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
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
70 #include <asm/irq_regs.h>
73 * Scheduler clock - returns current time in nanosec units.
74 * This is default implementation.
75 * Architectures and sub-architectures can override this.
77 unsigned long long __attribute__((weak
)) sched_clock(void)
79 return (unsigned long long)jiffies
* (NSEC_PER_SEC
/ HZ
);
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
119 * Since cpu_power is a 'constant', we can use a reciprocal divide.
121 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
123 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
127 * Each time a sched group cpu_power is changed,
128 * we must compute its reciprocal value
130 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
132 sg
->__cpu_power
+= val
;
133 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
137 static inline int rt_policy(int policy
)
139 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
144 static inline int task_has_rt_policy(struct task_struct
*p
)
146 return rt_policy(p
->policy
);
150 * This is the priority-queue data structure of the RT scheduling class:
152 struct rt_prio_array
{
153 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
154 struct list_head queue
[MAX_RT_PRIO
];
157 #ifdef CONFIG_FAIR_GROUP_SCHED
159 #include <linux/cgroup.h>
163 /* task group related information */
165 #ifdef CONFIG_FAIR_CGROUP_SCHED
166 struct cgroup_subsys_state css
;
168 /* schedulable entities of this group on each cpu */
169 struct sched_entity
**se
;
170 /* runqueue "owned" by this group on each cpu */
171 struct cfs_rq
**cfs_rq
;
174 * shares assigned to a task group governs how much of cpu bandwidth
175 * is allocated to the group. The more shares a group has, the more is
176 * the cpu bandwidth allocated to it.
178 * For ex, lets say that there are three task groups, A, B and C which
179 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
180 * cpu bandwidth allocated by the scheduler to task groups A, B and C
183 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
184 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
185 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
187 * The weight assigned to a task group's schedulable entities on every
188 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
189 * group's shares. For ex: lets say that task group A has been
190 * assigned shares of 1000 and there are two CPUs in a system. Then,
192 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
194 * Note: It's not necessary that each of a task's group schedulable
195 * entity have the same weight on all CPUs. If the group
196 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
197 * better distribution of weight could be:
199 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
200 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
202 * rebalance_shares() is responsible for distributing the shares of a
203 * task groups like this among the group's schedulable entities across
207 unsigned long shares
;
212 /* Default task group's sched entity on each cpu */
213 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
214 /* Default task group's cfs_rq on each cpu */
215 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
217 static struct sched_entity
*init_sched_entity_p
[NR_CPUS
];
218 static struct cfs_rq
*init_cfs_rq_p
[NR_CPUS
];
220 /* task_group_mutex serializes add/remove of task groups and also changes to
221 * a task group's cpu shares.
223 static DEFINE_MUTEX(task_group_mutex
);
225 /* doms_cur_mutex serializes access to doms_cur[] array */
226 static DEFINE_MUTEX(doms_cur_mutex
);
229 /* kernel thread that runs rebalance_shares() periodically */
230 static struct task_struct
*lb_monitor_task
;
231 static int load_balance_monitor(void *unused
);
234 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
);
236 /* Default task group.
237 * Every task in system belong to this group at bootup.
239 struct task_group init_task_group
= {
240 .se
= init_sched_entity_p
,
241 .cfs_rq
= init_cfs_rq_p
,
244 #ifdef CONFIG_FAIR_USER_SCHED
245 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
247 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
250 #define MIN_GROUP_SHARES 2
252 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
254 /* return group to which a task belongs */
255 static inline struct task_group
*task_group(struct task_struct
*p
)
257 struct task_group
*tg
;
259 #ifdef CONFIG_FAIR_USER_SCHED
261 #elif defined(CONFIG_FAIR_CGROUP_SCHED)
262 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
263 struct task_group
, css
);
265 tg
= &init_task_group
;
270 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
271 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
)
273 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
274 p
->se
.parent
= task_group(p
)->se
[cpu
];
277 static inline void lock_task_group_list(void)
279 mutex_lock(&task_group_mutex
);
282 static inline void unlock_task_group_list(void)
284 mutex_unlock(&task_group_mutex
);
287 static inline void lock_doms_cur(void)
289 mutex_lock(&doms_cur_mutex
);
292 static inline void unlock_doms_cur(void)
294 mutex_unlock(&doms_cur_mutex
);
299 static inline void set_task_cfs_rq(struct task_struct
*p
, unsigned int cpu
) { }
300 static inline void lock_task_group_list(void) { }
301 static inline void unlock_task_group_list(void) { }
302 static inline void lock_doms_cur(void) { }
303 static inline void unlock_doms_cur(void) { }
305 #endif /* CONFIG_FAIR_GROUP_SCHED */
307 /* CFS-related fields in a runqueue */
309 struct load_weight load
;
310 unsigned long nr_running
;
315 struct rb_root tasks_timeline
;
316 struct rb_node
*rb_leftmost
;
317 struct rb_node
*rb_load_balance_curr
;
318 /* 'curr' points to currently running entity on this cfs_rq.
319 * It is set to NULL otherwise (i.e when none are currently running).
321 struct sched_entity
*curr
;
323 unsigned long nr_spread_over
;
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
329 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
330 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
331 * (like users, containers etc.)
333 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
334 * list is used during load balance.
336 struct list_head leaf_cfs_rq_list
;
337 struct task_group
*tg
; /* group that "owns" this runqueue */
341 /* Real-Time classes' related field in a runqueue: */
343 struct rt_prio_array active
;
344 int rt_load_balance_idx
;
345 struct list_head
*rt_load_balance_head
, *rt_load_balance_curr
;
346 unsigned long rt_nr_running
;
347 unsigned long rt_nr_migratory
;
348 /* highest queued rt task prio */
356 * We add the notion of a root-domain which will be used to define per-domain
357 * variables. Each exclusive cpuset essentially defines an island domain by
358 * fully partitioning the member cpus from any other cpuset. Whenever a new
359 * exclusive cpuset is created, we also create and attach a new root-domain
369 * The "RT overload" flag: it gets set if a CPU has more than
370 * one runnable RT task.
377 * By default the system creates a single root-domain with all cpus as
378 * members (mimicking the global state we have today).
380 static struct root_domain def_root_domain
;
385 * This is the main, per-CPU runqueue data structure.
387 * Locking rule: those places that want to lock multiple runqueues
388 * (such as the load balancing or the thread migration code), lock
389 * acquire operations must be ordered by ascending &runqueue.
396 * nr_running and cpu_load should be in the same cacheline because
397 * remote CPUs use both these fields when doing load calculation.
399 unsigned long nr_running
;
400 #define CPU_LOAD_IDX_MAX 5
401 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
402 unsigned char idle_at_tick
;
404 unsigned char in_nohz_recently
;
406 /* capture load from *all* tasks on this cpu: */
407 struct load_weight load
;
408 unsigned long nr_load_updates
;
412 #ifdef CONFIG_FAIR_GROUP_SCHED
413 /* list of leaf cfs_rq on this cpu: */
414 struct list_head leaf_cfs_rq_list
;
419 * This is part of a global counter where only the total sum
420 * over all CPUs matters. A task can increase this counter on
421 * one CPU and if it got migrated afterwards it may decrease
422 * it on another CPU. Always updated under the runqueue lock:
424 unsigned long nr_uninterruptible
;
426 struct task_struct
*curr
, *idle
;
427 unsigned long next_balance
;
428 struct mm_struct
*prev_mm
;
430 u64 clock
, prev_clock_raw
;
433 unsigned int clock_warps
, clock_overflows
;
435 unsigned int clock_deep_idle_events
;
441 struct root_domain
*rd
;
442 struct sched_domain
*sd
;
444 /* For active balancing */
447 /* cpu of this runqueue: */
450 struct task_struct
*migration_thread
;
451 struct list_head migration_queue
;
454 #ifdef CONFIG_SCHEDSTATS
456 struct sched_info rq_sched_info
;
458 /* sys_sched_yield() stats */
459 unsigned int yld_exp_empty
;
460 unsigned int yld_act_empty
;
461 unsigned int yld_both_empty
;
462 unsigned int yld_count
;
464 /* schedule() stats */
465 unsigned int sched_switch
;
466 unsigned int sched_count
;
467 unsigned int sched_goidle
;
469 /* try_to_wake_up() stats */
470 unsigned int ttwu_count
;
471 unsigned int ttwu_local
;
474 unsigned int bkl_count
;
476 struct lock_class_key rq_lock_key
;
479 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
481 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
483 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
486 static inline int cpu_of(struct rq
*rq
)
496 * Update the per-runqueue clock, as finegrained as the platform can give
497 * us, but without assuming monotonicity, etc.:
499 static void __update_rq_clock(struct rq
*rq
)
501 u64 prev_raw
= rq
->prev_clock_raw
;
502 u64 now
= sched_clock();
503 s64 delta
= now
- prev_raw
;
504 u64 clock
= rq
->clock
;
506 #ifdef CONFIG_SCHED_DEBUG
507 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
510 * Protect against sched_clock() occasionally going backwards:
512 if (unlikely(delta
< 0)) {
517 * Catch too large forward jumps too:
519 if (unlikely(clock
+ delta
> rq
->tick_timestamp
+ TICK_NSEC
)) {
520 if (clock
< rq
->tick_timestamp
+ TICK_NSEC
)
521 clock
= rq
->tick_timestamp
+ TICK_NSEC
;
524 rq
->clock_overflows
++;
526 if (unlikely(delta
> rq
->clock_max_delta
))
527 rq
->clock_max_delta
= delta
;
532 rq
->prev_clock_raw
= now
;
536 static void update_rq_clock(struct rq
*rq
)
538 if (likely(smp_processor_id() == cpu_of(rq
)))
539 __update_rq_clock(rq
);
543 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
544 * See detach_destroy_domains: synchronize_sched for details.
546 * The domain tree of any CPU may only be accessed from within
547 * preempt-disabled sections.
549 #define for_each_domain(cpu, __sd) \
550 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
552 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
553 #define this_rq() (&__get_cpu_var(runqueues))
554 #define task_rq(p) cpu_rq(task_cpu(p))
555 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
558 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
560 #ifdef CONFIG_SCHED_DEBUG
561 # define const_debug __read_mostly
563 # define const_debug static const
567 * Debugging: various feature bits
570 SCHED_FEAT_NEW_FAIR_SLEEPERS
= 1,
571 SCHED_FEAT_WAKEUP_PREEMPT
= 2,
572 SCHED_FEAT_START_DEBIT
= 4,
573 SCHED_FEAT_TREE_AVG
= 8,
574 SCHED_FEAT_APPROX_AVG
= 16,
577 const_debug
unsigned int sysctl_sched_features
=
578 SCHED_FEAT_NEW_FAIR_SLEEPERS
* 1 |
579 SCHED_FEAT_WAKEUP_PREEMPT
* 1 |
580 SCHED_FEAT_START_DEBIT
* 1 |
581 SCHED_FEAT_TREE_AVG
* 0 |
582 SCHED_FEAT_APPROX_AVG
* 0;
584 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
587 * Number of tasks to iterate in a single balance run.
588 * Limited because this is done with IRQs disabled.
590 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
593 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
594 * clock constructed from sched_clock():
596 unsigned long long cpu_clock(int cpu
)
598 unsigned long long now
;
602 local_irq_save(flags
);
605 * Only call sched_clock() if the scheduler has already been
606 * initialized (some code might call cpu_clock() very early):
611 local_irq_restore(flags
);
615 EXPORT_SYMBOL_GPL(cpu_clock
);
617 #ifndef prepare_arch_switch
618 # define prepare_arch_switch(next) do { } while (0)
620 #ifndef finish_arch_switch
621 # define finish_arch_switch(prev) do { } while (0)
624 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
626 return rq
->curr
== p
;
629 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
630 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
632 return task_current(rq
, p
);
635 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
639 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
641 #ifdef CONFIG_DEBUG_SPINLOCK
642 /* this is a valid case when another task releases the spinlock */
643 rq
->lock
.owner
= current
;
646 * If we are tracking spinlock dependencies then we have to
647 * fix up the runqueue lock - which gets 'carried over' from
650 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
652 spin_unlock_irq(&rq
->lock
);
655 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
656 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
661 return task_current(rq
, p
);
665 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
669 * We can optimise this out completely for !SMP, because the
670 * SMP rebalancing from interrupt is the only thing that cares
675 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
676 spin_unlock_irq(&rq
->lock
);
678 spin_unlock(&rq
->lock
);
682 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
686 * After ->oncpu is cleared, the task can be moved to a different CPU.
687 * We must ensure this doesn't happen until the switch is completely
693 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
697 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
700 * __task_rq_lock - lock the runqueue a given task resides on.
701 * Must be called interrupts disabled.
703 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
707 struct rq
*rq
= task_rq(p
);
708 spin_lock(&rq
->lock
);
709 if (likely(rq
== task_rq(p
)))
711 spin_unlock(&rq
->lock
);
716 * task_rq_lock - lock the runqueue a given task resides on and disable
717 * interrupts. Note the ordering: we can safely lookup the task_rq without
718 * explicitly disabling preemption.
720 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
726 local_irq_save(*flags
);
728 spin_lock(&rq
->lock
);
729 if (likely(rq
== task_rq(p
)))
731 spin_unlock_irqrestore(&rq
->lock
, *flags
);
735 static void __task_rq_unlock(struct rq
*rq
)
738 spin_unlock(&rq
->lock
);
741 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
744 spin_unlock_irqrestore(&rq
->lock
, *flags
);
748 * this_rq_lock - lock this runqueue and disable interrupts.
750 static struct rq
*this_rq_lock(void)
757 spin_lock(&rq
->lock
);
763 * We are going deep-idle (irqs are disabled):
765 void sched_clock_idle_sleep_event(void)
767 struct rq
*rq
= cpu_rq(smp_processor_id());
769 spin_lock(&rq
->lock
);
770 __update_rq_clock(rq
);
771 spin_unlock(&rq
->lock
);
772 rq
->clock_deep_idle_events
++;
774 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
777 * We just idled delta nanoseconds (called with irqs disabled):
779 void sched_clock_idle_wakeup_event(u64 delta_ns
)
781 struct rq
*rq
= cpu_rq(smp_processor_id());
782 u64 now
= sched_clock();
784 touch_softlockup_watchdog();
785 rq
->idle_clock
+= delta_ns
;
787 * Override the previous timestamp and ignore all
788 * sched_clock() deltas that occured while we idled,
789 * and use the PM-provided delta_ns to advance the
792 spin_lock(&rq
->lock
);
793 rq
->prev_clock_raw
= now
;
794 rq
->clock
+= delta_ns
;
795 spin_unlock(&rq
->lock
);
797 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
800 * resched_task - mark a task 'to be rescheduled now'.
802 * On UP this means the setting of the need_resched flag, on SMP it
803 * might also involve a cross-CPU call to trigger the scheduler on
808 #ifndef tsk_is_polling
809 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
812 static void resched_task(struct task_struct
*p
)
816 assert_spin_locked(&task_rq(p
)->lock
);
818 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
821 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
824 if (cpu
== smp_processor_id())
827 /* NEED_RESCHED must be visible before we test polling */
829 if (!tsk_is_polling(p
))
830 smp_send_reschedule(cpu
);
833 static void resched_cpu(int cpu
)
835 struct rq
*rq
= cpu_rq(cpu
);
838 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
840 resched_task(cpu_curr(cpu
));
841 spin_unlock_irqrestore(&rq
->lock
, flags
);
844 static inline void resched_task(struct task_struct
*p
)
846 assert_spin_locked(&task_rq(p
)->lock
);
847 set_tsk_need_resched(p
);
851 #if BITS_PER_LONG == 32
852 # define WMULT_CONST (~0UL)
854 # define WMULT_CONST (1UL << 32)
857 #define WMULT_SHIFT 32
860 * Shift right and round:
862 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
865 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
866 struct load_weight
*lw
)
870 if (unlikely(!lw
->inv_weight
))
871 lw
->inv_weight
= (WMULT_CONST
- lw
->weight
/2) / lw
->weight
+ 1;
873 tmp
= (u64
)delta_exec
* weight
;
875 * Check whether we'd overflow the 64-bit multiplication:
877 if (unlikely(tmp
> WMULT_CONST
))
878 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
881 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
883 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
886 static inline unsigned long
887 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
889 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
892 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
897 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
903 * To aid in avoiding the subversion of "niceness" due to uneven distribution
904 * of tasks with abnormal "nice" values across CPUs the contribution that
905 * each task makes to its run queue's load is weighted according to its
906 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
907 * scaled version of the new time slice allocation that they receive on time
911 #define WEIGHT_IDLEPRIO 2
912 #define WMULT_IDLEPRIO (1 << 31)
915 * Nice levels are multiplicative, with a gentle 10% change for every
916 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
917 * nice 1, it will get ~10% less CPU time than another CPU-bound task
918 * that remained on nice 0.
920 * The "10% effect" is relative and cumulative: from _any_ nice level,
921 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
922 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
923 * If a task goes up by ~10% and another task goes down by ~10% then
924 * the relative distance between them is ~25%.)
926 static const int prio_to_weight
[40] = {
927 /* -20 */ 88761, 71755, 56483, 46273, 36291,
928 /* -15 */ 29154, 23254, 18705, 14949, 11916,
929 /* -10 */ 9548, 7620, 6100, 4904, 3906,
930 /* -5 */ 3121, 2501, 1991, 1586, 1277,
931 /* 0 */ 1024, 820, 655, 526, 423,
932 /* 5 */ 335, 272, 215, 172, 137,
933 /* 10 */ 110, 87, 70, 56, 45,
934 /* 15 */ 36, 29, 23, 18, 15,
938 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
940 * In cases where the weight does not change often, we can use the
941 * precalculated inverse to speed up arithmetics by turning divisions
942 * into multiplications:
944 static const u32 prio_to_wmult
[40] = {
945 /* -20 */ 48388, 59856, 76040, 92818, 118348,
946 /* -15 */ 147320, 184698, 229616, 287308, 360437,
947 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
948 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
949 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
950 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
951 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
952 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
955 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
958 * runqueue iterator, to support SMP load-balancing between different
959 * scheduling classes, without having to expose their internal data
960 * structures to the load-balancing proper:
964 struct task_struct
*(*start
)(void *);
965 struct task_struct
*(*next
)(void *);
970 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
971 unsigned long max_load_move
, struct sched_domain
*sd
,
972 enum cpu_idle_type idle
, int *all_pinned
,
973 int *this_best_prio
, struct rq_iterator
*iterator
);
976 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
977 struct sched_domain
*sd
, enum cpu_idle_type idle
,
978 struct rq_iterator
*iterator
);
981 #ifdef CONFIG_CGROUP_CPUACCT
982 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
984 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
987 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
989 update_load_add(&rq
->load
, load
);
992 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
994 update_load_sub(&rq
->load
, load
);
998 static unsigned long source_load(int cpu
, int type
);
999 static unsigned long target_load(int cpu
, int type
);
1000 static unsigned long cpu_avg_load_per_task(int cpu
);
1001 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1002 #endif /* CONFIG_SMP */
1004 #include "sched_stats.h"
1005 #include "sched_idletask.c"
1006 #include "sched_fair.c"
1007 #include "sched_rt.c"
1008 #ifdef CONFIG_SCHED_DEBUG
1009 # include "sched_debug.c"
1012 #define sched_class_highest (&rt_sched_class)
1014 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
1019 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
1024 static void set_load_weight(struct task_struct
*p
)
1026 if (task_has_rt_policy(p
)) {
1027 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1028 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1033 * SCHED_IDLE tasks get minimal weight:
1035 if (p
->policy
== SCHED_IDLE
) {
1036 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1037 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1041 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1042 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1045 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1047 sched_info_queued(p
);
1048 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1052 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1054 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1059 * __normal_prio - return the priority that is based on the static prio
1061 static inline int __normal_prio(struct task_struct
*p
)
1063 return p
->static_prio
;
1067 * Calculate the expected normal priority: i.e. priority
1068 * without taking RT-inheritance into account. Might be
1069 * boosted by interactivity modifiers. Changes upon fork,
1070 * setprio syscalls, and whenever the interactivity
1071 * estimator recalculates.
1073 static inline int normal_prio(struct task_struct
*p
)
1077 if (task_has_rt_policy(p
))
1078 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1080 prio
= __normal_prio(p
);
1085 * Calculate the current priority, i.e. the priority
1086 * taken into account by the scheduler. This value might
1087 * be boosted by RT tasks, or might be boosted by
1088 * interactivity modifiers. Will be RT if the task got
1089 * RT-boosted. If not then it returns p->normal_prio.
1091 static int effective_prio(struct task_struct
*p
)
1093 p
->normal_prio
= normal_prio(p
);
1095 * If we are RT tasks or we were boosted to RT priority,
1096 * keep the priority unchanged. Otherwise, update priority
1097 * to the normal priority:
1099 if (!rt_prio(p
->prio
))
1100 return p
->normal_prio
;
1105 * activate_task - move a task to the runqueue.
1107 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1109 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1110 rq
->nr_uninterruptible
--;
1112 enqueue_task(rq
, p
, wakeup
);
1113 inc_nr_running(p
, rq
);
1117 * deactivate_task - remove a task from the runqueue.
1119 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1121 if (p
->state
== TASK_UNINTERRUPTIBLE
)
1122 rq
->nr_uninterruptible
++;
1124 dequeue_task(rq
, p
, sleep
);
1125 dec_nr_running(p
, rq
);
1129 * task_curr - is this task currently executing on a CPU?
1130 * @p: the task in question.
1132 inline int task_curr(const struct task_struct
*p
)
1134 return cpu_curr(task_cpu(p
)) == p
;
1137 /* Used instead of source_load when we know the type == 0 */
1138 unsigned long weighted_cpuload(const int cpu
)
1140 return cpu_rq(cpu
)->load
.weight
;
1143 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1145 set_task_cfs_rq(p
, cpu
);
1148 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1149 * successfuly executed on another CPU. We must ensure that updates of
1150 * per-task data have been completed by this moment.
1153 task_thread_info(p
)->cpu
= cpu
;
1157 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1158 const struct sched_class
*prev_class
,
1159 int oldprio
, int running
)
1161 if (prev_class
!= p
->sched_class
) {
1162 if (prev_class
->switched_from
)
1163 prev_class
->switched_from(rq
, p
, running
);
1164 p
->sched_class
->switched_to(rq
, p
, running
);
1166 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1172 * Is this task likely cache-hot:
1175 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1179 if (p
->sched_class
!= &fair_sched_class
)
1182 if (sysctl_sched_migration_cost
== -1)
1184 if (sysctl_sched_migration_cost
== 0)
1187 delta
= now
- p
->se
.exec_start
;
1189 return delta
< (s64
)sysctl_sched_migration_cost
;
1193 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1195 int old_cpu
= task_cpu(p
);
1196 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1197 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1198 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1201 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1203 #ifdef CONFIG_SCHEDSTATS
1204 if (p
->se
.wait_start
)
1205 p
->se
.wait_start
-= clock_offset
;
1206 if (p
->se
.sleep_start
)
1207 p
->se
.sleep_start
-= clock_offset
;
1208 if (p
->se
.block_start
)
1209 p
->se
.block_start
-= clock_offset
;
1210 if (old_cpu
!= new_cpu
) {
1211 schedstat_inc(p
, se
.nr_migrations
);
1212 if (task_hot(p
, old_rq
->clock
, NULL
))
1213 schedstat_inc(p
, se
.nr_forced2_migrations
);
1216 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1217 new_cfsrq
->min_vruntime
;
1219 __set_task_cpu(p
, new_cpu
);
1222 struct migration_req
{
1223 struct list_head list
;
1225 struct task_struct
*task
;
1228 struct completion done
;
1232 * The task's runqueue lock must be held.
1233 * Returns true if you have to wait for migration thread.
1236 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1238 struct rq
*rq
= task_rq(p
);
1241 * If the task is not on a runqueue (and not running), then
1242 * it is sufficient to simply update the task's cpu field.
1244 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1245 set_task_cpu(p
, dest_cpu
);
1249 init_completion(&req
->done
);
1251 req
->dest_cpu
= dest_cpu
;
1252 list_add(&req
->list
, &rq
->migration_queue
);
1258 * wait_task_inactive - wait for a thread to unschedule.
1260 * The caller must ensure that the task *will* unschedule sometime soon,
1261 * else this function might spin for a *long* time. This function can't
1262 * be called with interrupts off, or it may introduce deadlock with
1263 * smp_call_function() if an IPI is sent by the same process we are
1264 * waiting to become inactive.
1266 void wait_task_inactive(struct task_struct
*p
)
1268 unsigned long flags
;
1274 * We do the initial early heuristics without holding
1275 * any task-queue locks at all. We'll only try to get
1276 * the runqueue lock when things look like they will
1282 * If the task is actively running on another CPU
1283 * still, just relax and busy-wait without holding
1286 * NOTE! Since we don't hold any locks, it's not
1287 * even sure that "rq" stays as the right runqueue!
1288 * But we don't care, since "task_running()" will
1289 * return false if the runqueue has changed and p
1290 * is actually now running somewhere else!
1292 while (task_running(rq
, p
))
1296 * Ok, time to look more closely! We need the rq
1297 * lock now, to be *sure*. If we're wrong, we'll
1298 * just go back and repeat.
1300 rq
= task_rq_lock(p
, &flags
);
1301 running
= task_running(rq
, p
);
1302 on_rq
= p
->se
.on_rq
;
1303 task_rq_unlock(rq
, &flags
);
1306 * Was it really running after all now that we
1307 * checked with the proper locks actually held?
1309 * Oops. Go back and try again..
1311 if (unlikely(running
)) {
1317 * It's not enough that it's not actively running,
1318 * it must be off the runqueue _entirely_, and not
1321 * So if it wa still runnable (but just not actively
1322 * running right now), it's preempted, and we should
1323 * yield - it could be a while.
1325 if (unlikely(on_rq
)) {
1326 schedule_timeout_uninterruptible(1);
1331 * Ahh, all good. It wasn't running, and it wasn't
1332 * runnable, which means that it will never become
1333 * running in the future either. We're all done!
1340 * kick_process - kick a running thread to enter/exit the kernel
1341 * @p: the to-be-kicked thread
1343 * Cause a process which is running on another CPU to enter
1344 * kernel-mode, without any delay. (to get signals handled.)
1346 * NOTE: this function doesnt have to take the runqueue lock,
1347 * because all it wants to ensure is that the remote task enters
1348 * the kernel. If the IPI races and the task has been migrated
1349 * to another CPU then no harm is done and the purpose has been
1352 void kick_process(struct task_struct
*p
)
1358 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1359 smp_send_reschedule(cpu
);
1364 * Return a low guess at the load of a migration-source cpu weighted
1365 * according to the scheduling class and "nice" value.
1367 * We want to under-estimate the load of migration sources, to
1368 * balance conservatively.
1370 static unsigned long source_load(int cpu
, int type
)
1372 struct rq
*rq
= cpu_rq(cpu
);
1373 unsigned long total
= weighted_cpuload(cpu
);
1378 return min(rq
->cpu_load
[type
-1], total
);
1382 * Return a high guess at the load of a migration-target cpu weighted
1383 * according to the scheduling class and "nice" value.
1385 static unsigned long target_load(int cpu
, int type
)
1387 struct rq
*rq
= cpu_rq(cpu
);
1388 unsigned long total
= weighted_cpuload(cpu
);
1393 return max(rq
->cpu_load
[type
-1], total
);
1397 * Return the average load per task on the cpu's run queue
1399 static unsigned long cpu_avg_load_per_task(int cpu
)
1401 struct rq
*rq
= cpu_rq(cpu
);
1402 unsigned long total
= weighted_cpuload(cpu
);
1403 unsigned long n
= rq
->nr_running
;
1405 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1409 * find_idlest_group finds and returns the least busy CPU group within the
1412 static struct sched_group
*
1413 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1415 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1416 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1417 int load_idx
= sd
->forkexec_idx
;
1418 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1421 unsigned long load
, avg_load
;
1425 /* Skip over this group if it has no CPUs allowed */
1426 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1429 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1431 /* Tally up the load of all CPUs in the group */
1434 for_each_cpu_mask(i
, group
->cpumask
) {
1435 /* Bias balancing toward cpus of our domain */
1437 load
= source_load(i
, load_idx
);
1439 load
= target_load(i
, load_idx
);
1444 /* Adjust by relative CPU power of the group */
1445 avg_load
= sg_div_cpu_power(group
,
1446 avg_load
* SCHED_LOAD_SCALE
);
1449 this_load
= avg_load
;
1451 } else if (avg_load
< min_load
) {
1452 min_load
= avg_load
;
1455 } while (group
= group
->next
, group
!= sd
->groups
);
1457 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1463 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1466 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1469 unsigned long load
, min_load
= ULONG_MAX
;
1473 /* Traverse only the allowed CPUs */
1474 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1476 for_each_cpu_mask(i
, tmp
) {
1477 load
= weighted_cpuload(i
);
1479 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1489 * sched_balance_self: balance the current task (running on cpu) in domains
1490 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1493 * Balance, ie. select the least loaded group.
1495 * Returns the target CPU number, or the same CPU if no balancing is needed.
1497 * preempt must be disabled.
1499 static int sched_balance_self(int cpu
, int flag
)
1501 struct task_struct
*t
= current
;
1502 struct sched_domain
*tmp
, *sd
= NULL
;
1504 for_each_domain(cpu
, tmp
) {
1506 * If power savings logic is enabled for a domain, stop there.
1508 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1510 if (tmp
->flags
& flag
)
1516 struct sched_group
*group
;
1517 int new_cpu
, weight
;
1519 if (!(sd
->flags
& flag
)) {
1525 group
= find_idlest_group(sd
, t
, cpu
);
1531 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1532 if (new_cpu
== -1 || new_cpu
== cpu
) {
1533 /* Now try balancing at a lower domain level of cpu */
1538 /* Now try balancing at a lower domain level of new_cpu */
1541 weight
= cpus_weight(span
);
1542 for_each_domain(cpu
, tmp
) {
1543 if (weight
<= cpus_weight(tmp
->span
))
1545 if (tmp
->flags
& flag
)
1548 /* while loop will break here if sd == NULL */
1554 #endif /* CONFIG_SMP */
1557 * try_to_wake_up - wake up a thread
1558 * @p: the to-be-woken-up thread
1559 * @state: the mask of task states that can be woken
1560 * @sync: do a synchronous wakeup?
1562 * Put it on the run-queue if it's not already there. The "current"
1563 * thread is always on the run-queue (except when the actual
1564 * re-schedule is in progress), and as such you're allowed to do
1565 * the simpler "current->state = TASK_RUNNING" to mark yourself
1566 * runnable without the overhead of this.
1568 * returns failure only if the task is already active.
1570 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1572 int cpu
, orig_cpu
, this_cpu
, success
= 0;
1573 unsigned long flags
;
1577 rq
= task_rq_lock(p
, &flags
);
1578 old_state
= p
->state
;
1579 if (!(old_state
& state
))
1587 this_cpu
= smp_processor_id();
1590 if (unlikely(task_running(rq
, p
)))
1593 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
1594 if (cpu
!= orig_cpu
) {
1595 set_task_cpu(p
, cpu
);
1596 task_rq_unlock(rq
, &flags
);
1597 /* might preempt at this point */
1598 rq
= task_rq_lock(p
, &flags
);
1599 old_state
= p
->state
;
1600 if (!(old_state
& state
))
1605 this_cpu
= smp_processor_id();
1609 #ifdef CONFIG_SCHEDSTATS
1610 schedstat_inc(rq
, ttwu_count
);
1611 if (cpu
== this_cpu
)
1612 schedstat_inc(rq
, ttwu_local
);
1614 struct sched_domain
*sd
;
1615 for_each_domain(this_cpu
, sd
) {
1616 if (cpu_isset(cpu
, sd
->span
)) {
1617 schedstat_inc(sd
, ttwu_wake_remote
);
1625 #endif /* CONFIG_SMP */
1626 schedstat_inc(p
, se
.nr_wakeups
);
1628 schedstat_inc(p
, se
.nr_wakeups_sync
);
1629 if (orig_cpu
!= cpu
)
1630 schedstat_inc(p
, se
.nr_wakeups_migrate
);
1631 if (cpu
== this_cpu
)
1632 schedstat_inc(p
, se
.nr_wakeups_local
);
1634 schedstat_inc(p
, se
.nr_wakeups_remote
);
1635 update_rq_clock(rq
);
1636 activate_task(rq
, p
, 1);
1637 check_preempt_curr(rq
, p
);
1641 p
->state
= TASK_RUNNING
;
1643 if (p
->sched_class
->task_wake_up
)
1644 p
->sched_class
->task_wake_up(rq
, p
);
1647 task_rq_unlock(rq
, &flags
);
1652 int fastcall
wake_up_process(struct task_struct
*p
)
1654 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1655 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1657 EXPORT_SYMBOL(wake_up_process
);
1659 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1661 return try_to_wake_up(p
, state
, 0);
1665 * Perform scheduler related setup for a newly forked process p.
1666 * p is forked by current.
1668 * __sched_fork() is basic setup used by init_idle() too:
1670 static void __sched_fork(struct task_struct
*p
)
1672 p
->se
.exec_start
= 0;
1673 p
->se
.sum_exec_runtime
= 0;
1674 p
->se
.prev_sum_exec_runtime
= 0;
1676 #ifdef CONFIG_SCHEDSTATS
1677 p
->se
.wait_start
= 0;
1678 p
->se
.sum_sleep_runtime
= 0;
1679 p
->se
.sleep_start
= 0;
1680 p
->se
.block_start
= 0;
1681 p
->se
.sleep_max
= 0;
1682 p
->se
.block_max
= 0;
1684 p
->se
.slice_max
= 0;
1688 INIT_LIST_HEAD(&p
->rt
.run_list
);
1691 #ifdef CONFIG_PREEMPT_NOTIFIERS
1692 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1696 * We mark the process as running here, but have not actually
1697 * inserted it onto the runqueue yet. This guarantees that
1698 * nobody will actually run it, and a signal or other external
1699 * event cannot wake it up and insert it on the runqueue either.
1701 p
->state
= TASK_RUNNING
;
1705 * fork()/clone()-time setup:
1707 void sched_fork(struct task_struct
*p
, int clone_flags
)
1709 int cpu
= get_cpu();
1714 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1716 set_task_cpu(p
, cpu
);
1719 * Make sure we do not leak PI boosting priority to the child:
1721 p
->prio
= current
->normal_prio
;
1722 if (!rt_prio(p
->prio
))
1723 p
->sched_class
= &fair_sched_class
;
1725 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1726 if (likely(sched_info_on()))
1727 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1729 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1732 #ifdef CONFIG_PREEMPT
1733 /* Want to start with kernel preemption disabled. */
1734 task_thread_info(p
)->preempt_count
= 1;
1740 * wake_up_new_task - wake up a newly created task for the first time.
1742 * This function will do some initial scheduler statistics housekeeping
1743 * that must be done for every newly created context, then puts the task
1744 * on the runqueue and wakes it.
1746 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1748 unsigned long flags
;
1751 rq
= task_rq_lock(p
, &flags
);
1752 BUG_ON(p
->state
!= TASK_RUNNING
);
1753 update_rq_clock(rq
);
1755 p
->prio
= effective_prio(p
);
1757 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
1758 activate_task(rq
, p
, 0);
1761 * Let the scheduling class do new task startup
1762 * management (if any):
1764 p
->sched_class
->task_new(rq
, p
);
1765 inc_nr_running(p
, rq
);
1767 check_preempt_curr(rq
, p
);
1769 if (p
->sched_class
->task_wake_up
)
1770 p
->sched_class
->task_wake_up(rq
, p
);
1772 task_rq_unlock(rq
, &flags
);
1775 #ifdef CONFIG_PREEMPT_NOTIFIERS
1778 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1779 * @notifier: notifier struct to register
1781 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1783 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1785 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1788 * preempt_notifier_unregister - no longer interested in preemption notifications
1789 * @notifier: notifier struct to unregister
1791 * This is safe to call from within a preemption notifier.
1793 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1795 hlist_del(¬ifier
->link
);
1797 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1799 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1801 struct preempt_notifier
*notifier
;
1802 struct hlist_node
*node
;
1804 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1805 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1809 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1810 struct task_struct
*next
)
1812 struct preempt_notifier
*notifier
;
1813 struct hlist_node
*node
;
1815 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1816 notifier
->ops
->sched_out(notifier
, next
);
1821 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1826 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1827 struct task_struct
*next
)
1834 * prepare_task_switch - prepare to switch tasks
1835 * @rq: the runqueue preparing to switch
1836 * @prev: the current task that is being switched out
1837 * @next: the task we are going to switch to.
1839 * This is called with the rq lock held and interrupts off. It must
1840 * be paired with a subsequent finish_task_switch after the context
1843 * prepare_task_switch sets up locking and calls architecture specific
1847 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1848 struct task_struct
*next
)
1850 fire_sched_out_preempt_notifiers(prev
, next
);
1851 prepare_lock_switch(rq
, next
);
1852 prepare_arch_switch(next
);
1856 * finish_task_switch - clean up after a task-switch
1857 * @rq: runqueue associated with task-switch
1858 * @prev: the thread we just switched away from.
1860 * finish_task_switch must be called after the context switch, paired
1861 * with a prepare_task_switch call before the context switch.
1862 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1863 * and do any other architecture-specific cleanup actions.
1865 * Note that we may have delayed dropping an mm in context_switch(). If
1866 * so, we finish that here outside of the runqueue lock. (Doing it
1867 * with the lock held can cause deadlocks; see schedule() for
1870 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1871 __releases(rq
->lock
)
1873 struct mm_struct
*mm
= rq
->prev_mm
;
1879 * A task struct has one reference for the use as "current".
1880 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1881 * schedule one last time. The schedule call will never return, and
1882 * the scheduled task must drop that reference.
1883 * The test for TASK_DEAD must occur while the runqueue locks are
1884 * still held, otherwise prev could be scheduled on another cpu, die
1885 * there before we look at prev->state, and then the reference would
1887 * Manfred Spraul <manfred@colorfullife.com>
1889 prev_state
= prev
->state
;
1890 finish_arch_switch(prev
);
1891 finish_lock_switch(rq
, prev
);
1893 if (current
->sched_class
->post_schedule
)
1894 current
->sched_class
->post_schedule(rq
);
1897 fire_sched_in_preempt_notifiers(current
);
1900 if (unlikely(prev_state
== TASK_DEAD
)) {
1902 * Remove function-return probe instances associated with this
1903 * task and put them back on the free list.
1905 kprobe_flush_task(prev
);
1906 put_task_struct(prev
);
1911 * schedule_tail - first thing a freshly forked thread must call.
1912 * @prev: the thread we just switched away from.
1914 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1915 __releases(rq
->lock
)
1917 struct rq
*rq
= this_rq();
1919 finish_task_switch(rq
, prev
);
1920 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1921 /* In this case, finish_task_switch does not reenable preemption */
1924 if (current
->set_child_tid
)
1925 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1929 * context_switch - switch to the new MM and the new
1930 * thread's register state.
1933 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1934 struct task_struct
*next
)
1936 struct mm_struct
*mm
, *oldmm
;
1938 prepare_task_switch(rq
, prev
, next
);
1940 oldmm
= prev
->active_mm
;
1942 * For paravirt, this is coupled with an exit in switch_to to
1943 * combine the page table reload and the switch backend into
1946 arch_enter_lazy_cpu_mode();
1948 if (unlikely(!mm
)) {
1949 next
->active_mm
= oldmm
;
1950 atomic_inc(&oldmm
->mm_count
);
1951 enter_lazy_tlb(oldmm
, next
);
1953 switch_mm(oldmm
, mm
, next
);
1955 if (unlikely(!prev
->mm
)) {
1956 prev
->active_mm
= NULL
;
1957 rq
->prev_mm
= oldmm
;
1960 * Since the runqueue lock will be released by the next
1961 * task (which is an invalid locking op but in the case
1962 * of the scheduler it's an obvious special-case), so we
1963 * do an early lockdep release here:
1965 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1966 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1969 /* Here we just switch the register state and the stack. */
1970 switch_to(prev
, next
, prev
);
1974 * this_rq must be evaluated again because prev may have moved
1975 * CPUs since it called schedule(), thus the 'rq' on its stack
1976 * frame will be invalid.
1978 finish_task_switch(this_rq(), prev
);
1982 * nr_running, nr_uninterruptible and nr_context_switches:
1984 * externally visible scheduler statistics: current number of runnable
1985 * threads, current number of uninterruptible-sleeping threads, total
1986 * number of context switches performed since bootup.
1988 unsigned long nr_running(void)
1990 unsigned long i
, sum
= 0;
1992 for_each_online_cpu(i
)
1993 sum
+= cpu_rq(i
)->nr_running
;
1998 unsigned long nr_uninterruptible(void)
2000 unsigned long i
, sum
= 0;
2002 for_each_possible_cpu(i
)
2003 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2006 * Since we read the counters lockless, it might be slightly
2007 * inaccurate. Do not allow it to go below zero though:
2009 if (unlikely((long)sum
< 0))
2015 unsigned long long nr_context_switches(void)
2018 unsigned long long sum
= 0;
2020 for_each_possible_cpu(i
)
2021 sum
+= cpu_rq(i
)->nr_switches
;
2026 unsigned long nr_iowait(void)
2028 unsigned long i
, sum
= 0;
2030 for_each_possible_cpu(i
)
2031 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2036 unsigned long nr_active(void)
2038 unsigned long i
, running
= 0, uninterruptible
= 0;
2040 for_each_online_cpu(i
) {
2041 running
+= cpu_rq(i
)->nr_running
;
2042 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2045 if (unlikely((long)uninterruptible
< 0))
2046 uninterruptible
= 0;
2048 return running
+ uninterruptible
;
2052 * Update rq->cpu_load[] statistics. This function is usually called every
2053 * scheduler tick (TICK_NSEC).
2055 static void update_cpu_load(struct rq
*this_rq
)
2057 unsigned long this_load
= this_rq
->load
.weight
;
2060 this_rq
->nr_load_updates
++;
2062 /* Update our load: */
2063 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2064 unsigned long old_load
, new_load
;
2066 /* scale is effectively 1 << i now, and >> i divides by scale */
2068 old_load
= this_rq
->cpu_load
[i
];
2069 new_load
= this_load
;
2071 * Round up the averaging division if load is increasing. This
2072 * prevents us from getting stuck on 9 if the load is 10, for
2075 if (new_load
> old_load
)
2076 new_load
+= scale
-1;
2077 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2084 * double_rq_lock - safely lock two runqueues
2086 * Note this does not disable interrupts like task_rq_lock,
2087 * you need to do so manually before calling.
2089 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2090 __acquires(rq1
->lock
)
2091 __acquires(rq2
->lock
)
2093 BUG_ON(!irqs_disabled());
2095 spin_lock(&rq1
->lock
);
2096 __acquire(rq2
->lock
); /* Fake it out ;) */
2099 spin_lock(&rq1
->lock
);
2100 spin_lock(&rq2
->lock
);
2102 spin_lock(&rq2
->lock
);
2103 spin_lock(&rq1
->lock
);
2106 update_rq_clock(rq1
);
2107 update_rq_clock(rq2
);
2111 * double_rq_unlock - safely unlock two runqueues
2113 * Note this does not restore interrupts like task_rq_unlock,
2114 * you need to do so manually after calling.
2116 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2117 __releases(rq1
->lock
)
2118 __releases(rq2
->lock
)
2120 spin_unlock(&rq1
->lock
);
2122 spin_unlock(&rq2
->lock
);
2124 __release(rq2
->lock
);
2128 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2130 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2131 __releases(this_rq
->lock
)
2132 __acquires(busiest
->lock
)
2133 __acquires(this_rq
->lock
)
2137 if (unlikely(!irqs_disabled())) {
2138 /* printk() doesn't work good under rq->lock */
2139 spin_unlock(&this_rq
->lock
);
2142 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2143 if (busiest
< this_rq
) {
2144 spin_unlock(&this_rq
->lock
);
2145 spin_lock(&busiest
->lock
);
2146 spin_lock(&this_rq
->lock
);
2149 spin_lock(&busiest
->lock
);
2155 * If dest_cpu is allowed for this process, migrate the task to it.
2156 * This is accomplished by forcing the cpu_allowed mask to only
2157 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2158 * the cpu_allowed mask is restored.
2160 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2162 struct migration_req req
;
2163 unsigned long flags
;
2166 rq
= task_rq_lock(p
, &flags
);
2167 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2168 || unlikely(cpu_is_offline(dest_cpu
)))
2171 /* force the process onto the specified CPU */
2172 if (migrate_task(p
, dest_cpu
, &req
)) {
2173 /* Need to wait for migration thread (might exit: take ref). */
2174 struct task_struct
*mt
= rq
->migration_thread
;
2176 get_task_struct(mt
);
2177 task_rq_unlock(rq
, &flags
);
2178 wake_up_process(mt
);
2179 put_task_struct(mt
);
2180 wait_for_completion(&req
.done
);
2185 task_rq_unlock(rq
, &flags
);
2189 * sched_exec - execve() is a valuable balancing opportunity, because at
2190 * this point the task has the smallest effective memory and cache footprint.
2192 void sched_exec(void)
2194 int new_cpu
, this_cpu
= get_cpu();
2195 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2197 if (new_cpu
!= this_cpu
)
2198 sched_migrate_task(current
, new_cpu
);
2202 * pull_task - move a task from a remote runqueue to the local runqueue.
2203 * Both runqueues must be locked.
2205 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2206 struct rq
*this_rq
, int this_cpu
)
2208 deactivate_task(src_rq
, p
, 0);
2209 set_task_cpu(p
, this_cpu
);
2210 activate_task(this_rq
, p
, 0);
2212 * Note that idle threads have a prio of MAX_PRIO, for this test
2213 * to be always true for them.
2215 check_preempt_curr(this_rq
, p
);
2219 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2222 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2223 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2227 * We do not migrate tasks that are:
2228 * 1) running (obviously), or
2229 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2230 * 3) are cache-hot on their current CPU.
2232 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2233 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2238 if (task_running(rq
, p
)) {
2239 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2244 * Aggressive migration if:
2245 * 1) task is cache cold, or
2246 * 2) too many balance attempts have failed.
2249 if (!task_hot(p
, rq
->clock
, sd
) ||
2250 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2251 #ifdef CONFIG_SCHEDSTATS
2252 if (task_hot(p
, rq
->clock
, sd
)) {
2253 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2254 schedstat_inc(p
, se
.nr_forced_migrations
);
2260 if (task_hot(p
, rq
->clock
, sd
)) {
2261 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2267 static unsigned long
2268 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2269 unsigned long max_load_move
, struct sched_domain
*sd
,
2270 enum cpu_idle_type idle
, int *all_pinned
,
2271 int *this_best_prio
, struct rq_iterator
*iterator
)
2273 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2274 struct task_struct
*p
;
2275 long rem_load_move
= max_load_move
;
2277 if (max_load_move
== 0)
2283 * Start the load-balancing iterator:
2285 p
= iterator
->start(iterator
->arg
);
2287 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2290 * To help distribute high priority tasks across CPUs we don't
2291 * skip a task if it will be the highest priority task (i.e. smallest
2292 * prio value) on its new queue regardless of its load weight
2294 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2295 SCHED_LOAD_SCALE_FUZZ
;
2296 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2297 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2298 p
= iterator
->next(iterator
->arg
);
2302 pull_task(busiest
, p
, this_rq
, this_cpu
);
2304 rem_load_move
-= p
->se
.load
.weight
;
2307 * We only want to steal up to the prescribed amount of weighted load.
2309 if (rem_load_move
> 0) {
2310 if (p
->prio
< *this_best_prio
)
2311 *this_best_prio
= p
->prio
;
2312 p
= iterator
->next(iterator
->arg
);
2317 * Right now, this is one of only two places pull_task() is called,
2318 * so we can safely collect pull_task() stats here rather than
2319 * inside pull_task().
2321 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2324 *all_pinned
= pinned
;
2326 return max_load_move
- rem_load_move
;
2330 * move_tasks tries to move up to max_load_move weighted load from busiest to
2331 * this_rq, as part of a balancing operation within domain "sd".
2332 * Returns 1 if successful and 0 otherwise.
2334 * Called with both runqueues locked.
2336 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2337 unsigned long max_load_move
,
2338 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2341 const struct sched_class
*class = sched_class_highest
;
2342 unsigned long total_load_moved
= 0;
2343 int this_best_prio
= this_rq
->curr
->prio
;
2347 class->load_balance(this_rq
, this_cpu
, busiest
,
2348 max_load_move
- total_load_moved
,
2349 sd
, idle
, all_pinned
, &this_best_prio
);
2350 class = class->next
;
2351 } while (class && max_load_move
> total_load_moved
);
2353 return total_load_moved
> 0;
2357 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2358 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2359 struct rq_iterator
*iterator
)
2361 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2365 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2366 pull_task(busiest
, p
, this_rq
, this_cpu
);
2368 * Right now, this is only the second place pull_task()
2369 * is called, so we can safely collect pull_task()
2370 * stats here rather than inside pull_task().
2372 schedstat_inc(sd
, lb_gained
[idle
]);
2376 p
= iterator
->next(iterator
->arg
);
2383 * move_one_task tries to move exactly one task from busiest to this_rq, as
2384 * part of active balancing operations within "domain".
2385 * Returns 1 if successful and 0 otherwise.
2387 * Called with both runqueues locked.
2389 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2390 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2392 const struct sched_class
*class;
2394 for (class = sched_class_highest
; class; class = class->next
)
2395 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2402 * find_busiest_group finds and returns the busiest CPU group within the
2403 * domain. It calculates and returns the amount of weighted load which
2404 * should be moved to restore balance via the imbalance parameter.
2406 static struct sched_group
*
2407 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2408 unsigned long *imbalance
, enum cpu_idle_type idle
,
2409 int *sd_idle
, cpumask_t
*cpus
, int *balance
)
2411 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2412 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2413 unsigned long max_pull
;
2414 unsigned long busiest_load_per_task
, busiest_nr_running
;
2415 unsigned long this_load_per_task
, this_nr_running
;
2416 int load_idx
, group_imb
= 0;
2417 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2418 int power_savings_balance
= 1;
2419 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2420 unsigned long min_nr_running
= ULONG_MAX
;
2421 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2424 max_load
= this_load
= total_load
= total_pwr
= 0;
2425 busiest_load_per_task
= busiest_nr_running
= 0;
2426 this_load_per_task
= this_nr_running
= 0;
2427 if (idle
== CPU_NOT_IDLE
)
2428 load_idx
= sd
->busy_idx
;
2429 else if (idle
== CPU_NEWLY_IDLE
)
2430 load_idx
= sd
->newidle_idx
;
2432 load_idx
= sd
->idle_idx
;
2435 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2438 int __group_imb
= 0;
2439 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2440 unsigned long sum_nr_running
, sum_weighted_load
;
2442 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2445 balance_cpu
= first_cpu(group
->cpumask
);
2447 /* Tally up the load of all CPUs in the group */
2448 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2450 min_cpu_load
= ~0UL;
2452 for_each_cpu_mask(i
, group
->cpumask
) {
2455 if (!cpu_isset(i
, *cpus
))
2460 if (*sd_idle
&& rq
->nr_running
)
2463 /* Bias balancing toward cpus of our domain */
2465 if (idle_cpu(i
) && !first_idle_cpu
) {
2470 load
= target_load(i
, load_idx
);
2472 load
= source_load(i
, load_idx
);
2473 if (load
> max_cpu_load
)
2474 max_cpu_load
= load
;
2475 if (min_cpu_load
> load
)
2476 min_cpu_load
= load
;
2480 sum_nr_running
+= rq
->nr_running
;
2481 sum_weighted_load
+= weighted_cpuload(i
);
2485 * First idle cpu or the first cpu(busiest) in this sched group
2486 * is eligible for doing load balancing at this and above
2487 * domains. In the newly idle case, we will allow all the cpu's
2488 * to do the newly idle load balance.
2490 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
2491 balance_cpu
!= this_cpu
&& balance
) {
2496 total_load
+= avg_load
;
2497 total_pwr
+= group
->__cpu_power
;
2499 /* Adjust by relative CPU power of the group */
2500 avg_load
= sg_div_cpu_power(group
,
2501 avg_load
* SCHED_LOAD_SCALE
);
2503 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
2506 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
2509 this_load
= avg_load
;
2511 this_nr_running
= sum_nr_running
;
2512 this_load_per_task
= sum_weighted_load
;
2513 } else if (avg_load
> max_load
&&
2514 (sum_nr_running
> group_capacity
|| __group_imb
)) {
2515 max_load
= avg_load
;
2517 busiest_nr_running
= sum_nr_running
;
2518 busiest_load_per_task
= sum_weighted_load
;
2519 group_imb
= __group_imb
;
2522 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2524 * Busy processors will not participate in power savings
2527 if (idle
== CPU_NOT_IDLE
||
2528 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2532 * If the local group is idle or completely loaded
2533 * no need to do power savings balance at this domain
2535 if (local_group
&& (this_nr_running
>= group_capacity
||
2537 power_savings_balance
= 0;
2540 * If a group is already running at full capacity or idle,
2541 * don't include that group in power savings calculations
2543 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2548 * Calculate the group which has the least non-idle load.
2549 * This is the group from where we need to pick up the load
2552 if ((sum_nr_running
< min_nr_running
) ||
2553 (sum_nr_running
== min_nr_running
&&
2554 first_cpu(group
->cpumask
) <
2555 first_cpu(group_min
->cpumask
))) {
2557 min_nr_running
= sum_nr_running
;
2558 min_load_per_task
= sum_weighted_load
/
2563 * Calculate the group which is almost near its
2564 * capacity but still has some space to pick up some load
2565 * from other group and save more power
2567 if (sum_nr_running
<= group_capacity
- 1) {
2568 if (sum_nr_running
> leader_nr_running
||
2569 (sum_nr_running
== leader_nr_running
&&
2570 first_cpu(group
->cpumask
) >
2571 first_cpu(group_leader
->cpumask
))) {
2572 group_leader
= group
;
2573 leader_nr_running
= sum_nr_running
;
2578 group
= group
->next
;
2579 } while (group
!= sd
->groups
);
2581 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2584 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2586 if (this_load
>= avg_load
||
2587 100*max_load
<= sd
->imbalance_pct
*this_load
)
2590 busiest_load_per_task
/= busiest_nr_running
;
2592 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
2595 * We're trying to get all the cpus to the average_load, so we don't
2596 * want to push ourselves above the average load, nor do we wish to
2597 * reduce the max loaded cpu below the average load, as either of these
2598 * actions would just result in more rebalancing later, and ping-pong
2599 * tasks around. Thus we look for the minimum possible imbalance.
2600 * Negative imbalances (*we* are more loaded than anyone else) will
2601 * be counted as no imbalance for these purposes -- we can't fix that
2602 * by pulling tasks to us. Be careful of negative numbers as they'll
2603 * appear as very large values with unsigned longs.
2605 if (max_load
<= busiest_load_per_task
)
2609 * In the presence of smp nice balancing, certain scenarios can have
2610 * max load less than avg load(as we skip the groups at or below
2611 * its cpu_power, while calculating max_load..)
2613 if (max_load
< avg_load
) {
2615 goto small_imbalance
;
2618 /* Don't want to pull so many tasks that a group would go idle */
2619 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2621 /* How much load to actually move to equalise the imbalance */
2622 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
2623 (avg_load
- this_load
) * this->__cpu_power
)
2627 * if *imbalance is less than the average load per runnable task
2628 * there is no gaurantee that any tasks will be moved so we'll have
2629 * a think about bumping its value to force at least one task to be
2632 if (*imbalance
< busiest_load_per_task
) {
2633 unsigned long tmp
, pwr_now
, pwr_move
;
2637 pwr_move
= pwr_now
= 0;
2639 if (this_nr_running
) {
2640 this_load_per_task
/= this_nr_running
;
2641 if (busiest_load_per_task
> this_load_per_task
)
2644 this_load_per_task
= SCHED_LOAD_SCALE
;
2646 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
2647 busiest_load_per_task
* imbn
) {
2648 *imbalance
= busiest_load_per_task
;
2653 * OK, we don't have enough imbalance to justify moving tasks,
2654 * however we may be able to increase total CPU power used by
2658 pwr_now
+= busiest
->__cpu_power
*
2659 min(busiest_load_per_task
, max_load
);
2660 pwr_now
+= this->__cpu_power
*
2661 min(this_load_per_task
, this_load
);
2662 pwr_now
/= SCHED_LOAD_SCALE
;
2664 /* Amount of load we'd subtract */
2665 tmp
= sg_div_cpu_power(busiest
,
2666 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2668 pwr_move
+= busiest
->__cpu_power
*
2669 min(busiest_load_per_task
, max_load
- tmp
);
2671 /* Amount of load we'd add */
2672 if (max_load
* busiest
->__cpu_power
<
2673 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2674 tmp
= sg_div_cpu_power(this,
2675 max_load
* busiest
->__cpu_power
);
2677 tmp
= sg_div_cpu_power(this,
2678 busiest_load_per_task
* SCHED_LOAD_SCALE
);
2679 pwr_move
+= this->__cpu_power
*
2680 min(this_load_per_task
, this_load
+ tmp
);
2681 pwr_move
/= SCHED_LOAD_SCALE
;
2683 /* Move if we gain throughput */
2684 if (pwr_move
> pwr_now
)
2685 *imbalance
= busiest_load_per_task
;
2691 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2692 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2695 if (this == group_leader
&& group_leader
!= group_min
) {
2696 *imbalance
= min_load_per_task
;
2706 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2709 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
2710 unsigned long imbalance
, cpumask_t
*cpus
)
2712 struct rq
*busiest
= NULL
, *rq
;
2713 unsigned long max_load
= 0;
2716 for_each_cpu_mask(i
, group
->cpumask
) {
2719 if (!cpu_isset(i
, *cpus
))
2723 wl
= weighted_cpuload(i
);
2725 if (rq
->nr_running
== 1 && wl
> imbalance
)
2728 if (wl
> max_load
) {
2738 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2739 * so long as it is large enough.
2741 #define MAX_PINNED_INTERVAL 512
2744 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2745 * tasks if there is an imbalance.
2747 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2748 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2751 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2752 struct sched_group
*group
;
2753 unsigned long imbalance
;
2755 cpumask_t cpus
= CPU_MASK_ALL
;
2756 unsigned long flags
;
2759 * When power savings policy is enabled for the parent domain, idle
2760 * sibling can pick up load irrespective of busy siblings. In this case,
2761 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2762 * portraying it as CPU_NOT_IDLE.
2764 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2765 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2768 schedstat_inc(sd
, lb_count
[idle
]);
2771 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2778 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2782 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2784 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2788 BUG_ON(busiest
== this_rq
);
2790 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2793 if (busiest
->nr_running
> 1) {
2795 * Attempt to move tasks. If find_busiest_group has found
2796 * an imbalance but busiest->nr_running <= 1, the group is
2797 * still unbalanced. ld_moved simply stays zero, so it is
2798 * correctly treated as an imbalance.
2800 local_irq_save(flags
);
2801 double_rq_lock(this_rq
, busiest
);
2802 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2803 imbalance
, sd
, idle
, &all_pinned
);
2804 double_rq_unlock(this_rq
, busiest
);
2805 local_irq_restore(flags
);
2808 * some other cpu did the load balance for us.
2810 if (ld_moved
&& this_cpu
!= smp_processor_id())
2811 resched_cpu(this_cpu
);
2813 /* All tasks on this runqueue were pinned by CPU affinity */
2814 if (unlikely(all_pinned
)) {
2815 cpu_clear(cpu_of(busiest
), cpus
);
2816 if (!cpus_empty(cpus
))
2823 schedstat_inc(sd
, lb_failed
[idle
]);
2824 sd
->nr_balance_failed
++;
2826 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2828 spin_lock_irqsave(&busiest
->lock
, flags
);
2830 /* don't kick the migration_thread, if the curr
2831 * task on busiest cpu can't be moved to this_cpu
2833 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2834 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2836 goto out_one_pinned
;
2839 if (!busiest
->active_balance
) {
2840 busiest
->active_balance
= 1;
2841 busiest
->push_cpu
= this_cpu
;
2844 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2846 wake_up_process(busiest
->migration_thread
);
2849 * We've kicked active balancing, reset the failure
2852 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2855 sd
->nr_balance_failed
= 0;
2857 if (likely(!active_balance
)) {
2858 /* We were unbalanced, so reset the balancing interval */
2859 sd
->balance_interval
= sd
->min_interval
;
2862 * If we've begun active balancing, start to back off. This
2863 * case may not be covered by the all_pinned logic if there
2864 * is only 1 task on the busy runqueue (because we don't call
2867 if (sd
->balance_interval
< sd
->max_interval
)
2868 sd
->balance_interval
*= 2;
2871 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2872 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2877 schedstat_inc(sd
, lb_balanced
[idle
]);
2879 sd
->nr_balance_failed
= 0;
2882 /* tune up the balancing interval */
2883 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2884 (sd
->balance_interval
< sd
->max_interval
))
2885 sd
->balance_interval
*= 2;
2887 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2888 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2894 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2895 * tasks if there is an imbalance.
2897 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2898 * this_rq is locked.
2901 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2903 struct sched_group
*group
;
2904 struct rq
*busiest
= NULL
;
2905 unsigned long imbalance
;
2909 cpumask_t cpus
= CPU_MASK_ALL
;
2912 * When power savings policy is enabled for the parent domain, idle
2913 * sibling can pick up load irrespective of busy siblings. In this case,
2914 * let the state of idle sibling percolate up as IDLE, instead of
2915 * portraying it as CPU_NOT_IDLE.
2917 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2918 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2921 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
2923 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
2924 &sd_idle
, &cpus
, NULL
);
2926 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
2930 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
,
2933 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
2937 BUG_ON(busiest
== this_rq
);
2939 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
2942 if (busiest
->nr_running
> 1) {
2943 /* Attempt to move tasks */
2944 double_lock_balance(this_rq
, busiest
);
2945 /* this_rq->clock is already updated */
2946 update_rq_clock(busiest
);
2947 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2948 imbalance
, sd
, CPU_NEWLY_IDLE
,
2950 spin_unlock(&busiest
->lock
);
2952 if (unlikely(all_pinned
)) {
2953 cpu_clear(cpu_of(busiest
), cpus
);
2954 if (!cpus_empty(cpus
))
2960 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
2961 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2962 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2965 sd
->nr_balance_failed
= 0;
2970 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
2971 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2972 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2974 sd
->nr_balance_failed
= 0;
2980 * idle_balance is called by schedule() if this_cpu is about to become
2981 * idle. Attempts to pull tasks from other CPUs.
2983 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2985 struct sched_domain
*sd
;
2986 int pulled_task
= -1;
2987 unsigned long next_balance
= jiffies
+ HZ
;
2989 for_each_domain(this_cpu
, sd
) {
2990 unsigned long interval
;
2992 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2995 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
2996 /* If we've pulled tasks over stop searching: */
2997 pulled_task
= load_balance_newidle(this_cpu
,
3000 interval
= msecs_to_jiffies(sd
->balance_interval
);
3001 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3002 next_balance
= sd
->last_balance
+ interval
;
3006 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3008 * We are going idle. next_balance may be set based on
3009 * a busy processor. So reset next_balance.
3011 this_rq
->next_balance
= next_balance
;
3016 * active_load_balance is run by migration threads. It pushes running tasks
3017 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3018 * running on each physical CPU where possible, and avoids physical /
3019 * logical imbalances.
3021 * Called with busiest_rq locked.
3023 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3025 int target_cpu
= busiest_rq
->push_cpu
;
3026 struct sched_domain
*sd
;
3027 struct rq
*target_rq
;
3029 /* Is there any task to move? */
3030 if (busiest_rq
->nr_running
<= 1)
3033 target_rq
= cpu_rq(target_cpu
);
3036 * This condition is "impossible", if it occurs
3037 * we need to fix it. Originally reported by
3038 * Bjorn Helgaas on a 128-cpu setup.
3040 BUG_ON(busiest_rq
== target_rq
);
3042 /* move a task from busiest_rq to target_rq */
3043 double_lock_balance(busiest_rq
, target_rq
);
3044 update_rq_clock(busiest_rq
);
3045 update_rq_clock(target_rq
);
3047 /* Search for an sd spanning us and the target CPU. */
3048 for_each_domain(target_cpu
, sd
) {
3049 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3050 cpu_isset(busiest_cpu
, sd
->span
))
3055 schedstat_inc(sd
, alb_count
);
3057 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3059 schedstat_inc(sd
, alb_pushed
);
3061 schedstat_inc(sd
, alb_failed
);
3063 spin_unlock(&target_rq
->lock
);
3068 atomic_t load_balancer
;
3070 } nohz ____cacheline_aligned
= {
3071 .load_balancer
= ATOMIC_INIT(-1),
3072 .cpu_mask
= CPU_MASK_NONE
,
3076 * This routine will try to nominate the ilb (idle load balancing)
3077 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3078 * load balancing on behalf of all those cpus. If all the cpus in the system
3079 * go into this tickless mode, then there will be no ilb owner (as there is
3080 * no need for one) and all the cpus will sleep till the next wakeup event
3083 * For the ilb owner, tick is not stopped. And this tick will be used
3084 * for idle load balancing. ilb owner will still be part of
3087 * While stopping the tick, this cpu will become the ilb owner if there
3088 * is no other owner. And will be the owner till that cpu becomes busy
3089 * or if all cpus in the system stop their ticks at which point
3090 * there is no need for ilb owner.
3092 * When the ilb owner becomes busy, it nominates another owner, during the
3093 * next busy scheduler_tick()
3095 int select_nohz_load_balancer(int stop_tick
)
3097 int cpu
= smp_processor_id();
3100 cpu_set(cpu
, nohz
.cpu_mask
);
3101 cpu_rq(cpu
)->in_nohz_recently
= 1;
3104 * If we are going offline and still the leader, give up!
3106 if (cpu_is_offline(cpu
) &&
3107 atomic_read(&nohz
.load_balancer
) == cpu
) {
3108 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3113 /* time for ilb owner also to sleep */
3114 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3115 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3116 atomic_set(&nohz
.load_balancer
, -1);
3120 if (atomic_read(&nohz
.load_balancer
) == -1) {
3121 /* make me the ilb owner */
3122 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3124 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3127 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3130 cpu_clear(cpu
, nohz
.cpu_mask
);
3132 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3133 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3140 static DEFINE_SPINLOCK(balancing
);
3143 * It checks each scheduling domain to see if it is due to be balanced,
3144 * and initiates a balancing operation if so.
3146 * Balancing parameters are set up in arch_init_sched_domains.
3148 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3151 struct rq
*rq
= cpu_rq(cpu
);
3152 unsigned long interval
;
3153 struct sched_domain
*sd
;
3154 /* Earliest time when we have to do rebalance again */
3155 unsigned long next_balance
= jiffies
+ 60*HZ
;
3156 int update_next_balance
= 0;
3158 for_each_domain(cpu
, sd
) {
3159 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3162 interval
= sd
->balance_interval
;
3163 if (idle
!= CPU_IDLE
)
3164 interval
*= sd
->busy_factor
;
3166 /* scale ms to jiffies */
3167 interval
= msecs_to_jiffies(interval
);
3168 if (unlikely(!interval
))
3170 if (interval
> HZ
*NR_CPUS
/10)
3171 interval
= HZ
*NR_CPUS
/10;
3174 if (sd
->flags
& SD_SERIALIZE
) {
3175 if (!spin_trylock(&balancing
))
3179 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3180 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
3182 * We've pulled tasks over so either we're no
3183 * longer idle, or one of our SMT siblings is
3186 idle
= CPU_NOT_IDLE
;
3188 sd
->last_balance
= jiffies
;
3190 if (sd
->flags
& SD_SERIALIZE
)
3191 spin_unlock(&balancing
);
3193 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3194 next_balance
= sd
->last_balance
+ interval
;
3195 update_next_balance
= 1;
3199 * Stop the load balance at this level. There is another
3200 * CPU in our sched group which is doing load balancing more
3208 * next_balance will be updated only when there is a need.
3209 * When the cpu is attached to null domain for ex, it will not be
3212 if (likely(update_next_balance
))
3213 rq
->next_balance
= next_balance
;
3217 * run_rebalance_domains is triggered when needed from the scheduler tick.
3218 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3219 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3221 static void run_rebalance_domains(struct softirq_action
*h
)
3223 int this_cpu
= smp_processor_id();
3224 struct rq
*this_rq
= cpu_rq(this_cpu
);
3225 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3226 CPU_IDLE
: CPU_NOT_IDLE
;
3228 rebalance_domains(this_cpu
, idle
);
3232 * If this cpu is the owner for idle load balancing, then do the
3233 * balancing on behalf of the other idle cpus whose ticks are
3236 if (this_rq
->idle_at_tick
&&
3237 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3238 cpumask_t cpus
= nohz
.cpu_mask
;
3242 cpu_clear(this_cpu
, cpus
);
3243 for_each_cpu_mask(balance_cpu
, cpus
) {
3245 * If this cpu gets work to do, stop the load balancing
3246 * work being done for other cpus. Next load
3247 * balancing owner will pick it up.
3252 rebalance_domains(balance_cpu
, CPU_IDLE
);
3254 rq
= cpu_rq(balance_cpu
);
3255 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3256 this_rq
->next_balance
= rq
->next_balance
;
3263 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3265 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3266 * idle load balancing owner or decide to stop the periodic load balancing,
3267 * if the whole system is idle.
3269 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3273 * If we were in the nohz mode recently and busy at the current
3274 * scheduler tick, then check if we need to nominate new idle
3277 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3278 rq
->in_nohz_recently
= 0;
3280 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3281 cpu_clear(cpu
, nohz
.cpu_mask
);
3282 atomic_set(&nohz
.load_balancer
, -1);
3285 if (atomic_read(&nohz
.load_balancer
) == -1) {
3287 * simple selection for now: Nominate the
3288 * first cpu in the nohz list to be the next
3291 * TBD: Traverse the sched domains and nominate
3292 * the nearest cpu in the nohz.cpu_mask.
3294 int ilb
= first_cpu(nohz
.cpu_mask
);
3302 * If this cpu is idle and doing idle load balancing for all the
3303 * cpus with ticks stopped, is it time for that to stop?
3305 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3306 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3312 * If this cpu is idle and the idle load balancing is done by
3313 * someone else, then no need raise the SCHED_SOFTIRQ
3315 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3316 cpu_isset(cpu
, nohz
.cpu_mask
))
3319 if (time_after_eq(jiffies
, rq
->next_balance
))
3320 raise_softirq(SCHED_SOFTIRQ
);
3323 #else /* CONFIG_SMP */
3326 * on UP we do not need to balance between CPUs:
3328 static inline void idle_balance(int cpu
, struct rq
*rq
)
3334 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3336 EXPORT_PER_CPU_SYMBOL(kstat
);
3339 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3340 * that have not yet been banked in case the task is currently running.
3342 unsigned long long task_sched_runtime(struct task_struct
*p
)
3344 unsigned long flags
;
3348 rq
= task_rq_lock(p
, &flags
);
3349 ns
= p
->se
.sum_exec_runtime
;
3350 if (task_current(rq
, p
)) {
3351 update_rq_clock(rq
);
3352 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3353 if ((s64
)delta_exec
> 0)
3356 task_rq_unlock(rq
, &flags
);
3362 * Account user cpu time to a process.
3363 * @p: the process that the cpu time gets accounted to
3364 * @cputime: the cpu time spent in user space since the last update
3366 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3368 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3371 p
->utime
= cputime_add(p
->utime
, cputime
);
3373 /* Add user time to cpustat. */
3374 tmp
= cputime_to_cputime64(cputime
);
3375 if (TASK_NICE(p
) > 0)
3376 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3378 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3382 * Account guest cpu time to a process.
3383 * @p: the process that the cpu time gets accounted to
3384 * @cputime: the cpu time spent in virtual machine since the last update
3386 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3389 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3391 tmp
= cputime_to_cputime64(cputime
);
3393 p
->utime
= cputime_add(p
->utime
, cputime
);
3394 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3396 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3397 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3401 * Account scaled user cpu time to a process.
3402 * @p: the process that the cpu time gets accounted to
3403 * @cputime: the cpu time spent in user space since the last update
3405 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3407 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3411 * Account system cpu time to a process.
3412 * @p: the process that the cpu time gets accounted to
3413 * @hardirq_offset: the offset to subtract from hardirq_count()
3414 * @cputime: the cpu time spent in kernel space since the last update
3416 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3419 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3420 struct rq
*rq
= this_rq();
3423 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0))
3424 return account_guest_time(p
, cputime
);
3426 p
->stime
= cputime_add(p
->stime
, cputime
);
3428 /* Add system time to cpustat. */
3429 tmp
= cputime_to_cputime64(cputime
);
3430 if (hardirq_count() - hardirq_offset
)
3431 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3432 else if (softirq_count())
3433 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3434 else if (p
!= rq
->idle
)
3435 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3436 else if (atomic_read(&rq
->nr_iowait
) > 0)
3437 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3439 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3440 /* Account for system time used */
3441 acct_update_integrals(p
);
3445 * Account scaled system cpu time to a process.
3446 * @p: the process that the cpu time gets accounted to
3447 * @hardirq_offset: the offset to subtract from hardirq_count()
3448 * @cputime: the cpu time spent in kernel space since the last update
3450 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3452 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
3456 * Account for involuntary wait time.
3457 * @p: the process from which the cpu time has been stolen
3458 * @steal: the cpu time spent in involuntary wait
3460 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3462 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3463 cputime64_t tmp
= cputime_to_cputime64(steal
);
3464 struct rq
*rq
= this_rq();
3466 if (p
== rq
->idle
) {
3467 p
->stime
= cputime_add(p
->stime
, steal
);
3468 if (atomic_read(&rq
->nr_iowait
) > 0)
3469 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3471 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3473 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3477 * This function gets called by the timer code, with HZ frequency.
3478 * We call it with interrupts disabled.
3480 * It also gets called by the fork code, when changing the parent's
3483 void scheduler_tick(void)
3485 int cpu
= smp_processor_id();
3486 struct rq
*rq
= cpu_rq(cpu
);
3487 struct task_struct
*curr
= rq
->curr
;
3488 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
3490 spin_lock(&rq
->lock
);
3491 __update_rq_clock(rq
);
3493 * Let rq->clock advance by at least TICK_NSEC:
3495 if (unlikely(rq
->clock
< next_tick
))
3496 rq
->clock
= next_tick
;
3497 rq
->tick_timestamp
= rq
->clock
;
3498 update_cpu_load(rq
);
3499 if (curr
!= rq
->idle
) /* FIXME: needed? */
3500 curr
->sched_class
->task_tick(rq
, curr
);
3501 spin_unlock(&rq
->lock
);
3504 rq
->idle_at_tick
= idle_cpu(cpu
);
3505 trigger_load_balance(rq
, cpu
);
3509 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3511 void fastcall
add_preempt_count(int val
)
3516 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3518 preempt_count() += val
;
3520 * Spinlock count overflowing soon?
3522 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3525 EXPORT_SYMBOL(add_preempt_count
);
3527 void fastcall
sub_preempt_count(int val
)
3532 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3535 * Is the spinlock portion underflowing?
3537 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3538 !(preempt_count() & PREEMPT_MASK
)))
3541 preempt_count() -= val
;
3543 EXPORT_SYMBOL(sub_preempt_count
);
3548 * Print scheduling while atomic bug:
3550 static noinline
void __schedule_bug(struct task_struct
*prev
)
3552 struct pt_regs
*regs
= get_irq_regs();
3554 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3555 prev
->comm
, prev
->pid
, preempt_count());
3557 debug_show_held_locks(prev
);
3558 if (irqs_disabled())
3559 print_irqtrace_events(prev
);
3568 * Various schedule()-time debugging checks and statistics:
3570 static inline void schedule_debug(struct task_struct
*prev
)
3573 * Test if we are atomic. Since do_exit() needs to call into
3574 * schedule() atomically, we ignore that path for now.
3575 * Otherwise, whine if we are scheduling when we should not be.
3577 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
3578 __schedule_bug(prev
);
3580 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3582 schedstat_inc(this_rq(), sched_count
);
3583 #ifdef CONFIG_SCHEDSTATS
3584 if (unlikely(prev
->lock_depth
>= 0)) {
3585 schedstat_inc(this_rq(), bkl_count
);
3586 schedstat_inc(prev
, sched_info
.bkl_count
);
3592 * Pick up the highest-prio task:
3594 static inline struct task_struct
*
3595 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3597 const struct sched_class
*class;
3598 struct task_struct
*p
;
3601 * Optimization: we know that if all tasks are in
3602 * the fair class we can call that function directly:
3604 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3605 p
= fair_sched_class
.pick_next_task(rq
);
3610 class = sched_class_highest
;
3612 p
= class->pick_next_task(rq
);
3616 * Will never be NULL as the idle class always
3617 * returns a non-NULL p:
3619 class = class->next
;
3624 * schedule() is the main scheduler function.
3626 asmlinkage
void __sched
schedule(void)
3628 struct task_struct
*prev
, *next
;
3635 cpu
= smp_processor_id();
3639 switch_count
= &prev
->nivcsw
;
3641 release_kernel_lock(prev
);
3642 need_resched_nonpreemptible
:
3644 schedule_debug(prev
);
3647 * Do the rq-clock update outside the rq lock:
3649 local_irq_disable();
3650 __update_rq_clock(rq
);
3651 spin_lock(&rq
->lock
);
3652 clear_tsk_need_resched(prev
);
3654 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3655 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3656 unlikely(signal_pending(prev
)))) {
3657 prev
->state
= TASK_RUNNING
;
3659 deactivate_task(rq
, prev
, 1);
3661 switch_count
= &prev
->nvcsw
;
3665 if (prev
->sched_class
->pre_schedule
)
3666 prev
->sched_class
->pre_schedule(rq
, prev
);
3669 if (unlikely(!rq
->nr_running
))
3670 idle_balance(cpu
, rq
);
3672 prev
->sched_class
->put_prev_task(rq
, prev
);
3673 next
= pick_next_task(rq
, prev
);
3675 sched_info_switch(prev
, next
);
3677 if (likely(prev
!= next
)) {
3682 context_switch(rq
, prev
, next
); /* unlocks the rq */
3684 spin_unlock_irq(&rq
->lock
);
3686 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3687 cpu
= smp_processor_id();
3689 goto need_resched_nonpreemptible
;
3691 preempt_enable_no_resched();
3692 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3695 EXPORT_SYMBOL(schedule
);
3697 #ifdef CONFIG_PREEMPT
3699 * this is the entry point to schedule() from in-kernel preemption
3700 * off of preempt_enable. Kernel preemptions off return from interrupt
3701 * occur there and call schedule directly.
3703 asmlinkage
void __sched
preempt_schedule(void)
3705 struct thread_info
*ti
= current_thread_info();
3706 #ifdef CONFIG_PREEMPT_BKL
3707 struct task_struct
*task
= current
;
3708 int saved_lock_depth
;
3711 * If there is a non-zero preempt_count or interrupts are disabled,
3712 * we do not want to preempt the current task. Just return..
3714 if (likely(ti
->preempt_count
|| irqs_disabled()))
3718 add_preempt_count(PREEMPT_ACTIVE
);
3721 * We keep the big kernel semaphore locked, but we
3722 * clear ->lock_depth so that schedule() doesnt
3723 * auto-release the semaphore:
3725 #ifdef CONFIG_PREEMPT_BKL
3726 saved_lock_depth
= task
->lock_depth
;
3727 task
->lock_depth
= -1;
3730 #ifdef CONFIG_PREEMPT_BKL
3731 task
->lock_depth
= saved_lock_depth
;
3733 sub_preempt_count(PREEMPT_ACTIVE
);
3736 * Check again in case we missed a preemption opportunity
3737 * between schedule and now.
3740 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3742 EXPORT_SYMBOL(preempt_schedule
);
3745 * this is the entry point to schedule() from kernel preemption
3746 * off of irq context.
3747 * Note, that this is called and return with irqs disabled. This will
3748 * protect us against recursive calling from irq.
3750 asmlinkage
void __sched
preempt_schedule_irq(void)
3752 struct thread_info
*ti
= current_thread_info();
3753 #ifdef CONFIG_PREEMPT_BKL
3754 struct task_struct
*task
= current
;
3755 int saved_lock_depth
;
3757 /* Catch callers which need to be fixed */
3758 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3761 add_preempt_count(PREEMPT_ACTIVE
);
3764 * We keep the big kernel semaphore locked, but we
3765 * clear ->lock_depth so that schedule() doesnt
3766 * auto-release the semaphore:
3768 #ifdef CONFIG_PREEMPT_BKL
3769 saved_lock_depth
= task
->lock_depth
;
3770 task
->lock_depth
= -1;
3774 local_irq_disable();
3775 #ifdef CONFIG_PREEMPT_BKL
3776 task
->lock_depth
= saved_lock_depth
;
3778 sub_preempt_count(PREEMPT_ACTIVE
);
3781 * Check again in case we missed a preemption opportunity
3782 * between schedule and now.
3785 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
3788 #endif /* CONFIG_PREEMPT */
3790 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3793 return try_to_wake_up(curr
->private, mode
, sync
);
3795 EXPORT_SYMBOL(default_wake_function
);
3798 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3799 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3800 * number) then we wake all the non-exclusive tasks and one exclusive task.
3802 * There are circumstances in which we can try to wake a task which has already
3803 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3804 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3806 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3807 int nr_exclusive
, int sync
, void *key
)
3809 wait_queue_t
*curr
, *next
;
3811 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3812 unsigned flags
= curr
->flags
;
3814 if (curr
->func(curr
, mode
, sync
, key
) &&
3815 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3821 * __wake_up - wake up threads blocked on a waitqueue.
3823 * @mode: which threads
3824 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3825 * @key: is directly passed to the wakeup function
3827 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3828 int nr_exclusive
, void *key
)
3830 unsigned long flags
;
3832 spin_lock_irqsave(&q
->lock
, flags
);
3833 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3834 spin_unlock_irqrestore(&q
->lock
, flags
);
3836 EXPORT_SYMBOL(__wake_up
);
3839 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3841 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3843 __wake_up_common(q
, mode
, 1, 0, NULL
);
3847 * __wake_up_sync - wake up threads blocked on a waitqueue.
3849 * @mode: which threads
3850 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3852 * The sync wakeup differs that the waker knows that it will schedule
3853 * away soon, so while the target thread will be woken up, it will not
3854 * be migrated to another CPU - ie. the two threads are 'synchronized'
3855 * with each other. This can prevent needless bouncing between CPUs.
3857 * On UP it can prevent extra preemption.
3860 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3862 unsigned long flags
;
3868 if (unlikely(!nr_exclusive
))
3871 spin_lock_irqsave(&q
->lock
, flags
);
3872 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3873 spin_unlock_irqrestore(&q
->lock
, flags
);
3875 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3877 void complete(struct completion
*x
)
3879 unsigned long flags
;
3881 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3883 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3885 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3887 EXPORT_SYMBOL(complete
);
3889 void complete_all(struct completion
*x
)
3891 unsigned long flags
;
3893 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3894 x
->done
+= UINT_MAX
/2;
3895 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3897 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3899 EXPORT_SYMBOL(complete_all
);
3901 static inline long __sched
3902 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3905 DECLARE_WAITQUEUE(wait
, current
);
3907 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3908 __add_wait_queue_tail(&x
->wait
, &wait
);
3910 if (state
== TASK_INTERRUPTIBLE
&&
3911 signal_pending(current
)) {
3912 __remove_wait_queue(&x
->wait
, &wait
);
3913 return -ERESTARTSYS
;
3915 __set_current_state(state
);
3916 spin_unlock_irq(&x
->wait
.lock
);
3917 timeout
= schedule_timeout(timeout
);
3918 spin_lock_irq(&x
->wait
.lock
);
3920 __remove_wait_queue(&x
->wait
, &wait
);
3924 __remove_wait_queue(&x
->wait
, &wait
);
3931 wait_for_common(struct completion
*x
, long timeout
, int state
)
3935 spin_lock_irq(&x
->wait
.lock
);
3936 timeout
= do_wait_for_common(x
, timeout
, state
);
3937 spin_unlock_irq(&x
->wait
.lock
);
3941 void __sched
wait_for_completion(struct completion
*x
)
3943 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3945 EXPORT_SYMBOL(wait_for_completion
);
3947 unsigned long __sched
3948 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3950 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3952 EXPORT_SYMBOL(wait_for_completion_timeout
);
3954 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3956 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3957 if (t
== -ERESTARTSYS
)
3961 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3963 unsigned long __sched
3964 wait_for_completion_interruptible_timeout(struct completion
*x
,
3965 unsigned long timeout
)
3967 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3969 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3972 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3974 unsigned long flags
;
3977 init_waitqueue_entry(&wait
, current
);
3979 __set_current_state(state
);
3981 spin_lock_irqsave(&q
->lock
, flags
);
3982 __add_wait_queue(q
, &wait
);
3983 spin_unlock(&q
->lock
);
3984 timeout
= schedule_timeout(timeout
);
3985 spin_lock_irq(&q
->lock
);
3986 __remove_wait_queue(q
, &wait
);
3987 spin_unlock_irqrestore(&q
->lock
, flags
);
3992 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3994 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3996 EXPORT_SYMBOL(interruptible_sleep_on
);
3999 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4001 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4003 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4005 void __sched
sleep_on(wait_queue_head_t
*q
)
4007 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4009 EXPORT_SYMBOL(sleep_on
);
4011 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4013 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4015 EXPORT_SYMBOL(sleep_on_timeout
);
4017 #ifdef CONFIG_RT_MUTEXES
4020 * rt_mutex_setprio - set the current priority of a task
4022 * @prio: prio value (kernel-internal form)
4024 * This function changes the 'effective' priority of a task. It does
4025 * not touch ->normal_prio like __setscheduler().
4027 * Used by the rt_mutex code to implement priority inheritance logic.
4029 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4031 unsigned long flags
;
4032 int oldprio
, on_rq
, running
;
4034 const struct sched_class
*prev_class
= p
->sched_class
;
4036 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4038 rq
= task_rq_lock(p
, &flags
);
4039 update_rq_clock(rq
);
4042 on_rq
= p
->se
.on_rq
;
4043 running
= task_current(rq
, p
);
4045 dequeue_task(rq
, p
, 0);
4047 p
->sched_class
->put_prev_task(rq
, p
);
4051 p
->sched_class
= &rt_sched_class
;
4053 p
->sched_class
= &fair_sched_class
;
4059 p
->sched_class
->set_curr_task(rq
);
4061 enqueue_task(rq
, p
, 0);
4063 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4065 task_rq_unlock(rq
, &flags
);
4070 void set_user_nice(struct task_struct
*p
, long nice
)
4072 int old_prio
, delta
, on_rq
;
4073 unsigned long flags
;
4076 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4079 * We have to be careful, if called from sys_setpriority(),
4080 * the task might be in the middle of scheduling on another CPU.
4082 rq
= task_rq_lock(p
, &flags
);
4083 update_rq_clock(rq
);
4085 * The RT priorities are set via sched_setscheduler(), but we still
4086 * allow the 'normal' nice value to be set - but as expected
4087 * it wont have any effect on scheduling until the task is
4088 * SCHED_FIFO/SCHED_RR:
4090 if (task_has_rt_policy(p
)) {
4091 p
->static_prio
= NICE_TO_PRIO(nice
);
4094 on_rq
= p
->se
.on_rq
;
4096 dequeue_task(rq
, p
, 0);
4098 p
->static_prio
= NICE_TO_PRIO(nice
);
4101 p
->prio
= effective_prio(p
);
4102 delta
= p
->prio
- old_prio
;
4105 enqueue_task(rq
, p
, 0);
4107 * If the task increased its priority or is running and
4108 * lowered its priority, then reschedule its CPU:
4110 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4111 resched_task(rq
->curr
);
4114 task_rq_unlock(rq
, &flags
);
4116 EXPORT_SYMBOL(set_user_nice
);
4119 * can_nice - check if a task can reduce its nice value
4123 int can_nice(const struct task_struct
*p
, const int nice
)
4125 /* convert nice value [19,-20] to rlimit style value [1,40] */
4126 int nice_rlim
= 20 - nice
;
4128 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4129 capable(CAP_SYS_NICE
));
4132 #ifdef __ARCH_WANT_SYS_NICE
4135 * sys_nice - change the priority of the current process.
4136 * @increment: priority increment
4138 * sys_setpriority is a more generic, but much slower function that
4139 * does similar things.
4141 asmlinkage
long sys_nice(int increment
)
4146 * Setpriority might change our priority at the same moment.
4147 * We don't have to worry. Conceptually one call occurs first
4148 * and we have a single winner.
4150 if (increment
< -40)
4155 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4161 if (increment
< 0 && !can_nice(current
, nice
))
4164 retval
= security_task_setnice(current
, nice
);
4168 set_user_nice(current
, nice
);
4175 * task_prio - return the priority value of a given task.
4176 * @p: the task in question.
4178 * This is the priority value as seen by users in /proc.
4179 * RT tasks are offset by -200. Normal tasks are centered
4180 * around 0, value goes from -16 to +15.
4182 int task_prio(const struct task_struct
*p
)
4184 return p
->prio
- MAX_RT_PRIO
;
4188 * task_nice - return the nice value of a given task.
4189 * @p: the task in question.
4191 int task_nice(const struct task_struct
*p
)
4193 return TASK_NICE(p
);
4195 EXPORT_SYMBOL_GPL(task_nice
);
4198 * idle_cpu - is a given cpu idle currently?
4199 * @cpu: the processor in question.
4201 int idle_cpu(int cpu
)
4203 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4207 * idle_task - return the idle task for a given cpu.
4208 * @cpu: the processor in question.
4210 struct task_struct
*idle_task(int cpu
)
4212 return cpu_rq(cpu
)->idle
;
4216 * find_process_by_pid - find a process with a matching PID value.
4217 * @pid: the pid in question.
4219 static struct task_struct
*find_process_by_pid(pid_t pid
)
4221 return pid
? find_task_by_vpid(pid
) : current
;
4224 /* Actually do priority change: must hold rq lock. */
4226 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4228 BUG_ON(p
->se
.on_rq
);
4231 switch (p
->policy
) {
4235 p
->sched_class
= &fair_sched_class
;
4239 p
->sched_class
= &rt_sched_class
;
4243 p
->rt_priority
= prio
;
4244 p
->normal_prio
= normal_prio(p
);
4245 /* we are holding p->pi_lock already */
4246 p
->prio
= rt_mutex_getprio(p
);
4251 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4252 * @p: the task in question.
4253 * @policy: new policy.
4254 * @param: structure containing the new RT priority.
4256 * NOTE that the task may be already dead.
4258 int sched_setscheduler(struct task_struct
*p
, int policy
,
4259 struct sched_param
*param
)
4261 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4262 unsigned long flags
;
4263 const struct sched_class
*prev_class
= p
->sched_class
;
4266 /* may grab non-irq protected spin_locks */
4267 BUG_ON(in_interrupt());
4269 /* double check policy once rq lock held */
4271 policy
= oldpolicy
= p
->policy
;
4272 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4273 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4274 policy
!= SCHED_IDLE
)
4277 * Valid priorities for SCHED_FIFO and SCHED_RR are
4278 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4279 * SCHED_BATCH and SCHED_IDLE is 0.
4281 if (param
->sched_priority
< 0 ||
4282 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4283 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4285 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4289 * Allow unprivileged RT tasks to decrease priority:
4291 if (!capable(CAP_SYS_NICE
)) {
4292 if (rt_policy(policy
)) {
4293 unsigned long rlim_rtprio
;
4295 if (!lock_task_sighand(p
, &flags
))
4297 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4298 unlock_task_sighand(p
, &flags
);
4300 /* can't set/change the rt policy */
4301 if (policy
!= p
->policy
&& !rlim_rtprio
)
4304 /* can't increase priority */
4305 if (param
->sched_priority
> p
->rt_priority
&&
4306 param
->sched_priority
> rlim_rtprio
)
4310 * Like positive nice levels, dont allow tasks to
4311 * move out of SCHED_IDLE either:
4313 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4316 /* can't change other user's priorities */
4317 if ((current
->euid
!= p
->euid
) &&
4318 (current
->euid
!= p
->uid
))
4322 retval
= security_task_setscheduler(p
, policy
, param
);
4326 * make sure no PI-waiters arrive (or leave) while we are
4327 * changing the priority of the task:
4329 spin_lock_irqsave(&p
->pi_lock
, flags
);
4331 * To be able to change p->policy safely, the apropriate
4332 * runqueue lock must be held.
4334 rq
= __task_rq_lock(p
);
4335 /* recheck policy now with rq lock held */
4336 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4337 policy
= oldpolicy
= -1;
4338 __task_rq_unlock(rq
);
4339 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4342 update_rq_clock(rq
);
4343 on_rq
= p
->se
.on_rq
;
4344 running
= task_current(rq
, p
);
4346 deactivate_task(rq
, p
, 0);
4348 p
->sched_class
->put_prev_task(rq
, p
);
4352 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4356 p
->sched_class
->set_curr_task(rq
);
4358 activate_task(rq
, p
, 0);
4360 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4362 __task_rq_unlock(rq
);
4363 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4365 rt_mutex_adjust_pi(p
);
4369 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4372 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4374 struct sched_param lparam
;
4375 struct task_struct
*p
;
4378 if (!param
|| pid
< 0)
4380 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4385 p
= find_process_by_pid(pid
);
4387 retval
= sched_setscheduler(p
, policy
, &lparam
);
4394 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4395 * @pid: the pid in question.
4396 * @policy: new policy.
4397 * @param: structure containing the new RT priority.
4400 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4402 /* negative values for policy are not valid */
4406 return do_sched_setscheduler(pid
, policy
, param
);
4410 * sys_sched_setparam - set/change the RT priority of a thread
4411 * @pid: the pid in question.
4412 * @param: structure containing the new RT priority.
4414 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4416 return do_sched_setscheduler(pid
, -1, param
);
4420 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4421 * @pid: the pid in question.
4423 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4425 struct task_struct
*p
;
4432 read_lock(&tasklist_lock
);
4433 p
= find_process_by_pid(pid
);
4435 retval
= security_task_getscheduler(p
);
4439 read_unlock(&tasklist_lock
);
4444 * sys_sched_getscheduler - get the RT priority of a thread
4445 * @pid: the pid in question.
4446 * @param: structure containing the RT priority.
4448 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4450 struct sched_param lp
;
4451 struct task_struct
*p
;
4454 if (!param
|| pid
< 0)
4457 read_lock(&tasklist_lock
);
4458 p
= find_process_by_pid(pid
);
4463 retval
= security_task_getscheduler(p
);
4467 lp
.sched_priority
= p
->rt_priority
;
4468 read_unlock(&tasklist_lock
);
4471 * This one might sleep, we cannot do it with a spinlock held ...
4473 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4478 read_unlock(&tasklist_lock
);
4482 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4484 cpumask_t cpus_allowed
;
4485 struct task_struct
*p
;
4489 read_lock(&tasklist_lock
);
4491 p
= find_process_by_pid(pid
);
4493 read_unlock(&tasklist_lock
);
4499 * It is not safe to call set_cpus_allowed with the
4500 * tasklist_lock held. We will bump the task_struct's
4501 * usage count and then drop tasklist_lock.
4504 read_unlock(&tasklist_lock
);
4507 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4508 !capable(CAP_SYS_NICE
))
4511 retval
= security_task_setscheduler(p
, 0, NULL
);
4515 cpus_allowed
= cpuset_cpus_allowed(p
);
4516 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4518 retval
= set_cpus_allowed(p
, new_mask
);
4521 cpus_allowed
= cpuset_cpus_allowed(p
);
4522 if (!cpus_subset(new_mask
, cpus_allowed
)) {
4524 * We must have raced with a concurrent cpuset
4525 * update. Just reset the cpus_allowed to the
4526 * cpuset's cpus_allowed
4528 new_mask
= cpus_allowed
;
4538 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4539 cpumask_t
*new_mask
)
4541 if (len
< sizeof(cpumask_t
)) {
4542 memset(new_mask
, 0, sizeof(cpumask_t
));
4543 } else if (len
> sizeof(cpumask_t
)) {
4544 len
= sizeof(cpumask_t
);
4546 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4550 * sys_sched_setaffinity - set the cpu affinity of a process
4551 * @pid: pid of the process
4552 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4553 * @user_mask_ptr: user-space pointer to the new cpu mask
4555 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4556 unsigned long __user
*user_mask_ptr
)
4561 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4565 return sched_setaffinity(pid
, new_mask
);
4569 * Represents all cpu's present in the system
4570 * In systems capable of hotplug, this map could dynamically grow
4571 * as new cpu's are detected in the system via any platform specific
4572 * method, such as ACPI for e.g.
4575 cpumask_t cpu_present_map __read_mostly
;
4576 EXPORT_SYMBOL(cpu_present_map
);
4579 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4580 EXPORT_SYMBOL(cpu_online_map
);
4582 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4583 EXPORT_SYMBOL(cpu_possible_map
);
4586 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4588 struct task_struct
*p
;
4592 read_lock(&tasklist_lock
);
4595 p
= find_process_by_pid(pid
);
4599 retval
= security_task_getscheduler(p
);
4603 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4606 read_unlock(&tasklist_lock
);
4613 * sys_sched_getaffinity - get the cpu affinity of a process
4614 * @pid: pid of the process
4615 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4616 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4618 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4619 unsigned long __user
*user_mask_ptr
)
4624 if (len
< sizeof(cpumask_t
))
4627 ret
= sched_getaffinity(pid
, &mask
);
4631 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4634 return sizeof(cpumask_t
);
4638 * sys_sched_yield - yield the current processor to other threads.
4640 * This function yields the current CPU to other tasks. If there are no
4641 * other threads running on this CPU then this function will return.
4643 asmlinkage
long sys_sched_yield(void)
4645 struct rq
*rq
= this_rq_lock();
4647 schedstat_inc(rq
, yld_count
);
4648 current
->sched_class
->yield_task(rq
);
4651 * Since we are going to call schedule() anyway, there's
4652 * no need to preempt or enable interrupts:
4654 __release(rq
->lock
);
4655 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4656 _raw_spin_unlock(&rq
->lock
);
4657 preempt_enable_no_resched();
4664 static void __cond_resched(void)
4666 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4667 __might_sleep(__FILE__
, __LINE__
);
4670 * The BKS might be reacquired before we have dropped
4671 * PREEMPT_ACTIVE, which could trigger a second
4672 * cond_resched() call.
4675 add_preempt_count(PREEMPT_ACTIVE
);
4677 sub_preempt_count(PREEMPT_ACTIVE
);
4678 } while (need_resched());
4681 int __sched
cond_resched(void)
4683 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4684 system_state
== SYSTEM_RUNNING
) {
4690 EXPORT_SYMBOL(cond_resched
);
4693 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4694 * call schedule, and on return reacquire the lock.
4696 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4697 * operations here to prevent schedule() from being called twice (once via
4698 * spin_unlock(), once by hand).
4700 int cond_resched_lock(spinlock_t
*lock
)
4704 if (need_lockbreak(lock
)) {
4710 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4711 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4712 _raw_spin_unlock(lock
);
4713 preempt_enable_no_resched();
4720 EXPORT_SYMBOL(cond_resched_lock
);
4722 int __sched
cond_resched_softirq(void)
4724 BUG_ON(!in_softirq());
4726 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4734 EXPORT_SYMBOL(cond_resched_softirq
);
4737 * yield - yield the current processor to other threads.
4739 * This is a shortcut for kernel-space yielding - it marks the
4740 * thread runnable and calls sys_sched_yield().
4742 void __sched
yield(void)
4744 set_current_state(TASK_RUNNING
);
4747 EXPORT_SYMBOL(yield
);
4750 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4751 * that process accounting knows that this is a task in IO wait state.
4753 * But don't do that if it is a deliberate, throttling IO wait (this task
4754 * has set its backing_dev_info: the queue against which it should throttle)
4756 void __sched
io_schedule(void)
4758 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4760 delayacct_blkio_start();
4761 atomic_inc(&rq
->nr_iowait
);
4763 atomic_dec(&rq
->nr_iowait
);
4764 delayacct_blkio_end();
4766 EXPORT_SYMBOL(io_schedule
);
4768 long __sched
io_schedule_timeout(long timeout
)
4770 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4773 delayacct_blkio_start();
4774 atomic_inc(&rq
->nr_iowait
);
4775 ret
= schedule_timeout(timeout
);
4776 atomic_dec(&rq
->nr_iowait
);
4777 delayacct_blkio_end();
4782 * sys_sched_get_priority_max - return maximum RT priority.
4783 * @policy: scheduling class.
4785 * this syscall returns the maximum rt_priority that can be used
4786 * by a given scheduling class.
4788 asmlinkage
long sys_sched_get_priority_max(int policy
)
4795 ret
= MAX_USER_RT_PRIO
-1;
4807 * sys_sched_get_priority_min - return minimum RT priority.
4808 * @policy: scheduling class.
4810 * this syscall returns the minimum rt_priority that can be used
4811 * by a given scheduling class.
4813 asmlinkage
long sys_sched_get_priority_min(int policy
)
4831 * sys_sched_rr_get_interval - return the default timeslice of a process.
4832 * @pid: pid of the process.
4833 * @interval: userspace pointer to the timeslice value.
4835 * this syscall writes the default timeslice value of a given process
4836 * into the user-space timespec buffer. A value of '0' means infinity.
4839 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4841 struct task_struct
*p
;
4842 unsigned int time_slice
;
4850 read_lock(&tasklist_lock
);
4851 p
= find_process_by_pid(pid
);
4855 retval
= security_task_getscheduler(p
);
4860 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4861 * tasks that are on an otherwise idle runqueue:
4864 if (p
->policy
== SCHED_RR
) {
4865 time_slice
= DEF_TIMESLICE
;
4867 struct sched_entity
*se
= &p
->se
;
4868 unsigned long flags
;
4871 rq
= task_rq_lock(p
, &flags
);
4872 if (rq
->cfs
.load
.weight
)
4873 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
4874 task_rq_unlock(rq
, &flags
);
4876 read_unlock(&tasklist_lock
);
4877 jiffies_to_timespec(time_slice
, &t
);
4878 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4882 read_unlock(&tasklist_lock
);
4886 static const char stat_nam
[] = "RSDTtZX";
4888 void sched_show_task(struct task_struct
*p
)
4890 unsigned long free
= 0;
4893 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4894 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
4895 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4896 #if BITS_PER_LONG == 32
4897 if (state
== TASK_RUNNING
)
4898 printk(KERN_CONT
" running ");
4900 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4902 if (state
== TASK_RUNNING
)
4903 printk(KERN_CONT
" running task ");
4905 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4907 #ifdef CONFIG_DEBUG_STACK_USAGE
4909 unsigned long *n
= end_of_stack(p
);
4912 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4915 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
4916 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
4918 if (state
!= TASK_RUNNING
)
4919 show_stack(p
, NULL
);
4922 void show_state_filter(unsigned long state_filter
)
4924 struct task_struct
*g
, *p
;
4926 #if BITS_PER_LONG == 32
4928 " task PC stack pid father\n");
4931 " task PC stack pid father\n");
4933 read_lock(&tasklist_lock
);
4934 do_each_thread(g
, p
) {
4936 * reset the NMI-timeout, listing all files on a slow
4937 * console might take alot of time:
4939 touch_nmi_watchdog();
4940 if (!state_filter
|| (p
->state
& state_filter
))
4942 } while_each_thread(g
, p
);
4944 touch_all_softlockup_watchdogs();
4946 #ifdef CONFIG_SCHED_DEBUG
4947 sysrq_sched_debug_show();
4949 read_unlock(&tasklist_lock
);
4951 * Only show locks if all tasks are dumped:
4953 if (state_filter
== -1)
4954 debug_show_all_locks();
4957 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4959 idle
->sched_class
= &idle_sched_class
;
4963 * init_idle - set up an idle thread for a given CPU
4964 * @idle: task in question
4965 * @cpu: cpu the idle task belongs to
4967 * NOTE: this function does not set the idle thread's NEED_RESCHED
4968 * flag, to make booting more robust.
4970 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4972 struct rq
*rq
= cpu_rq(cpu
);
4973 unsigned long flags
;
4976 idle
->se
.exec_start
= sched_clock();
4978 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4979 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4980 __set_task_cpu(idle
, cpu
);
4982 spin_lock_irqsave(&rq
->lock
, flags
);
4983 rq
->curr
= rq
->idle
= idle
;
4984 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4987 spin_unlock_irqrestore(&rq
->lock
, flags
);
4989 /* Set the preempt count _outside_ the spinlocks! */
4990 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4991 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4993 task_thread_info(idle
)->preempt_count
= 0;
4996 * The idle tasks have their own, simple scheduling class:
4998 idle
->sched_class
= &idle_sched_class
;
5002 * In a system that switches off the HZ timer nohz_cpu_mask
5003 * indicates which cpus entered this state. This is used
5004 * in the rcu update to wait only for active cpus. For system
5005 * which do not switch off the HZ timer nohz_cpu_mask should
5006 * always be CPU_MASK_NONE.
5008 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5011 * Increase the granularity value when there are more CPUs,
5012 * because with more CPUs the 'effective latency' as visible
5013 * to users decreases. But the relationship is not linear,
5014 * so pick a second-best guess by going with the log2 of the
5017 * This idea comes from the SD scheduler of Con Kolivas:
5019 static inline void sched_init_granularity(void)
5021 unsigned int factor
= 1 + ilog2(num_online_cpus());
5022 const unsigned long limit
= 200000000;
5024 sysctl_sched_min_granularity
*= factor
;
5025 if (sysctl_sched_min_granularity
> limit
)
5026 sysctl_sched_min_granularity
= limit
;
5028 sysctl_sched_latency
*= factor
;
5029 if (sysctl_sched_latency
> limit
)
5030 sysctl_sched_latency
= limit
;
5032 sysctl_sched_wakeup_granularity
*= factor
;
5033 sysctl_sched_batch_wakeup_granularity
*= factor
;
5038 * This is how migration works:
5040 * 1) we queue a struct migration_req structure in the source CPU's
5041 * runqueue and wake up that CPU's migration thread.
5042 * 2) we down() the locked semaphore => thread blocks.
5043 * 3) migration thread wakes up (implicitly it forces the migrated
5044 * thread off the CPU)
5045 * 4) it gets the migration request and checks whether the migrated
5046 * task is still in the wrong runqueue.
5047 * 5) if it's in the wrong runqueue then the migration thread removes
5048 * it and puts it into the right queue.
5049 * 6) migration thread up()s the semaphore.
5050 * 7) we wake up and the migration is done.
5054 * Change a given task's CPU affinity. Migrate the thread to a
5055 * proper CPU and schedule it away if the CPU it's executing on
5056 * is removed from the allowed bitmask.
5058 * NOTE: the caller must have a valid reference to the task, the
5059 * task must not exit() & deallocate itself prematurely. The
5060 * call is not atomic; no spinlocks may be held.
5062 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5064 struct migration_req req
;
5065 unsigned long flags
;
5069 rq
= task_rq_lock(p
, &flags
);
5070 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5075 if (p
->sched_class
->set_cpus_allowed
)
5076 p
->sched_class
->set_cpus_allowed(p
, &new_mask
);
5078 p
->cpus_allowed
= new_mask
;
5079 p
->nr_cpus_allowed
= cpus_weight(new_mask
);
5082 /* Can the task run on the task's current CPU? If so, we're done */
5083 if (cpu_isset(task_cpu(p
), new_mask
))
5086 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5087 /* Need help from migration thread: drop lock and wait. */
5088 task_rq_unlock(rq
, &flags
);
5089 wake_up_process(rq
->migration_thread
);
5090 wait_for_completion(&req
.done
);
5091 tlb_migrate_finish(p
->mm
);
5095 task_rq_unlock(rq
, &flags
);
5099 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5102 * Move (not current) task off this cpu, onto dest cpu. We're doing
5103 * this because either it can't run here any more (set_cpus_allowed()
5104 * away from this CPU, or CPU going down), or because we're
5105 * attempting to rebalance this task on exec (sched_exec).
5107 * So we race with normal scheduler movements, but that's OK, as long
5108 * as the task is no longer on this CPU.
5110 * Returns non-zero if task was successfully migrated.
5112 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5114 struct rq
*rq_dest
, *rq_src
;
5117 if (unlikely(cpu_is_offline(dest_cpu
)))
5120 rq_src
= cpu_rq(src_cpu
);
5121 rq_dest
= cpu_rq(dest_cpu
);
5123 double_rq_lock(rq_src
, rq_dest
);
5124 /* Already moved. */
5125 if (task_cpu(p
) != src_cpu
)
5127 /* Affinity changed (again). */
5128 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5131 on_rq
= p
->se
.on_rq
;
5133 deactivate_task(rq_src
, p
, 0);
5135 set_task_cpu(p
, dest_cpu
);
5137 activate_task(rq_dest
, p
, 0);
5138 check_preempt_curr(rq_dest
, p
);
5142 double_rq_unlock(rq_src
, rq_dest
);
5147 * migration_thread - this is a highprio system thread that performs
5148 * thread migration by bumping thread off CPU then 'pushing' onto
5151 static int migration_thread(void *data
)
5153 int cpu
= (long)data
;
5157 BUG_ON(rq
->migration_thread
!= current
);
5159 set_current_state(TASK_INTERRUPTIBLE
);
5160 while (!kthread_should_stop()) {
5161 struct migration_req
*req
;
5162 struct list_head
*head
;
5164 spin_lock_irq(&rq
->lock
);
5166 if (cpu_is_offline(cpu
)) {
5167 spin_unlock_irq(&rq
->lock
);
5171 if (rq
->active_balance
) {
5172 active_load_balance(rq
, cpu
);
5173 rq
->active_balance
= 0;
5176 head
= &rq
->migration_queue
;
5178 if (list_empty(head
)) {
5179 spin_unlock_irq(&rq
->lock
);
5181 set_current_state(TASK_INTERRUPTIBLE
);
5184 req
= list_entry(head
->next
, struct migration_req
, list
);
5185 list_del_init(head
->next
);
5187 spin_unlock(&rq
->lock
);
5188 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5191 complete(&req
->done
);
5193 __set_current_state(TASK_RUNNING
);
5197 /* Wait for kthread_stop */
5198 set_current_state(TASK_INTERRUPTIBLE
);
5199 while (!kthread_should_stop()) {
5201 set_current_state(TASK_INTERRUPTIBLE
);
5203 __set_current_state(TASK_RUNNING
);
5207 #ifdef CONFIG_HOTPLUG_CPU
5209 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5213 local_irq_disable();
5214 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5220 * Figure out where task on dead CPU should go, use force if necessary.
5221 * NOTE: interrupts should be disabled by the caller
5223 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5225 unsigned long flags
;
5232 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5233 cpus_and(mask
, mask
, p
->cpus_allowed
);
5234 dest_cpu
= any_online_cpu(mask
);
5236 /* On any allowed CPU? */
5237 if (dest_cpu
== NR_CPUS
)
5238 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5240 /* No more Mr. Nice Guy. */
5241 if (dest_cpu
== NR_CPUS
) {
5242 cpumask_t cpus_allowed
= cpuset_cpus_allowed_locked(p
);
5244 * Try to stay on the same cpuset, where the
5245 * current cpuset may be a subset of all cpus.
5246 * The cpuset_cpus_allowed_locked() variant of
5247 * cpuset_cpus_allowed() will not block. It must be
5248 * called within calls to cpuset_lock/cpuset_unlock.
5250 rq
= task_rq_lock(p
, &flags
);
5251 p
->cpus_allowed
= cpus_allowed
;
5252 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5253 task_rq_unlock(rq
, &flags
);
5256 * Don't tell them about moving exiting tasks or
5257 * kernel threads (both mm NULL), since they never
5260 if (p
->mm
&& printk_ratelimit()) {
5261 printk(KERN_INFO
"process %d (%s) no "
5262 "longer affine to cpu%d\n",
5263 task_pid_nr(p
), p
->comm
, dead_cpu
);
5266 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5270 * While a dead CPU has no uninterruptible tasks queued at this point,
5271 * it might still have a nonzero ->nr_uninterruptible counter, because
5272 * for performance reasons the counter is not stricly tracking tasks to
5273 * their home CPUs. So we just add the counter to another CPU's counter,
5274 * to keep the global sum constant after CPU-down:
5276 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5278 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5279 unsigned long flags
;
5281 local_irq_save(flags
);
5282 double_rq_lock(rq_src
, rq_dest
);
5283 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5284 rq_src
->nr_uninterruptible
= 0;
5285 double_rq_unlock(rq_src
, rq_dest
);
5286 local_irq_restore(flags
);
5289 /* Run through task list and migrate tasks from the dead cpu. */
5290 static void migrate_live_tasks(int src_cpu
)
5292 struct task_struct
*p
, *t
;
5294 read_lock(&tasklist_lock
);
5296 do_each_thread(t
, p
) {
5300 if (task_cpu(p
) == src_cpu
)
5301 move_task_off_dead_cpu(src_cpu
, p
);
5302 } while_each_thread(t
, p
);
5304 read_unlock(&tasklist_lock
);
5308 * Schedules idle task to be the next runnable task on current CPU.
5309 * It does so by boosting its priority to highest possible.
5310 * Used by CPU offline code.
5312 void sched_idle_next(void)
5314 int this_cpu
= smp_processor_id();
5315 struct rq
*rq
= cpu_rq(this_cpu
);
5316 struct task_struct
*p
= rq
->idle
;
5317 unsigned long flags
;
5319 /* cpu has to be offline */
5320 BUG_ON(cpu_online(this_cpu
));
5323 * Strictly not necessary since rest of the CPUs are stopped by now
5324 * and interrupts disabled on the current cpu.
5326 spin_lock_irqsave(&rq
->lock
, flags
);
5328 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5330 update_rq_clock(rq
);
5331 activate_task(rq
, p
, 0);
5333 spin_unlock_irqrestore(&rq
->lock
, flags
);
5337 * Ensures that the idle task is using init_mm right before its cpu goes
5340 void idle_task_exit(void)
5342 struct mm_struct
*mm
= current
->active_mm
;
5344 BUG_ON(cpu_online(smp_processor_id()));
5347 switch_mm(mm
, &init_mm
, current
);
5351 /* called under rq->lock with disabled interrupts */
5352 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5354 struct rq
*rq
= cpu_rq(dead_cpu
);
5356 /* Must be exiting, otherwise would be on tasklist. */
5357 BUG_ON(!p
->exit_state
);
5359 /* Cannot have done final schedule yet: would have vanished. */
5360 BUG_ON(p
->state
== TASK_DEAD
);
5365 * Drop lock around migration; if someone else moves it,
5366 * that's OK. No task can be added to this CPU, so iteration is
5369 spin_unlock_irq(&rq
->lock
);
5370 move_task_off_dead_cpu(dead_cpu
, p
);
5371 spin_lock_irq(&rq
->lock
);
5376 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5377 static void migrate_dead_tasks(unsigned int dead_cpu
)
5379 struct rq
*rq
= cpu_rq(dead_cpu
);
5380 struct task_struct
*next
;
5383 if (!rq
->nr_running
)
5385 update_rq_clock(rq
);
5386 next
= pick_next_task(rq
, rq
->curr
);
5389 migrate_dead(dead_cpu
, next
);
5393 #endif /* CONFIG_HOTPLUG_CPU */
5395 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5397 static struct ctl_table sd_ctl_dir
[] = {
5399 .procname
= "sched_domain",
5405 static struct ctl_table sd_ctl_root
[] = {
5407 .ctl_name
= CTL_KERN
,
5408 .procname
= "kernel",
5410 .child
= sd_ctl_dir
,
5415 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5417 struct ctl_table
*entry
=
5418 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5423 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5425 struct ctl_table
*entry
;
5428 * In the intermediate directories, both the child directory and
5429 * procname are dynamically allocated and could fail but the mode
5430 * will always be set. In the lowest directory the names are
5431 * static strings and all have proc handlers.
5433 for (entry
= *tablep
; entry
->mode
; entry
++) {
5435 sd_free_ctl_entry(&entry
->child
);
5436 if (entry
->proc_handler
== NULL
)
5437 kfree(entry
->procname
);
5445 set_table_entry(struct ctl_table
*entry
,
5446 const char *procname
, void *data
, int maxlen
,
5447 mode_t mode
, proc_handler
*proc_handler
)
5449 entry
->procname
= procname
;
5451 entry
->maxlen
= maxlen
;
5453 entry
->proc_handler
= proc_handler
;
5456 static struct ctl_table
*
5457 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5459 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
5464 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5465 sizeof(long), 0644, proc_doulongvec_minmax
);
5466 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5467 sizeof(long), 0644, proc_doulongvec_minmax
);
5468 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5469 sizeof(int), 0644, proc_dointvec_minmax
);
5470 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5471 sizeof(int), 0644, proc_dointvec_minmax
);
5472 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5473 sizeof(int), 0644, proc_dointvec_minmax
);
5474 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5475 sizeof(int), 0644, proc_dointvec_minmax
);
5476 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5477 sizeof(int), 0644, proc_dointvec_minmax
);
5478 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5479 sizeof(int), 0644, proc_dointvec_minmax
);
5480 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5481 sizeof(int), 0644, proc_dointvec_minmax
);
5482 set_table_entry(&table
[9], "cache_nice_tries",
5483 &sd
->cache_nice_tries
,
5484 sizeof(int), 0644, proc_dointvec_minmax
);
5485 set_table_entry(&table
[10], "flags", &sd
->flags
,
5486 sizeof(int), 0644, proc_dointvec_minmax
);
5487 /* &table[11] is terminator */
5492 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5494 struct ctl_table
*entry
, *table
;
5495 struct sched_domain
*sd
;
5496 int domain_num
= 0, i
;
5499 for_each_domain(cpu
, sd
)
5501 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5506 for_each_domain(cpu
, sd
) {
5507 snprintf(buf
, 32, "domain%d", i
);
5508 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5510 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5517 static struct ctl_table_header
*sd_sysctl_header
;
5518 static void register_sched_domain_sysctl(void)
5520 int i
, cpu_num
= num_online_cpus();
5521 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5524 WARN_ON(sd_ctl_dir
[0].child
);
5525 sd_ctl_dir
[0].child
= entry
;
5530 for_each_online_cpu(i
) {
5531 snprintf(buf
, 32, "cpu%d", i
);
5532 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5534 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5538 WARN_ON(sd_sysctl_header
);
5539 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5542 /* may be called multiple times per register */
5543 static void unregister_sched_domain_sysctl(void)
5545 if (sd_sysctl_header
)
5546 unregister_sysctl_table(sd_sysctl_header
);
5547 sd_sysctl_header
= NULL
;
5548 if (sd_ctl_dir
[0].child
)
5549 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5552 static void register_sched_domain_sysctl(void)
5555 static void unregister_sched_domain_sysctl(void)
5561 * migration_call - callback that gets triggered when a CPU is added.
5562 * Here we can start up the necessary migration thread for the new CPU.
5564 static int __cpuinit
5565 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5567 struct task_struct
*p
;
5568 int cpu
= (long)hcpu
;
5569 unsigned long flags
;
5574 case CPU_UP_PREPARE
:
5575 case CPU_UP_PREPARE_FROZEN
:
5576 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5579 kthread_bind(p
, cpu
);
5580 /* Must be high prio: stop_machine expects to yield to it. */
5581 rq
= task_rq_lock(p
, &flags
);
5582 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5583 task_rq_unlock(rq
, &flags
);
5584 cpu_rq(cpu
)->migration_thread
= p
;
5588 case CPU_ONLINE_FROZEN
:
5589 /* Strictly unnecessary, as first user will wake it. */
5590 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5592 /* Update our root-domain */
5594 spin_lock_irqsave(&rq
->lock
, flags
);
5596 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5597 cpu_set(cpu
, rq
->rd
->online
);
5599 spin_unlock_irqrestore(&rq
->lock
, flags
);
5602 #ifdef CONFIG_HOTPLUG_CPU
5603 case CPU_UP_CANCELED
:
5604 case CPU_UP_CANCELED_FROZEN
:
5605 if (!cpu_rq(cpu
)->migration_thread
)
5607 /* Unbind it from offline cpu so it can run. Fall thru. */
5608 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5609 any_online_cpu(cpu_online_map
));
5610 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5611 cpu_rq(cpu
)->migration_thread
= NULL
;
5615 case CPU_DEAD_FROZEN
:
5616 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5617 migrate_live_tasks(cpu
);
5619 kthread_stop(rq
->migration_thread
);
5620 rq
->migration_thread
= NULL
;
5621 /* Idle task back to normal (off runqueue, low prio) */
5622 spin_lock_irq(&rq
->lock
);
5623 update_rq_clock(rq
);
5624 deactivate_task(rq
, rq
->idle
, 0);
5625 rq
->idle
->static_prio
= MAX_PRIO
;
5626 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5627 rq
->idle
->sched_class
= &idle_sched_class
;
5628 migrate_dead_tasks(cpu
);
5629 spin_unlock_irq(&rq
->lock
);
5631 migrate_nr_uninterruptible(rq
);
5632 BUG_ON(rq
->nr_running
!= 0);
5635 * No need to migrate the tasks: it was best-effort if
5636 * they didn't take sched_hotcpu_mutex. Just wake up
5639 spin_lock_irq(&rq
->lock
);
5640 while (!list_empty(&rq
->migration_queue
)) {
5641 struct migration_req
*req
;
5643 req
= list_entry(rq
->migration_queue
.next
,
5644 struct migration_req
, list
);
5645 list_del_init(&req
->list
);
5646 complete(&req
->done
);
5648 spin_unlock_irq(&rq
->lock
);
5651 case CPU_DOWN_PREPARE
:
5652 /* Update our root-domain */
5654 spin_lock_irqsave(&rq
->lock
, flags
);
5656 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
5657 cpu_clear(cpu
, rq
->rd
->online
);
5659 spin_unlock_irqrestore(&rq
->lock
, flags
);
5666 /* Register at highest priority so that task migration (migrate_all_tasks)
5667 * happens before everything else.
5669 static struct notifier_block __cpuinitdata migration_notifier
= {
5670 .notifier_call
= migration_call
,
5674 void __init
migration_init(void)
5676 void *cpu
= (void *)(long)smp_processor_id();
5679 /* Start one for the boot CPU: */
5680 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5681 BUG_ON(err
== NOTIFY_BAD
);
5682 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5683 register_cpu_notifier(&migration_notifier
);
5689 /* Number of possible processor ids */
5690 int nr_cpu_ids __read_mostly
= NR_CPUS
;
5691 EXPORT_SYMBOL(nr_cpu_ids
);
5693 #ifdef CONFIG_SCHED_DEBUG
5695 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
)
5697 struct sched_group
*group
= sd
->groups
;
5698 cpumask_t groupmask
;
5701 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5702 cpus_clear(groupmask
);
5704 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5706 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5707 printk("does not load-balance\n");
5709 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5714 printk(KERN_CONT
"span %s\n", str
);
5716 if (!cpu_isset(cpu
, sd
->span
)) {
5717 printk(KERN_ERR
"ERROR: domain->span does not contain "
5720 if (!cpu_isset(cpu
, group
->cpumask
)) {
5721 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5725 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5729 printk(KERN_ERR
"ERROR: group is NULL\n");
5733 if (!group
->__cpu_power
) {
5734 printk(KERN_CONT
"\n");
5735 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5740 if (!cpus_weight(group
->cpumask
)) {
5741 printk(KERN_CONT
"\n");
5742 printk(KERN_ERR
"ERROR: empty group\n");
5746 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5747 printk(KERN_CONT
"\n");
5748 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5752 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5754 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5755 printk(KERN_CONT
" %s", str
);
5757 group
= group
->next
;
5758 } while (group
!= sd
->groups
);
5759 printk(KERN_CONT
"\n");
5761 if (!cpus_equal(sd
->span
, groupmask
))
5762 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5764 if (sd
->parent
&& !cpus_subset(groupmask
, sd
->parent
->span
))
5765 printk(KERN_ERR
"ERROR: parent span is not a superset "
5766 "of domain->span\n");
5770 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5775 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5779 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5782 if (sched_domain_debug_one(sd
, cpu
, level
))
5791 # define sched_domain_debug(sd, cpu) do { } while (0)
5794 static int sd_degenerate(struct sched_domain
*sd
)
5796 if (cpus_weight(sd
->span
) == 1)
5799 /* Following flags need at least 2 groups */
5800 if (sd
->flags
& (SD_LOAD_BALANCE
|
5801 SD_BALANCE_NEWIDLE
|
5805 SD_SHARE_PKG_RESOURCES
)) {
5806 if (sd
->groups
!= sd
->groups
->next
)
5810 /* Following flags don't use groups */
5811 if (sd
->flags
& (SD_WAKE_IDLE
|
5820 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5822 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5824 if (sd_degenerate(parent
))
5827 if (!cpus_equal(sd
->span
, parent
->span
))
5830 /* Does parent contain flags not in child? */
5831 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5832 if (cflags
& SD_WAKE_AFFINE
)
5833 pflags
&= ~SD_WAKE_BALANCE
;
5834 /* Flags needing groups don't count if only 1 group in parent */
5835 if (parent
->groups
== parent
->groups
->next
) {
5836 pflags
&= ~(SD_LOAD_BALANCE
|
5837 SD_BALANCE_NEWIDLE
|
5841 SD_SHARE_PKG_RESOURCES
);
5843 if (~cflags
& pflags
)
5849 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5851 unsigned long flags
;
5852 const struct sched_class
*class;
5854 spin_lock_irqsave(&rq
->lock
, flags
);
5857 struct root_domain
*old_rd
= rq
->rd
;
5859 for (class = sched_class_highest
; class; class = class->next
) {
5860 if (class->leave_domain
)
5861 class->leave_domain(rq
);
5864 cpu_clear(rq
->cpu
, old_rd
->span
);
5865 cpu_clear(rq
->cpu
, old_rd
->online
);
5867 if (atomic_dec_and_test(&old_rd
->refcount
))
5871 atomic_inc(&rd
->refcount
);
5874 cpu_set(rq
->cpu
, rd
->span
);
5875 if (cpu_isset(rq
->cpu
, cpu_online_map
))
5876 cpu_set(rq
->cpu
, rd
->online
);
5878 for (class = sched_class_highest
; class; class = class->next
) {
5879 if (class->join_domain
)
5880 class->join_domain(rq
);
5883 spin_unlock_irqrestore(&rq
->lock
, flags
);
5886 static void init_rootdomain(struct root_domain
*rd
)
5888 memset(rd
, 0, sizeof(*rd
));
5890 cpus_clear(rd
->span
);
5891 cpus_clear(rd
->online
);
5894 static void init_defrootdomain(void)
5896 init_rootdomain(&def_root_domain
);
5897 atomic_set(&def_root_domain
.refcount
, 1);
5900 static struct root_domain
*alloc_rootdomain(void)
5902 struct root_domain
*rd
;
5904 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5908 init_rootdomain(rd
);
5914 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5915 * hold the hotplug lock.
5918 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5920 struct rq
*rq
= cpu_rq(cpu
);
5921 struct sched_domain
*tmp
;
5923 /* Remove the sched domains which do not contribute to scheduling. */
5924 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5925 struct sched_domain
*parent
= tmp
->parent
;
5928 if (sd_parent_degenerate(tmp
, parent
)) {
5929 tmp
->parent
= parent
->parent
;
5931 parent
->parent
->child
= tmp
;
5935 if (sd
&& sd_degenerate(sd
)) {
5941 sched_domain_debug(sd
, cpu
);
5943 rq_attach_root(rq
, rd
);
5944 rcu_assign_pointer(rq
->sd
, sd
);
5947 /* cpus with isolated domains */
5948 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5950 /* Setup the mask of cpus configured for isolated domains */
5951 static int __init
isolated_cpu_setup(char *str
)
5953 int ints
[NR_CPUS
], i
;
5955 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5956 cpus_clear(cpu_isolated_map
);
5957 for (i
= 1; i
<= ints
[0]; i
++)
5958 if (ints
[i
] < NR_CPUS
)
5959 cpu_set(ints
[i
], cpu_isolated_map
);
5963 __setup("isolcpus=", isolated_cpu_setup
);
5966 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5967 * to a function which identifies what group(along with sched group) a CPU
5968 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5969 * (due to the fact that we keep track of groups covered with a cpumask_t).
5971 * init_sched_build_groups will build a circular linked list of the groups
5972 * covered by the given span, and will set each group's ->cpumask correctly,
5973 * and ->cpu_power to 0.
5976 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5977 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5978 struct sched_group
**sg
))
5980 struct sched_group
*first
= NULL
, *last
= NULL
;
5981 cpumask_t covered
= CPU_MASK_NONE
;
5984 for_each_cpu_mask(i
, span
) {
5985 struct sched_group
*sg
;
5986 int group
= group_fn(i
, cpu_map
, &sg
);
5989 if (cpu_isset(i
, covered
))
5992 sg
->cpumask
= CPU_MASK_NONE
;
5993 sg
->__cpu_power
= 0;
5995 for_each_cpu_mask(j
, span
) {
5996 if (group_fn(j
, cpu_map
, NULL
) != group
)
5999 cpu_set(j
, covered
);
6000 cpu_set(j
, sg
->cpumask
);
6011 #define SD_NODES_PER_DOMAIN 16
6016 * find_next_best_node - find the next node to include in a sched_domain
6017 * @node: node whose sched_domain we're building
6018 * @used_nodes: nodes already in the sched_domain
6020 * Find the next node to include in a given scheduling domain. Simply
6021 * finds the closest node not already in the @used_nodes map.
6023 * Should use nodemask_t.
6025 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6027 int i
, n
, val
, min_val
, best_node
= 0;
6031 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6032 /* Start at @node */
6033 n
= (node
+ i
) % MAX_NUMNODES
;
6035 if (!nr_cpus_node(n
))
6038 /* Skip already used nodes */
6039 if (test_bit(n
, used_nodes
))
6042 /* Simple min distance search */
6043 val
= node_distance(node
, n
);
6045 if (val
< min_val
) {
6051 set_bit(best_node
, used_nodes
);
6056 * sched_domain_node_span - get a cpumask for a node's sched_domain
6057 * @node: node whose cpumask we're constructing
6058 * @size: number of nodes to include in this span
6060 * Given a node, construct a good cpumask for its sched_domain to span. It
6061 * should be one that prevents unnecessary balancing, but also spreads tasks
6064 static cpumask_t
sched_domain_node_span(int node
)
6066 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6067 cpumask_t span
, nodemask
;
6071 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6073 nodemask
= node_to_cpumask(node
);
6074 cpus_or(span
, span
, nodemask
);
6075 set_bit(node
, used_nodes
);
6077 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6078 int next_node
= find_next_best_node(node
, used_nodes
);
6080 nodemask
= node_to_cpumask(next_node
);
6081 cpus_or(span
, span
, nodemask
);
6088 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6091 * SMT sched-domains:
6093 #ifdef CONFIG_SCHED_SMT
6094 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6095 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6098 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6101 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6107 * multi-core sched-domains:
6109 #ifdef CONFIG_SCHED_MC
6110 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6111 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6114 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6116 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6119 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6120 cpus_and(mask
, mask
, *cpu_map
);
6121 group
= first_cpu(mask
);
6123 *sg
= &per_cpu(sched_group_core
, group
);
6126 #elif defined(CONFIG_SCHED_MC)
6128 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6131 *sg
= &per_cpu(sched_group_core
, cpu
);
6136 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6137 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6140 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
)
6143 #ifdef CONFIG_SCHED_MC
6144 cpumask_t mask
= cpu_coregroup_map(cpu
);
6145 cpus_and(mask
, mask
, *cpu_map
);
6146 group
= first_cpu(mask
);
6147 #elif defined(CONFIG_SCHED_SMT)
6148 cpumask_t mask
= per_cpu(cpu_sibling_map
, cpu
);
6149 cpus_and(mask
, mask
, *cpu_map
);
6150 group
= first_cpu(mask
);
6155 *sg
= &per_cpu(sched_group_phys
, group
);
6161 * The init_sched_build_groups can't handle what we want to do with node
6162 * groups, so roll our own. Now each node has its own list of groups which
6163 * gets dynamically allocated.
6165 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6166 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6168 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6169 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6171 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6172 struct sched_group
**sg
)
6174 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6177 cpus_and(nodemask
, nodemask
, *cpu_map
);
6178 group
= first_cpu(nodemask
);
6181 *sg
= &per_cpu(sched_group_allnodes
, group
);
6185 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6187 struct sched_group
*sg
= group_head
;
6193 for_each_cpu_mask(j
, sg
->cpumask
) {
6194 struct sched_domain
*sd
;
6196 sd
= &per_cpu(phys_domains
, j
);
6197 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6199 * Only add "power" once for each
6205 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6208 } while (sg
!= group_head
);
6213 /* Free memory allocated for various sched_group structures */
6214 static void free_sched_groups(const cpumask_t
*cpu_map
)
6218 for_each_cpu_mask(cpu
, *cpu_map
) {
6219 struct sched_group
**sched_group_nodes
6220 = sched_group_nodes_bycpu
[cpu
];
6222 if (!sched_group_nodes
)
6225 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6226 cpumask_t nodemask
= node_to_cpumask(i
);
6227 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6229 cpus_and(nodemask
, nodemask
, *cpu_map
);
6230 if (cpus_empty(nodemask
))
6240 if (oldsg
!= sched_group_nodes
[i
])
6243 kfree(sched_group_nodes
);
6244 sched_group_nodes_bycpu
[cpu
] = NULL
;
6248 static void free_sched_groups(const cpumask_t
*cpu_map
)
6254 * Initialize sched groups cpu_power.
6256 * cpu_power indicates the capacity of sched group, which is used while
6257 * distributing the load between different sched groups in a sched domain.
6258 * Typically cpu_power for all the groups in a sched domain will be same unless
6259 * there are asymmetries in the topology. If there are asymmetries, group
6260 * having more cpu_power will pickup more load compared to the group having
6263 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6264 * the maximum number of tasks a group can handle in the presence of other idle
6265 * or lightly loaded groups in the same sched domain.
6267 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6269 struct sched_domain
*child
;
6270 struct sched_group
*group
;
6272 WARN_ON(!sd
|| !sd
->groups
);
6274 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6279 sd
->groups
->__cpu_power
= 0;
6282 * For perf policy, if the groups in child domain share resources
6283 * (for example cores sharing some portions of the cache hierarchy
6284 * or SMT), then set this domain groups cpu_power such that each group
6285 * can handle only one task, when there are other idle groups in the
6286 * same sched domain.
6288 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6290 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6291 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6296 * add cpu_power of each child group to this groups cpu_power
6298 group
= child
->groups
;
6300 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6301 group
= group
->next
;
6302 } while (group
!= child
->groups
);
6306 * Build sched domains for a given set of cpus and attach the sched domains
6307 * to the individual cpus
6309 static int build_sched_domains(const cpumask_t
*cpu_map
)
6312 struct root_domain
*rd
;
6314 struct sched_group
**sched_group_nodes
= NULL
;
6315 int sd_allnodes
= 0;
6318 * Allocate the per-node list of sched groups
6320 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6322 if (!sched_group_nodes
) {
6323 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6326 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6329 rd
= alloc_rootdomain();
6331 printk(KERN_WARNING
"Cannot alloc root domain\n");
6336 * Set up domains for cpus specified by the cpu_map.
6338 for_each_cpu_mask(i
, *cpu_map
) {
6339 struct sched_domain
*sd
= NULL
, *p
;
6340 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6342 cpus_and(nodemask
, nodemask
, *cpu_map
);
6345 if (cpus_weight(*cpu_map
) >
6346 SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6347 sd
= &per_cpu(allnodes_domains
, i
);
6348 *sd
= SD_ALLNODES_INIT
;
6349 sd
->span
= *cpu_map
;
6350 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6356 sd
= &per_cpu(node_domains
, i
);
6358 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6362 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6366 sd
= &per_cpu(phys_domains
, i
);
6368 sd
->span
= nodemask
;
6372 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6374 #ifdef CONFIG_SCHED_MC
6376 sd
= &per_cpu(core_domains
, i
);
6378 sd
->span
= cpu_coregroup_map(i
);
6379 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6382 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6385 #ifdef CONFIG_SCHED_SMT
6387 sd
= &per_cpu(cpu_domains
, i
);
6388 *sd
= SD_SIBLING_INIT
;
6389 sd
->span
= per_cpu(cpu_sibling_map
, i
);
6390 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6393 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6397 #ifdef CONFIG_SCHED_SMT
6398 /* Set up CPU (sibling) groups */
6399 for_each_cpu_mask(i
, *cpu_map
) {
6400 cpumask_t this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
6401 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6402 if (i
!= first_cpu(this_sibling_map
))
6405 init_sched_build_groups(this_sibling_map
, cpu_map
,
6410 #ifdef CONFIG_SCHED_MC
6411 /* Set up multi-core groups */
6412 for_each_cpu_mask(i
, *cpu_map
) {
6413 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6414 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6415 if (i
!= first_cpu(this_core_map
))
6417 init_sched_build_groups(this_core_map
, cpu_map
,
6418 &cpu_to_core_group
);
6422 /* Set up physical groups */
6423 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6424 cpumask_t nodemask
= node_to_cpumask(i
);
6426 cpus_and(nodemask
, nodemask
, *cpu_map
);
6427 if (cpus_empty(nodemask
))
6430 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6434 /* Set up node groups */
6436 init_sched_build_groups(*cpu_map
, cpu_map
,
6437 &cpu_to_allnodes_group
);
6439 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6440 /* Set up node groups */
6441 struct sched_group
*sg
, *prev
;
6442 cpumask_t nodemask
= node_to_cpumask(i
);
6443 cpumask_t domainspan
;
6444 cpumask_t covered
= CPU_MASK_NONE
;
6447 cpus_and(nodemask
, nodemask
, *cpu_map
);
6448 if (cpus_empty(nodemask
)) {
6449 sched_group_nodes
[i
] = NULL
;
6453 domainspan
= sched_domain_node_span(i
);
6454 cpus_and(domainspan
, domainspan
, *cpu_map
);
6456 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6458 printk(KERN_WARNING
"Can not alloc domain group for "
6462 sched_group_nodes
[i
] = sg
;
6463 for_each_cpu_mask(j
, nodemask
) {
6464 struct sched_domain
*sd
;
6466 sd
= &per_cpu(node_domains
, j
);
6469 sg
->__cpu_power
= 0;
6470 sg
->cpumask
= nodemask
;
6472 cpus_or(covered
, covered
, nodemask
);
6475 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6476 cpumask_t tmp
, notcovered
;
6477 int n
= (i
+ j
) % MAX_NUMNODES
;
6479 cpus_complement(notcovered
, covered
);
6480 cpus_and(tmp
, notcovered
, *cpu_map
);
6481 cpus_and(tmp
, tmp
, domainspan
);
6482 if (cpus_empty(tmp
))
6485 nodemask
= node_to_cpumask(n
);
6486 cpus_and(tmp
, tmp
, nodemask
);
6487 if (cpus_empty(tmp
))
6490 sg
= kmalloc_node(sizeof(struct sched_group
),
6494 "Can not alloc domain group for node %d\n", j
);
6497 sg
->__cpu_power
= 0;
6499 sg
->next
= prev
->next
;
6500 cpus_or(covered
, covered
, tmp
);
6507 /* Calculate CPU power for physical packages and nodes */
6508 #ifdef CONFIG_SCHED_SMT
6509 for_each_cpu_mask(i
, *cpu_map
) {
6510 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
6512 init_sched_groups_power(i
, sd
);
6515 #ifdef CONFIG_SCHED_MC
6516 for_each_cpu_mask(i
, *cpu_map
) {
6517 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
6519 init_sched_groups_power(i
, sd
);
6523 for_each_cpu_mask(i
, *cpu_map
) {
6524 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
6526 init_sched_groups_power(i
, sd
);
6530 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6531 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6534 struct sched_group
*sg
;
6536 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6537 init_numa_sched_groups_power(sg
);
6541 /* Attach the domains */
6542 for_each_cpu_mask(i
, *cpu_map
) {
6543 struct sched_domain
*sd
;
6544 #ifdef CONFIG_SCHED_SMT
6545 sd
= &per_cpu(cpu_domains
, i
);
6546 #elif defined(CONFIG_SCHED_MC)
6547 sd
= &per_cpu(core_domains
, i
);
6549 sd
= &per_cpu(phys_domains
, i
);
6551 cpu_attach_domain(sd
, rd
, i
);
6558 free_sched_groups(cpu_map
);
6563 static cpumask_t
*doms_cur
; /* current sched domains */
6564 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6567 * Special case: If a kmalloc of a doms_cur partition (array of
6568 * cpumask_t) fails, then fallback to a single sched domain,
6569 * as determined by the single cpumask_t fallback_doms.
6571 static cpumask_t fallback_doms
;
6574 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6575 * For now this just excludes isolated cpus, but could be used to
6576 * exclude other special cases in the future.
6578 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6583 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6585 doms_cur
= &fallback_doms
;
6586 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
6587 err
= build_sched_domains(doms_cur
);
6588 register_sched_domain_sysctl();
6593 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6595 free_sched_groups(cpu_map
);
6599 * Detach sched domains from a group of cpus specified in cpu_map
6600 * These cpus will now be attached to the NULL domain
6602 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6606 unregister_sched_domain_sysctl();
6608 for_each_cpu_mask(i
, *cpu_map
)
6609 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6610 synchronize_sched();
6611 arch_destroy_sched_domains(cpu_map
);
6615 * Partition sched domains as specified by the 'ndoms_new'
6616 * cpumasks in the array doms_new[] of cpumasks. This compares
6617 * doms_new[] to the current sched domain partitioning, doms_cur[].
6618 * It destroys each deleted domain and builds each new domain.
6620 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6621 * The masks don't intersect (don't overlap.) We should setup one
6622 * sched domain for each mask. CPUs not in any of the cpumasks will
6623 * not be load balanced. If the same cpumask appears both in the
6624 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6627 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6628 * ownership of it and will kfree it when done with it. If the caller
6629 * failed the kmalloc call, then it can pass in doms_new == NULL,
6630 * and partition_sched_domains() will fallback to the single partition
6633 * Call with hotplug lock held
6635 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
)
6641 /* always unregister in case we don't destroy any domains */
6642 unregister_sched_domain_sysctl();
6644 if (doms_new
== NULL
) {
6646 doms_new
= &fallback_doms
;
6647 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
6650 /* Destroy deleted domains */
6651 for (i
= 0; i
< ndoms_cur
; i
++) {
6652 for (j
= 0; j
< ndoms_new
; j
++) {
6653 if (cpus_equal(doms_cur
[i
], doms_new
[j
]))
6656 /* no match - a current sched domain not in new doms_new[] */
6657 detach_destroy_domains(doms_cur
+ i
);
6662 /* Build new domains */
6663 for (i
= 0; i
< ndoms_new
; i
++) {
6664 for (j
= 0; j
< ndoms_cur
; j
++) {
6665 if (cpus_equal(doms_new
[i
], doms_cur
[j
]))
6668 /* no match - add a new doms_new */
6669 build_sched_domains(doms_new
+ i
);
6674 /* Remember the new sched domains */
6675 if (doms_cur
!= &fallback_doms
)
6677 doms_cur
= doms_new
;
6678 ndoms_cur
= ndoms_new
;
6680 register_sched_domain_sysctl();
6685 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6686 static int arch_reinit_sched_domains(void)
6691 detach_destroy_domains(&cpu_online_map
);
6692 err
= arch_init_sched_domains(&cpu_online_map
);
6698 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6702 if (buf
[0] != '0' && buf
[0] != '1')
6706 sched_smt_power_savings
= (buf
[0] == '1');
6708 sched_mc_power_savings
= (buf
[0] == '1');
6710 ret
= arch_reinit_sched_domains();
6712 return ret
? ret
: count
;
6715 #ifdef CONFIG_SCHED_MC
6716 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6718 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6720 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6721 const char *buf
, size_t count
)
6723 return sched_power_savings_store(buf
, count
, 0);
6725 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6726 sched_mc_power_savings_store
);
6729 #ifdef CONFIG_SCHED_SMT
6730 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6732 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6734 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6735 const char *buf
, size_t count
)
6737 return sched_power_savings_store(buf
, count
, 1);
6739 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6740 sched_smt_power_savings_store
);
6743 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6747 #ifdef CONFIG_SCHED_SMT
6749 err
= sysfs_create_file(&cls
->kset
.kobj
,
6750 &attr_sched_smt_power_savings
.attr
);
6752 #ifdef CONFIG_SCHED_MC
6753 if (!err
&& mc_capable())
6754 err
= sysfs_create_file(&cls
->kset
.kobj
,
6755 &attr_sched_mc_power_savings
.attr
);
6762 * Force a reinitialization of the sched domains hierarchy. The domains
6763 * and groups cannot be updated in place without racing with the balancing
6764 * code, so we temporarily attach all running cpus to the NULL domain
6765 * which will prevent rebalancing while the sched domains are recalculated.
6767 static int update_sched_domains(struct notifier_block
*nfb
,
6768 unsigned long action
, void *hcpu
)
6771 case CPU_UP_PREPARE
:
6772 case CPU_UP_PREPARE_FROZEN
:
6773 case CPU_DOWN_PREPARE
:
6774 case CPU_DOWN_PREPARE_FROZEN
:
6775 detach_destroy_domains(&cpu_online_map
);
6778 case CPU_UP_CANCELED
:
6779 case CPU_UP_CANCELED_FROZEN
:
6780 case CPU_DOWN_FAILED
:
6781 case CPU_DOWN_FAILED_FROZEN
:
6783 case CPU_ONLINE_FROZEN
:
6785 case CPU_DEAD_FROZEN
:
6787 * Fall through and re-initialise the domains.
6794 /* The hotplug lock is already held by cpu_up/cpu_down */
6795 arch_init_sched_domains(&cpu_online_map
);
6800 void __init
sched_init_smp(void)
6802 cpumask_t non_isolated_cpus
;
6805 arch_init_sched_domains(&cpu_online_map
);
6806 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6807 if (cpus_empty(non_isolated_cpus
))
6808 cpu_set(smp_processor_id(), non_isolated_cpus
);
6810 /* XXX: Theoretical race here - CPU may be hotplugged now */
6811 hotcpu_notifier(update_sched_domains
, 0);
6813 /* Move init over to a non-isolated CPU */
6814 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6816 sched_init_granularity();
6818 #ifdef CONFIG_FAIR_GROUP_SCHED
6819 if (nr_cpu_ids
== 1)
6822 lb_monitor_task
= kthread_create(load_balance_monitor
, NULL
,
6824 if (!IS_ERR(lb_monitor_task
)) {
6825 lb_monitor_task
->flags
|= PF_NOFREEZE
;
6826 wake_up_process(lb_monitor_task
);
6828 printk(KERN_ERR
"Could not create load balance monitor thread"
6829 "(error = %ld) \n", PTR_ERR(lb_monitor_task
));
6834 void __init
sched_init_smp(void)
6836 sched_init_granularity();
6838 #endif /* CONFIG_SMP */
6840 int in_sched_functions(unsigned long addr
)
6842 return in_lock_functions(addr
) ||
6843 (addr
>= (unsigned long)__sched_text_start
6844 && addr
< (unsigned long)__sched_text_end
);
6847 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
6849 cfs_rq
->tasks_timeline
= RB_ROOT
;
6850 #ifdef CONFIG_FAIR_GROUP_SCHED
6853 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6856 void __init
sched_init(void)
6858 int highest_cpu
= 0;
6862 init_defrootdomain();
6865 for_each_possible_cpu(i
) {
6866 struct rt_prio_array
*array
;
6870 spin_lock_init(&rq
->lock
);
6871 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6874 init_cfs_rq(&rq
->cfs
, rq
);
6875 #ifdef CONFIG_FAIR_GROUP_SCHED
6876 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6878 struct cfs_rq
*cfs_rq
= &per_cpu(init_cfs_rq
, i
);
6879 struct sched_entity
*se
=
6880 &per_cpu(init_sched_entity
, i
);
6882 init_cfs_rq_p
[i
] = cfs_rq
;
6883 init_cfs_rq(cfs_rq
, rq
);
6884 cfs_rq
->tg
= &init_task_group
;
6885 list_add(&cfs_rq
->leaf_cfs_rq_list
,
6886 &rq
->leaf_cfs_rq_list
);
6888 init_sched_entity_p
[i
] = se
;
6889 se
->cfs_rq
= &rq
->cfs
;
6891 se
->load
.weight
= init_task_group_load
;
6892 se
->load
.inv_weight
=
6893 div64_64(1ULL<<32, init_task_group_load
);
6896 init_task_group
.shares
= init_task_group_load
;
6899 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6900 rq
->cpu_load
[j
] = 0;
6904 rq
->active_balance
= 0;
6905 rq
->next_balance
= jiffies
;
6908 rq
->migration_thread
= NULL
;
6909 INIT_LIST_HEAD(&rq
->migration_queue
);
6910 rq
->rt
.highest_prio
= MAX_RT_PRIO
;
6911 rq
->rt
.overloaded
= 0;
6912 rq_attach_root(rq
, &def_root_domain
);
6914 atomic_set(&rq
->nr_iowait
, 0);
6916 array
= &rq
->rt
.active
;
6917 for (j
= 0; j
< MAX_RT_PRIO
; j
++) {
6918 INIT_LIST_HEAD(array
->queue
+ j
);
6919 __clear_bit(j
, array
->bitmap
);
6922 /* delimiter for bitsearch: */
6923 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
6926 set_load_weight(&init_task
);
6928 #ifdef CONFIG_PREEMPT_NOTIFIERS
6929 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6933 nr_cpu_ids
= highest_cpu
+ 1;
6934 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6937 #ifdef CONFIG_RT_MUTEXES
6938 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6942 * The boot idle thread does lazy MMU switching as well:
6944 atomic_inc(&init_mm
.mm_count
);
6945 enter_lazy_tlb(&init_mm
, current
);
6948 * Make us the idle thread. Technically, schedule() should not be
6949 * called from this thread, however somewhere below it might be,
6950 * but because we are the idle thread, we just pick up running again
6951 * when this runqueue becomes "idle".
6953 init_idle(current
, smp_processor_id());
6955 * During early bootup we pretend to be a normal task:
6957 current
->sched_class
= &fair_sched_class
;
6960 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6961 void __might_sleep(char *file
, int line
)
6964 static unsigned long prev_jiffy
; /* ratelimiting */
6966 if ((in_atomic() || irqs_disabled()) &&
6967 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6968 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6970 prev_jiffy
= jiffies
;
6971 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6972 " context at %s:%d\n", file
, line
);
6973 printk("in_atomic():%d, irqs_disabled():%d\n",
6974 in_atomic(), irqs_disabled());
6975 debug_show_held_locks(current
);
6976 if (irqs_disabled())
6977 print_irqtrace_events(current
);
6982 EXPORT_SYMBOL(__might_sleep
);
6985 #ifdef CONFIG_MAGIC_SYSRQ
6986 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6989 update_rq_clock(rq
);
6990 on_rq
= p
->se
.on_rq
;
6992 deactivate_task(rq
, p
, 0);
6993 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6995 activate_task(rq
, p
, 0);
6996 resched_task(rq
->curr
);
7000 void normalize_rt_tasks(void)
7002 struct task_struct
*g
, *p
;
7003 unsigned long flags
;
7006 read_lock_irq(&tasklist_lock
);
7007 do_each_thread(g
, p
) {
7009 * Only normalize user tasks:
7014 p
->se
.exec_start
= 0;
7015 #ifdef CONFIG_SCHEDSTATS
7016 p
->se
.wait_start
= 0;
7017 p
->se
.sleep_start
= 0;
7018 p
->se
.block_start
= 0;
7020 task_rq(p
)->clock
= 0;
7024 * Renice negative nice level userspace
7027 if (TASK_NICE(p
) < 0 && p
->mm
)
7028 set_user_nice(p
, 0);
7032 spin_lock_irqsave(&p
->pi_lock
, flags
);
7033 rq
= __task_rq_lock(p
);
7035 normalize_task(rq
, p
);
7037 __task_rq_unlock(rq
);
7038 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
7039 } while_each_thread(g
, p
);
7041 read_unlock_irq(&tasklist_lock
);
7044 #endif /* CONFIG_MAGIC_SYSRQ */
7048 * These functions are only useful for the IA64 MCA handling.
7050 * They can only be called when the whole system has been
7051 * stopped - every CPU needs to be quiescent, and no scheduling
7052 * activity can take place. Using them for anything else would
7053 * be a serious bug, and as a result, they aren't even visible
7054 * under any other configuration.
7058 * curr_task - return the current task for a given cpu.
7059 * @cpu: the processor in question.
7061 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7063 struct task_struct
*curr_task(int cpu
)
7065 return cpu_curr(cpu
);
7069 * set_curr_task - set the current task for a given cpu.
7070 * @cpu: the processor in question.
7071 * @p: the task pointer to set.
7073 * Description: This function must only be used when non-maskable interrupts
7074 * are serviced on a separate stack. It allows the architecture to switch the
7075 * notion of the current task on a cpu in a non-blocking manner. This function
7076 * must be called with all CPU's synchronized, and interrupts disabled, the
7077 * and caller must save the original value of the current task (see
7078 * curr_task() above) and restore that value before reenabling interrupts and
7079 * re-starting the system.
7081 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7083 void set_curr_task(int cpu
, struct task_struct
*p
)
7090 #ifdef CONFIG_FAIR_GROUP_SCHED
7094 * distribute shares of all task groups among their schedulable entities,
7095 * to reflect load distrbution across cpus.
7097 static int rebalance_shares(struct sched_domain
*sd
, int this_cpu
)
7099 struct cfs_rq
*cfs_rq
;
7100 struct rq
*rq
= cpu_rq(this_cpu
);
7101 cpumask_t sdspan
= sd
->span
;
7104 /* Walk thr' all the task groups that we have */
7105 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
7107 unsigned long total_load
= 0, total_shares
;
7108 struct task_group
*tg
= cfs_rq
->tg
;
7110 /* Gather total task load of this group across cpus */
7111 for_each_cpu_mask(i
, sdspan
)
7112 total_load
+= tg
->cfs_rq
[i
]->load
.weight
;
7114 /* Nothing to do if this group has no load */
7119 * tg->shares represents the number of cpu shares the task group
7120 * is eligible to hold on a single cpu. On N cpus, it is
7121 * eligible to hold (N * tg->shares) number of cpu shares.
7123 total_shares
= tg
->shares
* cpus_weight(sdspan
);
7126 * redistribute total_shares across cpus as per the task load
7129 for_each_cpu_mask(i
, sdspan
) {
7130 unsigned long local_load
, local_shares
;
7132 local_load
= tg
->cfs_rq
[i
]->load
.weight
;
7133 local_shares
= (local_load
* total_shares
) / total_load
;
7135 local_shares
= MIN_GROUP_SHARES
;
7136 if (local_shares
== tg
->se
[i
]->load
.weight
)
7139 spin_lock_irq(&cpu_rq(i
)->lock
);
7140 set_se_shares(tg
->se
[i
], local_shares
);
7141 spin_unlock_irq(&cpu_rq(i
)->lock
);
7150 * How frequently should we rebalance_shares() across cpus?
7152 * The more frequently we rebalance shares, the more accurate is the fairness
7153 * of cpu bandwidth distribution between task groups. However higher frequency
7154 * also implies increased scheduling overhead.
7156 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7157 * consecutive calls to rebalance_shares() in the same sched domain.
7159 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7160 * consecutive calls to rebalance_shares() in the same sched domain.
7162 * These settings allows for the appropriate tradeoff between accuracy of
7163 * fairness and the associated overhead.
7167 /* default: 8ms, units: milliseconds */
7168 const_debug
unsigned int sysctl_sched_min_bal_int_shares
= 8;
7170 /* default: 128ms, units: milliseconds */
7171 const_debug
unsigned int sysctl_sched_max_bal_int_shares
= 128;
7173 /* kernel thread that runs rebalance_shares() periodically */
7174 static int load_balance_monitor(void *unused
)
7176 unsigned int timeout
= sysctl_sched_min_bal_int_shares
;
7177 struct sched_param schedparm
;
7181 * We don't want this thread's execution to be limited by the shares
7182 * assigned to default group (init_task_group). Hence make it run
7183 * as a SCHED_RR RT task at the lowest priority.
7185 schedparm
.sched_priority
= 1;
7186 ret
= sched_setscheduler(current
, SCHED_RR
, &schedparm
);
7188 printk(KERN_ERR
"Couldn't set SCHED_RR policy for load balance"
7189 " monitor thread (error = %d) \n", ret
);
7191 while (!kthread_should_stop()) {
7192 int i
, cpu
, balanced
= 1;
7194 /* Prevent cpus going down or coming up */
7196 /* lockout changes to doms_cur[] array */
7199 * Enter a rcu read-side critical section to safely walk rq->sd
7200 * chain on various cpus and to walk task group list
7201 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7205 for (i
= 0; i
< ndoms_cur
; i
++) {
7206 cpumask_t cpumap
= doms_cur
[i
];
7207 struct sched_domain
*sd
= NULL
, *sd_prev
= NULL
;
7209 cpu
= first_cpu(cpumap
);
7211 /* Find the highest domain at which to balance shares */
7212 for_each_domain(cpu
, sd
) {
7213 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7219 /* sd == NULL? No load balance reqd in this domain */
7223 balanced
&= rebalance_shares(sd
, cpu
);
7232 timeout
= sysctl_sched_min_bal_int_shares
;
7233 else if (timeout
< sysctl_sched_max_bal_int_shares
)
7236 msleep_interruptible(timeout
);
7241 #endif /* CONFIG_SMP */
7243 /* allocate runqueue etc for a new task group */
7244 struct task_group
*sched_create_group(void)
7246 struct task_group
*tg
;
7247 struct cfs_rq
*cfs_rq
;
7248 struct sched_entity
*se
;
7252 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7254 return ERR_PTR(-ENOMEM
);
7256 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * NR_CPUS
, GFP_KERNEL
);
7259 tg
->se
= kzalloc(sizeof(se
) * NR_CPUS
, GFP_KERNEL
);
7263 for_each_possible_cpu(i
) {
7266 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
), GFP_KERNEL
,
7271 se
= kmalloc_node(sizeof(struct sched_entity
), GFP_KERNEL
,
7276 memset(cfs_rq
, 0, sizeof(struct cfs_rq
));
7277 memset(se
, 0, sizeof(struct sched_entity
));
7279 tg
->cfs_rq
[i
] = cfs_rq
;
7280 init_cfs_rq(cfs_rq
, rq
);
7284 se
->cfs_rq
= &rq
->cfs
;
7286 se
->load
.weight
= NICE_0_LOAD
;
7287 se
->load
.inv_weight
= div64_64(1ULL<<32, NICE_0_LOAD
);
7291 tg
->shares
= NICE_0_LOAD
;
7293 lock_task_group_list();
7294 for_each_possible_cpu(i
) {
7296 cfs_rq
= tg
->cfs_rq
[i
];
7297 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7299 unlock_task_group_list();
7304 for_each_possible_cpu(i
) {
7306 kfree(tg
->cfs_rq
[i
]);
7314 return ERR_PTR(-ENOMEM
);
7317 /* rcu callback to free various structures associated with a task group */
7318 static void free_sched_group(struct rcu_head
*rhp
)
7320 struct task_group
*tg
= container_of(rhp
, struct task_group
, rcu
);
7321 struct cfs_rq
*cfs_rq
;
7322 struct sched_entity
*se
;
7325 /* now it should be safe to free those cfs_rqs */
7326 for_each_possible_cpu(i
) {
7327 cfs_rq
= tg
->cfs_rq
[i
];
7339 /* Destroy runqueue etc associated with a task group */
7340 void sched_destroy_group(struct task_group
*tg
)
7342 struct cfs_rq
*cfs_rq
= NULL
;
7345 lock_task_group_list();
7346 for_each_possible_cpu(i
) {
7347 cfs_rq
= tg
->cfs_rq
[i
];
7348 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7350 unlock_task_group_list();
7354 /* wait for possible concurrent references to cfs_rqs complete */
7355 call_rcu(&tg
->rcu
, free_sched_group
);
7358 /* change task's runqueue when it moves between groups.
7359 * The caller of this function should have put the task in its new group
7360 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7361 * reflect its new group.
7363 void sched_move_task(struct task_struct
*tsk
)
7366 unsigned long flags
;
7369 rq
= task_rq_lock(tsk
, &flags
);
7371 if (tsk
->sched_class
!= &fair_sched_class
) {
7372 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7376 update_rq_clock(rq
);
7378 running
= task_current(rq
, tsk
);
7379 on_rq
= tsk
->se
.on_rq
;
7382 dequeue_task(rq
, tsk
, 0);
7383 if (unlikely(running
))
7384 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7387 set_task_cfs_rq(tsk
, task_cpu(tsk
));
7390 if (unlikely(running
))
7391 tsk
->sched_class
->set_curr_task(rq
);
7392 enqueue_task(rq
, tsk
, 0);
7396 task_rq_unlock(rq
, &flags
);
7399 /* rq->lock to be locked by caller */
7400 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
7402 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
7403 struct rq
*rq
= cfs_rq
->rq
;
7407 shares
= MIN_GROUP_SHARES
;
7411 dequeue_entity(cfs_rq
, se
, 0);
7412 dec_cpu_load(rq
, se
->load
.weight
);
7415 se
->load
.weight
= shares
;
7416 se
->load
.inv_weight
= div64_64((1ULL<<32), shares
);
7419 enqueue_entity(cfs_rq
, se
, 0);
7420 inc_cpu_load(rq
, se
->load
.weight
);
7424 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7427 struct cfs_rq
*cfs_rq
;
7430 lock_task_group_list();
7431 if (tg
->shares
== shares
)
7434 if (shares
< MIN_GROUP_SHARES
)
7435 shares
= MIN_GROUP_SHARES
;
7438 * Prevent any load balance activity (rebalance_shares,
7439 * load_balance_fair) from referring to this group first,
7440 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7442 for_each_possible_cpu(i
) {
7443 cfs_rq
= tg
->cfs_rq
[i
];
7444 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
7447 /* wait for any ongoing reference to this group to finish */
7448 synchronize_sched();
7451 * Now we are free to modify the group's share on each cpu
7452 * w/o tripping rebalance_share or load_balance_fair.
7454 tg
->shares
= shares
;
7455 for_each_possible_cpu(i
) {
7456 spin_lock_irq(&cpu_rq(i
)->lock
);
7457 set_se_shares(tg
->se
[i
], shares
);
7458 spin_unlock_irq(&cpu_rq(i
)->lock
);
7462 * Enable load balance activity on this group, by inserting it back on
7463 * each cpu's rq->leaf_cfs_rq_list.
7465 for_each_possible_cpu(i
) {
7467 cfs_rq
= tg
->cfs_rq
[i
];
7468 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7471 unlock_task_group_list();
7475 unsigned long sched_group_shares(struct task_group
*tg
)
7480 #endif /* CONFIG_FAIR_GROUP_SCHED */
7482 #ifdef CONFIG_FAIR_CGROUP_SCHED
7484 /* return corresponding task_group object of a cgroup */
7485 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7487 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7488 struct task_group
, css
);
7491 static struct cgroup_subsys_state
*
7492 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7494 struct task_group
*tg
;
7496 if (!cgrp
->parent
) {
7497 /* This is early initialization for the top cgroup */
7498 init_task_group
.css
.cgroup
= cgrp
;
7499 return &init_task_group
.css
;
7502 /* we support only 1-level deep hierarchical scheduler atm */
7503 if (cgrp
->parent
->parent
)
7504 return ERR_PTR(-EINVAL
);
7506 tg
= sched_create_group();
7508 return ERR_PTR(-ENOMEM
);
7510 /* Bind the cgroup to task_group object we just created */
7511 tg
->css
.cgroup
= cgrp
;
7517 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7519 struct task_group
*tg
= cgroup_tg(cgrp
);
7521 sched_destroy_group(tg
);
7525 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7526 struct task_struct
*tsk
)
7528 /* We don't support RT-tasks being in separate groups */
7529 if (tsk
->sched_class
!= &fair_sched_class
)
7536 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7537 struct cgroup
*old_cont
, struct task_struct
*tsk
)
7539 sched_move_task(tsk
);
7542 static int cpu_shares_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7545 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
7548 static u64
cpu_shares_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7550 struct task_group
*tg
= cgroup_tg(cgrp
);
7552 return (u64
) tg
->shares
;
7555 static struct cftype cpu_files
[] = {
7558 .read_uint
= cpu_shares_read_uint
,
7559 .write_uint
= cpu_shares_write_uint
,
7563 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7565 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7568 struct cgroup_subsys cpu_cgroup_subsys
= {
7570 .create
= cpu_cgroup_create
,
7571 .destroy
= cpu_cgroup_destroy
,
7572 .can_attach
= cpu_cgroup_can_attach
,
7573 .attach
= cpu_cgroup_attach
,
7574 .populate
= cpu_cgroup_populate
,
7575 .subsys_id
= cpu_cgroup_subsys_id
,
7579 #endif /* CONFIG_FAIR_CGROUP_SCHED */
7581 #ifdef CONFIG_CGROUP_CPUACCT
7584 * CPU accounting code for task groups.
7586 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7587 * (balbir@in.ibm.com).
7590 /* track cpu usage of a group of tasks */
7592 struct cgroup_subsys_state css
;
7593 /* cpuusage holds pointer to a u64-type object on every cpu */
7597 struct cgroup_subsys cpuacct_subsys
;
7599 /* return cpu accounting group corresponding to this container */
7600 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cont
)
7602 return container_of(cgroup_subsys_state(cont
, cpuacct_subsys_id
),
7603 struct cpuacct
, css
);
7606 /* return cpu accounting group to which this task belongs */
7607 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
7609 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
7610 struct cpuacct
, css
);
7613 /* create a new cpu accounting group */
7614 static struct cgroup_subsys_state
*cpuacct_create(
7615 struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7617 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7620 return ERR_PTR(-ENOMEM
);
7622 ca
->cpuusage
= alloc_percpu(u64
);
7623 if (!ca
->cpuusage
) {
7625 return ERR_PTR(-ENOMEM
);
7631 /* destroy an existing cpu accounting group */
7633 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7635 struct cpuacct
*ca
= cgroup_ca(cont
);
7637 free_percpu(ca
->cpuusage
);
7641 /* return total cpu usage (in nanoseconds) of a group */
7642 static u64
cpuusage_read(struct cgroup
*cont
, struct cftype
*cft
)
7644 struct cpuacct
*ca
= cgroup_ca(cont
);
7645 u64 totalcpuusage
= 0;
7648 for_each_possible_cpu(i
) {
7649 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
7652 * Take rq->lock to make 64-bit addition safe on 32-bit
7655 spin_lock_irq(&cpu_rq(i
)->lock
);
7656 totalcpuusage
+= *cpuusage
;
7657 spin_unlock_irq(&cpu_rq(i
)->lock
);
7660 return totalcpuusage
;
7663 static struct cftype files
[] = {
7666 .read_uint
= cpuusage_read
,
7670 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7672 return cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
7676 * charge this task's execution time to its accounting group.
7678 * called with rq->lock held.
7680 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
7684 if (!cpuacct_subsys
.active
)
7689 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
7691 *cpuusage
+= cputime
;
7695 struct cgroup_subsys cpuacct_subsys
= {
7697 .create
= cpuacct_create
,
7698 .destroy
= cpuacct_destroy
,
7699 .populate
= cpuacct_populate
,
7700 .subsys_id
= cpuacct_subsys_id
,
7702 #endif /* CONFIG_CGROUP_CPUACCT */