK_RUNNING
;
1713 wakeup_balance_rt(rq
, p
);
1715 task_rq_unlock(rq
, &flags
);
1720 int fastcall
wake_up_process(struct task_struct
*p
)
1722 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1723 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1725 EXPORT_SYMBOL(wake_up_process
);
1727 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1729 return try_to_wake_up(p
, state
, 0);
1733 * Perform scheduler related setup for a newly forked process p.
1734 * p is forked by current.
1736 * __sched_fork() is basic setup used by init_idle() too:
1738 static void __sched_fork(struct task_struct
*p
)
1740 p
->se
.exec_start
= 0;
1741 p
->se
.sum_exec_runtime
= 0;
1742 p
->se
.prev_sum_exec_runtime
= 0;
1744 #ifdef CONFIG_SCHEDSTATS
1745 p
->se
.wait_start
= 0;
1746 p
->se
.sum_sleep_runtime
= 0;
1747 p
->se
.sleep_start
= 0;
1748 p
->se
.block_start
= 0;
1749 p
->se
.sleep_max
= 0;
1750 p
->se
.block_max
= 0;
1752 p
->se
.slice_max
= 0;
1756 INIT_LIST_HEAD(&p
->run_list
);
1759 #ifdef CONFIG_PREEMPT_NOTIFIERS
1760 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1764 * We mark the process as running here, but have not actually
1765 * inserted it onto the runqueue yet. This guarantees that
1766 * nobody will actually run it, and a signal or other external
1767 * event cannot wake it up and insert it on the runqueue either.
1769 p
->state
= TASK_RUNNING
;
1773 * fork()/clone()-time setup:
1775 void sched_fork(struct task_struct
*p
, int clone_flags
)
1777 int cpu
= get_cpu();
1782 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1784 set_task_cpu(p
, cpu
);
1787 * Make sure we do not leak PI boosting priority to the child:
1789 p
->prio
= current
->normal_prio
;
1790 if (!rt_prio(p
->prio
))
1791 p
->sched_class
= &fair_sched_class
;
1793 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1794 if (likely(sched_info_on()))
1795 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1797 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1800 #ifdef CONFIG_PREEMPT
1801 /* Want to start with kernel preemption disabled. */
1802 task_thread_info(p
)->preempt_count
= 1;
1808 * wake_up_new_task - wake up a newly created task for the first time.
1810 * This function will do some initial scheduler statistics housekeeping
1811 * that must be done for every newly created context, then puts the task
1812 * on the runqueue and wakes it.
1814 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1816 unsigned long flags
;
1819 rq
= task_rq_lock(p
, &flags
);
1820 BUG_ON(p
->state
!= TASK_RUNNING
);
1821 update_rq_clock(rq
);
1823 p
->prio
= effective_prio(p
);
1825 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
1826 activate_task(rq
, p
, 0);
1829 * Let the scheduling class do new task startup
1830 * management (if any):
1832 p
->sched_class
->task_new(rq
, p
);
1833 inc_nr_running(p
, rq
);
1835 check_preempt_curr(rq
, p
);
1836 task_rq_unlock(rq
, &flags
);
1839 #ifdef CONFIG_PREEMPT_NOTIFIERS
1842 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1843 * @notifier: notifier struct to register
1845 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1847 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1849 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1852 * preempt_notifier_unregister - no longer interested in preemption notifications
1853 * @notifier: notifier struct to unregister
1855 * This is safe to call from within a preemption notifier.
1857 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1859 hlist_del(¬ifier
->link
);
1861 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1863 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1865 struct preempt_notifier
*notifier
;
1866 struct hlist_node
*node
;
1868 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1869 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1873 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1874 struct task_struct
*next
)
1876 struct preempt_notifier
*notifier
;
1877 struct hlist_node
*node
;
1879 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1880 notifier
->ops
->sched_out(notifier
, next
);
1885 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1890 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1891 struct task_struct
*next
)
1898 * prepare_task_switch - prepare to switch tasks
1899 * @rq: the runqueue preparing to switch
1900 * @prev: the current task that is being switched out
1901 * @next: the task we are going to switch to.
1903 * This is called with the rq lock held and interrupts off. It must
1904 * be paired with a subsequent finish_task_switch after the context
1907 * prepare_task_switch sets up locking and calls architecture specific
1911 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1912 struct task_struct
*next
)
1914 fire_sched_out_preempt_notifiers(prev
, next
);
1915 prepare_lock_switch(rq
, next
);
1916 prepare_arch_switch(next
);
1920 * finish_task_switch - clean up after a task-switch
1921 * @rq: runqueue associated with task-switch
1922 * @prev: the thread we just switched away from.
1924 * finish_task_switch must be called after the context switch, paired
1925 * with a prepare_task_switch call before the context switch.
1926 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1927 * and do any other architecture-specific cleanup actions.
1929 * Note that we may have delayed dropping an mm in context_switch(). If
1930 * so, we finish that here outside of the runqueue lock. (Doing it
1931 * with the lock held can cause deadlocks; see schedule() for
1934 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1935 __releases(rq
->lock
)
1937 struct mm_struct
*mm
= rq
->prev_mm
;
1943 * A task struct has one reference for the use as "current".
1944 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1945 * schedule one last time. The schedule call will never return, and
1946 * the scheduled task must drop that reference.
1947 * The test for TASK_DEAD must occur while the runqueue locks are
1948 * still held, otherwise prev could be scheduled on another cpu, die
1949 * there before we look at prev->state, and then the reference would
1951 * Manfred Spraul <manfred@colorfullife.com>
1953 prev_state
= prev
->state
;
1954 finish_arch_switch(prev
);
1955 finish_lock_switch(rq
, prev
);
1956 schedule_tail_balance_rt(rq
);
1958 fire_sched_in_preempt_notifiers(current
);
1961 if (unlikely(prev_state
== TASK_DEAD
)) {
1963 * Remove function-return probe instances associated with this
1964 * task and put them back on the free list.
1966 kprobe_flush_task(prev
);
1967 put_task_struct(prev
);
1972 * schedule_tail - first thing a freshly forked thread must call.
1973 * @prev: the thread we just switched away from.
1975 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1976 __releases(rq
->lock
)
1978 struct rq
*rq
= this_rq();
1980 finish_task_switch(rq
, prev
);
1981 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1982 /* In this case, finish_task_switch does not reenable preemption */
1985 if (current
->set_child_tid
)
1986 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1990 * context_switch - switch to the new MM and the new
1991 * thread's register state.
1994 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1995 struct task_struct
*next
)
1997 struct mm_struct
*mm
, *oldmm
;
1999 prepare_task_switch(rq
, prev
, next
);
2001 oldmm
= prev
->active_mm
;
2003 * For paravirt, this is coupled with an exit in switch_to to
2004 * combine the page table reload and the switch backend into
2007 arch_enter_lazy_cpu_mode();
2009 if (unlikely(!mm
)) {
2010 next
->active_mm
= oldmm
;
2011 atomic_inc(&oldmm
->mm_count
);
2012 enter_lazy_tlb(oldmm
, next
);
2014 switch_mm(oldmm
, mm
, next
);
2016 if (unlikely(!prev
->mm
)) {
2017 prev
->active_mm
= NULL
;
2018 rq
->prev_mm
= oldmm
;
2021 * Since the runqueue lock will be released by the next
2022 * task (which is an invalid locking op but in the case
2023 * of the scheduler it's an obvious special-case), so we
2024 * do an early lockdep release here:
2026 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2027 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2030 /* Here we just switch the register state and the stack. */
2031 switch_to(prev
, next
, prev
);
2035 * this_rq must be evaluated again because prev may have moved
2036 * CPUs since it called schedule(), thus the 'rq' on its stack
2037 * frame will be invalid.
2039 finish_task_switch(this_rq(), prev
);
2043 * nr_running, nr_uninterruptible and nr_context_switches:
2045 * externally visible scheduler statistics: current number of runnable
2046 * threads, current number of uninterruptible-sleeping threads, total
2047 * number of context switches performed since bootup.
2049 unsigned long nr_running(void)
2051 unsigned long i
, sum
= 0;
2053 for_each_online_cpu(i
)
2054 sum
+= cpu_rq(i
)->nr_running
;
2059 unsigned long nr_uninterruptible(void)
2061 unsigned long i
, sum
= 0;
2063 for_each_possible_cpu(i
)
2064 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2067 * Since we read the counters lockless, it might be slightly
2068 * inaccurate. Do not allow it to go below zero though:
2070 if (unlikely((long)sum
< 0))
2076 unsigned long long nr_context_switches(void)
2079 unsigned long long sum
= 0;
2081 for_each_possible_cpu(i
)
2082 sum
+= cpu_rq(i
)->nr_switches
;
2087 unsigned long nr_iowait(void)
2089 unsigned long i
, sum
= 0;
2091 for_each_possible_cpu(i
)
2092 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2097 unsigned long nr_active(void)
2099 unsigned long i
, running
= 0, uninterruptible
= 0;
2101 for_each_online_cpu(i
) {
2102 running
+= cpu_rq(i
)->nr_running
;
2103 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2106 if (unlikely((long)uninterruptible
< 0))
2107 uninterruptible
= 0;
2109 return running
+ uninterruptible
;
2113 * Update rq->cpu_load[] statistics. This function is usually called every
2114 * scheduler tick (TICK_NSEC).
2116 static void update_cpu_load(struct rq
*this_rq
)
2118 unsigned long this_load
= this_rq
->load
.weight
;
2121 this_rq
->nr_load_updates
++;
2123 /* Update our load: */
2124 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2125 unsigned long old_load
, new_load
;
2127 /* scale is effectively 1 << i now, and >> i divides by scale */
2129 old_load
= this_rq
->cpu_load
[i
];
2130 new_load
= this_load
;
2132 * Round up the averaging division if load is increasing. This
2133 * prevents us from getting stuck on 9 if the load is 10, for
2136 if (new_load
> old_load
)
2137 new_load
+= scale
-1;
2138 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2145 * double_rq_lock - safely lock two runqueues
2147 * Note this does not disable interrupts like task_rq_lock,
2148 * you need to do so manually before calling.
2150 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2151 __acquires(rq1
->lock
)
2152 __acquires(rq2
->lock
)
2154 BUG_ON(!irqs_disabled());
2156 spin_lock(&rq1
->lock
);
2157 __acquire(rq2
->lock
); /* Fake it out ;) */
2160 spin_lock(&rq1
->lock
);
2161 spin_lock(&rq2
->lock
);
2163 spin_lock(&rq2
->lock
);
2164 spin_lock(&rq1
->lock
);
2167 update_rq_clock(rq1
);
2168 update_rq_clock(rq2
);
2172 * double_rq_unlock - safely unlock two runqueues
2174 * Note this does not restore interrupts like task_rq_unlock,
2175 * you need to do so manually after calling.
2177 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2178 __releases(rq1
->lock
)
2179 __releases(rq2
->lock
)
2181 spin_unlock(&rq1
->lock
);
2183 spin_unlock(&rq2
->lock
);
2185 __release(rq2
->lock
);
2189 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2191 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2192 __releases(this_rq
->lock
)
2193 __acquires(busiest
->lock
)
2194 __acquires(this_rq
->lock
)
2198 if (unlikely(!irqs_disabled())) {
2199 /* printk() doesn't work good under rq->lock */
2200 spin_unlock(&this_rq
->lock
);
2203 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2204 if (busiest
< this_rq
) {
2205 spin_unlock(&this_rq
->lock
);
2206 spin_lock(&busiest
->lock
);
2207 spin_lock(&this_rq
->lock
);
2210 spin_lock(&busiest
->lock
);
2216 * If dest_cpu is allowed for this process, migrate the task to it.
2217 * This is accomplished by forcing the cpu_allowed mask to only
2218 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2219 * the cpu_allowed mask is restored.
2221 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2223 struct migration_req req
;
2224 unsigned long flags
;
2227 rq
= task_rq_lock(p
, &flags
);
2228 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2229 || unlikely(cpu_is_offline(dest_cpu
)))
2232 /* force the process onto the specified CPU */
2233 if (migrate_task(p
, dest_cpu
, &req
)) {
2234 /* Need to wait for migration thread (might exit: take ref). */
2235 struct task_struct
*mt
= rq
->migration_thread
;
2237 get_task_struct(mt
);
2238 task_rq_unlock(rq
, &flags
);
2239 wake_up_process(mt
);
2240 put_task_struct(mt
);
2241 wait_for_completion(&req
.done
);
2246 task_rq_unlock(rq
, &flags
);
2250 * sched_exec - execve() is a valuable balancing opportunity, because at
2251 * this point the task has the smallest effective memory and cache footprint.
2253 void sched_exec(void)
2255 int new_cpu
, this_cpu
= get_cpu();
2256 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2258 if (new_cpu
!= this_cpu
)
2259 sched_migrate_task(current
, new_cpu
);
2263 * pull_task - move a task from a remote runqueue to the local runqueue.
2264 * Both runqueues must be locked.
2266 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2267 struct rq
*this_rq
, int this_cpu
)
2269 deactivate_task(src_rq
, p
, 0);
2270 set_task_cpu(p
, this_cpu
);
2271 activate_task(this_rq
, p
, 0);
2273 * Note that idle threads have a prio of MAX_PRIO, for this test
2274 * to be always true for them.
2276 check_preempt_curr(this_rq
, p
);
2280 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2283 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2284 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2288 * We do not migrate tasks that are:
2289 * 1) running (obviously), or
2290 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2291 * 3) are cache-hot on their current CPU.
2293 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2294 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2299 if (task_running(rq
, p
)) {
2300 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2305 * Aggressive migration if:
2306 * 1) task is cache cold, or
2307 * 2) too many balance attempts have failed.
2310 if (!task_hot(p
, rq
->clock
, sd
) ||
2311 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2312 #ifdef CONFIG_SCHEDSTATS
2313 if (task_hot(p
, rq
->clock
, sd
)) {
2314 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2315 schedstat_inc(p
, se
.nr_forced_migrations
);
2321 if (task_hot(p
, rq
->clock
, sd
)) {
2322 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2328 static unsigned long
2329 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2330 unsigned long max_load_move
, struct sched_domain
*sd
,
2331 enum cpu_idle_type idle
, int *all_pinned
,
2332 int *this_best_prio
, struct rq_iterator
*iterator
)
2334 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2335 struct task_struct
*p
;
2336 long rem_load_move
= max_load_move
;
2338 if (max_load_move
== 0)
2344 * Start the load-balancing iterator:
2346 p
= iterator
->start(iterator
->arg
);
2348 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2351 * To help distribute high priority tasks across CPUs we don't
2352 * skip a task if it will be the highest priority task (i.e. smallest
2353 * prio value) on its new queue regardless of its load weight
2355 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2356 SCHED_LOAD_SCALE_FUZZ
;
2357 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2358 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2359 p
= iterator
->next(iterator
->arg
);
2363 pull_task(busiest
, p
, this_rq
, this_cpu
);
2365 rem_load_move
-= p
->se
.load
.weight
;
2368 * We only want to steal up to the prescribed amount of weighted load.
2370 if (rem_load_move
> 0) {
2371 if (p
->prio
< *this_best_prio
)
2372 *this_best_prio
= p
->prio
;
2373 p
= iterator
->next(iterator
->arg
);
2378 * Right now, this is one of only two places pull_task() is called,
2379 * so we can safely collect pull_task() stats here rather than
2380 * inside pull_task().
2382 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2385 *all_pinned
= pinned
;
2387 return max_load_move
- rem_load_move
;
2391 * move_tasks tries to move up to max_load_move weighted load from busiest to
2392 * this_rq, as part of a balancing operation within domain "sd".
2393 * Returns 1 if successful and 0 otherwise.
2395 * Called with both runqueues locked.
2397 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2398 unsigned long max_load_move
,
2399 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2402 const struct sched_class
*class = sched_class_highest
;
2403 unsigned long total_load_moved
= 0;
2404 int this_best_prio
= this_rq
->curr
->prio
;
2408 class->load_balance(this_rq
, this_cpu
, busiest
,
2409 max_load_move
- total_load_moved
,
2410 sd
, idle
, all_pinned
, &this_best_prio
);
2411 class = class->next
;
2412 } while (class && max_load_move
> total_load_moved
);
2414 return total_load_moved
> 0;
2418 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2419 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2420 struct rq_iterator
*iterator
)
2422 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2426 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2427 pull_task(busiest
, p
, this_rq
, this_cpu
);
2429 * Right now, this is only the second place pull_task()
2430 * is called, so we can safely collect pull_task()
2431 * stats here rather than inside pull_task().
2433 schedstat_inc(sd
, lb_gained
[idle
]);
2437 p
= iterator
->next(iterator
->arg
);
2444 * move_one_task tries to move exactly one task from busiest to this_rq, as
2445 * part of active balancing operations within "domain".
2446 * Returns 1 if successful and 0 otherwise.
2448 * Called with both runqueues locked.
2450 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2451 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2453 const struct sched_class
*class;
2455 for (class = sched_class_highest
; class; class = class->next
)
2456 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2463 * find_busiest_group finds and returns the busiest CPU group within the
2464 * domain. It calculates and returns the amount of weighted load which
2465 * should be moved to restore balance via the imbalance parameter.
2467 static struct sched_group
*
2468 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2469 unsigned long *imbalance
, enum cpu_idle_type idle
,
2470 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2472 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2473 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2474 unsigned long max_pull
;
2475 unsigned long busiest_load_per_task
, busiest_nr_running
;
2476 unsigned long this_load_per_task
, this_nr_running
;
2477 int load_idx
, group_imb
= 0;
2478 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2479 int power_savings_balance
= 1;
2480 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2481 unsigned long min_nr_running
= ULONG_MAX
;
2482 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2485 max_load
= this_load
= total_load
= total_pwr
= 0;
2486 busiest_load_per_task
= busiest_nr_running
= 0;
2487 this_load_per_task
= this_nr_running
= 0;
2488 if (idle
== CPU_NOT_IDLE
)
2489 load_idx
= sd
->busy_idx
;
2490 else if (idle
== CPU_NEWLY_IDLE
)
2491 load_idx
= sd
->newidle_idx
;
2493 load_idx
= sd
->idle_idx
;
2496 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2499 int __group_imb
= 0;
2500 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2501 unsigned long sum_nr_running
, sum_weighted_load
;
2503 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2506 balance_cpu
= first_cpu(group
->cpumask
);
2508 /* Tally up the load of all CPUs in the group */
2509 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2511 min_cpu_load
= ~0UL;
2513 for_each_cpu_mask(i
, group
->cpumask
) {
2516 if (!cpu_isset(i
, *cpus
))
2521 if (*sd_idle
&& rq
->nr_running
)
2524 /* Bias balancing toward cpus of our domain */
2526 if (idle_cpu(i
) && !first_idle_cpu
) {
2531 load
= target_load(i
, load_idx
);
2533 load
= source_load(i
, load_idx
);
2534 if (load
> max_cpu_load
)
2535 max_cpu_load
= load
;
2536 if (min_cpu_load
> load
)
2537 min_cpu_load
= load
;
2541 sum_nr_running
+= rq
->nr_running
;
2542 sum_weighted_load
+= weighted_cpuload(i
);
2546 * First idle cpu or the first cpu(busiest) in this sched group
2547 * is eligible for doing load balancing at this and above
2548 * domains. In the newly idle case, we will allow all the cpu's
2549 * to do the newly idle load balance.
2551 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2552 balance_cpu
!= this_cpu
&& balance
) {
2557 total_load
+= avg_load
;
2558 total_pwr
+= group
->__cpu_power
;
2560 /* Adjust by relative CPU power of the group */
2561 avg_load
= sg_div_cpu_power(group
,
2562 avg_load
* SCHED_LOAD_SCALE
);
2564 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2567 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2570 this_load
= avg_load
;
2572 this_nr_running
= sum_nr_running
;
2573 this_load_per_task
= sum_weighted_load
;
2574 } else if (avg_load
> max_load
&&
2575 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2576 max_load
= avg_load
;
2578 busiest_nr_running
= sum_nr_running
;
2579 busiest_load_per_task
= sum_weighted_load
;
2580 group_imb
= __group_imb
;
2583 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2585 * Busy processors will not participate in power savings
2588 if (idle
== CPU_NOT_IDLE
||
2589 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2593 * If the local group is idle or completely loaded
2594 * no need to do power savings balance at this domain
2596 if (local_group
&& (this_nr_running
>= group_capacity
||
2598 power_savings_balance
= 0;
2601 * If a group is already running at full capacity or idle,
2602 * don't include that group in power savings calculations
2604 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2609 * Calculate the group which has the least non-idle load.
2610 * This is the group from where we need to pick up the load
2613 if ((sum_nr_running
< min_nr_running
) ||
2614 (sum_nr_running
== min_nr_running
&&
2615 first_cpu(group
->cpumask
) <
2616 first_cpu(group_min
->cpumask
))) {
2618 min_nr_running
= sum_nr_running
;
2619 min_load_per_task
= sum_weighted_load
/
2624 * Calculate the group which is almost near its
2625 * capacity but still has some space to pick up some load
2626 * from other group and save more power
2628 if (sum_nr_running
<= group_capacity
- 1) {
2629 if (sum_nr_running
> leader_nr_running
||
2630 (sum_nr_running
== leader_nr_running
&&
2631 first_cpu(group
->cpumask
) >
2632 first_cpu(group_leader
->cpumask
))) {
2633 group_leader
= group
;
2634 leader_nr_running
= sum_nr_running
;
2639 group
= group
->next
;
2640 } while (group
!= sd
->groups
);
2642 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2645 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2647 if (this_load
>= avg_load
||
2648 100*max_load
<= sd
->imbalance_pct
*this_load
)
2651 busiest_load_per_task
/= busiest_nr_running
;
2653 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2656 * We're trying to get all the cpus to the average_load, so we don't
2657 * want to push ourselves above the average load, nor do we wish to
2658 * reduce the max loaded cpu below the average load, as either of these
2659 * actions would just result in more rebalancing later, and ping-pong
2660 * tasks around. Thus we look for the minimum possible imbalance.
2661 * Negative imbalances (*we* are more loaded than anyone else) will
2662 * be counted as no imbalance for these purposes -- we can't fix that
2663 * by pulling tasks to us. Be careful of negative numbers as they'll
2664 * appear as very large values with unsigned longs.
2666 if (max_load
<= busiest_load_per_task
)
2670 * In the presence of smp nice balancing, certain scenarios can have
2671 * max load less than avg load(as we skip the groups at or below
2672 * its cpu_power, while calculating max_load..)
2674 if (max_load
< avg_load
) {
2676 goto small_imbalance
;
2679 /* Don't want to pull so many tasks that a group would go idle */
2680 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2682 /* How much load to actually move to equalise the imbalance */
2683 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2684 (avg_load
- this_load
) * this->__cpu_power
)
2688 * if *imbalance is less than the average load per runnable task
2689 * there is no gaurantee that any tasks will be moved so we'll have
2690 * a think about bumping its value to force at least one task to be
2693 if (*imbalance
< busiest_load_per_task
) {
2694 unsigned long tmp
, pwr_now
, pwr_move
;
2698 pwr_move
= pwr_now
= 0;
2700 if (this_nr_running
) {
2701 this_load_per_task
/= this_nr_running
;
2702 if (busiest_load_per_task
> this_load_per_task
)
2705 this_load_per_task
= SCHED_LOAD_SCALE
;
2707 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2708 busiest_load_per_task
* imbn
) {
2709 *imbalance
= busiest_load_per_task
;
2714 * OK, we don't have enough imbalance to justify moving tasks,
2715 * however we may be able to increase total CPU power used by
2719 pwr_now
+= busiest
->__cpu_power
*
2720 min(busiest_load_per_task
, max_load
);
2721 pwr_now
+= this->__cpu_power
*
2722 min(this_load_per_task
, this_load
);
2723 pwr_now
/= SCHED_LOAD_SCALE
;
2725 /* Amount of load we'd subtract */
2726 tmp
= sg_div_cpu_power(busiest
,
2727 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2729 pwr_move
+= busiest
->__cpu_power
*
2730 min(busiest_load_per_task
, max_load
- tmp
);
2732 /* Amount of load we'd add */
2733 if (max_load
* busiest
->__cpu_power
<
2734 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2735 tmp
= sg_div_cpu_power(this,
2736 max_load
* busiest
->__cpu_power
);
2738 tmp
= sg_div_cpu_power(this,
2739 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2740 pwr_move
+= this->__cpu_power
*
2741 min(this_load_per_task
, this_load
+ tmp
);
2742 pwr_move
/= SCHED_LOAD_SCALE
;
2744 /* Move if we gain throughput */
2745 if (pwr_move
> pwr_now
)
2746 *imbalance
= busiest_load_per_task
;
2752 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2753 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2756 if (this == group_leader
&& group_leader
!= group_min
) {
2757 *imbalance
= min_load_per_task
;
2767 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2770 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2771 unsigned long imbalance
, cpumask_t
*cpus
)
2773 struct rq
*busiest
= NULL
, *rq
;
2774 unsigned long max_load
= 0;
2777 for_each_cpu_mask(i
, group
->cpumask
) {
2780 if (!cpu_isset(i
, *cpus
))
2784 wl
= weighted_cpuload(i
);
2786 if (rq
->nr_running
== 1 && wl
> imbalance
)
2789 if (wl
> max_load
) {
2799 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2800 * so long as it is large enough.
2802 #define MAX_PINNED_INTERVAL 512
2805 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2806 * tasks if there is an imbalance.
2808 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2809 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2812 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2813 struct sched_group
*group
;
2814 unsigned long imbalance
;
2816 cpumask_t cpus
= CPU_MASK_ALL
;
2817 unsigned long flags
;
2820 * When power savings policy is enabled for the parent domain, idle
2821 * sibling can pick up load irrespective of busy siblings. In this case,
2822 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2823 * portraying it as CPU_NOT_IDLE.
2825 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2826 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2829 schedstat_inc(sd
, lb_count
[idle
]);
2832 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2839 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2843 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2845 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2849 BUG_ON(busiest
== this_rq
);
2851 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2854 if (busiest
->nr_running
> 1) {
2856 * Attempt to move tasks. If find_busiest_group has found
2857 * an imbalance but busiest->nr_running <= 1, the group is
2858 * still unbalanced. ld_moved simply stays zero, so it is
2859 * correctly treated as an imbalance.
2861 local_irq_save(flags
);
2862 double_rq_lock(this_rq
, busiest
);
2863 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2864 imbalance
, sd
, idle
, &all_pinned
);
2865 double_rq_unlock(this_rq
, busiest
);
2866 local_irq_restore(flags
);
2869 * some other cpu did the load balance for us.
2871 if (ld_moved
&& this_cpu
!= smp_processor_id())
2872 resched_cpu(this_cpu
);
2874 /* All tasks on this runqueue were pinned by CPU affinity */
2875 if (unlikely(all_pinned
)) {
2876 cpu_clear(cpu_of(busiest
), cpus
);
2877 if (!cpus_empty(cpus
))
2884 schedstat_inc(sd
, lb_failed
[idle
]);
2885 sd
->nr_balance_failed
++;
2887 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2889 spin_lock_irqsave(&busiest
->lock
, flags
);
2891 /* don't kick the migration_thread, if the curr
2892 * task on busiest cpu can't be moved to this_cpu
2894 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2895 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2897 goto out_one_pinned
;
2900 if (!busiest
->active_balance
) {
2901 busiest
->active_balance
= 1;
2902 busiest
->push_cpu
= this_cpu
;
2905 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2907 wake_up_process(busiest
->migration_thread
);
2910 * We've kicked active balancing, reset the failure
2913 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2916 sd
->nr_balance_failed
= 0;
2918 if (likely(!active_balance
)) {
2919 /* We were unbalanced, so reset the balancing interval */
2920 sd
->balance_interval
= sd
->min_interval
;
2923 * If we've begun active balancing, start to back off. This
2924 * case may not be covered by the all_pinned logic if there
2925 * is only 1 task on the busy runqueue (because we don't call
2928 if (sd
->balance_interval
< sd
->max_interval
)
2929 sd
->balance_interval
*= 2;
2932 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2933 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2938 schedstat_inc(sd
, lb_balanced
[idle
]);
2940 sd
->nr_balance_failed
= 0;
2943 /* tune up the balancing interval */
2944 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2945 (sd
->balance_interval
< sd
->max_interval
))
2946 sd
->balance_interval
*= 2;
2948 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2949 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2955 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2956 * tasks if there is an imbalance.
2958 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2959 * this_rq is locked.
2962 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2964 struct sched_group
*group
;
2965 struct rq
*busiest
= NULL
;
2966 unsigned long imbalance
;
2970 cpumask_t cpus
= CPU_MASK_ALL
;
2973 * When power savings policy is enabled for the parent domain, idle
2974 * sibling can pick up load irrespective of busy siblings. In this case,
2975 * let the state of idle sibling percolate up as IDLE, instead of
2976 * portraying it as CPU_NOT_IDLE.
2978 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2979 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2982 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
2984 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2985 &sd_idle
, &cpus
, NULL
);
2987 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2991 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2994 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2998 BUG_ON(busiest
== this_rq
);
3000 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3003 if (busiest
->nr_running
> 1) {
3004 /* Attempt to move tasks */
3005 double_lock_balance(this_rq
, busiest
);
3006 /* this_rq->clock is already updated */
3007 update_rq_clock(busiest
);
3008 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3009 imbalance
, sd
, CPU_NEWLY_IDLE
,
3011 spin_unlock(&busiest
->lock
);
3013 if (unlikely(all_pinned
)) {
3014 cpu_clear(cpu_of(busiest
), cpus
);
3015 if (!cpus_empty(cpus
))
3021 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3022 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3023 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3026 sd
->nr_balance_failed
= 0;
3031 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3032 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3033 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3035 sd
->nr_balance_failed
= 0;
3041 * idle_balance is called by schedule() if this_cpu is about to become
3042 * idle. Attempts to pull tasks from other CPUs.
3044 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3046 struct sched_domain
*sd
;
3047 int pulled_task
= -1;
3048 unsigned long next_balance
= jiffies
+ HZ
;
3050 for_each_domain(this_cpu
, sd
) {
3051 unsigned long interval
;
3053 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3056 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3057 /* If we've pulled tasks over stop searching: */
3058 pulled_task
= load_balance_newidle(this_cpu
,
3061 interval
= msecs_to_jiffies(sd
->balance_interval
);
3062 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3063 next_balance
= sd
->last_balance
+ interval
;
3067 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3069 * We are going idle. next_balance may be set based on
3070 * a busy processor. So reset next_balance.
3072 this_rq
->next_balance
= next_balance
;
3077 * active_load_balance is run by migration threads. It pushes running tasks
3078 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3079 * running on each physical CPU where possible, and avoids physical /
3080 * logical imbalances.
3082 * Called with busiest_rq locked.
3084 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3086 int target_cpu
= busiest_rq
->push_cpu
;
3087 struct sched_domain
*sd
;
3088 struct rq
*target_rq
;
3090 /* Is there any task to move? */
3091 if (busiest_rq
->nr_running
<= 1)
3094 target_rq
= cpu_rq(target_cpu
);
3097 * This condition is "impossible", if it occurs
3098 * we need to fix it. Originally reported by
3099 * Bjorn Helgaas on a 128-cpu setup.
3101 BUG_ON(busiest_rq
== target_rq
);
3103 /* move a task from busiest_rq to target_rq */
3104 double_lock_balance(busiest_rq
, target_rq
);
3105 update_rq_clock(busiest_rq
);
3106 update_rq_clock(target_rq
);
3108 /* Search for an sd spanning us and the target CPU. */
3109 for_each_domain(target_cpu
, sd
) {
3110 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3111 cpu_isset(busiest_cpu
, sd
->span
))
3116 schedstat_inc(sd
, alb_count
);
3118 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3120 schedstat_inc(sd
, alb_pushed
);
3122 schedstat_inc(sd
, alb_failed
);
3124 spin_unlock(&target_rq
->lock
);
3129 atomic_t load_balancer
;
3131 } nohz ____cacheline_aligned
= {
3132 .load_balancer
= ATOMIC_INIT(-1),
3133 .cpu_mask
= CPU_MASK_NONE
,
3137 * This routine will try to nominate the ilb (idle load balancing)
3138 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3139 * load balancing on behalf of all those cpus. If all the cpus in the system
3140 * go into this tickless mode, then there will be no ilb owner (as there is
3141 * no need for one) and all the cpus will sleep till the next wakeup event
3144 * For the ilb owner, tick is not stopped. And this tick will be used
3145 * for idle load balancing. ilb owner will still be part of
3148 * While stopping the tick, this cpu will become the ilb owner if there
3149 * is no other owner. And will be the owner till that cpu becomes busy
3150 * or if all cpus in the system stop their ticks at which point
3151 * there is no need for ilb owner.
3153 * When the ilb owner becomes busy, it nominates another owner, during the
3154 * next busy scheduler_tick()
3156 int select_nohz_load_balancer(int stop_tick
)
3158 int cpu
= smp_processor_id();
3161 cpu_set(cpu
, nohz
.cpu_mask
);
3162 cpu_rq(cpu
)->in_nohz_recently
= 1;
3165 * If we are going offline and still the leader, give up!
3167 if (cpu_is_offline(cpu
) &&
3168 atomic_read(&nohz
.load_balancer
) == cpu
) {
3169 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3174 /* time for ilb owner also to sleep */
3175 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3176 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3177 atomic_set(&nohz
.load_balancer
, -1);
3181 if (atomic_read(&nohz
.load_balancer
) == -1) {
3182 /* make me the ilb owner */
3183 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3185 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3188 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3191 cpu_clear(cpu
, nohz
.cpu_mask
);
3193 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3194 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3201 static DEFINE_SPINLOCK(balancing
);
3204 * It checks each scheduling domain to see if it is due to be balanced,
3205 * and initiates a balancing operation if so.
3207 * Balancing parameters are set up in arch_init_sched_domains.
3209 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3212 struct rq
*rq
= cpu_rq(cpu
);
3213 unsigned long interval
;
3214 struct sched_domain
*sd
;
3215 /* Earliest time when we have to do rebalance again */
3216 unsigned long next_balance
= jiffies
+ 60*HZ
;
3217 int update_next_balance
= 0;
3219 for_each_domain(cpu
, sd
) {
3220 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3223 interval
= sd
->balance_interval
;
3224 if (idle
!= CPU_IDLE
)
3225 interval
*= sd
->busy_factor
;
3227 /* scale ms to jiffies */
3228 interval
= msecs_to_jiffies(interval
);
3229 if (unlikely(!interval
))
3231 if (interval
> HZ
*NR_CPUS
/10)
3232 interval
= HZ
*NR_CPUS
/10;
3235 if (sd
->flags
& SD_SERIALIZE
) {
3236 if (!spin_trylock(&balancing
))
3240 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3241 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3243 * We've pulled tasks over so either we're no
3244 * longer idle, or one of our SMT siblings is
3247 idle
= CPU_NOT_IDLE
;
3249 sd
->last_balance
= jiffies
;
3251 if (sd
->flags
& SD_SERIALIZE
)
3252 spin_unlock(&balancing
);
3254 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3255 next_balance
= sd
->last_balance
+ interval
;
3256 update_next_balance
= 1;
3260 * Stop the load balance at this level. There is another
3261 * CPU in our sched group which is doing load balancing more
3269 * next_balance will be updated only when there is a need.
3270 * When the cpu is attached to null domain for ex, it will not be
3273 if (likely(update_next_balance
))
3274 rq
->next_balance
= next_balance
;
3278 * run_rebalance_domains is triggered when needed from the scheduler tick.
3279 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3280 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3282 static void run_rebalance_domains(struct softirq_action
*h
)
3284 int this_cpu
= smp_processor_id();
3285 struct rq
*this_rq
= cpu_rq(this_cpu
);
3286 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3287 CPU_IDLE
: CPU_NOT_IDLE
;
3289 rebalance_domains(this_cpu
, idle
);
3293 * If this cpu is the owner for idle load balancing, then do the
3294 * balancing on behalf of the other idle cpus whose ticks are
3297 if (this_rq
->idle_at_tick
&&
3298 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3299 cpumask_t cpus
= nohz
.cpu_mask
;
3303 cpu_clear(this_cpu
, cpus
);
3304 for_each_cpu_mask(balance_cpu
, cpus
) {
3306 * If this cpu gets work to do, stop the load balancing
3307 * work being done for other cpus. Next load
3308 * balancing owner will pick it up.
3313 rebalance_domains(balance_cpu
, CPU_IDLE
);
3315 rq
= cpu_rq(balance_cpu
);
3316 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3317 this_rq
->next_balance
= rq
->next_balance
;
3324 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3326 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3327 * idle load balancing owner or decide to stop the periodic load balancing,
3328 * if the whole system is idle.
3330 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3334 * If we were in the nohz mode recently and busy at the current
3335 * scheduler tick, then check if we need to nominate new idle
3338 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3339 rq
->in_nohz_recently
= 0;
3341 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3342 cpu_clear(cpu
, nohz
.cpu_mask
);
3343 atomic_set(&nohz
.load_balancer
, -1);
3346 if (atomic_read(&nohz
.load_balancer
) == -1) {
3348 * simple selection for now: Nominate the
3349 * first cpu in the nohz list to be the next
3352 * TBD: Traverse the sched domains and nominate
3353 * the nearest cpu in the nohz.cpu_mask.
3355 int ilb
= first_cpu(nohz
.cpu_mask
);
3363 * If this cpu is idle and doing idle load balancing for all the
3364 * cpus with ticks stopped, is it time for that to stop?
3366 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3367 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3373 * If this cpu is idle and the idle load balancing is done by
3374 * someone else, then no need raise the SCHED_SOFTIRQ
3376 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3377 cpu_isset(cpu
, nohz
.cpu_mask
))
3380 if (time_after_eq(jiffies
, rq
->next_balance
))
3381 raise_softirq(SCHED_SOFTIRQ
);
3384 #else /* CONFIG_SMP */
3387 * on UP we do not need to balance between CPUs:
3389 static inline void idle_balance(int cpu
, struct rq
*rq
)
3395 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3397 EXPORT_PER_CPU_SYMBOL(kstat
);
3400 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3401 * that have not yet been banked in case the task is currently running.
3403 unsigned long long task_sched_runtime(struct task_struct
*p
)
3405 unsigned long flags
;
3409 rq
= task_rq_lock(p
, &flags
);
3410 ns
= p
->se
.sum_exec_runtime
;
3411 if (task_current(rq
, p
)) {
3412 update_rq_clock(rq
);
3413 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3414 if ((s64
)delta_exec
> 0)
3417 task_rq_unlock(rq
, &flags
);
3423 * Account user cpu time to a process.
3424 * @p: the process that the cpu time gets accounted to
3425 * @cputime: the cpu time spent in user space since the last update
3427 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3429 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3432 p
->utime
= cputime_add(p
->utime
, cputime
);
3434 /* Add user time to cpustat. */
3435 tmp
= cputime_to_cputime64(cputime
);
3436 if (TASK_NICE(p
) > 0)
3437 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3439 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3443 * Account guest cpu time to a process.
3444 * @p: the process that the cpu time gets accounted to
3445 * @cputime: the cpu time spent in virtual machine since the last update
3447 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3450 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3452 tmp
= cputime_to_cputime64(cputime
);
3454 p
->utime
= cputime_add(p
->utime
, cputime
);
3455 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3457 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3458 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3462 * Account scaled user cpu time to a process.
3463 * @p: the process that the cpu time gets accounted to
3464 * @cputime: the cpu time spent in user space since the last update
3466 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3468 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3472 * Account system cpu time to a process.
3473 * @p: the process that the cpu time gets accounted to
3474 * @hardirq_offset: the offset to subtract from hardirq_count()
3475 * @cputime: the cpu time spent in kernel space since the last update
3477 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3480 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3481 struct rq
*rq
= this_rq();
3484 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3485 return account_guest_time(p
, cputime
);
3487 p
->stime
= cputime_add(p
->stime
, cputime
);
3489 /* Add system time to cpustat. */
3490 tmp
= cputime_to_cputime64(cputime
);
3491 if (hardirq_count() - hardirq_offset
)
3492 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3493 else if (softirq_count())
3494 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3495 else if (p
!= rq
->idle
)
3496 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3497 else if (atomic_read(&rq
->nr_iowait
) > 0)
3498 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3500 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3501 /* Account for system time used */
3502 acct_update_integrals(p
);
3506 * Account scaled system cpu time to a process.
3507 * @p: the process that the cpu time gets accounted to
3508 * @hardirq_offset: the offset to subtract from hardirq_count()
3509 * @cputime: the cpu time spent in kernel space since the last update
3511 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3513 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3517 * Account for involuntary wait time.
3518 * @p: the process from which the cpu time has been stolen
3519 * @steal: the cpu time spent in involuntary wait
3521 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3523 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3524 cputime64_t tmp
= cputime_to_cputime64(steal
);
3525 struct rq
*rq
= this_rq();
3527 if (p
== rq
->idle
) {
3528 p
->stime
= cputime_add(p
->stime
, steal
);
3529 if (atomic_read(&rq
->nr_iowait
) > 0)
3530 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3532 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3534 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3538 * This function gets called by the timer code, with HZ frequency.
3539 * We call it with interrupts disabled.
3541 * It also gets called by the fork code, when changing the parent's
3544 void scheduler_tick(void)
3546 int cpu
= smp_processor_id();
3547 struct rq
*rq
= cpu_rq(cpu
);
3548 struct task_struct
*curr
= rq
->curr
;
3549 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3551 spin_lock(&rq
->lock
);
3552 __update_rq_clock(rq
);
3554 * Let rq->clock advance by at least TICK_NSEC:
3556 if (unlikely(rq
->clock
< next_tick
))
3557 rq
->clock
= next_tick
;
3558 rq
->tick_timestamp
= rq
->clock
;
3559 update_cpu_load(rq
);
3560 if (curr
!= rq
->idle
) /* FIXME: needed? */
3561 curr
->sched_class
->task_tick(rq
, curr
);
3562 spin_unlock(&rq
->lock
);
3565 rq
->idle_at_tick
= idle_cpu(cpu
);
3566 trigger_load_balance(rq
, cpu
);
3570 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3572 void fastcall
add_preempt_count(int val
)
3577 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3579 preempt_count() += val
;
3581 * Spinlock count overflowing soon?
3583 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3586 EXPORT_SYMBOL(add_preempt_count
);
3588 void fastcall
sub_preempt_count(int val
)
3593 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3596 * Is the spinlock portion underflowing?
3598 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3599 !(preempt_count() & PREEMPT_MASK
)))
3602 preempt_count() -= val
;
3604 EXPORT_SYMBOL(sub_preempt_count
);
3609 * Print scheduling while atomic bug:
3611 static noinline
void __schedule_bug(struct task_struct
*prev
)
3613 struct pt_regs
*regs
= get_irq_regs();
3615 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3616 prev
->comm
, prev
->pid
, preempt_count());
3618 debug_show_held_locks(prev
);
3619 if (irqs_disabled())
3620 print_irqtrace_events(prev
);
3629 * Various schedule()-time debugging checks and statistics:
3631 static inline void schedule_debug(struct task_struct
*prev
)
3634 * Test if we are atomic. Since do_exit() needs to call into
3635 * schedule() atomically, we ignore that path for now.
3636 * Otherwise, whine if we are scheduling when we should not be.
3638 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3639 __schedule_bug(prev
);
3641 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3643 schedstat_inc(this_rq(), sched_count
);
3644 #ifdef CONFIG_SCHEDSTATS
3645 if (unlikely(prev
->lock_depth
>= 0)) {
3646 schedstat_inc(this_rq(), bkl_count
);
3647 schedstat_inc(prev
, sched_info
.bkl_count
);
3653 * Pick up the highest-prio task:
3655 static inline struct task_struct
*
3656 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3658 const struct sched_class
*class;
3659 struct task_struct
*p
;
3662 * Optimization: we know that if all tasks are in
3663 * the fair class we can call that function directly:
3665 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3666 p
= fair_sched_class
.pick_next_task(rq
);
3671 class = sched_class_highest
;
3673 p
= class->pick_next_task(rq
);
3677 * Will never be NULL as the idle class always
3678 * returns a non-NULL p:
3680 class = class->next
;
3685 * schedule() is the main scheduler function.
3687 asmlinkage
void __sched
schedule(void)
3689 struct task_struct
*prev
, *next
;
3696 cpu
= smp_processor_id();
3700 switch_count
= &prev
->nivcsw
;
3702 release_kernel_lock(prev
);
3703 need_resched_nonpreemptible
:
3705 schedule_debug(prev
);
3708 * Do the rq-clock update outside the rq lock:
3710 local_irq_disable();
3711 __update_rq_clock(rq
);
3712 spin_lock(&rq
->lock
);
3713 clear_tsk_need_resched(prev
);
3715 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3716 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3717 unlikely(signal_pending(prev
)))) {
3718 prev
->state
= TASK_RUNNING
;
3720 deactivate_task(rq
, prev
, 1);
3722 switch_count
= &prev
->nvcsw
;
3725 schedule_balance_rt(rq
, prev
);
3727 if (unlikely(!rq
->nr_running
))
3728 idle_balance(cpu
, rq
);
3730 prev
->sched_class
->put_prev_task(rq
, prev
);
3731 next
= pick_next_task(rq
, prev
);
3733 sched_info_switch(prev
, next
);
3735 if (likely(prev
!= next
)) {
3740 context_switch(rq
, prev
, next
); /* unlocks the rq */
3742 spin_unlock_irq(&rq
->lock
);
3744 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3745 cpu
= smp_processor_id();
3747 goto need_resched_nonpreemptible
;
3749 preempt_enable_no_resched();
3750 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3753 EXPORT_SYMBOL(schedule
);
3755 #ifdef CONFIG_PREEMPT
3757 * this is the entry point to schedule() from in-kernel preemption
3758 * off of preempt_enable. Kernel preemptions off return from interrupt
3759 * occur there and call schedule directly.
3761 asmlinkage
void __sched
preempt_schedule(void)
3763 struct thread_info
*ti
= current_thread_info();
3764 #ifdef CONFIG_PREEMPT_BKL
3765 struct task_struct
*task
= current
;
3766 int saved_lock_depth
;
3769 * If there is a non-zero preempt_count or interrupts are disabled,
3770 * we do not want to preempt the current task. Just return..
3772 if (likely(ti
->preempt_count
|| irqs_disabled()))
3776 add_preempt_count(PREEMPT_ACTIVE
);
3779 * We keep the big kernel semaphore locked, but we
3780 * clear ->lock_depth so that schedule() doesnt
3781 * auto-release the semaphore:
3783 #ifdef CONFIG_PREEMPT_BKL
3784 saved_lock_depth
= task
->lock_depth
;
3785 task
->lock_depth
= -1;
3788 #ifdef CONFIG_PREEMPT_BKL
3789 task
->lock_depth
= saved_lock_depth
;
3791 sub_preempt_count(PREEMPT_ACTIVE
);
3794 * Check again in case we missed a preemption opportunity
3795 * between schedule and now.
3798 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3800 EXPORT_SYMBOL(preempt_schedule
);
3803 * this is the entry point to schedule() from kernel preemption
3804 * off of irq context.
3805 * Note, that this is called and return with irqs disabled. This will
3806 * protect us against recursive calling from irq.
3808 asmlinkage
void __sched
preempt_schedule_irq(void)
3810 struct thread_info
*ti
= current_thread_info();
3811 #ifdef CONFIG_PREEMPT_BKL
3812 struct task_struct
*task
= current
;
3813 int saved_lock_depth
;
3815 /* Catch callers which need to be fixed */
3816 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3819 add_preempt_count(PREEMPT_ACTIVE
);
3822 * We keep the big kernel semaphore locked, but we
3823 * clear ->lock_depth so that schedule() doesnt
3824 * auto-release the semaphore:
3826 #ifdef CONFIG_PREEMPT_BKL
3827 saved_lock_depth
= task
->lock_depth
;
3828 task
->lock_depth
= -1;
3832 local_irq_disable();
3833 #ifdef CONFIG_PREEMPT_BKL
3834 task
->lock_depth
= saved_lock_depth
;
3836 sub_preempt_count(PREEMPT_ACTIVE
);
3839 * Check again in case we missed a preemption opportunity
3840 * between schedule and now.
3843 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3846 #endif /* CONFIG_PREEMPT */
3848 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3851 return try_to_wake_up(curr
->private, mode
, sync
);
3853 EXPORT_SYMBOL(default_wake_function
);
3856 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3857 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3858 * number) then we wake all the non-exclusive tasks and one exclusive task.
3860 * There are circumstances in which we can try to wake a task which has already
3861 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3862 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3864 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3865 int nr_exclusive
, int sync
, void *key
)
3867 wait_queue_t
*curr
, *next
;
3869 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3870 unsigned flags
= curr
->flags
;
3872 if (curr
->func(curr
, mode
, sync
, key
) &&
3873 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3879 * __wake_up - wake up threads blocked on a waitqueue.
3881 * @mode: which threads
3882 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3883 * @key: is directly passed to the wakeup function
3885 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3886 int nr_exclusive
, void *key
)
3888 unsigned long flags
;
3890 spin_lock_irqsave(&q
->lock
, flags
);
3891 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3892 spin_unlock_irqrestore(&q
->lock
, flags
);
3894 EXPORT_SYMBOL(__wake_up
);
3897 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3899 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3901 __wake_up_common(q
, mode
, 1, 0, NULL
);
3905 * __wake_up_sync - wake up threads blocked on a waitqueue.
3907 * @mode: which threads
3908 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3910 * The sync wakeup differs that the waker knows that it will schedule
3911 * away soon, so while the target thread will be woken up, it will not
3912 * be migrated to another CPU - ie. the two threads are 'synchronized'
3913 * with each other. This can prevent needless bouncing between CPUs.
3915 * On UP it can prevent extra preemption.
3918 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3920 unsigned long flags
;
3926 if (unlikely(!nr_exclusive
))
3929 spin_lock_irqsave(&q
->lock
, flags
);
3930 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3931 spin_unlock_irqrestore(&q
->lock
, flags
);
3933 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3935 void complete(struct completion
*x
)
3937 unsigned long flags
;
3939 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3941 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3943 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3945 EXPORT_SYMBOL(complete
);
3947 void complete_all(struct completion
*x
)
3949 unsigned long flags
;
3951 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3952 x
->done
+= UINT_MAX
/2;
3953 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3955 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3957 EXPORT_SYMBOL(complete_all
);
3959 static inline long __sched
3960 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3963 DECLARE_WAITQUEUE(wait
, current
);
3965 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3966 __add_wait_queue_tail(&x
->wait
, &wait
);
3968 if (state
== TASK_INTERRUPTIBLE
&&
3969 signal_pending(current
)) {
3970 __remove_wait_queue(&x
->wait
, &wait
);
3971 return -ERESTARTSYS
;
3973 __set_current_state(state
);
3974 spin_unlock_irq(&x
->wait
.lock
);
3975 timeout
= schedule_timeout(timeout
);
3976 spin_lock_irq(&x
->wait
.lock
);
3978 __remove_wait_queue(&x
->wait
, &wait
);
3982 __remove_wait_queue(&x
->wait
, &wait
);
3989 wait_for_common(struct completion
*x
, long timeout
, int state
)
3993 spin_lock_irq(&x
->wait
.lock
);
3994 timeout
= do_wait_for_common(x
, timeout
, state
);
3995 spin_unlock_irq(&x
->wait
.lock
);
3999 void __sched
wait_for_completion(struct completion
*x
)
4001 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4003 EXPORT_SYMBOL(wait_for_completion
);
4005 unsigned long __sched
4006 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4008 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4010 EXPORT_SYMBOL(wait_for_completion_timeout
);
4012 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4014 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4015 if (t
== -ERESTARTSYS
)
4019 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4021 unsigned long __sched
4022 wait_for_completion_interruptible_timeout(struct completion
*x
,
4023 unsigned long timeout
)
4025 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4027 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4030 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4032 unsigned long flags
;
4035 init_waitqueue_entry(&wait
, current
);
4037 __set_current_state(state
);
4039 spin_lock_irqsave(&q
->lock
, flags
);
4040 __add_wait_queue(q
, &wait
);
4041 spin_unlock(&q
->lock
);
4042 timeout
= schedule_timeout(timeout
);
4043 spin_lock_irq(&q
->lock
);
4044 __remove_wait_queue(q
, &wait
);
4045 spin_unlock_irqrestore(&q
->lock
, flags
);
4050 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4052 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4054 EXPORT_SYMBOL(interruptible_sleep_on
);
4057 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4059 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4061 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4063 void __sched
sleep_on(wait_queue_head_t
*q
)
4065 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4067 EXPORT_SYMBOL(sleep_on
);
4069 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4071 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4073 EXPORT_SYMBOL(sleep_on_timeout
);
4075 #ifdef CONFIG_RT_MUTEXES
4078 * rt_mutex_setprio - set the current priority of a task
4080 * @prio: prio value (kernel-internal form)
4082 * This function changes the 'effective' priority of a task. It does
4083 * not touch ->normal_prio like __setscheduler().
4085 * Used by the rt_mutex code to implement priority inheritance logic.
4087 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4089 unsigned long flags
;
4090 int oldprio
, on_rq
, running
;
4093 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4095 rq
= task_rq_lock(p
, &flags
);
4096 update_rq_clock(rq
);
4099 on_rq
= p
->se
.on_rq
;
4100 running
= task_current(rq
, p
);
4102 dequeue_task(rq
, p
, 0);
4104 p
->sched_class
->put_prev_task(rq
, p
);
4108 p
->sched_class
= &rt_sched_class
;
4110 p
->sched_class
= &fair_sched_class
;
4116 p
->sched_class
->set_curr_task(rq
);
4117 enqueue_task(rq
, p
, 0);
4119 * Reschedule if we are currently running on this runqueue and
4120 * our priority decreased, or if we are not currently running on
4121 * this runqueue and our priority is higher than the current's
4124 if (p
->prio
> oldprio
)
4125 resched_task(rq
->curr
);
4127 check_preempt_curr(rq
, p
);
4130 task_rq_unlock(rq
, &flags
);
4135 void set_user_nice(struct task_struct
*p
, long nice
)
4137 int old_prio
, delta
, on_rq
;
4138 unsigned long flags
;
4141 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4144 * We have to be careful, if called from sys_setpriority(),
4145 * the task might be in the middle of scheduling on another CPU.
4147 rq
= task_rq_lock(p
, &flags
);
4148 update_rq_clock(rq
);
4150 * The RT priorities are set via sched_setscheduler(), but we still
4151 * allow the 'normal' nice value to be set - but as expected
4152 * it wont have any effect on scheduling until the task is
4153 * SCHED_FIFO/SCHED_RR:
4155 if (task_has_rt_policy(p
)) {
4156 p
->static_prio
= NICE_TO_PRIO(nice
);
4159 on_rq
= p
->se
.on_rq
;
4161 dequeue_task(rq
, p
, 0);
4163 p
->static_prio
= NICE_TO_PRIO(nice
);
4166 p
->prio
= effective_prio(p
);
4167 delta
= p
->prio
- old_prio
;
4170 enqueue_task(rq
, p
, 0);
4172 * If the task increased its priority or is running and
4173 * lowered its priority, then reschedule its CPU:
4175 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4176 resched_task(rq
->curr
);
4179 task_rq_unlock(rq
, &flags
);
4181 EXPORT_SYMBOL(set_user_nice
);
4184 * can_nice - check if a task can reduce its nice value
4188 int can_nice(const struct task_struct
*p
, const int nice
)
4190 /* convert nice value [19,-20] to rlimit style value [1,40] */
4191 int nice_rlim
= 20 - nice
;
4193 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4194 capable(CAP_SYS_NICE
));
4197 #ifdef __ARCH_WANT_SYS_NICE
4200 * sys_nice - change the priority of the current process.
4201 * @increment: priority increment
4203 * sys_setpriority is a more generic, but much slower function that
4204 * does similar things.
4206 asmlinkage
long sys_nice(int increment
)
4211 * Setpriority might change our priority at the same moment.
4212 * We don't have to worry. Conceptually one call occurs first
4213 * and we have a single winner.
4215 if (increment
< -40)
4220 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4226 if (increment
< 0 && !can_nice(current
, nice
))
4229 retval
= security_task_setnice(current
, nice
);
4233 set_user_nice(current
, nice
);
4240 * task_prio - return the priority value of a given task.
4241 * @p: the task in question.
4243 * This is the priority value as seen by users in /proc.
4244 * RT tasks are offset by -200. Normal tasks are centered
4245 * around 0, value goes from -16 to +15.
4247 int task_prio(const struct task_struct
*p
)
4249 return p
->prio
- MAX_RT_PRIO
;
4253 * task_nice - return the nice value of a given task.
4254 * @p: the task in question.
4256 int task_nice(const struct task_struct
*p
)
4258 return TASK_NICE(p
);
4260 EXPORT_SYMBOL_GPL(task_nice
);
4263 * idle_cpu - is a given cpu idle currently?
4264 * @cpu: the processor in question.
4266 int idle_cpu(int cpu
)
4268 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4272 * idle_task - return the idle task for a given cpu.
4273 * @cpu: the processor in question.
4275 struct task_struct
*idle_task(int cpu
)
4277 return cpu_rq(cpu
)->idle
;
4281 * find_process_by_pid - find a process with a matching PID value.
4282 * @pid: the pid in question.
4284 static struct task_struct
*find_process_by_pid(pid_t pid
)
4286 return pid
? find_task_by_vpid(pid
) : current
;
4289 /* Actually do priority change: must hold rq lock. */
4291 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4293 BUG_ON(p
->se
.on_rq
);
4296 switch (p
->policy
) {
4300 p
->sched_class
= &fair_sched_class
;
4304 p
->sched_class
= &rt_sched_class
;
4308 p
->rt_priority
= prio
;
4309 p
->normal_prio
= normal_prio(p
);
4310 /* we are holding p->pi_lock already */
4311 p
->prio
= rt_mutex_getprio(p
);
4316 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4317 * @p: the task in question.
4318 * @policy: new policy.
4319 * @param: structure containing the new RT priority.
4321 * NOTE that the task may be already dead.
4323 int sched_setscheduler(struct task_struct
*p
, int policy
,
4324 struct sched_param
*param
)
4326 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4327 unsigned long flags
;
4330 /* may grab non-irq protected spin_locks */
4331 BUG_ON(in_interrupt());
4333 /* double check policy once rq lock held */
4335 policy
= oldpolicy
= p
->policy
;
4336 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4337 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4338 policy
!= SCHED_IDLE
)
4341 * Valid priorities for SCHED_FIFO and SCHED_RR are
4342 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4343 * SCHED_BATCH and SCHED_IDLE is 0.
4345 if (param
->sched_priority
< 0 ||
4346 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4347 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4349 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4353 * Allow unprivileged RT tasks to decrease priority:
4355 if (!capable(CAP_SYS_NICE
)) {
4356 if (rt_policy(policy
)) {
4357 unsigned long rlim_rtprio
;
4359 if (!lock_task_sighand(p
, &flags
))
4361 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4362 unlock_task_sighand(p
, &flags
);
4364 /* can't set/change the rt policy */
4365 if (policy
!= p
->policy
&& !rlim_rtprio
)
4368 /* can't increase priority */
4369 if (param
->sched_priority
> p
->rt_priority
&&
4370 param
->sched_priority
> rlim_rtprio
)
4374 * Like positive nice levels, dont allow tasks to
4375 * move out of SCHED_IDLE either:
4377 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4380 /* can't change other user's priorities */
4381 if ((current
->euid
!= p
->euid
) &&
4382 (current
->euid
!= p
->uid
))
4386 retval
= security_task_setscheduler(p
, policy
, param
);
4390 * make sure no PI-waiters arrive (or leave) while we are
4391 * changing the priority of the task:
4393 spin_lock_irqsave(&p
->pi_lock
, flags
);
4395 * To be able to change p->policy safely, the apropriate
4396 * runqueue lock must be held.
4398 rq
= __task_rq_lock(p
);
4399 /* recheck policy now with rq lock held */
4400 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4401 policy
= oldpolicy
= -1;
4402 __task_rq_unlock(rq
);
4403 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4406 update_rq_clock(rq
);
4407 on_rq
= p
->se
.on_rq
;
4408 running
= task_current(rq
, p
);
4410 deactivate_task(rq
, p
, 0);
4412 p
->sched_class
->put_prev_task(rq
, p
);
4416 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4420 p
->sched_class
->set_curr_task(rq
);
4421 activate_task(rq
, p
, 0);
4423 * Reschedule if we are currently running on this runqueue and
4424 * our priority decreased, or if we are not currently running on
4425 * this runqueue and our priority is higher than the current's
4428 if (p
->prio
> oldprio
)
4429 resched_task(rq
->curr
);
4431 check_preempt_curr(rq
, p
);
4434 __task_rq_unlock(rq
);
4435 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4437 rt_mutex_adjust_pi(p
);
4441 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4444 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4446 struct sched_param lparam
;
4447 struct task_struct
*p
;
4450 if (!param
|| pid
< 0)
4452 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4457 p
= find_process_by_pid(pid
);
4459 retval
= sched_setscheduler(p
, policy
, &lparam
);
4466 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4467 * @pid: the pid in question.
4468 * @policy: new policy.
4469 * @param: structure containing the new RT priority.
4472 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4474 /* negative values for policy are not valid */
4478 return do_sched_setscheduler(pid
, policy
, param
);
4482 * sys_sched_setparam - set/change the RT priority of a thread
4483 * @pid: the pid in question.
4484 * @param: structure containing the new RT priority.
4486 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4488 return do_sched_setscheduler(pid
, -1, param
);
4492 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4493 * @pid: the pid in question.
4495 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4497 struct task_struct
*p
;
4504 read_lock(&tasklist_lock
);
4505 p
= find_process_by_pid(pid
);
4507 retval
= security_task_getscheduler(p
);
4511 read_unlock(&tasklist_lock
);
4516 * sys_sched_getscheduler - get the RT priority of a thread
4517 * @pid: the pid in question.
4518 * @param: structure containing the RT priority.
4520 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4522 struct sched_param lp
;
4523 struct task_struct
*p
;
4526 if (!param
|| pid
< 0)
4529 read_lock(&tasklist_lock
);
4530 p
= find_process_by_pid(pid
);
4535 retval
= security_task_getscheduler(p
);
4539 lp
.sched_priority
= p
->rt_priority
;
4540 read_unlock(&tasklist_lock
);
4543 * This one might sleep, we cannot do it with a spinlock held ...
4545 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4550 read_unlock(&tasklist_lock
);
4554 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4556 cpumask_t cpus_allowed
;
4557 struct task_struct
*p
;
4561 read_lock(&tasklist_lock
);
4563 p
= find_process_by_pid(pid
);
4565 read_unlock(&tasklist_lock
);
4571 * It is not safe to call set_cpus_allowed with the
4572 * tasklist_lock held. We will bump the task_struct's
4573 * usage count and then drop tasklist_lock.
4576 read_unlock(&tasklist_lock
);
4579 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4580 !capable(CAP_SYS_NICE
))
4583 retval
= security_task_setscheduler(p
, 0, NULL
);
4587 cpus_allowed
= cpuset_cpus_allowed(p
);
4588 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4590 retval
= set_cpus_allowed(p
, new_mask
);
4593 cpus_allowed
= cpuset_cpus_allowed(p
);
4594 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4596 * We must have raced with a concurrent cpuset
4597 * update. Just reset the cpus_allowed to the
4598 * cpuset's cpus_allowed
4600 new_mask
= cpus_allowed
;
4610 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4611 cpumask_t
*new_mask
)
4613 if (len
< sizeof(cpumask_t
)) {
4614 memset(new_mask
, 0, sizeof(cpumask_t
));
4615 } else if (len
> sizeof(cpumask_t
)) {
4616 len
= sizeof(cpumask_t
);
4618 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4622 * sys_sched_setaffinity - set the cpu affinity of a process
4623 * @pid: pid of the process
4624 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4625 * @user_mask_ptr: user-space pointer to the new cpu mask
4627 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4628 unsigned long __user
*user_mask_ptr
)
4633 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4637 return sched_setaffinity(pid
, new_mask
);
4641 * Represents all cpu's present in the system
4642 * In systems capable of hotplug, this map could dynamically grow
4643 * as new cpu's are detected in the system via any platform specific
4644 * method, such as ACPI for e.g.
4647 cpumask_t cpu_present_map __read_mostly
;
4648 EXPORT_SYMBOL(cpu_present_map
);
4651 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4652 EXPORT_SYMBOL(cpu_online_map
);
4654 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4655 EXPORT_SYMBOL(cpu_possible_map
);
4658 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4660 struct task_struct
*p
;
4664 read_lock(&tasklist_lock
);
4667 p
= find_process_by_pid(pid
);
4671 retval
= security_task_getscheduler(p
);
4675 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4678 read_unlock(&tasklist_lock
);
4685 * sys_sched_getaffinity - get the cpu affinity of a process
4686 * @pid: pid of the process
4687 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4688 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4690 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4691 unsigned long __user
*user_mask_ptr
)
4696 if (len
< sizeof(cpumask_t
))
4699 ret
= sched_getaffinity(pid
, &mask
);
4703 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4706 return sizeof(cpumask_t
);
4710 * sys_sched_yield - yield the current processor to other threads.
4712 * This function yields the current CPU to other tasks. If there are no
4713 * other threads running on this CPU then this function will return.
4715 asmlinkage
long sys_sched_yield(void)
4717 struct rq
*rq
= this_rq_lock();
4719 schedstat_inc(rq
, yld_count
);
4720 current
->sched_class
->yield_task(rq
);
4723 * Since we are going to call schedule() anyway, there's
4724 * no need to preempt or enable interrupts:
4726 __release(rq
->lock
);
4727 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4728 _raw_spin_unlock(&rq
->lock
);
4729 preempt_enable_no_resched();
4736 static void __cond_resched(void)
4738 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4739 __might_sleep(__FILE__
, __LINE__
);
4742 * The BKS might be reacquired before we have dropped
4743 * PREEMPT_ACTIVE, which could trigger a second
4744 * cond_resched() call.
4747 add_preempt_count(PREEMPT_ACTIVE
);
4749 sub_preempt_count(PREEMPT_ACTIVE
);
4750 } while (need_resched());
4753 int __sched
cond_resched(void)
4755 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4756 system_state
== SYSTEM_RUNNING
) {
4762 EXPORT_SYMBOL(cond_resched
);
4765 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4766 * call schedule, and on return reacquire the lock.
4768 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4769 * operations here to prevent schedule() from being called twice (once via
4770 * spin_unlock(), once by hand).
4772 int cond_resched_lock(spinlock_t
*lock
)
4776 if (need_lockbreak(lock
)) {
4782 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4783 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4784 _raw_spin_unlock(lock
);
4785 preempt_enable_no_resched();
4792 EXPORT_SYMBOL(cond_resched_lock
);
4794 int __sched
cond_resched_softirq(void)
4796 BUG_ON(!in_softirq());
4798 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4806 EXPORT_SYMBOL(cond_resched_softirq
);
4809 * yield - yield the current processor to other threads.
4811 * This is a shortcut for kernel-space yielding - it marks the
4812 * thread runnable and calls sys_sched_yield().
4814 void __sched
yield(void)
4816 set_current_state(TASK_RUNNING
);
4819 EXPORT_SYMBOL(yield
);
4822 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4823 * that process accounting knows that this is a task in IO wait state.
4825 * But don't do that if it is a deliberate, throttling IO wait (this task
4826 * has set its backing_dev_info: the queue against which it should throttle)
4828 void __sched
io_schedule(void)
4830 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4832 delayacct_blkio_start();
4833 atomic_inc(&rq
->nr_iowait
);
4835 atomic_dec(&rq
->nr_iowait
);
4836 delayacct_blkio_end();
4838 EXPORT_SYMBOL(io_schedule
);
4840 long __sched
io_schedule_timeout(long timeout
)
4842 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4845 delayacct_blkio_start();
4846 atomic_inc(&rq
->nr_iowait
);
4847 ret
= schedule_timeout(timeout
);
4848 atomic_dec(&rq
->nr_iowait
);
4849 delayacct_blkio_end();
4854 * sys_sched_get_priority_max - return maximum RT priority.
4855 * @policy: scheduling class.
4857 * this syscall returns the maximum rt_priority that can be used
4858 * by a given scheduling class.
4860 asmlinkage
long sys_sched_get_priority_max(int policy
)
4867 ret
= MAX_USER_RT_PRIO
-1;
4879 * sys_sched_get_priority_min - return minimum RT priority.
4880 * @policy: scheduling class.
4882 * this syscall returns the minimum rt_priority that can be used
4883 * by a given scheduling class.
4885 asmlinkage
long sys_sched_get_priority_min(int policy
)
4903 * sys_sched_rr_get_interval - return the default timeslice of a process.
4904 * @pid: pid of the process.
4905 * @interval: userspace pointer to the timeslice value.
4907 * this syscall writes the default timeslice value of a given process
4908 * into the user-space timespec buffer. A value of '0' means infinity.
4911 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4913 struct task_struct
*p
;
4914 unsigned int time_slice
;
4922 read_lock(&tasklist_lock
);
4923 p
= find_process_by_pid(pid
);
4927 retval
= security_task_getscheduler(p
);
4932 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4933 * tasks that are on an otherwise idle runqueue:
4936 if (p
->policy
== SCHED_RR
) {
4937 time_slice
= DEF_TIMESLICE
;
4939 struct sched_entity
*se
= &p
->se
;
4940 unsigned long flags
;
4943 rq
= task_rq_lock(p
, &flags
);
4944 if (rq
->cfs
.load
.weight
)
4945 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
4946 task_rq_unlock(rq
, &flags
);
4948 read_unlock(&tasklist_lock
);
4949 jiffies_to_timespec(time_slice
, &t
);
4950 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4954 read_unlock(&tasklist_lock
);
4958 static const char stat_nam
[] = "RSDTtZX";
4960 void sched_show_task(struct task_struct
*p
)
4962 unsigned long free
= 0;
4965 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4966 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
4967 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4968 #if BITS_PER_LONG == 32
4969 if (state
== TASK_RUNNING
)
4970 printk(KERN_CONT
" running ");
4972 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4974 if (state
== TASK_RUNNING
)
4975 printk(KERN_CONT
" running task ");
4977 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4979 #ifdef CONFIG_DEBUG_STACK_USAGE
4981 unsigned long *n
= end_of_stack(p
);
4984 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4987 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
4988 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
4990 if (state
!= TASK_RUNNING
)
4991 show_stack(p
, NULL
);
4994 void show_state_filter(unsigned long state_filter
)
4996 struct task_struct
*g
, *p
;
4998 #if BITS_PER_LONG == 32
5000 " task PC stack pid father\n");
5003 " task PC stack pid father\n");
5005 read_lock(&tasklist_lock
);
5006 do_each_thread(g
, p
) {
5008 * reset the NMI-timeout, listing all files on a slow
5009 * console might take alot of time:
5011 touch_nmi_watchdog();
5012 if (!state_filter
|| (p
->state
& state_filter
))
5014 } while_each_thread(g
, p
);
5016 touch_all_softlockup_watchdogs();
5018 #ifdef CONFIG_SCHED_DEBUG
5019 sysrq_sched_debug_show();
5021 read_unlock(&tasklist_lock
);
5023 * Only show locks if all tasks are dumped:
5025 if (state_filter
== -1)
5026 debug_show_all_locks();
5029 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5031 idle
->sched_class
= &idle_sched_class
;
5035 * init_idle - set up an idle thread for a given CPU
5036 * @idle: task in question
5037 * @cpu: cpu the idle task belongs to
5039 * NOTE: this function does not set the idle thread's NEED_RESCHED
5040 * flag, to make booting more robust.
5042 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5044 struct rq
*rq
= cpu_rq(cpu
);
5045 unsigned long flags
;
5048 idle
->se
.exec_start
= sched_clock();
5050 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5051 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5052 __set_task_cpu(idle
, cpu
);
5054 spin_lock_irqsave(&rq
->lock
, flags
);
5055 rq
->curr
= rq
->idle
= idle
;
5056 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5059 spin_unlock_irqrestore(&rq
->lock
, flags
);
5061 /* Set the preempt count _outside_ the spinlocks! */
5062 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
5063 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5065 task_thread_info(idle
)->preempt_count
= 0;
5068 * The idle tasks have their own, simple scheduling class:
5070 idle
->sched_class
= &idle_sched_class
;
5074 * In a system that switches off the HZ timer nohz_cpu_mask
5075 * indicates which cpus entered this state. This is used
5076 * in the rcu update to wait only for active cpus. For system
5077 * which do not switch off the HZ timer nohz_cpu_mask should
5078 * always be CPU_MASK_NONE.
5080 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5083 * Increase the granularity value when there are more CPUs,
5084 * because with more CPUs the 'effective latency' as visible
5085 * to users decreases. But the relationship is not linear,
5086 * so pick a second-best guess by going with the log2 of the
5089 * This idea comes from the SD scheduler of Con Kolivas:
5091 static inline void sched_init_granularity(void)
5093 unsigned int factor
= 1 + ilog2(num_online_cpus());
5094 const unsigned long limit
= 200000000;
5096 sysctl_sched_min_granularity
*= factor
;
5097 if (sysctl_sched_min_granularity
> limit
)
5098 sysctl_sched_min_granularity
= limit
;
5100 sysctl_sched_latency
*= factor
;
5101 if (sysctl_sched_latency
> limit
)
5102 sysctl_sched_latency
= limit
;
5104 sysctl_sched_wakeup_granularity
*= factor
;
5105 sysctl_sched_batch_wakeup_granularity
*= factor
;
5110 * This is how migration works:
5112 * 1) we queue a struct migration_req structure in the source CPU's
5113 * runqueue and wake up that CPU's migration thread.
5114 * 2) we down() the locked semaphore => thread blocks.
5115 * 3) migration thread wakes up (implicitly it forces the migrated
5116 * thread off the CPU)
5117 * 4) it gets the migration request and checks whether the migrated
5118 * task is still in the wrong runqueue.
5119 * 5) if it's in the wrong runqueue then the migration thread removes
5120 * it and puts it into the right queue.
5121 * 6) migration thread up()s the semaphore.
5122 * 7) we wake up and the migration is done.
5126 * Change a given task's CPU affinity. Migrate the thread to a
5127 * proper CPU and schedule it away if the CPU it's executing on
5128 * is removed from the allowed bitmask.
5130 * NOTE: the caller must have a valid reference to the task, the
5131 * task must not exit() & deallocate itself prematurely. The
5132 * call is not atomic; no spinlocks may be held.
5134 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5136 struct migration_req req
;
5137 unsigned long flags
;
5141 rq
= task_rq_lock(p
, &flags
);
5142 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5147 p
->cpus_allowed
= new_mask
;
5148 /* Can the task run on the task's current CPU? If so, we're done */
5149 if (cpu_isset(task_cpu(p
), new_mask
))
5152 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5153 /* Need help from migration thread: drop lock and wait. */
5154 task_rq_unlock(rq
, &flags
);
5155 wake_up_process(rq
->migration_thread
);
5156 wait_for_completion(&req
.done
);
5157 tlb_migrate_finish(p
->mm
);
5161 task_rq_unlock(rq
, &flags
);
5165 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5168 * Move (not current) task off this cpu, onto dest cpu. We're doing
5169 * this because either it can't run here any more (set_cpus_allowed()
5170 * away from this CPU, or CPU going down), or because we're
5171 * attempting to rebalance this task on exec (sched_exec).
5173 * So we race with normal scheduler movements, but that's OK, as long
5174 * as the task is no longer on this CPU.
5176 * Returns non-zero if task was successfully migrated.
5178 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5180 struct rq
*rq_dest
, *rq_src
;
5183 if (unlikely(cpu_is_offline(dest_cpu
)))
5186 rq_src
= cpu_rq(src_cpu
);
5187 rq_dest
= cpu_rq(dest_cpu
);
5189 double_rq_lock(rq_src
, rq_dest
);
5190 /* Already moved. */
5191 if (task_cpu(p
) != src_cpu
)
5193 /* Affinity changed (again). */
5194 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5197 on_rq
= p
->se
.on_rq
;
5199 deactivate_task(rq_src
, p
, 0);
5201 set_task_cpu(p
, dest_cpu
);
5203 activate_task(rq_dest
, p
, 0);
5204 check_preempt_curr(rq_dest
, p
);
5208 double_rq_unlock(rq_src
, rq_dest
);
5213 * migration_thread - this is a highprio system thread that performs
5214 * thread migration by bumping thread off CPU then 'pushing' onto
5217 static int migration_thread(void *data
)
5219 int cpu
= (long)data
;
5223 BUG_ON(rq
->migration_thread
!= current
);
5225 set_current_state(TASK_INTERRUPTIBLE
);
5226 while (!kthread_should_stop()) {
5227 struct migration_req
*req
;
5228 struct list_head
*head
;
5230 spin_lock_irq(&rq
->lock
);
5232 if (cpu_is_offline(cpu
)) {
5233 spin_unlock_irq(&rq
->lock
);
5237 if (rq
->active_balance
) {
5238 active_load_balance(rq
, cpu
);
5239 rq
->active_balance
= 0;
5242 head
= &rq
->migration_queue
;
5244 if (list_empty(head
)) {
5245 spin_unlock_irq(&rq
->lock
);
5247 set_current_state(TASK_INTERRUPTIBLE
);
5250 req
= list_entry(head
->next
, struct migration_req
, list
);
5251 list_del_init(head
->next
);
5253 spin_unlock(&rq
->lock
);
5254 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5257 complete(&req
->done
);
5259 __set_current_state(TASK_RUNNING
);
5263 /* Wait for kthread_stop */
5264 set_current_state(TASK_INTERRUPTIBLE
);
5265 while (!kthread_should_stop()) {
5267 set_current_state(TASK_INTERRUPTIBLE
);
5269 __set_current_state(TASK_RUNNING
);
5273 #ifdef CONFIG_HOTPLUG_CPU
5275 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5279 local_irq_disable();
5280 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5286 * Figure out where task on dead CPU should go, use force if necessary.
5287 * NOTE: interrupts should be disabled by the caller
5289 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5291 unsigned long flags
;
5298 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5299 cpus_and(mask
, mask
, p
->cpus_allowed
);
5300 dest_cpu
= any_online_cpu(mask
);
5302 /* On any allowed CPU? */
5303 if (dest_cpu
== NR_CPUS
)
5304 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5306 /* No more Mr. Nice Guy. */
5307 if (dest_cpu
== NR_CPUS
) {
5308 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5310 * Try to stay on the same cpuset, where the
5311 * current cpuset may be a subset of all cpus.
5312 * The cpuset_cpus_allowed_locked() variant of
5313 * cpuset_cpus_allowed() will not block. It must be
5314 * called within calls to cpuset_lock/cpuset_unlock.
5316 rq
= task_rq_lock(p
, &flags
);
5317 p
->cpus_allowed
= cpus_allowed
;
5318 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5319 task_rq_unlock(rq
, &flags
);
5322 * Don't tell them about moving exiting tasks or
5323 * kernel threads (both mm NULL), since they never
5326 if (p
->mm
&& printk_ratelimit()) {
5327 printk(KERN_INFO
"process %d (%s) no "
5328 "longer affine to cpu%d\n",
5329 task_pid_nr(p
), p
->comm
, dead_cpu
);
5332 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5336 * While a dead CPU has no uninterruptible tasks queued at this point,
5337 * it might still have a nonzero ->nr_uninterruptible counter, because
5338 * for performance reasons the counter is not stricly tracking tasks to
5339 * their home CPUs. So we just add the counter to another CPU's counter,
5340 * to keep the global sum constant after CPU-down:
5342 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5344 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5345 unsigned long flags
;
5347 local_irq_save(flags
);
5348 double_rq_lock(rq_src
, rq_dest
);
5349 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5350 rq_src
->nr_uninterruptible
= 0;
5351 double_rq_unlock(rq_src
, rq_dest
);
5352 local_irq_restore(flags
);
5355 /* Run through task list and migrate tasks from the dead cpu. */
5356 static void migrate_live_tasks(int src_cpu
)
5358 struct task_struct
*p
, *t
;
5360 read_lock(&tasklist_lock
);
5362 do_each_thread(t
, p
) {
5366 if (task_cpu(p
) == src_cpu
)
5367 move_task_off_dead_cpu(src_cpu
, p
);
5368 } while_each_thread(t
, p
);
5370 read_unlock(&tasklist_lock
);
5374 * Schedules idle task to be the next runnable task on current CPU.
5375 * It does so by boosting its priority to highest possible.
5376 * Used by CPU offline code.
5378 void sched_idle_next(void)
5380 int this_cpu
= smp_processor_id();
5381 struct rq
*rq
= cpu_rq(this_cpu
);
5382 struct task_struct
*p
= rq
->idle
;
5383 unsigned long flags
;
5385 /* cpu has to be offline */
5386 BUG_ON(cpu_online(this_cpu
));
5389 * Strictly not necessary since rest of the CPUs are stopped by now
5390 * and interrupts disabled on the current cpu.
5392 spin_lock_irqsave(&rq
->lock
, flags
);
5394 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5396 update_rq_clock(rq
);
5397 activate_task(rq
, p
, 0);
5399 spin_unlock_irqrestore(&rq
->lock
, flags
);
5403 * Ensures that the idle task is using init_mm right before its cpu goes
5406 void idle_task_exit(void)
5408 struct mm_struct
*mm
= current
->active_mm
;
5410 BUG_ON(cpu_online(smp_processor_id()));
5413 switch_mm(mm
, &init_mm
, current
);
5417 /* called under rq->lock with disabled interrupts */
5418 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5420 struct rq
*rq
= cpu_rq(dead_cpu
);
5422 /* Must be exiting, otherwise would be on tasklist. */
5423 BUG_ON(!p
->exit_state
);
5425 /* Cannot have done final schedule yet: would have vanished. */
5426 BUG_ON(p
->state
== TASK_DEAD
);
5431 * Drop lock around migration; if someone else moves it,
5432 * that's OK. No task can be added to this CPU, so iteration is
5435 spin_unlock_irq(&rq
->lock
);
5436 move_task_off_dead_cpu(dead_cpu
, p
);
5437 spin_lock_irq(&rq
->lock
);
5442 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5443 static void migrate_dead_tasks(unsigned int dead_cpu
)
5445 struct rq
*rq
= cpu_rq(dead_cpu
);
5446 struct task_struct
*next
;
5449 if (!rq
->nr_running
)
5451 update_rq_clock(rq
);
5452 next
= pick_next_task(rq
, rq
->curr
);
5455 migrate_dead(dead_cpu
, next
);
5459 #endif /* CONFIG_HOTPLUG_CPU */
5461 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5463 static struct ctl_table sd_ctl_dir
[] = {
5465 .procname
= "sched_domain",
5471 static struct ctl_table sd_ctl_root
[] = {
5473 .ctl_name
= CTL_KERN
,
5474 .procname
= "kernel",
5476 .child
= sd_ctl_dir
,
5481 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5483 struct ctl_table
*entry
=
5484 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5489 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5491 struct ctl_table
*entry
;
5494 * In the intermediate directories, both the child directory and
5495 * procname are dynamically allocated and could fail but the mode
5496 * will always be set. In the lowest directory the names are
5497 * static strings and all have proc handlers.
5499 for (entry
= *tablep
; entry
->mode
; entry
++) {
5501 sd_free_ctl_entry(&entry
->child
);
5502 if (entry
->proc_handler
== NULL
)
5503 kfree(entry
->procname
);
5511 set_table_entry(struct ctl_table
*entry
,
5512 const char *procname
, void *data
, int maxlen
,
5513 mode_t mode
, proc_handler
*proc_handler
)
5515 entry
->procname
= procname
;
5517 entry
->maxlen
= maxlen
;
5519 entry
->proc_handler
= proc_handler
;
5522 static struct ctl_table
*
5523 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5525 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5530 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5531 sizeof(long), 0644, proc_doulongvec_minmax
);
5532 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5533 sizeof(long), 0644, proc_doulongvec_minmax
);
5534 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5535 sizeof(int), 0644, proc_dointvec_minmax
);
5536 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5537 sizeof(int), 0644, proc_dointvec_minmax
);
5538 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5539 sizeof(int), 0644, proc_dointvec_minmax
);
5540 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5541 sizeof(int), 0644, proc_dointvec_minmax
);
5542 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5543 sizeof(int), 0644, proc_dointvec_minmax
);
5544 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5545 sizeof(int), 0644, proc_dointvec_minmax
);
5546 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5547 sizeof(int), 0644, proc_dointvec_minmax
);
5548 set_table_entry(&table
[9], "cache_nice_tries",
5549 &sd
->cache_nice_tries
,
5550 sizeof(int), 0644, proc_dointvec_minmax
);
5551 set_table_entry(&table
[10], "flags", &sd
->flags
,
5552 sizeof(int), 0644, proc_dointvec_minmax
);
5553 /* &table[11] is terminator */
5558 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5560 struct ctl_table
*entry
, *table
;
5561 struct sched_domain
*sd
;
5562 int domain_num
= 0, i
;
5565 for_each_domain(cpu
, sd
)
5567 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5572 for_each_domain(cpu
, sd
) {
5573 snprintf(buf
, 32, "domain%d", i
);
5574 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5576 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5583 static struct ctl_table_header
*sd_sysctl_header
;
5584 static void register_sched_domain_sysctl(void)
5586 int i
, cpu_num
= num_online_cpus();
5587 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5590 WARN_ON(sd_ctl_dir
[0].child
);
5591 sd_ctl_dir
[0].child
= entry
;
5596 for_each_online_cpu(i
) {
5597 snprintf(buf
, 32, "cpu%d", i
);
5598 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5600 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5604 WARN_ON(sd_sysctl_header
);
5605 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5608 /* may be called multiple times per register */
5609 static void unregister_sched_domain_sysctl(void)
5611 if (sd_sysctl_header
)
5612 unregister_sysctl_table(sd_sysctl_header
);
5613 sd_sysctl_header
= NULL
;
5614 if (sd_ctl_dir
[0].child
)
5615 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5618 static void register_sched_domain_sysctl(void)
5621 static void unregister_sched_domain_sysctl(void)
5627 * migration_call - callback that gets triggered when a CPU is added.
5628 * Here we can start up the necessary migration thread for the new CPU.
5630 static int __cpuinit
5631 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5633 struct task_struct
*p
;
5634 int cpu
= (long)hcpu
;
5635 unsigned long flags
;
5640 case CPU_UP_PREPARE
:
5641 case CPU_UP_PREPARE_FROZEN
:
5642 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5645 kthread_bind(p
, cpu
);
5646 /* Must be high prio: stop_machine expects to yield to it. */
5647 rq
= task_rq_lock(p
, &flags
);
5648 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5649 task_rq_unlock(rq
, &flags
);
5650 cpu_rq(cpu
)->migration_thread
= p
;
5654 case CPU_ONLINE_FROZEN
:
5655 /* Strictly unnecessary, as first user will wake it. */
5656 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5659 #ifdef CONFIG_HOTPLUG_CPU
5660 case CPU_UP_CANCELED
:
5661 case CPU_UP_CANCELED_FROZEN
:
5662 if (!cpu_rq(cpu
)->migration_thread
)
5664 /* Unbind it from offline cpu so it can run. Fall thru. */
5665 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5666 any_online_cpu(cpu_online_map
));
5667 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5668 cpu_rq(cpu
)->migration_thread
= NULL
;
5672 case CPU_DEAD_FROZEN
:
5673 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5674 migrate_live_tasks(cpu
);
5676 kthread_stop(rq
->migration_thread
);
5677 rq
->migration_thread
= NULL
;
5678 /* Idle task back to normal (off runqueue, low prio) */
5679 spin_lock_irq(&rq
->lock
);
5680 update_rq_clock(rq
);
5681 deactivate_task(rq
, rq
->idle
, 0);
5682 rq
->idle
->static_prio
= MAX_PRIO
;
5683 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5684 rq
->idle
->sched_class
= &idle_sched_class
;
5685 migrate_dead_tasks(cpu
);
5686 spin_unlock_irq(&rq
->lock
);
5688 migrate_nr_uninterruptible(rq
);
5689 BUG_ON(rq
->nr_running
!= 0);
5692 * No need to migrate the tasks: it was best-effort if
5693 * they didn't take sched_hotcpu_mutex. Just wake up
5696 spin_lock_irq(&rq
->lock
);
5697 while (!list_empty(&rq
->migration_queue
)) {
5698 struct migration_req
*req
;
5700 req
= list_entry(rq
->migration_queue
.next
,
5701 struct migration_req
, list
);
5702 list_del_init(&req
->list
);
5703 complete(&req
->done
);
5705 spin_unlock_irq(&rq
->lock
);
5712 /* Register at highest priority so that task migration (migrate_all_tasks)
5713 * happens before everything else.
5715 static struct notifier_block __cpuinitdata migration_notifier
= {
5716 .notifier_call
= migration_call
,
5720 void __init
migration_init(void)
5722 void *cpu
= (void *)(long)smp_processor_id();
5725 /* Start one for the boot CPU: */
5726 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5727 BUG_ON(err
== NOTIFY_BAD
);
5728 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5729 register_cpu_notifier(&migration_notifier
);
5735 /* Number of possible processor ids */
5736 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5737 EXPORT_SYMBOL(nr_cpu_ids
);
5739 #ifdef CONFIG_SCHED_DEBUG
5741 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
5743 struct sched_group
*group
= sd
->groups
;
5744 cpumask_t groupmask
;
5747 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5748 cpus_clear(groupmask
);
5750 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5752 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5753 printk("does not load-balance\n");
5755 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5760 printk(KERN_CONT
"span %s\n", str
);
5762 if (!cpu_isset(cpu
, sd
->span
)) {
5763 printk(KERN_ERR
"ERROR: domain->span does not contain "
5766 if (!cpu_isset(cpu
, group
->cpumask
)) {
5767 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5771 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5775 printk(KERN_ERR
"ERROR: group is NULL\n");
5779 if (!group
->__cpu_power
) {
5780 printk(KERN_CONT
"\n");
5781 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5786 if (!cpus_weight(group
->cpumask
)) {
5787 printk(KERN_CONT
"\n");
5788 printk(KERN_ERR
"ERROR: empty group\n");
5792 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5793 printk(KERN_CONT
"\n");
5794 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5798 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5800 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5801 printk(KERN_CONT
" %s", str
);
5803 group
= group
->next
;
5804 } while (group
!= sd
->groups
);
5805 printk(KERN_CONT
"\n");
5807 if (!cpus_equal(sd
->span
, groupmask
))
5808 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5810 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
5811 printk(KERN_ERR
"ERROR: parent span is not a superset "
5812 "of domain->span\n");
5816 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5821 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5825 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5828 if (sched_domain_debug_one(sd
, cpu
, level
))
5837 # define sched_domain_debug(sd, cpu) do { } while (0)
5840 static int sd_degenerate(struct sched_domain
*sd
)
5842 if (cpus_weight(sd
->span
) == 1)
5845 /* Following flags need at least 2 groups */
5846 if (sd
->flags
& (SD_LOAD_BALANCE
|
5847 SD_BALANCE_NEWIDLE
|
5851 SD_SHARE_PKG_RESOURCES
)) {
5852 if (sd
->groups
!= sd
->groups
->next
)
5856 /* Following flags don't use groups */
5857 if (sd
->flags
& (SD_WAKE_IDLE
|
5866 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5868 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5870 if (sd_degenerate(parent
))
5873 if (!cpus_equal(sd
->span
, parent
->span
))
5876 /* Does parent contain flags not in child? */
5877 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5878 if (cflags
& SD_WAKE_AFFINE
)
5879 pflags
&= ~SD_WAKE_BALANCE
;
5880 /* Flags needing groups don't count if only 1 group in parent */
5881 if (parent
->groups
== parent
->groups
->next
) {
5882 pflags
&= ~(SD_LOAD_BALANCE
|
5883 SD_BALANCE_NEWIDLE
|
5887 SD_SHARE_PKG_RESOURCES
);
5889 if (~cflags
& pflags
)
5896 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5897 * hold the hotplug lock.
5899 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5901 struct rq
*rq
= cpu_rq(cpu
);
5902 struct sched_domain
*tmp
;
5904 /* Remove the sched domains which do not contribute to scheduling. */
5905 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5906 struct sched_domain
*parent
= tmp
->parent
;
5909 if (sd_parent_degenerate(tmp
, parent
)) {
5910 tmp
->parent
= parent
->parent
;
5912 parent
->parent
->child
= tmp
;
5916 if (sd
&& sd_degenerate(sd
)) {
5922 sched_domain_debug(sd
, cpu
);
5924 rcu_assign_pointer(rq
->sd
, sd
);
5927 /* cpus with isolated domains */
5928 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5930 /* Setup the mask of cpus configured for isolated domains */
5931 static int __init
isolated_cpu_setup(char *str
)
5933 int ints
[NR_CPUS
], i
;
5935 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5936 cpus_clear(cpu_isolated_map
);
5937 for (i
= 1; i
<= ints
[0]; i
++)
5938 if (ints
[i
] < NR_CPUS
)
5939 cpu_set(ints
[i
], cpu_isolated_map
);
5943 __setup("isolcpus=", isolated_cpu_setup
);
5946 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5947 * to a function which identifies what group(along with sched group) a CPU
5948 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5949 * (due to the fact that we keep track of groups covered with a cpumask_t).
5951 * init_sched_build_groups will build a circular linked list of the groups
5952 * covered by the given span, and will set each group's ->cpumask correctly,
5953 * and ->cpu_power to 0.
5956 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5957 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5958 struct sched_group
**sg
))
5960 struct sched_group
*first
= NULL
, *last
= NULL
;
5961 cpumask_t covered
= CPU_MASK_NONE
;
5964 for_each_cpu_mask(i
, span
) {
5965 struct sched_group
*sg
;
5966 int group
= group_fn(i
, cpu_map
, &sg
);
5969 if (cpu_isset(i
, covered
))
5972 sg
->cpumask
= CPU_MASK_NONE
;
5973 sg
->__cpu_power
= 0;
5975 for_each_cpu_mask(j
, span
) {
5976 if (group_fn(j
, cpu_map
, NULL
) != group
)
5979 cpu_set(j
, covered
);
5980 cpu_set(j
, sg
->cpumask
);
5991 #define SD_NODES_PER_DOMAIN 16
5996 * find_next_best_node - find the next node to include in a sched_domain
5997 * @node: node whose sched_domain we're building
5998 * @used_nodes: nodes already in the sched_domain
6000 * Find the next node to include in a given scheduling domain. Simply
6001 * finds the closest node not already in the @used_nodes map.
6003 * Should use nodemask_t.
6005 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6007 int i
, n
, val
, min_val
, best_node
= 0;
6011 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6012 /* Start at @node */
6013 n
= (node
+ i
) % MAX_NUMNODES
;
6015 if (!nr_cpus_node(n
))
6018 /* Skip already used nodes */
6019 if (test_bit(n
, used_nodes
))
6022 /* Simple min distance search */
6023 val
= node_distance(node
, n
);
6025 if (val
< min_val
) {
6031 set_bit(best_node
, used_nodes
);
6036 * sched_domain_node_span - get a cpumask for a node's sched_domain
6037 * @node: node whose cpumask we're constructing
6038 * @size: number of nodes to include in this span
6040 * Given a node, construct a good cpumask for its sched_domain to span. It
6041 * should be one that prevents unnecessary balancing, but also spreads tasks
6044 static cpumask_t
sched_domain_node_span(int node
)
6046 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6047 cpumask_t span
, nodemask
;
6051 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6053 nodemask
= node_to_cpumask(node
);
6054 cpus_or(span
, span
, nodemask
);
6055 set_bit(node
, used_nodes
);
6057 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6058 int next_node
= find_next_best_node(node
, used_nodes
);
6060 nodemask
= node_to_cpumask(next_node
);
6061 cpus_or(span
, span
, nodemask
);
6068 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6071 * SMT sched-domains:
6073 #ifdef CONFIG_SCHED_SMT
6074 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6075 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6078 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6081 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6087 * multi-core sched-domains:
6089 #ifdef CONFIG_SCHED_MC
6090 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6091 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6094 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6096 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6099 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6100 cpus_and(mask
, mask
, *cpu_map
);
6101 group
= first_cpu(mask
);
6103 *sg
= &per_cpu(sched_group_core
, group
);
6106 #elif defined(CONFIG_SCHED_MC)
6108 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6111 *sg
= &per_cpu(sched_group_core
, cpu
);
6116 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6117 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6120 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6123 #ifdef CONFIG_SCHED_MC
6124 cpumask_t mask
= cpu_coregroup_map(cpu
);
6125 cpus_and(mask
, mask
, *cpu_map
);
6126 group
= first_cpu(mask
);
6127 #elif defined(CONFIG_SCHED_SMT)
6128 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6129 cpus_and(mask
, mask
, *cpu_map
);
6130 group
= first_cpu(mask
);
6135 *sg
= &per_cpu(sched_group_phys
, group
);
6141 * The init_sched_build_groups can't handle what we want to do with node
6142 * groups, so roll our own. Now each node has its own list of groups which
6143 * gets dynamically allocated.
6145 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6146 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6148 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6149 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6151 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6152 struct sched_group
**sg
)
6154 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6157 cpus_and(nodemask
, nodemask
, *cpu_map
);
6158 group
= first_cpu(nodemask
);
6161 *sg
= &per_cpu(sched_group_allnodes
, group
);
6165 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6167 struct sched_group
*sg
= group_head
;
6173 for_each_cpu_mask(j
, sg
->cpumask
) {
6174 struct sched_domain
*sd
;
6176 sd
= &per_cpu(phys_domains
, j
);
6177 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6179 * Only add "power" once for each
6185 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6188 } while (sg
!= group_head
);
6193 /* Free memory allocated for various sched_group structures */
6194 static void free_sched_groups(const cpumask_t
*cpu_map
)
6198 for_each_cpu_mask(cpu
, *cpu_map
) {
6199 struct sched_group
**sched_group_nodes
6200 = sched_group_nodes_bycpu
[cpu
];
6202 if (!sched_group_nodes
)
6205 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6206 cpumask_t nodemask
= node_to_cpumask(i
);
6207 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6209 cpus_and(nodemask
, nodemask
, *cpu_map
);
6210 if (cpus_empty(nodemask
))
6220 if (oldsg
!= sched_group_nodes
[i
])
6223 kfree(sched_group_nodes
);
6224 sched_group_nodes_bycpu
[cpu
] = NULL
;
6228 static void free_sched_groups(const cpumask_t
*cpu_map
)
6234 * Initialize sched groups cpu_power.
6236 * cpu_power indicates the capacity of sched group, which is used while
6237 * distributing the load between different sched groups in a sched domain.
6238 * Typically cpu_power for all the groups in a sched domain will be same unless
6239 * there are asymmetries in the topology. If there are asymmetries, group
6240 * having more cpu_power will pickup more load compared to the group having
6243 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6244 * the maximum number of tasks a group can handle in the presence of other idle
6245 * or lightly loaded groups in the same sched domain.
6247 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6249 struct sched_domain
*child
;
6250 struct sched_group
*group
;
6252 WARN_ON(!sd
|| !sd
->groups
);
6254 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6259 sd
->groups
->__cpu_power
= 0;
6262 * For perf policy, if the groups in child domain share resources
6263 * (for example cores sharing some portions of the cache hierarchy
6264 * or SMT), then set this domain groups cpu_power such that each group
6265 * can handle only one task, when there are other idle groups in the
6266 * same sched domain.
6268 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6270 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6271 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6276 * add cpu_power of each child group to this groups cpu_power
6278 group
= child
->groups
;
6280 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6281 group
= group
->next
;
6282 } while (group
!= child
->groups
);
6286 * Build sched domains for a given set of cpus and attach the sched domains
6287 * to the individual cpus
6289 static int build_sched_domains(const cpumask_t
*cpu_map
)
6293 struct sched_group
**sched_group_nodes
= NULL
;
6294 int sd_allnodes
= 0;
6297 * Allocate the per-node list of sched groups
6299 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6301 if (!sched_group_nodes
) {
6302 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6305 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6309 * Set up domains for cpus specified by the cpu_map.
6311 for_each_cpu_mask(i
, *cpu_map
) {
6312 struct sched_domain
*sd
= NULL
, *p
;
6313 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6315 cpus_and(nodemask
, nodemask
, *cpu_map
);
6318 if (cpus_weight(*cpu_map
) >
6319 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6320 sd
= &per_cpu(allnodes_domains
, i
);
6321 *sd
= SD_ALLNODES_INIT
;
6322 sd
->span
= *cpu_map
;
6323 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6329 sd
= &per_cpu(node_domains
, i
);
6331 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6335 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6339 sd
= &per_cpu(phys_domains
, i
);
6341 sd
->span
= nodemask
;
6345 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6347 #ifdef CONFIG_SCHED_MC
6349 sd
= &per_cpu(core_domains
, i
);
6351 sd
->span
= cpu_coregroup_map(i
);
6352 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6355 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6358 #ifdef CONFIG_SCHED_SMT
6360 sd
= &per_cpu(cpu_domains
, i
);
6361 *sd
= SD_SIBLING_INIT
;
6362 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6363 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6366 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6370 #ifdef CONFIG_SCHED_SMT
6371 /* Set up CPU (sibling) groups */
6372 for_each_cpu_mask(i
, *cpu_map
) {
6373 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6374 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6375 if (i
!= first_cpu(this_sibling_map
))
6378 init_sched_build_groups(this_sibling_map
, cpu_map
,
6383 #ifdef CONFIG_SCHED_MC
6384 /* Set up multi-core groups */
6385 for_each_cpu_mask(i
, *cpu_map
) {
6386 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6387 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6388 if (i
!= first_cpu(this_core_map
))
6390 init_sched_build_groups(this_core_map
, cpu_map
,
6391 &cpu_to_core_group
);
6395 /* Set up physical groups */
6396 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6397 cpumask_t nodemask
= node_to_cpumask(i
);
6399 cpus_and(nodemask
, nodemask
, *cpu_map
);
6400 if (cpus_empty(nodemask
))
6403 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6407 /* Set up node groups */
6409 init_sched_build_groups(*cpu_map
, cpu_map
,
6410 &cpu_to_allnodes_group
);
6412 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6413 /* Set up node groups */
6414 struct sched_group
*sg
, *prev
;
6415 cpumask_t nodemask
= node_to_cpumask(i
);
6416 cpumask_t domainspan
;
6417 cpumask_t covered
= CPU_MASK_NONE
;
6420 cpus_and(nodemask
, nodemask
, *cpu_map
);
6421 if (cpus_empty(nodemask
)) {
6422 sched_group_nodes
[i
] = NULL
;
6426 domainspan
= sched_domain_node_span(i
);
6427 cpus_and(domainspan
, domainspan
, *cpu_map
);
6429 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6431 printk(KERN_WARNING
"Can not alloc domain group for "
6435 sched_group_nodes
[i
] = sg
;
6436 for_each_cpu_mask(j
, nodemask
) {
6437 struct sched_domain
*sd
;
6439 sd
= &per_cpu(node_domains
, j
);
6442 sg
->__cpu_power
= 0;
6443 sg
->cpumask
= nodemask
;
6445 cpus_or(covered
, covered
, nodemask
);
6448 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6449 cpumask_t tmp
, notcovered
;
6450 int n
= (i
+ j
) % MAX_NUMNODES
;
6452 cpus_complement(notcovered
, covered
);
6453 cpus_and(tmp
, notcovered
, *cpu_map
);
6454 cpus_and(tmp
, tmp
, domainspan
);
6455 if (cpus_empty(tmp
))
6458 nodemask
= node_to_cpumask(n
);
6459 cpus_and(tmp
, tmp
, nodemask
);
6460 if (cpus_empty(tmp
))
6463 sg
= kmalloc_node(sizeof(struct sched_group
),
6467 "Can not alloc domain group for node %d\n", j
);
6470 sg
->__cpu_power
= 0;
6472 sg
->next
= prev
->next
;
6473 cpus_or(covered
, covered
, tmp
);
6480 /* Calculate CPU power for physical packages and nodes */
6481 #ifdef CONFIG_SCHED_SMT
6482 for_each_cpu_mask(i
, *cpu_map
) {
6483 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6485 init_sched_groups_power(i
, sd
);
6488 #ifdef CONFIG_SCHED_MC
6489 for_each_cpu_mask(i
, *cpu_map
) {
6490 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6492 init_sched_groups_power(i
, sd
);
6496 for_each_cpu_mask(i
, *cpu_map
) {
6497 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6499 init_sched_groups_power(i
, sd
);
6503 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6504 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6507 struct sched_group
*sg
;
6509 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6510 init_numa_sched_groups_power(sg
);
6514 /* Attach the domains */
6515 for_each_cpu_mask(i
, *cpu_map
) {
6516 struct sched_domain
*sd
;
6517 #ifdef CONFIG_SCHED_SMT
6518 sd
= &per_cpu(cpu_domains
, i
);
6519 #elif defined(CONFIG_SCHED_MC)
6520 sd
= &per_cpu(core_domains
, i
);
6522 sd
= &per_cpu(phys_domains
, i
);
6524 cpu_attach_domain(sd
, i
);
6531 free_sched_groups(cpu_map
);
6536 static cpumask_t
*doms_cur
; /* current sched domains */
6537 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6540 * Special case: If a kmalloc of a doms_cur partition (array of
6541 * cpumask_t) fails, then fallback to a single sched domain,
6542 * as determined by the single cpumask_t fallback_doms.
6544 static cpumask_t fallback_doms
;
6547 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6548 * For now this just excludes isolated cpus, but could be used to
6549 * exclude other special cases in the future.
6551 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6556 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6558 doms_cur
= &fallback_doms
;
6559 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6560 err
= build_sched_domains(doms_cur
);
6561 register_sched_domain_sysctl();
6566 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6568 free_sched_groups(cpu_map
);
6572 * Detach sched domains from a group of cpus specified in cpu_map
6573 * These cpus will now be attached to the NULL domain
6575 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6579 unregister_sched_domain_sysctl();
6581 for_each_cpu_mask(i
, *cpu_map
)
6582 cpu_attach_domain(NULL
, i
);
6583 synchronize_sched();
6584 arch_destroy_sched_domains(cpu_map
);
6588 * Partition sched domains as specified by the 'ndoms_new'
6589 * cpumasks in the array doms_new[] of cpumasks. This compares
6590 * doms_new[] to the current sched domain partitioning, doms_cur[].
6591 * It destroys each deleted domain and builds each new domain.
6593 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6594 * The masks don't intersect (don't overlap.) We should setup one
6595 * sched domain for each mask. CPUs not in any of the cpumasks will
6596 * not be load balanced. If the same cpumask appears both in the
6597 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6600 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6601 * ownership of it and will kfree it when done with it. If the caller
6602 * failed the kmalloc call, then it can pass in doms_new == NULL,
6603 * and partition_sched_domains() will fallback to the single partition
6606 * Call with hotplug lock held
6608 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6614 /* always unregister in case we don't destroy any domains */
6615 unregister_sched_domain_sysctl();
6617 if (doms_new
== NULL
) {
6619 doms_new
= &fallback_doms
;
6620 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6623 /* Destroy deleted domains */
6624 for (i
= 0; i
< ndoms_cur
; i
++) {
6625 for (j
= 0; j
< ndoms_new
; j
++) {
6626 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
6629 /* no match - a current sched domain not in new doms_new[] */
6630 detach_destroy_domains(doms_cur
+ i
);
6635 /* Build new domains */
6636 for (i
= 0; i
< ndoms_new
; i
++) {
6637 for (j
= 0; j
< ndoms_cur
; j
++) {
6638 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
6641 /* no match - add a new doms_new */
6642 build_sched_domains(doms_new
+ i
);
6647 /* Remember the new sched domains */
6648 if (doms_cur
!= &fallback_doms
)
6650 doms_cur
= doms_new
;
6651 ndoms_cur
= ndoms_new
;
6653 register_sched_domain_sysctl();
6658 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6659 static int arch_reinit_sched_domains(void)
6664 detach_destroy_domains(&cpu_online_map
);
6665 err
= arch_init_sched_domains(&cpu_online_map
);
6671 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6675 if (buf
[0] != '0' && buf
[0] != '1')
6679 sched_smt_power_savings
= (buf
[0] == '1');
6681 sched_mc_power_savings
= (buf
[0] == '1');
6683 ret
= arch_reinit_sched_domains();
6685 return ret
? ret
: count
;
6688 #ifdef CONFIG_SCHED_MC
6689 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6691 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6693 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6694 const char *buf
, size_t count
)
6696 return sched_power_savings_store(buf
, count
, 0);
6698 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6699 sched_mc_power_savings_store
);
6702 #ifdef CONFIG_SCHED_SMT
6703 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6705 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6707 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6708 const char *buf
, size_t count
)
6710 return sched_power_savings_store(buf
, count
, 1);
6712 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6713 sched_smt_power_savings_store
);
6716 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6720 #ifdef CONFIG_SCHED_SMT
6722 err
= sysfs_create_file(&cls
->kset
.kobj
,
6723 &attr_sched_smt_power_savings
.attr
);
6725 #ifdef CONFIG_SCHED_MC
6726 if (!err
&& mc_capable())
6727 err
= sysfs_create_file(&cls
->kset
.kobj
,
6728 &attr_sched_mc_power_savings
.attr
);
6735 * Force a reinitialization of the sched domains hierarchy. The domains
6736 * and groups cannot be updated in place without racing with the balancing
6737 * code, so we temporarily attach all running cpus to the NULL domain
6738 * which will prevent rebalancing while the sched domains are recalculated.
6740 static int update_sched_domains(struct notifier_block
*nfb
,
6741 unsigned long action
, void *hcpu
)
6744 case CPU_UP_PREPARE
:
6745 case CPU_UP_PREPARE_FROZEN
:
6746 case CPU_DOWN_PREPARE
:
6747 case CPU_DOWN_PREPARE_FROZEN
:
6748 detach_destroy_domains(&cpu_online_map
);
6751 case CPU_UP_CANCELED
:
6752 case CPU_UP_CANCELED_FROZEN
:
6753 case CPU_DOWN_FAILED
:
6754 case CPU_DOWN_FAILED_FROZEN
:
6756 case CPU_ONLINE_FROZEN
:
6758 case CPU_DEAD_FROZEN
:
6760 * Fall through and re-initialise the domains.
6767 /* The hotplug lock is already held by cpu_up/cpu_down */
6768 arch_init_sched_domains(&cpu_online_map
);
6773 void __init
sched_init_smp(void)
6775 cpumask_t non_isolated_cpus
;
6778 arch_init_sched_domains(&cpu_online_map
);
6779 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6780 if (cpus_empty(non_isolated_cpus
))
6781 cpu_set(smp_processor_id(), non_isolated_cpus
);
6783 /* XXX: Theoretical race here - CPU may be hotplugged now */
6784 hotcpu_notifier(update_sched_domains
, 0);
6786 /* Move init over to a non-isolated CPU */
6787 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6789 sched_init_granularity();
6791 #ifdef CONFIG_FAIR_GROUP_SCHED
6792 if (nr_cpu_ids
== 1)
6795 lb_monitor_task
= kthread_create(load_balance_monitor
, NULL
,
6797 if (!IS_ERR(lb_monitor_task
)) {
6798 lb_monitor_task
->flags
|= PF_NOFREEZE
;
6799 wake_up_process(lb_monitor_task
);
6801 printk(KERN_ERR
"Could not create load balance monitor thread"
6802 "(error = %ld) \n", PTR_ERR(lb_monitor_task
));
6807 void __init
sched_init_smp(void)
6809 sched_init_granularity();
6811 #endif /* CONFIG_SMP */
6813 int in_sched_functions(unsigned long addr
)
6815 return in_lock_functions(addr
) ||
6816 (addr
>= (unsigned long)__sched_text_start
6817 && addr
< (unsigned long)__sched_text_end
);
6820 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6822 cfs_rq
->tasks_timeline
= RB_ROOT
;
6823 #ifdef CONFIG_FAIR_GROUP_SCHED
6826 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6829 void __init
sched_init(void)
6831 int highest_cpu
= 0;
6834 for_each_possible_cpu(i
) {
6835 struct rt_prio_array
*array
;
6839 spin_lock_init(&rq
->lock
);
6840 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6843 init_cfs_rq(&rq
->cfs
, rq
);
6844 #ifdef CONFIG_FAIR_GROUP_SCHED
6845 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6847 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6848 struct sched_entity
*se
=
6849 &per_cpu(init_sched_entity
, i
);
6851 init_cfs_rq_p
[i
] = cfs_rq
;
6852 init_cfs_rq(cfs_rq
, rq
);
6853 cfs_rq
->tg
= &init_task_group
;
6854 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6855 &rq
->leaf_cfs_rq_list
);
6857 init_sched_entity_p
[i
] = se
;
6858 se
->cfs_rq
= &rq
->cfs
;
6860 se
->load
.weight
= init_task_group_load
;
6861 se
->load
.inv_weight
=
6862 div64_64(1ULL<<32, init_task_group_load
);
6865 init_task_group
.shares
= init_task_group_load
;
6868 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6869 rq
->cpu_load
[j
] = 0;
6872 rq
->active_balance
= 0;
6873 rq
->next_balance
= jiffies
;
6876 rq
->migration_thread
= NULL
;
6877 INIT_LIST_HEAD(&rq
->migration_queue
);
6878 rq
->rt
.highest_prio
= MAX_RT_PRIO
;
6880 atomic_set(&rq
->nr_iowait
, 0);
6882 array
= &rq
->rt
.active
;
6883 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6884 INIT_LIST_HEAD(array
->queue
+ j
);
6885 __clear_bit(j
, array
->bitmap
);
6888 /* delimiter for bitsearch: */
6889 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6892 set_load_weight(&init_task
);
6894 #ifdef CONFIG_PREEMPT_NOTIFIERS
6895 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6899 nr_cpu_ids
= highest_cpu
+ 1;
6900 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6903 #ifdef CONFIG_RT_MUTEXES
6904 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6908 * The boot idle thread does lazy MMU switching as well:
6910 atomic_inc(&init_mm
.mm_count
);
6911 enter_lazy_tlb(&init_mm
, current
);
6914 * Make us the idle thread. Technically, schedule() should not be
6915 * called from this thread, however somewhere below it might be,
6916 * but because we are the idle thread, we just pick up running again
6917 * when this runqueue becomes "idle".
6919 init_idle(current
, smp_processor_id());
6921 * During early bootup we pretend to be a normal task:
6923 current
->sched_class
= &fair_sched_class
;
6926 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6927 void __might_sleep(char *file
, int line
)
6930 static unsigned long prev_jiffy
; /* ratelimiting */
6932 if ((in_atomic() || irqs_disabled()) &&
6933 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6934 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6936 prev_jiffy
= jiffies
;
6937 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6938 " context at %s:%d\n", file
, line
);
6939 printk("in_atomic():%d, irqs_disabled():%d\n",
6940 in_atomic(), irqs_disabled());
6941 debug_show_held_locks(current
);
6942 if (irqs_disabled())
6943 print_irqtrace_events(current
);
6948 EXPORT_SYMBOL(__might_sleep
);
6951 #ifdef CONFIG_MAGIC_SYSRQ
6952 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6955 update_rq_clock(rq
);
6956 on_rq
= p
->se
.on_rq
;
6958 deactivate_task(rq
, p
, 0);
6959 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6961 activate_task(rq
, p
, 0);
6962 resched_task(rq
->curr
);
6966 void normalize_rt_tasks(void)
6968 struct task_struct
*g
, *p
;
6969 unsigned long flags
;
6972 read_lock_irq(&tasklist_lock
);
6973 do_each_thread(g
, p
) {
6975 * Only normalize user tasks:
6980 p
->se
.exec_start
= 0;
6981 #ifdef CONFIG_SCHEDSTATS
6982 p
->se
.wait_start
= 0;
6983 p
->se
.sleep_start
= 0;
6984 p
->se
.block_start
= 0;
6986 task_rq(p
)->clock
= 0;
6990 * Renice negative nice level userspace
6993 if (TASK_NICE(p
) < 0 && p
->mm
)
6994 set_user_nice(p
, 0);
6998 spin_lock_irqsave(&p
->pi_lock
, flags
);
6999 rq
= __task_rq_lock(p
);
7001 normalize_task(rq
, p
);
7003 __task_rq_unlock(rq
);
7004 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
7005 } while_each_thread(g
, p
);
7007 read_unlock_irq(&tasklist_lock
);
7010 #endif /* CONFIG_MAGIC_SYSRQ */
7014 * These functions are only useful for the IA64 MCA handling.
7016 * They can only be called when the whole system has been
7017 * stopped - every CPU needs to be quiescent, and no scheduling
7018 * activity can take place. Using them for anything else would
7019 * be a serious bug, and as a result, they aren't even visible
7020 * under any other configuration.
7024 * curr_task - return the current task for a given cpu.
7025 * @cpu: the processor in question.
7027 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7029 struct task_struct
*curr_task(int cpu
)
7031 return cpu_curr(cpu
);
7035 * set_curr_task - set the current task for a given cpu.
7036 * @cpu: the processor in question.
7037 * @p: the task pointer to set.
7039 * Description: This function must only be used when non-maskable interrupts
7040 * are serviced on a separate stack. It allows the architecture to switch the
7041 * notion of the current task on a cpu in a non-blocking manner. This function
7042 * must be called with all CPU's synchronized, and interrupts disabled, the
7043 * and caller must save the original value of the current task (see
7044 * curr_task() above) and restore that value before reenabling interrupts and
7045 * re-starting the system.
7047 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7049 void set_curr_task(int cpu
, struct task_struct
*p
)
7056 #ifdef CONFIG_FAIR_GROUP_SCHED
7060 * distribute shares of all task groups among their schedulable entities,
7061 * to reflect load distrbution across cpus.
7063 static int rebalance_shares(struct sched_domain
*sd
, int this_cpu
)
7065 struct cfs_rq
*cfs_rq
;
7066 struct rq
*rq
= cpu_rq(this_cpu
);
7067 cpumask_t sdspan
= sd
->span
;
7070 /* Walk thr' all the task groups that we have */
7071 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
7073 unsigned long total_load
= 0, total_shares
;
7074 struct task_group
*tg
= cfs_rq
->tg
;
7076 /* Gather total task load of this group across cpus */
7077 for_each_cpu_mask(i
, sdspan
)
7078 total_load
+= tg
->cfs_rq
[i
]->load
.weight
;
7080 /* Nothing to do if this group has no load */
7085 * tg->shares represents the number of cpu shares the task group
7086 * is eligible to hold on a single cpu. On N cpus, it is
7087 * eligible to hold (N * tg->shares) number of cpu shares.
7089 total_shares
= tg
->shares
* cpus_weight(sdspan
);
7092 * redistribute total_shares across cpus as per the task load
7095 for_each_cpu_mask(i
, sdspan
) {
7096 unsigned long local_load
, local_shares
;
7098 local_load
= tg
->cfs_rq
[i
]->load
.weight
;
7099 local_shares
= (local_load
* total_shares
) / total_load
;
7101 local_shares
= MIN_GROUP_SHARES
;
7102 if (local_shares
== tg
->se
[i
]->load
.weight
)
7105 spin_lock_irq(&cpu_rq(i
)->lock
);
7106 set_se_shares(tg
->se
[i
], local_shares
);
7107 spin_unlock_irq(&cpu_rq(i
)->lock
);
7116 * How frequently should we rebalance_shares() across cpus?
7118 * The more frequently we rebalance shares, the more accurate is the fairness
7119 * of cpu bandwidth distribution between task groups. However higher frequency
7120 * also implies increased scheduling overhead.
7122 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7123 * consecutive calls to rebalance_shares() in the same sched domain.
7125 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7126 * consecutive calls to rebalance_shares() in the same sched domain.
7128 * These settings allows for the appropriate tradeoff between accuracy of
7129 * fairness and the associated overhead.
7133 /* default: 8ms, units: milliseconds */
7134 const_debug
unsigned int sysctl_sched_min_bal_int_shares
= 8;
7136 /* default: 128ms, units: milliseconds */
7137 const_debug
unsigned int sysctl_sched_max_bal_int_shares
= 128;
7139 /* kernel thread that runs rebalance_shares() periodically */
7140 static int load_balance_monitor(void *unused
)
7142 unsigned int timeout
= sysctl_sched_min_bal_int_shares
;
7143 struct sched_param schedparm
;
7147 * We don't want this thread's execution to be limited by the shares
7148 * assigned to default group (init_task_group). Hence make it run
7149 * as a SCHED_RR RT task at the lowest priority.
7151 schedparm
.sched_priority
= 1;
7152 ret
= sched_setscheduler(current
, SCHED_RR
, &schedparm
);
7154 printk(KERN_ERR
"Couldn't set SCHED_RR policy for load balance"
7155 " monitor thread (error = %d) \n", ret
);
7157 while (!kthread_should_stop()) {
7158 int i
, cpu
, balanced
= 1;
7160 /* Prevent cpus going down or coming up */
7162 /* lockout changes to doms_cur[] array */
7165 * Enter a rcu read-side critical section to safely walk rq->sd
7166 * chain on various cpus and to walk task group list
7167 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7171 for (i
= 0; i
< ndoms_cur
; i
++) {
7172 cpumask_t cpumap
= doms_cur
[i
];
7173 struct sched_domain
*sd
= NULL
, *sd_prev
= NULL
;
7175 cpu
= first_cpu(cpumap
);
7177 /* Find the highest domain at which to balance shares */
7178 for_each_domain(cpu
, sd
) {
7179 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7185 /* sd == NULL? No load balance reqd in this domain */
7189 balanced
&= rebalance_shares(sd
, cpu
);
7198 timeout
= sysctl_sched_min_bal_int_shares
;
7199 else if (timeout
< sysctl_sched_max_bal_int_shares
)
7202 msleep_interruptible(timeout
);
7207 #endif /* CONFIG_SMP */
7209 /* allocate runqueue etc for a new task group */
7210 struct task_group
*sched_create_group(void)
7212 struct task_group
*tg
;
7213 struct cfs_rq
*cfs_rq
;
7214 struct sched_entity
*se
;
7218 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7220 return ERR_PTR(-ENOMEM
);
7222 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
7225 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
7229 for_each_possible_cpu(i
) {
7232 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
7237 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
7242 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
7243 memset(se
, 0, sizeof(struct sched_entity
));
7245 tg
->cfs_rq
[i
] = cfs_rq
;
7246 init_cfs_rq(cfs_rq
, rq
);
7250 se
->cfs_rq
= &rq
->cfs
;
7252 se
->load
.weight
= NICE_0_LOAD
;
7253 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
7257 tg
->shares
= NICE_0_LOAD
;
7259 lock_task_group_list();
7260 for_each_possible_cpu(i
) {
7262 cfs_rq
= tg
->cfs_rq
[i
];
7263 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7265 unlock_task_group_list();
7270 for_each_possible_cpu(i
) {
7272 kfree(tg
->cfs_rq
[i
]);
7280 return ERR_PTR(-ENOMEM
);
7283 /* rcu callback to free various structures associated with a task group */
7284 static void free_sched_group(struct rcu_head
*rhp
)
7286 struct task_group
*tg
= container_of(rhp
, struct task_group
, rcu
);
7287 struct cfs_rq
*cfs_rq
;
7288 struct sched_entity
*se
;
7291 /* now it should be safe to free those cfs_rqs */
7292 for_each_possible_cpu(i
) {
7293 cfs_rq
= tg
->cfs_rq
[i
];
7305 /* Destroy runqueue etc associated with a task group */
7306 void sched_destroy_group(struct task_group
*tg
)
7308 struct cfs_rq
*cfs_rq
= NULL
;
7311 lock_task_group_list();
7312 for_each_possible_cpu(i
) {
7313 cfs_rq
= tg
->cfs_rq
[i
];
7314 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7316 unlock_task_group_list();
7320 /* wait for possible concurrent references to cfs_rqs complete */
7321 call_rcu(&tg
->rcu
, free_sched_group
);
7324 /* change task's runqueue when it moves between groups.
7325 * The caller of this function should have put the task in its new group
7326 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7327 * reflect its new group.
7329 void sched_move_task(struct task_struct
*tsk
)
7332 unsigned long flags
;
7335 rq
= task_rq_lock(tsk
, &flags
);
7337 if (tsk
->sched_class
!= &fair_sched_class
) {
7338 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7342 update_rq_clock(rq
);
7344 running
= task_current(rq
, tsk
);
7345 on_rq
= tsk
->se
.on_rq
;
7348 dequeue_task(rq
, tsk
, 0);
7349 if (unlikely(running
))
7350 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7353 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7356 if (unlikely(running
))
7357 tsk
->sched_class
->set_curr_task(rq
);
7358 enqueue_task(rq
, tsk
, 0);
7362 task_rq_unlock(rq
, &flags
);
7365 /* rq->lock to be locked by caller */
7366 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7368 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7369 struct rq
*rq
= cfs_rq
->rq
;
7373 shares
= MIN_GROUP_SHARES
;
7377 dequeue_entity(cfs_rq
, se
, 0);
7378 dec_cpu_load(rq
, se
->load
.weight
);
7381 se
->load
.weight
= shares
;
7382 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7385 enqueue_entity(cfs_rq
, se
, 0);
7386 inc_cpu_load(rq
, se
->load
.weight
);
7390 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7393 struct cfs_rq
*cfs_rq
;
7396 lock_task_group_list();
7397 if (tg
->shares
== shares
)
7400 if (shares
< MIN_GROUP_SHARES
)
7401 shares
= MIN_GROUP_SHARES
;
7404 * Prevent any load balance activity (rebalance_shares,
7405 * load_balance_fair) from referring to this group first,
7406 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7408 for_each_possible_cpu(i
) {
7409 cfs_rq
= tg
->cfs_rq
[i
];
7410 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7413 /* wait for any ongoing reference to this group to finish */
7414 synchronize_sched();
7417 * Now we are free to modify the group's share on each cpu
7418 * w/o tripping rebalance_share or load_balance_fair.
7420 tg
->shares
= shares
;
7421 for_each_possible_cpu(i
) {
7422 spin_lock_irq(&cpu_rq(i
)->lock
);
7423 set_se_shares(tg
->se
[i
], shares
);
7424 spin_unlock_irq(&cpu_rq(i
)->lock
);
7428 * Enable load balance activity on this group, by inserting it back on
7429 * each cpu's rq->leaf_cfs_rq_list.
7431 for_each_possible_cpu(i
) {
7433 cfs_rq
= tg
->cfs_rq
[i
];
7434 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7437 unlock_task_group_list();
7441 unsigned long sched_group_shares(struct task_group
*tg
)
7446 #endif /* CONFIG_FAIR_GROUP_SCHED */
7448 #ifdef CONFIG_FAIR_CGROUP_SCHED
7450 /* return corresponding task_group object of a cgroup */
7451 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7453 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7454 struct task_group
, css
);
7457 static struct cgroup_subsys_state
*
7458 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7460 struct task_group
*tg
;
7462 if (!cgrp
->parent
) {
7463 /* This is early initialization for the top cgroup */
7464 init_task_group
.css
.cgroup
= cgrp
;
7465 return &init_task_group
.css
;
7468 /* we support only 1-level deep hierarchical scheduler atm */
7469 if (cgrp
->parent
->parent
)
7470 return ERR_PTR(-EINVAL
);
7472 tg
= sched_create_group();
7474 return ERR_PTR(-ENOMEM
);
7476 /* Bind the cgroup to task_group object we just created */
7477 tg
->css
.cgroup
= cgrp
;
7483 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7485 struct task_group
*tg
= cgroup_tg(cgrp
);
7487 sched_destroy_group(tg
);
7491 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7492 struct task_struct
*tsk
)
7494 /* We don't support RT-tasks being in separate groups */
7495 if (tsk
->sched_class
!= &fair_sched_class
)
7502 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7503 struct cgroup
*old_cont
, struct task_struct
*tsk
)
7505 sched_move_task(tsk
);
7508 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7511 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7514 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7516 struct task_group
*tg
= cgroup_tg(cgrp
);
7518 return (u64
) tg
->shares
;
7521 static struct cftype cpu_files
[] = {
7524 .read_uint
= cpu_shares_read_uint
,
7525 .write_uint
= cpu_shares_write_uint
,
7529 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7531 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7534 struct cgroup_subsys cpu_cgroup_subsys
= {
7536 .create
= cpu_cgroup_create
,
7537 .destroy
= cpu_cgroup_destroy
,
7538 .can_attach
= cpu_cgroup_can_attach
,
7539 .attach
= cpu_cgroup_attach
,
7540 .populate
= cpu_cgroup_populate
,
7541 .subsys_id
= cpu_cgroup_subsys_id
,
7545 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7547 #ifdef CONFIG_CGROUP_CPUACCT
7550 * CPU accounting code for task groups.
7552 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7553 * (balbir@in.ibm.com).
7556 /* track cpu usage of a group of tasks */
7558 struct cgroup_subsys_state css
;
7559 /* cpuusage holds pointer to a u64-type object on every cpu */
7563 struct cgroup_subsys cpuacct_subsys
;
7565 /* return cpu accounting group corresponding to this container */
7566 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
7568 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
7569 struct cpuacct
, css
);
7572 /* return cpu accounting group to which this task belongs */
7573 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
7575 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
7576 struct cpuacct
, css
);
7579 /* create a new cpu accounting group */
7580 static struct cgroup_subsys_state
*cpuacct_create(
7581 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7583 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7586 return ERR_PTR(-ENOMEM
);
7588 ca
->cpuusage
= alloc_percpu(u64
);
7589 if (!ca
->cpuusage
) {
7591 return ERR_PTR(-ENOMEM
);
7597 /* destroy an existing cpu accounting group */
7599 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7601 struct cpuacct
*ca
= cgroup_ca(cont
);
7603 free_percpu(ca
->cpuusage
);
7607 /* return total cpu usage (in nanoseconds) of a group */
7608 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
7610 struct cpuacct
*ca
= cgroup_ca(cont
);
7611 u64 totalcpuusage
= 0;
7614 for_each_possible_cpu(i
) {
7615 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
7618 * Take rq->lock to make 64-bit addition safe on 32-bit
7621 spin_lock_irq(&cpu_rq(i
)->lock
);
7622 totalcpuusage
+= *cpuusage
;
7623 spin_unlock_irq(&cpu_rq(i
)->lock
);
7626 return totalcpuusage
;
7629 static struct cftype files
[] = {
7632 .read_uint
= cpuusage_read
,
7636 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7638 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
7642 * charge this task's execution time to its accounting group.
7644 * called with rq->lock held.
7646 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
7650 if (!cpuacct_subsys
.active
)
7655 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
7657 *cpuusage
+= cputime
;
7661 struct cgroup_subsys cpuacct_subsys
= {
7663 .create
= cpuacct_create
,
7664 .destroy
= cpuacct_destroy
,
7665 .populate
= cpuacct_populate
,
7666 .subsys_id
= cpuacct_subsys_id
,
7668 #endif /* CONFIG_CGROUP_CPUACCT */