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>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
75 #include <asm/irq_regs.h>
77 #include "sched_cpupri.h"
80 * Convert user-nice values [ -20 ... 0 ... 19 ]
81 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
84 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
85 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
86 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
89 * 'User priority' is the nice value converted to something we
90 * can work with better when scaling various scheduler parameters,
91 * it's a [ 0 ... 39 ] range.
93 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
94 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
95 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
98 * Helpers for converting nanosecond timing to jiffy resolution
100 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
114 * single value that denotes runtime == period, ie unlimited time.
116 #define RUNTIME_INF ((u64)~0ULL)
120 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
121 * Since cpu_power is a 'constant', we can use a reciprocal divide.
123 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
125 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
129 * Each time a sched group cpu_power is changed,
130 * we must compute its reciprocal value
132 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
134 sg
->__cpu_power
+= val
;
135 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
139 static inline int rt_policy(int policy
)
141 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
146 static inline int task_has_rt_policy(struct task_struct
*p
)
148 return rt_policy(p
->policy
);
152 * This is the priority-queue data structure of the RT scheduling class:
154 struct rt_prio_array
{
155 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
156 struct list_head queue
[MAX_RT_PRIO
];
159 struct rt_bandwidth
{
160 /* nests inside the rq lock: */
161 spinlock_t rt_runtime_lock
;
164 struct hrtimer rt_period_timer
;
167 static struct rt_bandwidth def_rt_bandwidth
;
169 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
171 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
173 struct rt_bandwidth
*rt_b
=
174 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
180 now
= hrtimer_cb_get_time(timer
);
181 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
186 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
189 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
193 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
195 rt_b
->rt_period
= ns_to_ktime(period
);
196 rt_b
->rt_runtime
= runtime
;
198 spin_lock_init(&rt_b
->rt_runtime_lock
);
200 hrtimer_init(&rt_b
->rt_period_timer
,
201 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
202 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
203 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
206 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
210 if (rt_b
->rt_runtime
== RUNTIME_INF
)
213 if (hrtimer_active(&rt_b
->rt_period_timer
))
216 spin_lock(&rt_b
->rt_runtime_lock
);
218 if (hrtimer_active(&rt_b
->rt_period_timer
))
221 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
222 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
223 hrtimer_start(&rt_b
->rt_period_timer
,
224 rt_b
->rt_period_timer
.expires
,
227 spin_unlock(&rt_b
->rt_runtime_lock
);
230 #ifdef CONFIG_RT_GROUP_SCHED
231 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
233 hrtimer_cancel(&rt_b
->rt_period_timer
);
238 * sched_domains_mutex serializes calls to arch_init_sched_domains,
239 * detach_destroy_domains and partition_sched_domains.
241 static DEFINE_MUTEX(sched_domains_mutex
);
243 #ifdef CONFIG_GROUP_SCHED
245 #include <linux/cgroup.h>
249 static LIST_HEAD(task_groups
);
251 /* task group related information */
253 #ifdef CONFIG_CGROUP_SCHED
254 struct cgroup_subsys_state css
;
257 #ifdef CONFIG_FAIR_GROUP_SCHED
258 /* schedulable entities of this group on each cpu */
259 struct sched_entity
**se
;
260 /* runqueue "owned" by this group on each cpu */
261 struct cfs_rq
**cfs_rq
;
262 unsigned long shares
;
265 #ifdef CONFIG_RT_GROUP_SCHED
266 struct sched_rt_entity
**rt_se
;
267 struct rt_rq
**rt_rq
;
269 struct rt_bandwidth rt_bandwidth
;
273 struct list_head list
;
275 struct task_group
*parent
;
276 struct list_head siblings
;
277 struct list_head children
;
280 #ifdef CONFIG_USER_SCHED
284 * Every UID task group (including init_task_group aka UID-0) will
285 * be a child to this group.
287 struct task_group root_task_group
;
289 #ifdef CONFIG_FAIR_GROUP_SCHED
290 /* Default task group's sched entity on each cpu */
291 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
292 /* Default task group's cfs_rq on each cpu */
293 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
294 #endif /* CONFIG_FAIR_GROUP_SCHED */
296 #ifdef CONFIG_RT_GROUP_SCHED
297 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
298 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
299 #endif /* CONFIG_RT_GROUP_SCHED */
300 #else /* !CONFIG_FAIR_GROUP_SCHED */
301 #define root_task_group init_task_group
302 #endif /* CONFIG_FAIR_GROUP_SCHED */
304 /* task_group_lock serializes add/remove of task groups and also changes to
305 * a task group's cpu shares.
307 static DEFINE_SPINLOCK(task_group_lock
);
309 #ifdef CONFIG_FAIR_GROUP_SCHED
310 #ifdef CONFIG_USER_SCHED
311 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 #else /* !CONFIG_USER_SCHED */
313 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
314 #endif /* CONFIG_USER_SCHED */
317 * A weight of 0 or 1 can cause arithmetics problems.
318 * A weight of a cfs_rq is the sum of weights of which entities
319 * are queued on this cfs_rq, so a weight of a entity should not be
320 * too large, so as the shares value of a task group.
321 * (The default weight is 1024 - so there's no practical
322 * limitation from this.)
325 #define MAX_SHARES (1UL << 18)
327 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
330 /* Default task group.
331 * Every task in system belong to this group at bootup.
333 struct task_group init_task_group
;
335 /* return group to which a task belongs */
336 static inline struct task_group
*task_group(struct task_struct
*p
)
338 struct task_group
*tg
;
340 #ifdef CONFIG_USER_SCHED
342 #elif defined(CONFIG_CGROUP_SCHED)
343 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
344 struct task_group
, css
);
346 tg
= &init_task_group
;
351 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
352 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
354 #ifdef CONFIG_FAIR_GROUP_SCHED
355 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
356 p
->se
.parent
= task_group(p
)->se
[cpu
];
359 #ifdef CONFIG_RT_GROUP_SCHED
360 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
361 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
367 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
369 #endif /* CONFIG_GROUP_SCHED */
371 /* CFS-related fields in a runqueue */
373 struct load_weight load
;
374 unsigned long nr_running
;
380 struct rb_root tasks_timeline
;
381 struct rb_node
*rb_leftmost
;
383 struct list_head tasks
;
384 struct list_head
*balance_iterator
;
387 * 'curr' points to currently running entity on this cfs_rq.
388 * It is set to NULL otherwise (i.e when none are currently running).
390 struct sched_entity
*curr
, *next
;
392 unsigned long nr_spread_over
;
394 #ifdef CONFIG_FAIR_GROUP_SCHED
395 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
398 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
399 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
400 * (like users, containers etc.)
402 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
403 * list is used during load balance.
405 struct list_head leaf_cfs_rq_list
;
406 struct task_group
*tg
; /* group that "owns" this runqueue */
410 * the part of load.weight contributed by tasks
412 unsigned long task_weight
;
415 * h_load = weight * f(tg)
417 * Where f(tg) is the recursive weight fraction assigned to
420 unsigned long h_load
;
423 * this cpu's part of tg->shares
425 unsigned long shares
;
430 /* Real-Time classes' related field in a runqueue: */
432 struct rt_prio_array active
;
433 unsigned long rt_nr_running
;
434 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
435 int highest_prio
; /* highest queued rt task prio */
438 unsigned long rt_nr_migratory
;
444 /* Nests inside the rq lock: */
445 spinlock_t rt_runtime_lock
;
447 #ifdef CONFIG_RT_GROUP_SCHED
448 unsigned long rt_nr_boosted
;
451 struct list_head leaf_rt_rq_list
;
452 struct task_group
*tg
;
453 struct sched_rt_entity
*rt_se
;
460 * We add the notion of a root-domain which will be used to define per-domain
461 * variables. Each exclusive cpuset essentially defines an island domain by
462 * fully partitioning the member cpus from any other cpuset. Whenever a new
463 * exclusive cpuset is created, we also create and attach a new root-domain
473 * The "RT overload" flag: it gets set if a CPU has more than
474 * one runnable RT task.
479 struct cpupri cpupri
;
484 * By default the system creates a single root-domain with all cpus as
485 * members (mimicking the global state we have today).
487 static struct root_domain def_root_domain
;
492 * This is the main, per-CPU runqueue data structure.
494 * Locking rule: those places that want to lock multiple runqueues
495 * (such as the load balancing or the thread migration code), lock
496 * acquire operations must be ordered by ascending &runqueue.
503 * nr_running and cpu_load should be in the same cacheline because
504 * remote CPUs use both these fields when doing load calculation.
506 unsigned long nr_running
;
507 #define CPU_LOAD_IDX_MAX 5
508 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
509 unsigned char idle_at_tick
;
511 unsigned long last_tick_seen
;
512 unsigned char in_nohz_recently
;
514 /* capture load from *all* tasks on this cpu: */
515 struct load_weight load
;
516 unsigned long nr_load_updates
;
522 #ifdef CONFIG_FAIR_GROUP_SCHED
523 /* list of leaf cfs_rq on this cpu: */
524 struct list_head leaf_cfs_rq_list
;
526 #ifdef CONFIG_RT_GROUP_SCHED
527 struct list_head leaf_rt_rq_list
;
531 * This is part of a global counter where only the total sum
532 * over all CPUs matters. A task can increase this counter on
533 * one CPU and if it got migrated afterwards it may decrease
534 * it on another CPU. Always updated under the runqueue lock:
536 unsigned long nr_uninterruptible
;
538 struct task_struct
*curr
, *idle
;
539 unsigned long next_balance
;
540 struct mm_struct
*prev_mm
;
547 struct root_domain
*rd
;
548 struct sched_domain
*sd
;
550 /* For active balancing */
553 /* cpu of this runqueue: */
557 unsigned long avg_load_per_task
;
559 struct task_struct
*migration_thread
;
560 struct list_head migration_queue
;
563 #ifdef CONFIG_SCHED_HRTICK
564 unsigned long hrtick_flags
;
565 ktime_t hrtick_expire
;
566 struct hrtimer hrtick_timer
;
569 #ifdef CONFIG_SCHEDSTATS
571 struct sched_info rq_sched_info
;
573 /* sys_sched_yield() stats */
574 unsigned int yld_exp_empty
;
575 unsigned int yld_act_empty
;
576 unsigned int yld_both_empty
;
577 unsigned int yld_count
;
579 /* schedule() stats */
580 unsigned int sched_switch
;
581 unsigned int sched_count
;
582 unsigned int sched_goidle
;
584 /* try_to_wake_up() stats */
585 unsigned int ttwu_count
;
586 unsigned int ttwu_local
;
589 unsigned int bkl_count
;
591 struct lock_class_key rq_lock_key
;
594 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
596 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
598 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
601 static inline int cpu_of(struct rq
*rq
)
611 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
612 * See detach_destroy_domains: synchronize_sched for details.
614 * The domain tree of any CPU may only be accessed from within
615 * preempt-disabled sections.
617 #define for_each_domain(cpu, __sd) \
618 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
620 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
621 #define this_rq() (&__get_cpu_var(runqueues))
622 #define task_rq(p) cpu_rq(task_cpu(p))
623 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
625 static inline void update_rq_clock(struct rq
*rq
)
627 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
631 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
633 #ifdef CONFIG_SCHED_DEBUG
634 # define const_debug __read_mostly
636 # define const_debug static const
640 * Debugging: various feature bits
643 #define SCHED_FEAT(name, enabled) \
644 __SCHED_FEAT_##name ,
647 #include "sched_features.h"
652 #define SCHED_FEAT(name, enabled) \
653 (1UL << __SCHED_FEAT_##name) * enabled |
655 const_debug
unsigned int sysctl_sched_features
=
656 #include "sched_features.h"
661 #ifdef CONFIG_SCHED_DEBUG
662 #define SCHED_FEAT(name, enabled) \
665 static __read_mostly
char *sched_feat_names
[] = {
666 #include "sched_features.h"
672 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
674 filp
->private_data
= inode
->i_private
;
679 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
680 size_t cnt
, loff_t
*ppos
)
687 for (i
= 0; sched_feat_names
[i
]; i
++) {
688 len
+= strlen(sched_feat_names
[i
]);
692 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
696 for (i
= 0; sched_feat_names
[i
]; i
++) {
697 if (sysctl_sched_features
& (1UL << i
))
698 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
700 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
703 r
+= sprintf(buf
+ r
, "\n");
704 WARN_ON(r
>= len
+ 2);
706 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
714 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
715 size_t cnt
, loff_t
*ppos
)
725 if (copy_from_user(&buf
, ubuf
, cnt
))
730 if (strncmp(buf
, "NO_", 3) == 0) {
735 for (i
= 0; sched_feat_names
[i
]; i
++) {
736 int len
= strlen(sched_feat_names
[i
]);
738 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
740 sysctl_sched_features
&= ~(1UL << i
);
742 sysctl_sched_features
|= (1UL << i
);
747 if (!sched_feat_names
[i
])
755 static struct file_operations sched_feat_fops
= {
756 .open
= sched_feat_open
,
757 .read
= sched_feat_read
,
758 .write
= sched_feat_write
,
761 static __init
int sched_init_debug(void)
763 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
768 late_initcall(sched_init_debug
);
772 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
775 * Number of tasks to iterate in a single balance run.
776 * Limited because this is done with IRQs disabled.
778 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
781 * period over which we measure -rt task cpu usage in us.
784 unsigned int sysctl_sched_rt_period
= 1000000;
786 static __read_mostly
int scheduler_running
;
789 * part of the period that we allow rt tasks to run in us.
792 int sysctl_sched_rt_runtime
= 950000;
794 static inline u64
global_rt_period(void)
796 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
799 static inline u64
global_rt_runtime(void)
801 if (sysctl_sched_rt_period
< 0)
804 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
807 #ifndef prepare_arch_switch
808 # define prepare_arch_switch(next) do { } while (0)
810 #ifndef finish_arch_switch
811 # define finish_arch_switch(prev) do { } while (0)
814 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
816 return rq
->curr
== p
;
819 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
820 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
822 return task_current(rq
, p
);
825 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
829 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
831 #ifdef CONFIG_DEBUG_SPINLOCK
832 /* this is a valid case when another task releases the spinlock */
833 rq
->lock
.owner
= current
;
836 * If we are tracking spinlock dependencies then we have to
837 * fix up the runqueue lock - which gets 'carried over' from
840 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
842 spin_unlock_irq(&rq
->lock
);
845 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
846 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
851 return task_current(rq
, p
);
855 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
859 * We can optimise this out completely for !SMP, because the
860 * SMP rebalancing from interrupt is the only thing that cares
865 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
866 spin_unlock_irq(&rq
->lock
);
868 spin_unlock(&rq
->lock
);
872 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
876 * After ->oncpu is cleared, the task can be moved to a different CPU.
877 * We must ensure this doesn't happen until the switch is completely
883 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
887 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
890 * __task_rq_lock - lock the runqueue a given task resides on.
891 * Must be called interrupts disabled.
893 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
897 struct rq
*rq
= task_rq(p
);
898 spin_lock(&rq
->lock
);
899 if (likely(rq
== task_rq(p
)))
901 spin_unlock(&rq
->lock
);
906 * task_rq_lock - lock the runqueue a given task resides on and disable
907 * interrupts. Note the ordering: we can safely lookup the task_rq without
908 * explicitly disabling preemption.
910 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
916 local_irq_save(*flags
);
918 spin_lock(&rq
->lock
);
919 if (likely(rq
== task_rq(p
)))
921 spin_unlock_irqrestore(&rq
->lock
, *flags
);
925 static void __task_rq_unlock(struct rq
*rq
)
928 spin_unlock(&rq
->lock
);
931 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
934 spin_unlock_irqrestore(&rq
->lock
, *flags
);
938 * this_rq_lock - lock this runqueue and disable interrupts.
940 static struct rq
*this_rq_lock(void)
947 spin_lock(&rq
->lock
);
952 static void __resched_task(struct task_struct
*p
, int tif_bit
);
954 static inline void resched_task(struct task_struct
*p
)
956 __resched_task(p
, TIF_NEED_RESCHED
);
959 #ifdef CONFIG_SCHED_HRTICK
961 * Use HR-timers to deliver accurate preemption points.
963 * Its all a bit involved since we cannot program an hrt while holding the
964 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
967 * When we get rescheduled we reprogram the hrtick_timer outside of the
970 static inline void resched_hrt(struct task_struct
*p
)
972 __resched_task(p
, TIF_HRTICK_RESCHED
);
975 static inline void resched_rq(struct rq
*rq
)
979 spin_lock_irqsave(&rq
->lock
, flags
);
980 resched_task(rq
->curr
);
981 spin_unlock_irqrestore(&rq
->lock
, flags
);
985 HRTICK_SET
, /* re-programm hrtick_timer */
986 HRTICK_RESET
, /* not a new slice */
987 HRTICK_BLOCK
, /* stop hrtick operations */
992 * - enabled by features
993 * - hrtimer is actually high res
995 static inline int hrtick_enabled(struct rq
*rq
)
997 if (!sched_feat(HRTICK
))
999 if (unlikely(test_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
)))
1001 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1005 * Called to set the hrtick timer state.
1007 * called with rq->lock held and irqs disabled
1009 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
1011 assert_spin_locked(&rq
->lock
);
1014 * preempt at: now + delay
1017 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
1019 * indicate we need to program the timer
1021 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1023 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1026 * New slices are called from the schedule path and don't need a
1027 * forced reschedule.
1030 resched_hrt(rq
->curr
);
1033 static void hrtick_clear(struct rq
*rq
)
1035 if (hrtimer_active(&rq
->hrtick_timer
))
1036 hrtimer_cancel(&rq
->hrtick_timer
);
1040 * Update the timer from the possible pending state.
1042 static void hrtick_set(struct rq
*rq
)
1046 unsigned long flags
;
1048 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1050 spin_lock_irqsave(&rq
->lock
, flags
);
1051 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1052 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1053 time
= rq
->hrtick_expire
;
1054 clear_thread_flag(TIF_HRTICK_RESCHED
);
1055 spin_unlock_irqrestore(&rq
->lock
, flags
);
1058 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
1059 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
1066 * High-resolution timer tick.
1067 * Runs from hardirq context with interrupts disabled.
1069 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1071 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1073 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1075 spin_lock(&rq
->lock
);
1076 update_rq_clock(rq
);
1077 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1078 spin_unlock(&rq
->lock
);
1080 return HRTIMER_NORESTART
;
1084 static void hotplug_hrtick_disable(int cpu
)
1086 struct rq
*rq
= cpu_rq(cpu
);
1087 unsigned long flags
;
1089 spin_lock_irqsave(&rq
->lock
, flags
);
1090 rq
->hrtick_flags
= 0;
1091 __set_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1092 spin_unlock_irqrestore(&rq
->lock
, flags
);
1097 static void hotplug_hrtick_enable(int cpu
)
1099 struct rq
*rq
= cpu_rq(cpu
);
1100 unsigned long flags
;
1102 spin_lock_irqsave(&rq
->lock
, flags
);
1103 __clear_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1104 spin_unlock_irqrestore(&rq
->lock
, flags
);
1108 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1110 int cpu
= (int)(long)hcpu
;
1113 case CPU_UP_CANCELED
:
1114 case CPU_UP_CANCELED_FROZEN
:
1115 case CPU_DOWN_PREPARE
:
1116 case CPU_DOWN_PREPARE_FROZEN
:
1118 case CPU_DEAD_FROZEN
:
1119 hotplug_hrtick_disable(cpu
);
1122 case CPU_UP_PREPARE
:
1123 case CPU_UP_PREPARE_FROZEN
:
1124 case CPU_DOWN_FAILED
:
1125 case CPU_DOWN_FAILED_FROZEN
:
1127 case CPU_ONLINE_FROZEN
:
1128 hotplug_hrtick_enable(cpu
);
1135 static void init_hrtick(void)
1137 hotcpu_notifier(hotplug_hrtick
, 0);
1139 #endif /* CONFIG_SMP */
1141 static void init_rq_hrtick(struct rq
*rq
)
1143 rq
->hrtick_flags
= 0;
1144 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1145 rq
->hrtick_timer
.function
= hrtick
;
1146 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1149 void hrtick_resched(void)
1152 unsigned long flags
;
1154 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
1157 local_irq_save(flags
);
1158 rq
= cpu_rq(smp_processor_id());
1160 local_irq_restore(flags
);
1163 static inline void hrtick_clear(struct rq
*rq
)
1167 static inline void hrtick_set(struct rq
*rq
)
1171 static inline void init_rq_hrtick(struct rq
*rq
)
1175 void hrtick_resched(void)
1179 static inline void init_hrtick(void)
1185 * resched_task - mark a task 'to be rescheduled now'.
1187 * On UP this means the setting of the need_resched flag, on SMP it
1188 * might also involve a cross-CPU call to trigger the scheduler on
1193 #ifndef tsk_is_polling
1194 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1197 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1201 assert_spin_locked(&task_rq(p
)->lock
);
1203 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1206 set_tsk_thread_flag(p
, tif_bit
);
1209 if (cpu
== smp_processor_id())
1212 /* NEED_RESCHED must be visible before we test polling */
1214 if (!tsk_is_polling(p
))
1215 smp_send_reschedule(cpu
);
1218 static void resched_cpu(int cpu
)
1220 struct rq
*rq
= cpu_rq(cpu
);
1221 unsigned long flags
;
1223 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1225 resched_task(cpu_curr(cpu
));
1226 spin_unlock_irqrestore(&rq
->lock
, flags
);
1231 * When add_timer_on() enqueues a timer into the timer wheel of an
1232 * idle CPU then this timer might expire before the next timer event
1233 * which is scheduled to wake up that CPU. In case of a completely
1234 * idle system the next event might even be infinite time into the
1235 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1236 * leaves the inner idle loop so the newly added timer is taken into
1237 * account when the CPU goes back to idle and evaluates the timer
1238 * wheel for the next timer event.
1240 void wake_up_idle_cpu(int cpu
)
1242 struct rq
*rq
= cpu_rq(cpu
);
1244 if (cpu
== smp_processor_id())
1248 * This is safe, as this function is called with the timer
1249 * wheel base lock of (cpu) held. When the CPU is on the way
1250 * to idle and has not yet set rq->curr to idle then it will
1251 * be serialized on the timer wheel base lock and take the new
1252 * timer into account automatically.
1254 if (rq
->curr
!= rq
->idle
)
1258 * We can set TIF_RESCHED on the idle task of the other CPU
1259 * lockless. The worst case is that the other CPU runs the
1260 * idle task through an additional NOOP schedule()
1262 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1264 /* NEED_RESCHED must be visible before we test polling */
1266 if (!tsk_is_polling(rq
->idle
))
1267 smp_send_reschedule(cpu
);
1269 #endif /* CONFIG_NO_HZ */
1271 #else /* !CONFIG_SMP */
1272 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1274 assert_spin_locked(&task_rq(p
)->lock
);
1275 set_tsk_thread_flag(p
, tif_bit
);
1277 #endif /* CONFIG_SMP */
1279 #if BITS_PER_LONG == 32
1280 # define WMULT_CONST (~0UL)
1282 # define WMULT_CONST (1UL << 32)
1285 #define WMULT_SHIFT 32
1288 * Shift right and round:
1290 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1293 * delta *= weight / lw
1295 static unsigned long
1296 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1297 struct load_weight
*lw
)
1301 if (!lw
->inv_weight
) {
1302 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1305 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1309 tmp
= (u64
)delta_exec
* weight
;
1311 * Check whether we'd overflow the 64-bit multiplication:
1313 if (unlikely(tmp
> WMULT_CONST
))
1314 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1317 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1319 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1322 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1328 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1335 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1336 * of tasks with abnormal "nice" values across CPUs the contribution that
1337 * each task makes to its run queue's load is weighted according to its
1338 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1339 * scaled version of the new time slice allocation that they receive on time
1343 #define WEIGHT_IDLEPRIO 2
1344 #define WMULT_IDLEPRIO (1 << 31)
1347 * Nice levels are multiplicative, with a gentle 10% change for every
1348 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1349 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1350 * that remained on nice 0.
1352 * The "10% effect" is relative and cumulative: from _any_ nice level,
1353 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1354 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1355 * If a task goes up by ~10% and another task goes down by ~10% then
1356 * the relative distance between them is ~25%.)
1358 static const int prio_to_weight
[40] = {
1359 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1360 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1361 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1362 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1363 /* 0 */ 1024, 820, 655, 526, 423,
1364 /* 5 */ 335, 272, 215, 172, 137,
1365 /* 10 */ 110, 87, 70, 56, 45,
1366 /* 15 */ 36, 29, 23, 18, 15,
1370 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1372 * In cases where the weight does not change often, we can use the
1373 * precalculated inverse to speed up arithmetics by turning divisions
1374 * into multiplications:
1376 static const u32 prio_to_wmult
[40] = {
1377 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1378 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1379 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1380 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1381 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1382 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1383 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1384 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1387 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1390 * runqueue iterator, to support SMP load-balancing between different
1391 * scheduling classes, without having to expose their internal data
1392 * structures to the load-balancing proper:
1394 struct rq_iterator
{
1396 struct task_struct
*(*start
)(void *);
1397 struct task_struct
*(*next
)(void *);
1401 static unsigned long
1402 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1403 unsigned long max_load_move
, struct sched_domain
*sd
,
1404 enum cpu_idle_type idle
, int *all_pinned
,
1405 int *this_best_prio
, struct rq_iterator
*iterator
);
1408 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1409 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1410 struct rq_iterator
*iterator
);
1413 #ifdef CONFIG_CGROUP_CPUACCT
1414 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1416 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1419 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1421 update_load_add(&rq
->load
, load
);
1424 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1426 update_load_sub(&rq
->load
, load
);
1430 static unsigned long source_load(int cpu
, int type
);
1431 static unsigned long target_load(int cpu
, int type
);
1432 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1434 static unsigned long cpu_avg_load_per_task(int cpu
)
1436 struct rq
*rq
= cpu_rq(cpu
);
1439 rq
->avg_load_per_task
= rq
->load
.weight
/ rq
->nr_running
;
1441 return rq
->avg_load_per_task
;
1444 #ifdef CONFIG_FAIR_GROUP_SCHED
1446 typedef void (*tg_visitor
)(struct task_group
*, int, struct sched_domain
*);
1449 * Iterate the full tree, calling @down when first entering a node and @up when
1450 * leaving it for the final time.
1453 walk_tg_tree(tg_visitor down
, tg_visitor up
, int cpu
, struct sched_domain
*sd
)
1455 struct task_group
*parent
, *child
;
1458 parent
= &root_task_group
;
1460 (*down
)(parent
, cpu
, sd
);
1461 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1468 (*up
)(parent
, cpu
, sd
);
1471 parent
= parent
->parent
;
1477 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1480 * Calculate and set the cpu's group shares.
1483 __update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1484 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1487 unsigned long shares
;
1488 unsigned long rq_weight
;
1493 rq_weight
= tg
->cfs_rq
[cpu
]->load
.weight
;
1496 * If there are currently no tasks on the cpu pretend there is one of
1497 * average load so that when a new task gets to run here it will not
1498 * get delayed by group starvation.
1502 rq_weight
= NICE_0_LOAD
;
1505 if (unlikely(rq_weight
> sd_rq_weight
))
1506 rq_weight
= sd_rq_weight
;
1509 * \Sum shares * rq_weight
1510 * shares = -----------------------
1514 shares
= (sd_shares
* rq_weight
) / (sd_rq_weight
+ 1);
1517 * record the actual number of shares, not the boosted amount.
1519 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1521 if (shares
< MIN_SHARES
)
1522 shares
= MIN_SHARES
;
1523 else if (shares
> MAX_SHARES
)
1524 shares
= MAX_SHARES
;
1526 __set_se_shares(tg
->se
[cpu
], shares
);
1530 * Re-compute the task group their per cpu shares over the given domain.
1531 * This needs to be done in a bottom-up fashion because the rq weight of a
1532 * parent group depends on the shares of its child groups.
1535 tg_shares_up(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1537 unsigned long rq_weight
= 0;
1538 unsigned long shares
= 0;
1541 for_each_cpu_mask(i
, sd
->span
) {
1542 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1543 shares
+= tg
->cfs_rq
[i
]->shares
;
1546 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1547 shares
= tg
->shares
;
1549 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1550 shares
= tg
->shares
;
1552 for_each_cpu_mask(i
, sd
->span
) {
1553 struct rq
*rq
= cpu_rq(i
);
1554 unsigned long flags
;
1556 spin_lock_irqsave(&rq
->lock
, flags
);
1557 __update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1558 spin_unlock_irqrestore(&rq
->lock
, flags
);
1563 * Compute the cpu's hierarchical load factor for each task group.
1564 * This needs to be done in a top-down fashion because the load of a child
1565 * group is a fraction of its parents load.
1568 tg_load_down(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1573 load
= cpu_rq(cpu
)->load
.weight
;
1575 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1576 load
*= tg
->cfs_rq
[cpu
]->shares
;
1577 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1580 tg
->cfs_rq
[cpu
]->h_load
= load
;
1584 tg_nop(struct task_group
*tg
, int cpu
, struct sched_domain
*sd
)
1588 static void update_shares(struct sched_domain
*sd
)
1590 walk_tg_tree(tg_nop
, tg_shares_up
, 0, sd
);
1593 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1595 spin_unlock(&rq
->lock
);
1597 spin_lock(&rq
->lock
);
1600 static void update_h_load(int cpu
)
1602 walk_tg_tree(tg_load_down
, tg_nop
, cpu
, NULL
);
1605 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1607 cfs_rq
->shares
= shares
;
1612 static inline void update_shares(struct sched_domain
*sd
)
1616 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1624 #include "sched_stats.h"
1625 #include "sched_idletask.c"
1626 #include "sched_fair.c"
1627 #include "sched_rt.c"
1628 #ifdef CONFIG_SCHED_DEBUG
1629 # include "sched_debug.c"
1632 #define sched_class_highest (&rt_sched_class)
1633 #define for_each_class(class) \
1634 for (class = sched_class_highest; class; class = class->next)
1636 static void inc_nr_running(struct rq
*rq
)
1641 static void dec_nr_running(struct rq
*rq
)
1646 static void set_load_weight(struct task_struct
*p
)
1648 if (task_has_rt_policy(p
)) {
1649 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1650 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1655 * SCHED_IDLE tasks get minimal weight:
1657 if (p
->policy
== SCHED_IDLE
) {
1658 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1659 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1663 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1664 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1667 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1669 sched_info_queued(p
);
1670 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1674 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1676 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1681 * __normal_prio - return the priority that is based on the static prio
1683 static inline int __normal_prio(struct task_struct
*p
)
1685 return p
->static_prio
;
1689 * Calculate the expected normal priority: i.e. priority
1690 * without taking RT-inheritance into account. Might be
1691 * boosted by interactivity modifiers. Changes upon fork,
1692 * setprio syscalls, and whenever the interactivity
1693 * estimator recalculates.
1695 static inline int normal_prio(struct task_struct
*p
)
1699 if (task_has_rt_policy(p
))
1700 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1702 prio
= __normal_prio(p
);
1707 * Calculate the current priority, i.e. the priority
1708 * taken into account by the scheduler. This value might
1709 * be boosted by RT tasks, or might be boosted by
1710 * interactivity modifiers. Will be RT if the task got
1711 * RT-boosted. If not then it returns p->normal_prio.
1713 static int effective_prio(struct task_struct
*p
)
1715 p
->normal_prio
= normal_prio(p
);
1717 * If we are RT tasks or we were boosted to RT priority,
1718 * keep the priority unchanged. Otherwise, update priority
1719 * to the normal priority:
1721 if (!rt_prio(p
->prio
))
1722 return p
->normal_prio
;
1727 * activate_task - move a task to the runqueue.
1729 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1731 if (task_contributes_to_load(p
))
1732 rq
->nr_uninterruptible
--;
1734 enqueue_task(rq
, p
, wakeup
);
1739 * deactivate_task - remove a task from the runqueue.
1741 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1743 if (task_contributes_to_load(p
))
1744 rq
->nr_uninterruptible
++;
1746 dequeue_task(rq
, p
, sleep
);
1751 * task_curr - is this task currently executing on a CPU?
1752 * @p: the task in question.
1754 inline int task_curr(const struct task_struct
*p
)
1756 return cpu_curr(task_cpu(p
)) == p
;
1759 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1761 set_task_rq(p
, cpu
);
1764 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1765 * successfuly executed on another CPU. We must ensure that updates of
1766 * per-task data have been completed by this moment.
1769 task_thread_info(p
)->cpu
= cpu
;
1773 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1774 const struct sched_class
*prev_class
,
1775 int oldprio
, int running
)
1777 if (prev_class
!= p
->sched_class
) {
1778 if (prev_class
->switched_from
)
1779 prev_class
->switched_from(rq
, p
, running
);
1780 p
->sched_class
->switched_to(rq
, p
, running
);
1782 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1787 /* Used instead of source_load when we know the type == 0 */
1788 static unsigned long weighted_cpuload(const int cpu
)
1790 return cpu_rq(cpu
)->load
.weight
;
1794 * Is this task likely cache-hot:
1797 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1802 * Buddy candidates are cache hot:
1804 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
1807 if (p
->sched_class
!= &fair_sched_class
)
1810 if (sysctl_sched_migration_cost
== -1)
1812 if (sysctl_sched_migration_cost
== 0)
1815 delta
= now
- p
->se
.exec_start
;
1817 return delta
< (s64
)sysctl_sched_migration_cost
;
1821 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1823 int old_cpu
= task_cpu(p
);
1824 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1825 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1826 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1829 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1831 #ifdef CONFIG_SCHEDSTATS
1832 if (p
->se
.wait_start
)
1833 p
->se
.wait_start
-= clock_offset
;
1834 if (p
->se
.sleep_start
)
1835 p
->se
.sleep_start
-= clock_offset
;
1836 if (p
->se
.block_start
)
1837 p
->se
.block_start
-= clock_offset
;
1838 if (old_cpu
!= new_cpu
) {
1839 schedstat_inc(p
, se
.nr_migrations
);
1840 if (task_hot(p
, old_rq
->clock
, NULL
))
1841 schedstat_inc(p
, se
.nr_forced2_migrations
);
1844 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1845 new_cfsrq
->min_vruntime
;
1847 __set_task_cpu(p
, new_cpu
);
1850 struct migration_req
{
1851 struct list_head list
;
1853 struct task_struct
*task
;
1856 struct completion done
;
1860 * The task's runqueue lock must be held.
1861 * Returns true if you have to wait for migration thread.
1864 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1866 struct rq
*rq
= task_rq(p
);
1869 * If the task is not on a runqueue (and not running), then
1870 * it is sufficient to simply update the task's cpu field.
1872 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1873 set_task_cpu(p
, dest_cpu
);
1877 init_completion(&req
->done
);
1879 req
->dest_cpu
= dest_cpu
;
1880 list_add(&req
->list
, &rq
->migration_queue
);
1886 * wait_task_inactive - wait for a thread to unschedule.
1888 * The caller must ensure that the task *will* unschedule sometime soon,
1889 * else this function might spin for a *long* time. This function can't
1890 * be called with interrupts off, or it may introduce deadlock with
1891 * smp_call_function() if an IPI is sent by the same process we are
1892 * waiting to become inactive.
1894 void wait_task_inactive(struct task_struct
*p
)
1896 unsigned long flags
;
1902 * We do the initial early heuristics without holding
1903 * any task-queue locks at all. We'll only try to get
1904 * the runqueue lock when things look like they will
1910 * If the task is actively running on another CPU
1911 * still, just relax and busy-wait without holding
1914 * NOTE! Since we don't hold any locks, it's not
1915 * even sure that "rq" stays as the right runqueue!
1916 * But we don't care, since "task_running()" will
1917 * return false if the runqueue has changed and p
1918 * is actually now running somewhere else!
1920 while (task_running(rq
, p
))
1924 * Ok, time to look more closely! We need the rq
1925 * lock now, to be *sure*. If we're wrong, we'll
1926 * just go back and repeat.
1928 rq
= task_rq_lock(p
, &flags
);
1929 running
= task_running(rq
, p
);
1930 on_rq
= p
->se
.on_rq
;
1931 task_rq_unlock(rq
, &flags
);
1934 * Was it really running after all now that we
1935 * checked with the proper locks actually held?
1937 * Oops. Go back and try again..
1939 if (unlikely(running
)) {
1945 * It's not enough that it's not actively running,
1946 * it must be off the runqueue _entirely_, and not
1949 * So if it wa still runnable (but just not actively
1950 * running right now), it's preempted, and we should
1951 * yield - it could be a while.
1953 if (unlikely(on_rq
)) {
1954 schedule_timeout_uninterruptible(1);
1959 * Ahh, all good. It wasn't running, and it wasn't
1960 * runnable, which means that it will never become
1961 * running in the future either. We're all done!
1968 * kick_process - kick a running thread to enter/exit the kernel
1969 * @p: the to-be-kicked thread
1971 * Cause a process which is running on another CPU to enter
1972 * kernel-mode, without any delay. (to get signals handled.)
1974 * NOTE: this function doesnt have to take the runqueue lock,
1975 * because all it wants to ensure is that the remote task enters
1976 * the kernel. If the IPI races and the task has been migrated
1977 * to another CPU then no harm is done and the purpose has been
1980 void kick_process(struct task_struct
*p
)
1986 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1987 smp_send_reschedule(cpu
);
1992 * Return a low guess at the load of a migration-source cpu weighted
1993 * according to the scheduling class and "nice" value.
1995 * We want to under-estimate the load of migration sources, to
1996 * balance conservatively.
1998 static unsigned long source_load(int cpu
, int type
)
2000 struct rq
*rq
= cpu_rq(cpu
);
2001 unsigned long total
= weighted_cpuload(cpu
);
2006 return min(rq
->cpu_load
[type
-1], total
);
2010 * Return a high guess at the load of a migration-target cpu weighted
2011 * according to the scheduling class and "nice" value.
2013 static unsigned long target_load(int cpu
, int type
)
2015 struct rq
*rq
= cpu_rq(cpu
);
2016 unsigned long total
= weighted_cpuload(cpu
);
2021 return max(rq
->cpu_load
[type
-1], total
);
2025 * find_idlest_group finds and returns the least busy CPU group within the
2028 static struct sched_group
*
2029 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2031 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2032 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2033 int load_idx
= sd
->forkexec_idx
;
2034 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2037 unsigned long load
, avg_load
;
2041 /* Skip over this group if it has no CPUs allowed */
2042 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2045 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2047 /* Tally up the load of all CPUs in the group */
2050 for_each_cpu_mask(i
, group
->cpumask
) {
2051 /* Bias balancing toward cpus of our domain */
2053 load
= source_load(i
, load_idx
);
2055 load
= target_load(i
, load_idx
);
2060 /* Adjust by relative CPU power of the group */
2061 avg_load
= sg_div_cpu_power(group
,
2062 avg_load
* SCHED_LOAD_SCALE
);
2065 this_load
= avg_load
;
2067 } else if (avg_load
< min_load
) {
2068 min_load
= avg_load
;
2071 } while (group
= group
->next
, group
!= sd
->groups
);
2073 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2079 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2082 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2085 unsigned long load
, min_load
= ULONG_MAX
;
2089 /* Traverse only the allowed CPUs */
2090 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2092 for_each_cpu_mask(i
, *tmp
) {
2093 load
= weighted_cpuload(i
);
2095 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2105 * sched_balance_self: balance the current task (running on cpu) in domains
2106 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2109 * Balance, ie. select the least loaded group.
2111 * Returns the target CPU number, or the same CPU if no balancing is needed.
2113 * preempt must be disabled.
2115 static int sched_balance_self(int cpu
, int flag
)
2117 struct task_struct
*t
= current
;
2118 struct sched_domain
*tmp
, *sd
= NULL
;
2120 for_each_domain(cpu
, tmp
) {
2122 * If power savings logic is enabled for a domain, stop there.
2124 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2126 if (tmp
->flags
& flag
)
2134 cpumask_t span
, tmpmask
;
2135 struct sched_group
*group
;
2136 int new_cpu
, weight
;
2138 if (!(sd
->flags
& flag
)) {
2144 group
= find_idlest_group(sd
, t
, cpu
);
2150 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2151 if (new_cpu
== -1 || new_cpu
== cpu
) {
2152 /* Now try balancing at a lower domain level of cpu */
2157 /* Now try balancing at a lower domain level of new_cpu */
2160 weight
= cpus_weight(span
);
2161 for_each_domain(cpu
, tmp
) {
2162 if (weight
<= cpus_weight(tmp
->span
))
2164 if (tmp
->flags
& flag
)
2167 /* while loop will break here if sd == NULL */
2173 #endif /* CONFIG_SMP */
2176 * try_to_wake_up - wake up a thread
2177 * @p: the to-be-woken-up thread
2178 * @state: the mask of task states that can be woken
2179 * @sync: do a synchronous wakeup?
2181 * Put it on the run-queue if it's not already there. The "current"
2182 * thread is always on the run-queue (except when the actual
2183 * re-schedule is in progress), and as such you're allowed to do
2184 * the simpler "current->state = TASK_RUNNING" to mark yourself
2185 * runnable without the overhead of this.
2187 * returns failure only if the task is already active.
2189 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2191 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2192 unsigned long flags
;
2196 if (!sched_feat(SYNC_WAKEUPS
))
2200 rq
= task_rq_lock(p
, &flags
);
2201 old_state
= p
->state
;
2202 if (!(old_state
& state
))
2210 this_cpu
= smp_processor_id();
2213 if (unlikely(task_running(rq
, p
)))
2216 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2217 if (cpu
!= orig_cpu
) {
2218 set_task_cpu(p
, cpu
);
2219 task_rq_unlock(rq
, &flags
);
2220 /* might preempt at this point */
2221 rq
= task_rq_lock(p
, &flags
);
2222 old_state
= p
->state
;
2223 if (!(old_state
& state
))
2228 this_cpu
= smp_processor_id();
2232 #ifdef CONFIG_SCHEDSTATS
2233 schedstat_inc(rq
, ttwu_count
);
2234 if (cpu
== this_cpu
)
2235 schedstat_inc(rq
, ttwu_local
);
2237 struct sched_domain
*sd
;
2238 for_each_domain(this_cpu
, sd
) {
2239 if (cpu_isset(cpu
, sd
->span
)) {
2240 schedstat_inc(sd
, ttwu_wake_remote
);
2245 #endif /* CONFIG_SCHEDSTATS */
2248 #endif /* CONFIG_SMP */
2249 schedstat_inc(p
, se
.nr_wakeups
);
2251 schedstat_inc(p
, se
.nr_wakeups_sync
);
2252 if (orig_cpu
!= cpu
)
2253 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2254 if (cpu
== this_cpu
)
2255 schedstat_inc(p
, se
.nr_wakeups_local
);
2257 schedstat_inc(p
, se
.nr_wakeups_remote
);
2258 update_rq_clock(rq
);
2259 activate_task(rq
, p
, 1);
2263 check_preempt_curr(rq
, p
);
2265 p
->state
= TASK_RUNNING
;
2267 if (p
->sched_class
->task_wake_up
)
2268 p
->sched_class
->task_wake_up(rq
, p
);
2271 task_rq_unlock(rq
, &flags
);
2276 int wake_up_process(struct task_struct
*p
)
2278 return try_to_wake_up(p
, TASK_ALL
, 0);
2280 EXPORT_SYMBOL(wake_up_process
);
2282 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2284 return try_to_wake_up(p
, state
, 0);
2288 * Perform scheduler related setup for a newly forked process p.
2289 * p is forked by current.
2291 * __sched_fork() is basic setup used by init_idle() too:
2293 static void __sched_fork(struct task_struct
*p
)
2295 p
->se
.exec_start
= 0;
2296 p
->se
.sum_exec_runtime
= 0;
2297 p
->se
.prev_sum_exec_runtime
= 0;
2298 p
->se
.last_wakeup
= 0;
2299 p
->se
.avg_overlap
= 0;
2301 #ifdef CONFIG_SCHEDSTATS
2302 p
->se
.wait_start
= 0;
2303 p
->se
.sum_sleep_runtime
= 0;
2304 p
->se
.sleep_start
= 0;
2305 p
->se
.block_start
= 0;
2306 p
->se
.sleep_max
= 0;
2307 p
->se
.block_max
= 0;
2309 p
->se
.slice_max
= 0;
2313 INIT_LIST_HEAD(&p
->rt
.run_list
);
2315 INIT_LIST_HEAD(&p
->se
.group_node
);
2317 #ifdef CONFIG_PREEMPT_NOTIFIERS
2318 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2322 * We mark the process as running here, but have not actually
2323 * inserted it onto the runqueue yet. This guarantees that
2324 * nobody will actually run it, and a signal or other external
2325 * event cannot wake it up and insert it on the runqueue either.
2327 p
->state
= TASK_RUNNING
;
2331 * fork()/clone()-time setup:
2333 void sched_fork(struct task_struct
*p
, int clone_flags
)
2335 int cpu
= get_cpu();
2340 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2342 set_task_cpu(p
, cpu
);
2345 * Make sure we do not leak PI boosting priority to the child:
2347 p
->prio
= current
->normal_prio
;
2348 if (!rt_prio(p
->prio
))
2349 p
->sched_class
= &fair_sched_class
;
2351 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2352 if (likely(sched_info_on()))
2353 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2355 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2358 #ifdef CONFIG_PREEMPT
2359 /* Want to start with kernel preemption disabled. */
2360 task_thread_info(p
)->preempt_count
= 1;
2366 * wake_up_new_task - wake up a newly created task for the first time.
2368 * This function will do some initial scheduler statistics housekeeping
2369 * that must be done for every newly created context, then puts the task
2370 * on the runqueue and wakes it.
2372 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2374 unsigned long flags
;
2377 rq
= task_rq_lock(p
, &flags
);
2378 BUG_ON(p
->state
!= TASK_RUNNING
);
2379 update_rq_clock(rq
);
2381 p
->prio
= effective_prio(p
);
2383 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2384 activate_task(rq
, p
, 0);
2387 * Let the scheduling class do new task startup
2388 * management (if any):
2390 p
->sched_class
->task_new(rq
, p
);
2393 check_preempt_curr(rq
, p
);
2395 if (p
->sched_class
->task_wake_up
)
2396 p
->sched_class
->task_wake_up(rq
, p
);
2398 task_rq_unlock(rq
, &flags
);
2401 #ifdef CONFIG_PREEMPT_NOTIFIERS
2404 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2405 * @notifier: notifier struct to register
2407 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2409 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2411 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2414 * preempt_notifier_unregister - no longer interested in preemption notifications
2415 * @notifier: notifier struct to unregister
2417 * This is safe to call from within a preemption notifier.
2419 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2421 hlist_del(¬ifier
->link
);
2423 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2425 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2427 struct preempt_notifier
*notifier
;
2428 struct hlist_node
*node
;
2430 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2431 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2435 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2436 struct task_struct
*next
)
2438 struct preempt_notifier
*notifier
;
2439 struct hlist_node
*node
;
2441 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2442 notifier
->ops
->sched_out(notifier
, next
);
2445 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2447 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2452 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2453 struct task_struct
*next
)
2457 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2460 * prepare_task_switch - prepare to switch tasks
2461 * @rq: the runqueue preparing to switch
2462 * @prev: the current task that is being switched out
2463 * @next: the task we are going to switch to.
2465 * This is called with the rq lock held and interrupts off. It must
2466 * be paired with a subsequent finish_task_switch after the context
2469 * prepare_task_switch sets up locking and calls architecture specific
2473 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2474 struct task_struct
*next
)
2476 fire_sched_out_preempt_notifiers(prev
, next
);
2477 prepare_lock_switch(rq
, next
);
2478 prepare_arch_switch(next
);
2482 * finish_task_switch - clean up after a task-switch
2483 * @rq: runqueue associated with task-switch
2484 * @prev: the thread we just switched away from.
2486 * finish_task_switch must be called after the context switch, paired
2487 * with a prepare_task_switch call before the context switch.
2488 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2489 * and do any other architecture-specific cleanup actions.
2491 * Note that we may have delayed dropping an mm in context_switch(). If
2492 * so, we finish that here outside of the runqueue lock. (Doing it
2493 * with the lock held can cause deadlocks; see schedule() for
2496 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2497 __releases(rq
->lock
)
2499 struct mm_struct
*mm
= rq
->prev_mm
;
2505 * A task struct has one reference for the use as "current".
2506 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2507 * schedule one last time. The schedule call will never return, and
2508 * the scheduled task must drop that reference.
2509 * The test for TASK_DEAD must occur while the runqueue locks are
2510 * still held, otherwise prev could be scheduled on another cpu, die
2511 * there before we look at prev->state, and then the reference would
2513 * Manfred Spraul <manfred@colorfullife.com>
2515 prev_state
= prev
->state
;
2516 finish_arch_switch(prev
);
2517 finish_lock_switch(rq
, prev
);
2519 if (current
->sched_class
->post_schedule
)
2520 current
->sched_class
->post_schedule(rq
);
2523 fire_sched_in_preempt_notifiers(current
);
2526 if (unlikely(prev_state
== TASK_DEAD
)) {
2528 * Remove function-return probe instances associated with this
2529 * task and put them back on the free list.
2531 kprobe_flush_task(prev
);
2532 put_task_struct(prev
);
2537 * schedule_tail - first thing a freshly forked thread must call.
2538 * @prev: the thread we just switched away from.
2540 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2541 __releases(rq
->lock
)
2543 struct rq
*rq
= this_rq();
2545 finish_task_switch(rq
, prev
);
2546 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2547 /* In this case, finish_task_switch does not reenable preemption */
2550 if (current
->set_child_tid
)
2551 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2555 * context_switch - switch to the new MM and the new
2556 * thread's register state.
2559 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2560 struct task_struct
*next
)
2562 struct mm_struct
*mm
, *oldmm
;
2564 prepare_task_switch(rq
, prev
, next
);
2566 oldmm
= prev
->active_mm
;
2568 * For paravirt, this is coupled with an exit in switch_to to
2569 * combine the page table reload and the switch backend into
2572 arch_enter_lazy_cpu_mode();
2574 if (unlikely(!mm
)) {
2575 next
->active_mm
= oldmm
;
2576 atomic_inc(&oldmm
->mm_count
);
2577 enter_lazy_tlb(oldmm
, next
);
2579 switch_mm(oldmm
, mm
, next
);
2581 if (unlikely(!prev
->mm
)) {
2582 prev
->active_mm
= NULL
;
2583 rq
->prev_mm
= oldmm
;
2586 * Since the runqueue lock will be released by the next
2587 * task (which is an invalid locking op but in the case
2588 * of the scheduler it's an obvious special-case), so we
2589 * do an early lockdep release here:
2591 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2592 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2595 /* Here we just switch the register state and the stack. */
2596 switch_to(prev
, next
, prev
);
2600 * this_rq must be evaluated again because prev may have moved
2601 * CPUs since it called schedule(), thus the 'rq' on its stack
2602 * frame will be invalid.
2604 finish_task_switch(this_rq(), prev
);
2608 * nr_running, nr_uninterruptible and nr_context_switches:
2610 * externally visible scheduler statistics: current number of runnable
2611 * threads, current number of uninterruptible-sleeping threads, total
2612 * number of context switches performed since bootup.
2614 unsigned long nr_running(void)
2616 unsigned long i
, sum
= 0;
2618 for_each_online_cpu(i
)
2619 sum
+= cpu_rq(i
)->nr_running
;
2624 unsigned long nr_uninterruptible(void)
2626 unsigned long i
, sum
= 0;
2628 for_each_possible_cpu(i
)
2629 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2632 * Since we read the counters lockless, it might be slightly
2633 * inaccurate. Do not allow it to go below zero though:
2635 if (unlikely((long)sum
< 0))
2641 unsigned long long nr_context_switches(void)
2644 unsigned long long sum
= 0;
2646 for_each_possible_cpu(i
)
2647 sum
+= cpu_rq(i
)->nr_switches
;
2652 unsigned long nr_iowait(void)
2654 unsigned long i
, sum
= 0;
2656 for_each_possible_cpu(i
)
2657 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2662 unsigned long nr_active(void)
2664 unsigned long i
, running
= 0, uninterruptible
= 0;
2666 for_each_online_cpu(i
) {
2667 running
+= cpu_rq(i
)->nr_running
;
2668 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2671 if (unlikely((long)uninterruptible
< 0))
2672 uninterruptible
= 0;
2674 return running
+ uninterruptible
;
2678 * Update rq->cpu_load[] statistics. This function is usually called every
2679 * scheduler tick (TICK_NSEC).
2681 static void update_cpu_load(struct rq
*this_rq
)
2683 unsigned long this_load
= this_rq
->load
.weight
;
2686 this_rq
->nr_load_updates
++;
2688 /* Update our load: */
2689 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2690 unsigned long old_load
, new_load
;
2692 /* scale is effectively 1 << i now, and >> i divides by scale */
2694 old_load
= this_rq
->cpu_load
[i
];
2695 new_load
= this_load
;
2697 * Round up the averaging division if load is increasing. This
2698 * prevents us from getting stuck on 9 if the load is 10, for
2701 if (new_load
> old_load
)
2702 new_load
+= scale
-1;
2703 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2710 * double_rq_lock - safely lock two runqueues
2712 * Note this does not disable interrupts like task_rq_lock,
2713 * you need to do so manually before calling.
2715 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2716 __acquires(rq1
->lock
)
2717 __acquires(rq2
->lock
)
2719 BUG_ON(!irqs_disabled());
2721 spin_lock(&rq1
->lock
);
2722 __acquire(rq2
->lock
); /* Fake it out ;) */
2725 spin_lock(&rq1
->lock
);
2726 spin_lock(&rq2
->lock
);
2728 spin_lock(&rq2
->lock
);
2729 spin_lock(&rq1
->lock
);
2732 update_rq_clock(rq1
);
2733 update_rq_clock(rq2
);
2737 * double_rq_unlock - safely unlock two runqueues
2739 * Note this does not restore interrupts like task_rq_unlock,
2740 * you need to do so manually after calling.
2742 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2743 __releases(rq1
->lock
)
2744 __releases(rq2
->lock
)
2746 spin_unlock(&rq1
->lock
);
2748 spin_unlock(&rq2
->lock
);
2750 __release(rq2
->lock
);
2754 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2756 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2757 __releases(this_rq
->lock
)
2758 __acquires(busiest
->lock
)
2759 __acquires(this_rq
->lock
)
2763 if (unlikely(!irqs_disabled())) {
2764 /* printk() doesn't work good under rq->lock */
2765 spin_unlock(&this_rq
->lock
);
2768 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2769 if (busiest
< this_rq
) {
2770 spin_unlock(&this_rq
->lock
);
2771 spin_lock(&busiest
->lock
);
2772 spin_lock(&this_rq
->lock
);
2775 spin_lock(&busiest
->lock
);
2781 * If dest_cpu is allowed for this process, migrate the task to it.
2782 * This is accomplished by forcing the cpu_allowed mask to only
2783 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2784 * the cpu_allowed mask is restored.
2786 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2788 struct migration_req req
;
2789 unsigned long flags
;
2792 rq
= task_rq_lock(p
, &flags
);
2793 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2794 || unlikely(cpu_is_offline(dest_cpu
)))
2797 /* force the process onto the specified CPU */
2798 if (migrate_task(p
, dest_cpu
, &req
)) {
2799 /* Need to wait for migration thread (might exit: take ref). */
2800 struct task_struct
*mt
= rq
->migration_thread
;
2802 get_task_struct(mt
);
2803 task_rq_unlock(rq
, &flags
);
2804 wake_up_process(mt
);
2805 put_task_struct(mt
);
2806 wait_for_completion(&req
.done
);
2811 task_rq_unlock(rq
, &flags
);
2815 * sched_exec - execve() is a valuable balancing opportunity, because at
2816 * this point the task has the smallest effective memory and cache footprint.
2818 void sched_exec(void)
2820 int new_cpu
, this_cpu
= get_cpu();
2821 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2823 if (new_cpu
!= this_cpu
)
2824 sched_migrate_task(current
, new_cpu
);
2828 * pull_task - move a task from a remote runqueue to the local runqueue.
2829 * Both runqueues must be locked.
2831 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2832 struct rq
*this_rq
, int this_cpu
)
2834 deactivate_task(src_rq
, p
, 0);
2835 set_task_cpu(p
, this_cpu
);
2836 activate_task(this_rq
, p
, 0);
2838 * Note that idle threads have a prio of MAX_PRIO, for this test
2839 * to be always true for them.
2841 check_preempt_curr(this_rq
, p
);
2845 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2848 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2849 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2853 * We do not migrate tasks that are:
2854 * 1) running (obviously), or
2855 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2856 * 3) are cache-hot on their current CPU.
2858 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2859 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2864 if (task_running(rq
, p
)) {
2865 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2870 * Aggressive migration if:
2871 * 1) task is cache cold, or
2872 * 2) too many balance attempts have failed.
2875 if (!task_hot(p
, rq
->clock
, sd
) ||
2876 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2877 #ifdef CONFIG_SCHEDSTATS
2878 if (task_hot(p
, rq
->clock
, sd
)) {
2879 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2880 schedstat_inc(p
, se
.nr_forced_migrations
);
2886 if (task_hot(p
, rq
->clock
, sd
)) {
2887 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2893 static unsigned long
2894 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2895 unsigned long max_load_move
, struct sched_domain
*sd
,
2896 enum cpu_idle_type idle
, int *all_pinned
,
2897 int *this_best_prio
, struct rq_iterator
*iterator
)
2899 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2900 struct task_struct
*p
;
2901 long rem_load_move
= max_load_move
;
2903 if (max_load_move
== 0)
2909 * Start the load-balancing iterator:
2911 p
= iterator
->start(iterator
->arg
);
2913 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2916 * To help distribute high priority tasks across CPUs we don't
2917 * skip a task if it will be the highest priority task (i.e. smallest
2918 * prio value) on its new queue regardless of its load weight
2920 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2921 SCHED_LOAD_SCALE_FUZZ
;
2922 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2923 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2924 p
= iterator
->next(iterator
->arg
);
2928 pull_task(busiest
, p
, this_rq
, this_cpu
);
2930 rem_load_move
-= p
->se
.load
.weight
;
2933 * We only want to steal up to the prescribed amount of weighted load.
2935 if (rem_load_move
> 0) {
2936 if (p
->prio
< *this_best_prio
)
2937 *this_best_prio
= p
->prio
;
2938 p
= iterator
->next(iterator
->arg
);
2943 * Right now, this is one of only two places pull_task() is called,
2944 * so we can safely collect pull_task() stats here rather than
2945 * inside pull_task().
2947 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2950 *all_pinned
= pinned
;
2952 return max_load_move
- rem_load_move
;
2956 * move_tasks tries to move up to max_load_move weighted load from busiest to
2957 * this_rq, as part of a balancing operation within domain "sd".
2958 * Returns 1 if successful and 0 otherwise.
2960 * Called with both runqueues locked.
2962 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2963 unsigned long max_load_move
,
2964 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2967 const struct sched_class
*class = sched_class_highest
;
2968 unsigned long total_load_moved
= 0;
2969 int this_best_prio
= this_rq
->curr
->prio
;
2973 class->load_balance(this_rq
, this_cpu
, busiest
,
2974 max_load_move
- total_load_moved
,
2975 sd
, idle
, all_pinned
, &this_best_prio
);
2976 class = class->next
;
2977 } while (class && max_load_move
> total_load_moved
);
2979 return total_load_moved
> 0;
2983 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2984 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2985 struct rq_iterator
*iterator
)
2987 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2991 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2992 pull_task(busiest
, p
, this_rq
, this_cpu
);
2994 * Right now, this is only the second place pull_task()
2995 * is called, so we can safely collect pull_task()
2996 * stats here rather than inside pull_task().
2998 schedstat_inc(sd
, lb_gained
[idle
]);
3002 p
= iterator
->next(iterator
->arg
);
3009 * move_one_task tries to move exactly one task from busiest to this_rq, as
3010 * part of active balancing operations within "domain".
3011 * Returns 1 if successful and 0 otherwise.
3013 * Called with both runqueues locked.
3015 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3016 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3018 const struct sched_class
*class;
3020 for (class = sched_class_highest
; class; class = class->next
)
3021 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3028 * find_busiest_group finds and returns the busiest CPU group within the
3029 * domain. It calculates and returns the amount of weighted load which
3030 * should be moved to restore balance via the imbalance parameter.
3032 static struct sched_group
*
3033 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3034 unsigned long *imbalance
, enum cpu_idle_type idle
,
3035 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3037 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3038 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3039 unsigned long max_pull
;
3040 unsigned long busiest_load_per_task
, busiest_nr_running
;
3041 unsigned long this_load_per_task
, this_nr_running
;
3042 int load_idx
, group_imb
= 0;
3043 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3044 int power_savings_balance
= 1;
3045 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3046 unsigned long min_nr_running
= ULONG_MAX
;
3047 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3050 max_load
= this_load
= total_load
= total_pwr
= 0;
3051 busiest_load_per_task
= busiest_nr_running
= 0;
3052 this_load_per_task
= this_nr_running
= 0;
3053 if (idle
== CPU_NOT_IDLE
)
3054 load_idx
= sd
->busy_idx
;
3055 else if (idle
== CPU_NEWLY_IDLE
)
3056 load_idx
= sd
->newidle_idx
;
3058 load_idx
= sd
->idle_idx
;
3061 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3064 int __group_imb
= 0;
3065 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3066 unsigned long sum_nr_running
, sum_weighted_load
;
3068 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3071 balance_cpu
= first_cpu(group
->cpumask
);
3073 /* Tally up the load of all CPUs in the group */
3074 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3076 min_cpu_load
= ~0UL;
3078 for_each_cpu_mask(i
, group
->cpumask
) {
3081 if (!cpu_isset(i
, *cpus
))
3086 if (*sd_idle
&& rq
->nr_running
)
3089 /* Bias balancing toward cpus of our domain */
3091 if (idle_cpu(i
) && !first_idle_cpu
) {
3096 load
= target_load(i
, load_idx
);
3098 load
= source_load(i
, load_idx
);
3099 if (load
> max_cpu_load
)
3100 max_cpu_load
= load
;
3101 if (min_cpu_load
> load
)
3102 min_cpu_load
= load
;
3106 sum_nr_running
+= rq
->nr_running
;
3107 sum_weighted_load
+= weighted_cpuload(i
);
3111 * First idle cpu or the first cpu(busiest) in this sched group
3112 * is eligible for doing load balancing at this and above
3113 * domains. In the newly idle case, we will allow all the cpu's
3114 * to do the newly idle load balance.
3116 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3117 balance_cpu
!= this_cpu
&& balance
) {
3122 total_load
+= avg_load
;
3123 total_pwr
+= group
->__cpu_power
;
3125 /* Adjust by relative CPU power of the group */
3126 avg_load
= sg_div_cpu_power(group
,
3127 avg_load
* SCHED_LOAD_SCALE
);
3129 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
3132 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3135 this_load
= avg_load
;
3137 this_nr_running
= sum_nr_running
;
3138 this_load_per_task
= sum_weighted_load
;
3139 } else if (avg_load
> max_load
&&
3140 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3141 max_load
= avg_load
;
3143 busiest_nr_running
= sum_nr_running
;
3144 busiest_load_per_task
= sum_weighted_load
;
3145 group_imb
= __group_imb
;
3148 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3150 * Busy processors will not participate in power savings
3153 if (idle
== CPU_NOT_IDLE
||
3154 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3158 * If the local group is idle or completely loaded
3159 * no need to do power savings balance at this domain
3161 if (local_group
&& (this_nr_running
>= group_capacity
||
3163 power_savings_balance
= 0;
3166 * If a group is already running at full capacity or idle,
3167 * don't include that group in power savings calculations
3169 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3174 * Calculate the group which has the least non-idle load.
3175 * This is the group from where we need to pick up the load
3178 if ((sum_nr_running
< min_nr_running
) ||
3179 (sum_nr_running
== min_nr_running
&&
3180 first_cpu(group
->cpumask
) <
3181 first_cpu(group_min
->cpumask
))) {
3183 min_nr_running
= sum_nr_running
;
3184 min_load_per_task
= sum_weighted_load
/
3189 * Calculate the group which is almost near its
3190 * capacity but still has some space to pick up some load
3191 * from other group and save more power
3193 if (sum_nr_running
<= group_capacity
- 1) {
3194 if (sum_nr_running
> leader_nr_running
||
3195 (sum_nr_running
== leader_nr_running
&&
3196 first_cpu(group
->cpumask
) >
3197 first_cpu(group_leader
->cpumask
))) {
3198 group_leader
= group
;
3199 leader_nr_running
= sum_nr_running
;
3204 group
= group
->next
;
3205 } while (group
!= sd
->groups
);
3207 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3210 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3212 if (this_load
>= avg_load
||
3213 100*max_load
<= sd
->imbalance_pct
*this_load
)
3216 busiest_load_per_task
/= busiest_nr_running
;
3218 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3221 * We're trying to get all the cpus to the average_load, so we don't
3222 * want to push ourselves above the average load, nor do we wish to
3223 * reduce the max loaded cpu below the average load, as either of these
3224 * actions would just result in more rebalancing later, and ping-pong
3225 * tasks around. Thus we look for the minimum possible imbalance.
3226 * Negative imbalances (*we* are more loaded than anyone else) will
3227 * be counted as no imbalance for these purposes -- we can't fix that
3228 * by pulling tasks to us. Be careful of negative numbers as they'll
3229 * appear as very large values with unsigned longs.
3231 if (max_load
<= busiest_load_per_task
)
3235 * In the presence of smp nice balancing, certain scenarios can have
3236 * max load less than avg load(as we skip the groups at or below
3237 * its cpu_power, while calculating max_load..)
3239 if (max_load
< avg_load
) {
3241 goto small_imbalance
;
3244 /* Don't want to pull so many tasks that a group would go idle */
3245 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3247 /* How much load to actually move to equalise the imbalance */
3248 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3249 (avg_load
- this_load
) * this->__cpu_power
)
3253 * if *imbalance is less than the average load per runnable task
3254 * there is no gaurantee that any tasks will be moved so we'll have
3255 * a think about bumping its value to force at least one task to be
3258 if (*imbalance
< busiest_load_per_task
) {
3259 unsigned long tmp
, pwr_now
, pwr_move
;
3263 pwr_move
= pwr_now
= 0;
3265 if (this_nr_running
) {
3266 this_load_per_task
/= this_nr_running
;
3267 if (busiest_load_per_task
> this_load_per_task
)
3270 this_load_per_task
= SCHED_LOAD_SCALE
;
3272 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
3273 busiest_load_per_task
* imbn
) {
3274 *imbalance
= busiest_load_per_task
;
3279 * OK, we don't have enough imbalance to justify moving tasks,
3280 * however we may be able to increase total CPU power used by
3284 pwr_now
+= busiest
->__cpu_power
*
3285 min(busiest_load_per_task
, max_load
);
3286 pwr_now
+= this->__cpu_power
*
3287 min(this_load_per_task
, this_load
);
3288 pwr_now
/= SCHED_LOAD_SCALE
;
3290 /* Amount of load we'd subtract */
3291 tmp
= sg_div_cpu_power(busiest
,
3292 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3294 pwr_move
+= busiest
->__cpu_power
*
3295 min(busiest_load_per_task
, max_load
- tmp
);
3297 /* Amount of load we'd add */
3298 if (max_load
* busiest
->__cpu_power
<
3299 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3300 tmp
= sg_div_cpu_power(this,
3301 max_load
* busiest
->__cpu_power
);
3303 tmp
= sg_div_cpu_power(this,
3304 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3305 pwr_move
+= this->__cpu_power
*
3306 min(this_load_per_task
, this_load
+ tmp
);
3307 pwr_move
/= SCHED_LOAD_SCALE
;
3309 /* Move if we gain throughput */
3310 if (pwr_move
> pwr_now
)
3311 *imbalance
= busiest_load_per_task
;
3317 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3318 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3321 if (this == group_leader
&& group_leader
!= group_min
) {
3322 *imbalance
= min_load_per_task
;
3332 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3335 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3336 unsigned long imbalance
, const cpumask_t
*cpus
)
3338 struct rq
*busiest
= NULL
, *rq
;
3339 unsigned long max_load
= 0;
3342 for_each_cpu_mask(i
, group
->cpumask
) {
3345 if (!cpu_isset(i
, *cpus
))
3349 wl
= weighted_cpuload(i
);
3351 if (rq
->nr_running
== 1 && wl
> imbalance
)
3354 if (wl
> max_load
) {
3364 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3365 * so long as it is large enough.
3367 #define MAX_PINNED_INTERVAL 512
3370 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3371 * tasks if there is an imbalance.
3373 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3374 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3375 int *balance
, cpumask_t
*cpus
)
3377 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3378 struct sched_group
*group
;
3379 unsigned long imbalance
;
3381 unsigned long flags
;
3386 * When power savings policy is enabled for the parent domain, idle
3387 * sibling can pick up load irrespective of busy siblings. In this case,
3388 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3389 * portraying it as CPU_NOT_IDLE.
3391 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3392 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3395 schedstat_inc(sd
, lb_count
[idle
]);
3399 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3406 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3410 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3412 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3416 BUG_ON(busiest
== this_rq
);
3418 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3421 if (busiest
->nr_running
> 1) {
3423 * Attempt to move tasks. If find_busiest_group has found
3424 * an imbalance but busiest->nr_running <= 1, the group is
3425 * still unbalanced. ld_moved simply stays zero, so it is
3426 * correctly treated as an imbalance.
3428 local_irq_save(flags
);
3429 double_rq_lock(this_rq
, busiest
);
3430 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3431 imbalance
, sd
, idle
, &all_pinned
);
3432 double_rq_unlock(this_rq
, busiest
);
3433 local_irq_restore(flags
);
3436 * some other cpu did the load balance for us.
3438 if (ld_moved
&& this_cpu
!= smp_processor_id())
3439 resched_cpu(this_cpu
);
3441 /* All tasks on this runqueue were pinned by CPU affinity */
3442 if (unlikely(all_pinned
)) {
3443 cpu_clear(cpu_of(busiest
), *cpus
);
3444 if (!cpus_empty(*cpus
))
3451 schedstat_inc(sd
, lb_failed
[idle
]);
3452 sd
->nr_balance_failed
++;
3454 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3456 spin_lock_irqsave(&busiest
->lock
, flags
);
3458 /* don't kick the migration_thread, if the curr
3459 * task on busiest cpu can't be moved to this_cpu
3461 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3462 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3464 goto out_one_pinned
;
3467 if (!busiest
->active_balance
) {
3468 busiest
->active_balance
= 1;
3469 busiest
->push_cpu
= this_cpu
;
3472 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3474 wake_up_process(busiest
->migration_thread
);
3477 * We've kicked active balancing, reset the failure
3480 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3483 sd
->nr_balance_failed
= 0;
3485 if (likely(!active_balance
)) {
3486 /* We were unbalanced, so reset the balancing interval */
3487 sd
->balance_interval
= sd
->min_interval
;
3490 * If we've begun active balancing, start to back off. This
3491 * case may not be covered by the all_pinned logic if there
3492 * is only 1 task on the busy runqueue (because we don't call
3495 if (sd
->balance_interval
< sd
->max_interval
)
3496 sd
->balance_interval
*= 2;
3499 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3500 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3506 schedstat_inc(sd
, lb_balanced
[idle
]);
3508 sd
->nr_balance_failed
= 0;
3511 /* tune up the balancing interval */
3512 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3513 (sd
->balance_interval
< sd
->max_interval
))
3514 sd
->balance_interval
*= 2;
3516 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3517 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3528 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3529 * tasks if there is an imbalance.
3531 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3532 * this_rq is locked.
3535 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3538 struct sched_group
*group
;
3539 struct rq
*busiest
= NULL
;
3540 unsigned long imbalance
;
3548 * When power savings policy is enabled for the parent domain, idle
3549 * sibling can pick up load irrespective of busy siblings. In this case,
3550 * let the state of idle sibling percolate up as IDLE, instead of
3551 * portraying it as CPU_NOT_IDLE.
3553 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3554 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3557 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3559 update_shares_locked(this_rq
, sd
);
3560 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3561 &sd_idle
, cpus
, NULL
);
3563 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3567 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3569 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3573 BUG_ON(busiest
== this_rq
);
3575 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3578 if (busiest
->nr_running
> 1) {
3579 /* Attempt to move tasks */
3580 double_lock_balance(this_rq
, busiest
);
3581 /* this_rq->clock is already updated */
3582 update_rq_clock(busiest
);
3583 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3584 imbalance
, sd
, CPU_NEWLY_IDLE
,
3586 spin_unlock(&busiest
->lock
);
3588 if (unlikely(all_pinned
)) {
3589 cpu_clear(cpu_of(busiest
), *cpus
);
3590 if (!cpus_empty(*cpus
))
3596 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3597 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3598 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3601 sd
->nr_balance_failed
= 0;
3603 update_shares_locked(this_rq
, sd
);
3607 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3608 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3609 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3611 sd
->nr_balance_failed
= 0;
3617 * idle_balance is called by schedule() if this_cpu is about to become
3618 * idle. Attempts to pull tasks from other CPUs.
3620 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3622 struct sched_domain
*sd
;
3623 int pulled_task
= -1;
3624 unsigned long next_balance
= jiffies
+ HZ
;
3627 for_each_domain(this_cpu
, sd
) {
3628 unsigned long interval
;
3630 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3633 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3634 /* If we've pulled tasks over stop searching: */
3635 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3638 interval
= msecs_to_jiffies(sd
->balance_interval
);
3639 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3640 next_balance
= sd
->last_balance
+ interval
;
3644 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3646 * We are going idle. next_balance may be set based on
3647 * a busy processor. So reset next_balance.
3649 this_rq
->next_balance
= next_balance
;
3654 * active_load_balance is run by migration threads. It pushes running tasks
3655 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3656 * running on each physical CPU where possible, and avoids physical /
3657 * logical imbalances.
3659 * Called with busiest_rq locked.
3661 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3663 int target_cpu
= busiest_rq
->push_cpu
;
3664 struct sched_domain
*sd
;
3665 struct rq
*target_rq
;
3667 /* Is there any task to move? */
3668 if (busiest_rq
->nr_running
<= 1)
3671 target_rq
= cpu_rq(target_cpu
);
3674 * This condition is "impossible", if it occurs
3675 * we need to fix it. Originally reported by
3676 * Bjorn Helgaas on a 128-cpu setup.
3678 BUG_ON(busiest_rq
== target_rq
);
3680 /* move a task from busiest_rq to target_rq */
3681 double_lock_balance(busiest_rq
, target_rq
);
3682 update_rq_clock(busiest_rq
);
3683 update_rq_clock(target_rq
);
3685 /* Search for an sd spanning us and the target CPU. */
3686 for_each_domain(target_cpu
, sd
) {
3687 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3688 cpu_isset(busiest_cpu
, sd
->span
))
3693 schedstat_inc(sd
, alb_count
);
3695 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3697 schedstat_inc(sd
, alb_pushed
);
3699 schedstat_inc(sd
, alb_failed
);
3701 spin_unlock(&target_rq
->lock
);
3706 atomic_t load_balancer
;
3708 } nohz ____cacheline_aligned
= {
3709 .load_balancer
= ATOMIC_INIT(-1),
3710 .cpu_mask
= CPU_MASK_NONE
,
3714 * This routine will try to nominate the ilb (idle load balancing)
3715 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3716 * load balancing on behalf of all those cpus. If all the cpus in the system
3717 * go into this tickless mode, then there will be no ilb owner (as there is
3718 * no need for one) and all the cpus will sleep till the next wakeup event
3721 * For the ilb owner, tick is not stopped. And this tick will be used
3722 * for idle load balancing. ilb owner will still be part of
3725 * While stopping the tick, this cpu will become the ilb owner if there
3726 * is no other owner. And will be the owner till that cpu becomes busy
3727 * or if all cpus in the system stop their ticks at which point
3728 * there is no need for ilb owner.
3730 * When the ilb owner becomes busy, it nominates another owner, during the
3731 * next busy scheduler_tick()
3733 int select_nohz_load_balancer(int stop_tick
)
3735 int cpu
= smp_processor_id();
3738 cpu_set(cpu
, nohz
.cpu_mask
);
3739 cpu_rq(cpu
)->in_nohz_recently
= 1;
3742 * If we are going offline and still the leader, give up!
3744 if (cpu_is_offline(cpu
) &&
3745 atomic_read(&nohz
.load_balancer
) == cpu
) {
3746 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3751 /* time for ilb owner also to sleep */
3752 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3753 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3754 atomic_set(&nohz
.load_balancer
, -1);
3758 if (atomic_read(&nohz
.load_balancer
) == -1) {
3759 /* make me the ilb owner */
3760 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3762 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3765 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3768 cpu_clear(cpu
, nohz
.cpu_mask
);
3770 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3771 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3778 static DEFINE_SPINLOCK(balancing
);
3781 * It checks each scheduling domain to see if it is due to be balanced,
3782 * and initiates a balancing operation if so.
3784 * Balancing parameters are set up in arch_init_sched_domains.
3786 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3789 struct rq
*rq
= cpu_rq(cpu
);
3790 unsigned long interval
;
3791 struct sched_domain
*sd
;
3792 /* Earliest time when we have to do rebalance again */
3793 unsigned long next_balance
= jiffies
+ 60*HZ
;
3794 int update_next_balance
= 0;
3798 for_each_domain(cpu
, sd
) {
3799 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3802 interval
= sd
->balance_interval
;
3803 if (idle
!= CPU_IDLE
)
3804 interval
*= sd
->busy_factor
;
3806 /* scale ms to jiffies */
3807 interval
= msecs_to_jiffies(interval
);
3808 if (unlikely(!interval
))
3810 if (interval
> HZ
*NR_CPUS
/10)
3811 interval
= HZ
*NR_CPUS
/10;
3813 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3815 if (need_serialize
) {
3816 if (!spin_trylock(&balancing
))
3820 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3821 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3823 * We've pulled tasks over so either we're no
3824 * longer idle, or one of our SMT siblings is
3827 idle
= CPU_NOT_IDLE
;
3829 sd
->last_balance
= jiffies
;
3832 spin_unlock(&balancing
);
3834 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3835 next_balance
= sd
->last_balance
+ interval
;
3836 update_next_balance
= 1;
3840 * Stop the load balance at this level. There is another
3841 * CPU in our sched group which is doing load balancing more
3849 * next_balance will be updated only when there is a need.
3850 * When the cpu is attached to null domain for ex, it will not be
3853 if (likely(update_next_balance
))
3854 rq
->next_balance
= next_balance
;
3858 * run_rebalance_domains is triggered when needed from the scheduler tick.
3859 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3860 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3862 static void run_rebalance_domains(struct softirq_action
*h
)
3864 int this_cpu
= smp_processor_id();
3865 struct rq
*this_rq
= cpu_rq(this_cpu
);
3866 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3867 CPU_IDLE
: CPU_NOT_IDLE
;
3869 rebalance_domains(this_cpu
, idle
);
3873 * If this cpu is the owner for idle load balancing, then do the
3874 * balancing on behalf of the other idle cpus whose ticks are
3877 if (this_rq
->idle_at_tick
&&
3878 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3879 cpumask_t cpus
= nohz
.cpu_mask
;
3883 cpu_clear(this_cpu
, cpus
);
3884 for_each_cpu_mask(balance_cpu
, cpus
) {
3886 * If this cpu gets work to do, stop the load balancing
3887 * work being done for other cpus. Next load
3888 * balancing owner will pick it up.
3893 rebalance_domains(balance_cpu
, CPU_IDLE
);
3895 rq
= cpu_rq(balance_cpu
);
3896 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3897 this_rq
->next_balance
= rq
->next_balance
;
3904 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3906 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3907 * idle load balancing owner or decide to stop the periodic load balancing,
3908 * if the whole system is idle.
3910 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3914 * If we were in the nohz mode recently and busy at the current
3915 * scheduler tick, then check if we need to nominate new idle
3918 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3919 rq
->in_nohz_recently
= 0;
3921 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3922 cpu_clear(cpu
, nohz
.cpu_mask
);
3923 atomic_set(&nohz
.load_balancer
, -1);
3926 if (atomic_read(&nohz
.load_balancer
) == -1) {
3928 * simple selection for now: Nominate the
3929 * first cpu in the nohz list to be the next
3932 * TBD: Traverse the sched domains and nominate
3933 * the nearest cpu in the nohz.cpu_mask.
3935 int ilb
= first_cpu(nohz
.cpu_mask
);
3937 if (ilb
< nr_cpu_ids
)
3943 * If this cpu is idle and doing idle load balancing for all the
3944 * cpus with ticks stopped, is it time for that to stop?
3946 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3947 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3953 * If this cpu is idle and the idle load balancing is done by
3954 * someone else, then no need raise the SCHED_SOFTIRQ
3956 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3957 cpu_isset(cpu
, nohz
.cpu_mask
))
3960 if (time_after_eq(jiffies
, rq
->next_balance
))
3961 raise_softirq(SCHED_SOFTIRQ
);
3964 #else /* CONFIG_SMP */
3967 * on UP we do not need to balance between CPUs:
3969 static inline void idle_balance(int cpu
, struct rq
*rq
)
3975 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3977 EXPORT_PER_CPU_SYMBOL(kstat
);
3980 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3981 * that have not yet been banked in case the task is currently running.
3983 unsigned long long task_sched_runtime(struct task_struct
*p
)
3985 unsigned long flags
;
3989 rq
= task_rq_lock(p
, &flags
);
3990 ns
= p
->se
.sum_exec_runtime
;
3991 if (task_current(rq
, p
)) {
3992 update_rq_clock(rq
);
3993 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3994 if ((s64
)delta_exec
> 0)
3997 task_rq_unlock(rq
, &flags
);
4003 * Account user cpu time to a process.
4004 * @p: the process that the cpu time gets accounted to
4005 * @cputime: the cpu time spent in user space since the last update
4007 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4009 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4012 p
->utime
= cputime_add(p
->utime
, cputime
);
4014 /* Add user time to cpustat. */
4015 tmp
= cputime_to_cputime64(cputime
);
4016 if (TASK_NICE(p
) > 0)
4017 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4019 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4023 * Account guest cpu time to a process.
4024 * @p: the process that the cpu time gets accounted to
4025 * @cputime: the cpu time spent in virtual machine since the last update
4027 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4030 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4032 tmp
= cputime_to_cputime64(cputime
);
4034 p
->utime
= cputime_add(p
->utime
, cputime
);
4035 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4037 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4038 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4042 * Account scaled user cpu time to a process.
4043 * @p: the process that the cpu time gets accounted to
4044 * @cputime: the cpu time spent in user space since the last update
4046 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4048 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4052 * Account system cpu time to a process.
4053 * @p: the process that the cpu time gets accounted to
4054 * @hardirq_offset: the offset to subtract from hardirq_count()
4055 * @cputime: the cpu time spent in kernel space since the last update
4057 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4060 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4061 struct rq
*rq
= this_rq();
4064 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4065 account_guest_time(p
, cputime
);
4069 p
->stime
= cputime_add(p
->stime
, cputime
);
4071 /* Add system time to cpustat. */
4072 tmp
= cputime_to_cputime64(cputime
);
4073 if (hardirq_count() - hardirq_offset
)
4074 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4075 else if (softirq_count())
4076 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4077 else if (p
!= rq
->idle
)
4078 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4079 else if (atomic_read(&rq
->nr_iowait
) > 0)
4080 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4082 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4083 /* Account for system time used */
4084 acct_update_integrals(p
);
4088 * Account scaled system cpu time to a process.
4089 * @p: the process that the cpu time gets accounted to
4090 * @hardirq_offset: the offset to subtract from hardirq_count()
4091 * @cputime: the cpu time spent in kernel space since the last update
4093 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4095 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4099 * Account for involuntary wait time.
4100 * @p: the process from which the cpu time has been stolen
4101 * @steal: the cpu time spent in involuntary wait
4103 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4105 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4106 cputime64_t tmp
= cputime_to_cputime64(steal
);
4107 struct rq
*rq
= this_rq();
4109 if (p
== rq
->idle
) {
4110 p
->stime
= cputime_add(p
->stime
, steal
);
4111 if (atomic_read(&rq
->nr_iowait
) > 0)
4112 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4114 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4116 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4120 * This function gets called by the timer code, with HZ frequency.
4121 * We call it with interrupts disabled.
4123 * It also gets called by the fork code, when changing the parent's
4126 void scheduler_tick(void)
4128 int cpu
= smp_processor_id();
4129 struct rq
*rq
= cpu_rq(cpu
);
4130 struct task_struct
*curr
= rq
->curr
;
4134 spin_lock(&rq
->lock
);
4135 update_rq_clock(rq
);
4136 update_cpu_load(rq
);
4137 curr
->sched_class
->task_tick(rq
, curr
, 0);
4138 spin_unlock(&rq
->lock
);
4141 rq
->idle_at_tick
= idle_cpu(cpu
);
4142 trigger_load_balance(rq
, cpu
);
4146 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4148 void __kprobes
add_preempt_count(int val
)
4153 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4155 preempt_count() += val
;
4157 * Spinlock count overflowing soon?
4159 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4162 EXPORT_SYMBOL(add_preempt_count
);
4164 void __kprobes
sub_preempt_count(int val
)
4169 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4172 * Is the spinlock portion underflowing?
4174 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4175 !(preempt_count() & PREEMPT_MASK
)))
4178 preempt_count() -= val
;
4180 EXPORT_SYMBOL(sub_preempt_count
);
4185 * Print scheduling while atomic bug:
4187 static noinline
void __schedule_bug(struct task_struct
*prev
)
4189 struct pt_regs
*regs
= get_irq_regs();
4191 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4192 prev
->comm
, prev
->pid
, preempt_count());
4194 debug_show_held_locks(prev
);
4196 if (irqs_disabled())
4197 print_irqtrace_events(prev
);
4206 * Various schedule()-time debugging checks and statistics:
4208 static inline void schedule_debug(struct task_struct
*prev
)
4211 * Test if we are atomic. Since do_exit() needs to call into
4212 * schedule() atomically, we ignore that path for now.
4213 * Otherwise, whine if we are scheduling when we should not be.
4215 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4216 __schedule_bug(prev
);
4218 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4220 schedstat_inc(this_rq(), sched_count
);
4221 #ifdef CONFIG_SCHEDSTATS
4222 if (unlikely(prev
->lock_depth
>= 0)) {
4223 schedstat_inc(this_rq(), bkl_count
);
4224 schedstat_inc(prev
, sched_info
.bkl_count
);
4230 * Pick up the highest-prio task:
4232 static inline struct task_struct
*
4233 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4235 const struct sched_class
*class;
4236 struct task_struct
*p
;
4239 * Optimization: we know that if all tasks are in
4240 * the fair class we can call that function directly:
4242 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4243 p
= fair_sched_class
.pick_next_task(rq
);
4248 class = sched_class_highest
;
4250 p
= class->pick_next_task(rq
);
4254 * Will never be NULL as the idle class always
4255 * returns a non-NULL p:
4257 class = class->next
;
4262 * schedule() is the main scheduler function.
4264 asmlinkage
void __sched
schedule(void)
4266 struct task_struct
*prev
, *next
;
4267 unsigned long *switch_count
;
4269 int cpu
, hrtick
= sched_feat(HRTICK
);
4273 cpu
= smp_processor_id();
4277 switch_count
= &prev
->nivcsw
;
4279 release_kernel_lock(prev
);
4280 need_resched_nonpreemptible
:
4282 schedule_debug(prev
);
4288 * Do the rq-clock update outside the rq lock:
4290 local_irq_disable();
4291 update_rq_clock(rq
);
4292 spin_lock(&rq
->lock
);
4293 clear_tsk_need_resched(prev
);
4295 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4296 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4297 prev
->state
= TASK_RUNNING
;
4299 deactivate_task(rq
, prev
, 1);
4300 switch_count
= &prev
->nvcsw
;
4304 if (prev
->sched_class
->pre_schedule
)
4305 prev
->sched_class
->pre_schedule(rq
, prev
);
4308 if (unlikely(!rq
->nr_running
))
4309 idle_balance(cpu
, rq
);
4311 prev
->sched_class
->put_prev_task(rq
, prev
);
4312 next
= pick_next_task(rq
, prev
);
4314 if (likely(prev
!= next
)) {
4315 sched_info_switch(prev
, next
);
4321 context_switch(rq
, prev
, next
); /* unlocks the rq */
4323 * the context switch might have flipped the stack from under
4324 * us, hence refresh the local variables.
4326 cpu
= smp_processor_id();
4329 spin_unlock_irq(&rq
->lock
);
4334 if (unlikely(reacquire_kernel_lock(current
) < 0))
4335 goto need_resched_nonpreemptible
;
4337 preempt_enable_no_resched();
4338 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4341 EXPORT_SYMBOL(schedule
);
4343 #ifdef CONFIG_PREEMPT
4345 * this is the entry point to schedule() from in-kernel preemption
4346 * off of preempt_enable. Kernel preemptions off return from interrupt
4347 * occur there and call schedule directly.
4349 asmlinkage
void __sched
preempt_schedule(void)
4351 struct thread_info
*ti
= current_thread_info();
4354 * If there is a non-zero preempt_count or interrupts are disabled,
4355 * we do not want to preempt the current task. Just return..
4357 if (likely(ti
->preempt_count
|| irqs_disabled()))
4361 add_preempt_count(PREEMPT_ACTIVE
);
4363 sub_preempt_count(PREEMPT_ACTIVE
);
4366 * Check again in case we missed a preemption opportunity
4367 * between schedule and now.
4370 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4372 EXPORT_SYMBOL(preempt_schedule
);
4375 * this is the entry point to schedule() from kernel preemption
4376 * off of irq context.
4377 * Note, that this is called and return with irqs disabled. This will
4378 * protect us against recursive calling from irq.
4380 asmlinkage
void __sched
preempt_schedule_irq(void)
4382 struct thread_info
*ti
= current_thread_info();
4384 /* Catch callers which need to be fixed */
4385 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4388 add_preempt_count(PREEMPT_ACTIVE
);
4391 local_irq_disable();
4392 sub_preempt_count(PREEMPT_ACTIVE
);
4395 * Check again in case we missed a preemption opportunity
4396 * between schedule and now.
4399 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4402 #endif /* CONFIG_PREEMPT */
4404 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4407 return try_to_wake_up(curr
->private, mode
, sync
);
4409 EXPORT_SYMBOL(default_wake_function
);
4412 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4413 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4414 * number) then we wake all the non-exclusive tasks and one exclusive task.
4416 * There are circumstances in which we can try to wake a task which has already
4417 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4418 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4420 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4421 int nr_exclusive
, int sync
, void *key
)
4423 wait_queue_t
*curr
, *next
;
4425 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4426 unsigned flags
= curr
->flags
;
4428 if (curr
->func(curr
, mode
, sync
, key
) &&
4429 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4435 * __wake_up - wake up threads blocked on a waitqueue.
4437 * @mode: which threads
4438 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4439 * @key: is directly passed to the wakeup function
4441 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4442 int nr_exclusive
, void *key
)
4444 unsigned long flags
;
4446 spin_lock_irqsave(&q
->lock
, flags
);
4447 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4448 spin_unlock_irqrestore(&q
->lock
, flags
);
4450 EXPORT_SYMBOL(__wake_up
);
4453 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4455 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4457 __wake_up_common(q
, mode
, 1, 0, NULL
);
4461 * __wake_up_sync - wake up threads blocked on a waitqueue.
4463 * @mode: which threads
4464 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4466 * The sync wakeup differs that the waker knows that it will schedule
4467 * away soon, so while the target thread will be woken up, it will not
4468 * be migrated to another CPU - ie. the two threads are 'synchronized'
4469 * with each other. This can prevent needless bouncing between CPUs.
4471 * On UP it can prevent extra preemption.
4474 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4476 unsigned long flags
;
4482 if (unlikely(!nr_exclusive
))
4485 spin_lock_irqsave(&q
->lock
, flags
);
4486 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4487 spin_unlock_irqrestore(&q
->lock
, flags
);
4489 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4491 void complete(struct completion
*x
)
4493 unsigned long flags
;
4495 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4497 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4498 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4500 EXPORT_SYMBOL(complete
);
4502 void complete_all(struct completion
*x
)
4504 unsigned long flags
;
4506 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4507 x
->done
+= UINT_MAX
/2;
4508 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4509 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4511 EXPORT_SYMBOL(complete_all
);
4513 static inline long __sched
4514 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4517 DECLARE_WAITQUEUE(wait
, current
);
4519 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4520 __add_wait_queue_tail(&x
->wait
, &wait
);
4522 if ((state
== TASK_INTERRUPTIBLE
&&
4523 signal_pending(current
)) ||
4524 (state
== TASK_KILLABLE
&&
4525 fatal_signal_pending(current
))) {
4526 timeout
= -ERESTARTSYS
;
4529 __set_current_state(state
);
4530 spin_unlock_irq(&x
->wait
.lock
);
4531 timeout
= schedule_timeout(timeout
);
4532 spin_lock_irq(&x
->wait
.lock
);
4533 } while (!x
->done
&& timeout
);
4534 __remove_wait_queue(&x
->wait
, &wait
);
4539 return timeout
?: 1;
4543 wait_for_common(struct completion
*x
, long timeout
, int state
)
4547 spin_lock_irq(&x
->wait
.lock
);
4548 timeout
= do_wait_for_common(x
, timeout
, state
);
4549 spin_unlock_irq(&x
->wait
.lock
);
4553 void __sched
wait_for_completion(struct completion
*x
)
4555 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4557 EXPORT_SYMBOL(wait_for_completion
);
4559 unsigned long __sched
4560 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4562 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4564 EXPORT_SYMBOL(wait_for_completion_timeout
);
4566 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4568 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4569 if (t
== -ERESTARTSYS
)
4573 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4575 unsigned long __sched
4576 wait_for_completion_interruptible_timeout(struct completion
*x
,
4577 unsigned long timeout
)
4579 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4581 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4583 int __sched
wait_for_completion_killable(struct completion
*x
)
4585 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4586 if (t
== -ERESTARTSYS
)
4590 EXPORT_SYMBOL(wait_for_completion_killable
);
4593 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4595 unsigned long flags
;
4598 init_waitqueue_entry(&wait
, current
);
4600 __set_current_state(state
);
4602 spin_lock_irqsave(&q
->lock
, flags
);
4603 __add_wait_queue(q
, &wait
);
4604 spin_unlock(&q
->lock
);
4605 timeout
= schedule_timeout(timeout
);
4606 spin_lock_irq(&q
->lock
);
4607 __remove_wait_queue(q
, &wait
);
4608 spin_unlock_irqrestore(&q
->lock
, flags
);
4613 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4615 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4617 EXPORT_SYMBOL(interruptible_sleep_on
);
4620 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4622 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4624 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4626 void __sched
sleep_on(wait_queue_head_t
*q
)
4628 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4630 EXPORT_SYMBOL(sleep_on
);
4632 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4634 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4636 EXPORT_SYMBOL(sleep_on_timeout
);
4638 #ifdef CONFIG_RT_MUTEXES
4641 * rt_mutex_setprio - set the current priority of a task
4643 * @prio: prio value (kernel-internal form)
4645 * This function changes the 'effective' priority of a task. It does
4646 * not touch ->normal_prio like __setscheduler().
4648 * Used by the rt_mutex code to implement priority inheritance logic.
4650 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4652 unsigned long flags
;
4653 int oldprio
, on_rq
, running
;
4655 const struct sched_class
*prev_class
= p
->sched_class
;
4657 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4659 rq
= task_rq_lock(p
, &flags
);
4660 update_rq_clock(rq
);
4663 on_rq
= p
->se
.on_rq
;
4664 running
= task_current(rq
, p
);
4666 dequeue_task(rq
, p
, 0);
4668 p
->sched_class
->put_prev_task(rq
, p
);
4671 p
->sched_class
= &rt_sched_class
;
4673 p
->sched_class
= &fair_sched_class
;
4678 p
->sched_class
->set_curr_task(rq
);
4680 enqueue_task(rq
, p
, 0);
4682 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4684 task_rq_unlock(rq
, &flags
);
4689 void set_user_nice(struct task_struct
*p
, long nice
)
4691 int old_prio
, delta
, on_rq
;
4692 unsigned long flags
;
4695 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4698 * We have to be careful, if called from sys_setpriority(),
4699 * the task might be in the middle of scheduling on another CPU.
4701 rq
= task_rq_lock(p
, &flags
);
4702 update_rq_clock(rq
);
4704 * The RT priorities are set via sched_setscheduler(), but we still
4705 * allow the 'normal' nice value to be set - but as expected
4706 * it wont have any effect on scheduling until the task is
4707 * SCHED_FIFO/SCHED_RR:
4709 if (task_has_rt_policy(p
)) {
4710 p
->static_prio
= NICE_TO_PRIO(nice
);
4713 on_rq
= p
->se
.on_rq
;
4715 dequeue_task(rq
, p
, 0);
4717 p
->static_prio
= NICE_TO_PRIO(nice
);
4720 p
->prio
= effective_prio(p
);
4721 delta
= p
->prio
- old_prio
;
4724 enqueue_task(rq
, p
, 0);
4726 * If the task increased its priority or is running and
4727 * lowered its priority, then reschedule its CPU:
4729 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4730 resched_task(rq
->curr
);
4733 task_rq_unlock(rq
, &flags
);
4735 EXPORT_SYMBOL(set_user_nice
);
4738 * can_nice - check if a task can reduce its nice value
4742 int can_nice(const struct task_struct
*p
, const int nice
)
4744 /* convert nice value [19,-20] to rlimit style value [1,40] */
4745 int nice_rlim
= 20 - nice
;
4747 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4748 capable(CAP_SYS_NICE
));
4751 #ifdef __ARCH_WANT_SYS_NICE
4754 * sys_nice - change the priority of the current process.
4755 * @increment: priority increment
4757 * sys_setpriority is a more generic, but much slower function that
4758 * does similar things.
4760 asmlinkage
long sys_nice(int increment
)
4765 * Setpriority might change our priority at the same moment.
4766 * We don't have to worry. Conceptually one call occurs first
4767 * and we have a single winner.
4769 if (increment
< -40)
4774 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4780 if (increment
< 0 && !can_nice(current
, nice
))
4783 retval
= security_task_setnice(current
, nice
);
4787 set_user_nice(current
, nice
);
4794 * task_prio - return the priority value of a given task.
4795 * @p: the task in question.
4797 * This is the priority value as seen by users in /proc.
4798 * RT tasks are offset by -200. Normal tasks are centered
4799 * around 0, value goes from -16 to +15.
4801 int task_prio(const struct task_struct
*p
)
4803 return p
->prio
- MAX_RT_PRIO
;
4807 * task_nice - return the nice value of a given task.
4808 * @p: the task in question.
4810 int task_nice(const struct task_struct
*p
)
4812 return TASK_NICE(p
);
4814 EXPORT_SYMBOL(task_nice
);
4817 * idle_cpu - is a given cpu idle currently?
4818 * @cpu: the processor in question.
4820 int idle_cpu(int cpu
)
4822 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4826 * idle_task - return the idle task for a given cpu.
4827 * @cpu: the processor in question.
4829 struct task_struct
*idle_task(int cpu
)
4831 return cpu_rq(cpu
)->idle
;
4835 * find_process_by_pid - find a process with a matching PID value.
4836 * @pid: the pid in question.
4838 static struct task_struct
*find_process_by_pid(pid_t pid
)
4840 return pid
? find_task_by_vpid(pid
) : current
;
4843 /* Actually do priority change: must hold rq lock. */
4845 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4847 BUG_ON(p
->se
.on_rq
);
4850 switch (p
->policy
) {
4854 p
->sched_class
= &fair_sched_class
;
4858 p
->sched_class
= &rt_sched_class
;
4862 p
->rt_priority
= prio
;
4863 p
->normal_prio
= normal_prio(p
);
4864 /* we are holding p->pi_lock already */
4865 p
->prio
= rt_mutex_getprio(p
);
4870 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4871 * @p: the task in question.
4872 * @policy: new policy.
4873 * @param: structure containing the new RT priority.
4875 * NOTE that the task may be already dead.
4877 int sched_setscheduler(struct task_struct
*p
, int policy
,
4878 struct sched_param
*param
)
4880 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4881 unsigned long flags
;
4882 const struct sched_class
*prev_class
= p
->sched_class
;
4885 /* may grab non-irq protected spin_locks */
4886 BUG_ON(in_interrupt());
4888 /* double check policy once rq lock held */
4890 policy
= oldpolicy
= p
->policy
;
4891 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4892 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4893 policy
!= SCHED_IDLE
)
4896 * Valid priorities for SCHED_FIFO and SCHED_RR are
4897 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4898 * SCHED_BATCH and SCHED_IDLE is 0.
4900 if (param
->sched_priority
< 0 ||
4901 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4902 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4904 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4908 * Allow unprivileged RT tasks to decrease priority:
4910 if (!capable(CAP_SYS_NICE
)) {
4911 if (rt_policy(policy
)) {
4912 unsigned long rlim_rtprio
;
4914 if (!lock_task_sighand(p
, &flags
))
4916 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4917 unlock_task_sighand(p
, &flags
);
4919 /* can't set/change the rt policy */
4920 if (policy
!= p
->policy
&& !rlim_rtprio
)
4923 /* can't increase priority */
4924 if (param
->sched_priority
> p
->rt_priority
&&
4925 param
->sched_priority
> rlim_rtprio
)
4929 * Like positive nice levels, dont allow tasks to
4930 * move out of SCHED_IDLE either:
4932 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4935 /* can't change other user's priorities */
4936 if ((current
->euid
!= p
->euid
) &&
4937 (current
->euid
!= p
->uid
))
4941 #ifdef CONFIG_RT_GROUP_SCHED
4943 * Do not allow realtime tasks into groups that have no runtime
4946 if (rt_policy(policy
) && task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4950 retval
= security_task_setscheduler(p
, policy
, param
);
4954 * make sure no PI-waiters arrive (or leave) while we are
4955 * changing the priority of the task:
4957 spin_lock_irqsave(&p
->pi_lock
, flags
);
4959 * To be able to change p->policy safely, the apropriate
4960 * runqueue lock must be held.
4962 rq
= __task_rq_lock(p
);
4963 /* recheck policy now with rq lock held */
4964 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4965 policy
= oldpolicy
= -1;
4966 __task_rq_unlock(rq
);
4967 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4970 update_rq_clock(rq
);
4971 on_rq
= p
->se
.on_rq
;
4972 running
= task_current(rq
, p
);
4974 deactivate_task(rq
, p
, 0);
4976 p
->sched_class
->put_prev_task(rq
, p
);
4979 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4982 p
->sched_class
->set_curr_task(rq
);
4984 activate_task(rq
, p
, 0);
4986 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4988 __task_rq_unlock(rq
);
4989 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4991 rt_mutex_adjust_pi(p
);
4995 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4998 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5000 struct sched_param lparam
;
5001 struct task_struct
*p
;
5004 if (!param
|| pid
< 0)
5006 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5011 p
= find_process_by_pid(pid
);
5013 retval
= sched_setscheduler(p
, policy
, &lparam
);
5020 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5021 * @pid: the pid in question.
5022 * @policy: new policy.
5023 * @param: structure containing the new RT priority.
5026 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5028 /* negative values for policy are not valid */
5032 return do_sched_setscheduler(pid
, policy
, param
);
5036 * sys_sched_setparam - set/change the RT priority of a thread
5037 * @pid: the pid in question.
5038 * @param: structure containing the new RT priority.
5040 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5042 return do_sched_setscheduler(pid
, -1, param
);
5046 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5047 * @pid: the pid in question.
5049 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5051 struct task_struct
*p
;
5058 read_lock(&tasklist_lock
);
5059 p
= find_process_by_pid(pid
);
5061 retval
= security_task_getscheduler(p
);
5065 read_unlock(&tasklist_lock
);
5070 * sys_sched_getscheduler - get the RT priority of a thread
5071 * @pid: the pid in question.
5072 * @param: structure containing the RT priority.
5074 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5076 struct sched_param lp
;
5077 struct task_struct
*p
;
5080 if (!param
|| pid
< 0)
5083 read_lock(&tasklist_lock
);
5084 p
= find_process_by_pid(pid
);
5089 retval
= security_task_getscheduler(p
);
5093 lp
.sched_priority
= p
->rt_priority
;
5094 read_unlock(&tasklist_lock
);
5097 * This one might sleep, we cannot do it with a spinlock held ...
5099 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5104 read_unlock(&tasklist_lock
);
5108 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5110 cpumask_t cpus_allowed
;
5111 cpumask_t new_mask
= *in_mask
;
5112 struct task_struct
*p
;
5116 read_lock(&tasklist_lock
);
5118 p
= find_process_by_pid(pid
);
5120 read_unlock(&tasklist_lock
);
5126 * It is not safe to call set_cpus_allowed with the
5127 * tasklist_lock held. We will bump the task_struct's
5128 * usage count and then drop tasklist_lock.
5131 read_unlock(&tasklist_lock
);
5134 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5135 !capable(CAP_SYS_NICE
))
5138 retval
= security_task_setscheduler(p
, 0, NULL
);
5142 cpuset_cpus_allowed(p
, &cpus_allowed
);
5143 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5145 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5148 cpuset_cpus_allowed(p
, &cpus_allowed
);
5149 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5151 * We must have raced with a concurrent cpuset
5152 * update. Just reset the cpus_allowed to the
5153 * cpuset's cpus_allowed
5155 new_mask
= cpus_allowed
;
5165 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5166 cpumask_t
*new_mask
)
5168 if (len
< sizeof(cpumask_t
)) {
5169 memset(new_mask
, 0, sizeof(cpumask_t
));
5170 } else if (len
> sizeof(cpumask_t
)) {
5171 len
= sizeof(cpumask_t
);
5173 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5177 * sys_sched_setaffinity - set the cpu affinity of a process
5178 * @pid: pid of the process
5179 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5180 * @user_mask_ptr: user-space pointer to the new cpu mask
5182 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5183 unsigned long __user
*user_mask_ptr
)
5188 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5192 return sched_setaffinity(pid
, &new_mask
);
5195 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5197 struct task_struct
*p
;
5201 read_lock(&tasklist_lock
);
5204 p
= find_process_by_pid(pid
);
5208 retval
= security_task_getscheduler(p
);
5212 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5215 read_unlock(&tasklist_lock
);
5222 * sys_sched_getaffinity - get the cpu affinity of a process
5223 * @pid: pid of the process
5224 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5225 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5227 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5228 unsigned long __user
*user_mask_ptr
)
5233 if (len
< sizeof(cpumask_t
))
5236 ret
= sched_getaffinity(pid
, &mask
);
5240 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5243 return sizeof(cpumask_t
);
5247 * sys_sched_yield - yield the current processor to other threads.
5249 * This function yields the current CPU to other tasks. If there are no
5250 * other threads running on this CPU then this function will return.
5252 asmlinkage
long sys_sched_yield(void)
5254 struct rq
*rq
= this_rq_lock();
5256 schedstat_inc(rq
, yld_count
);
5257 current
->sched_class
->yield_task(rq
);
5260 * Since we are going to call schedule() anyway, there's
5261 * no need to preempt or enable interrupts:
5263 __release(rq
->lock
);
5264 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5265 _raw_spin_unlock(&rq
->lock
);
5266 preempt_enable_no_resched();
5273 static void __cond_resched(void)
5275 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5276 __might_sleep(__FILE__
, __LINE__
);
5279 * The BKS might be reacquired before we have dropped
5280 * PREEMPT_ACTIVE, which could trigger a second
5281 * cond_resched() call.
5284 add_preempt_count(PREEMPT_ACTIVE
);
5286 sub_preempt_count(PREEMPT_ACTIVE
);
5287 } while (need_resched());
5290 int __sched
_cond_resched(void)
5292 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5293 system_state
== SYSTEM_RUNNING
) {
5299 EXPORT_SYMBOL(_cond_resched
);
5302 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5303 * call schedule, and on return reacquire the lock.
5305 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5306 * operations here to prevent schedule() from being called twice (once via
5307 * spin_unlock(), once by hand).
5309 int cond_resched_lock(spinlock_t
*lock
)
5311 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5314 if (spin_needbreak(lock
) || resched
) {
5316 if (resched
&& need_resched())
5325 EXPORT_SYMBOL(cond_resched_lock
);
5327 int __sched
cond_resched_softirq(void)
5329 BUG_ON(!in_softirq());
5331 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5339 EXPORT_SYMBOL(cond_resched_softirq
);
5342 * yield - yield the current processor to other threads.
5344 * This is a shortcut for kernel-space yielding - it marks the
5345 * thread runnable and calls sys_sched_yield().
5347 void __sched
yield(void)
5349 set_current_state(TASK_RUNNING
);
5352 EXPORT_SYMBOL(yield
);
5355 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5356 * that process accounting knows that this is a task in IO wait state.
5358 * But don't do that if it is a deliberate, throttling IO wait (this task
5359 * has set its backing_dev_info: the queue against which it should throttle)
5361 void __sched
io_schedule(void)
5363 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5365 delayacct_blkio_start();
5366 atomic_inc(&rq
->nr_iowait
);
5368 atomic_dec(&rq
->nr_iowait
);
5369 delayacct_blkio_end();
5371 EXPORT_SYMBOL(io_schedule
);
5373 long __sched
io_schedule_timeout(long timeout
)
5375 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5378 delayacct_blkio_start();
5379 atomic_inc(&rq
->nr_iowait
);
5380 ret
= schedule_timeout(timeout
);
5381 atomic_dec(&rq
->nr_iowait
);
5382 delayacct_blkio_end();
5387 * sys_sched_get_priority_max - return maximum RT priority.
5388 * @policy: scheduling class.
5390 * this syscall returns the maximum rt_priority that can be used
5391 * by a given scheduling class.
5393 asmlinkage
long sys_sched_get_priority_max(int policy
)
5400 ret
= MAX_USER_RT_PRIO
-1;
5412 * sys_sched_get_priority_min - return minimum RT priority.
5413 * @policy: scheduling class.
5415 * this syscall returns the minimum rt_priority that can be used
5416 * by a given scheduling class.
5418 asmlinkage
long sys_sched_get_priority_min(int policy
)
5436 * sys_sched_rr_get_interval - return the default timeslice of a process.
5437 * @pid: pid of the process.
5438 * @interval: userspace pointer to the timeslice value.
5440 * this syscall writes the default timeslice value of a given process
5441 * into the user-space timespec buffer. A value of '0' means infinity.
5444 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5446 struct task_struct
*p
;
5447 unsigned int time_slice
;
5455 read_lock(&tasklist_lock
);
5456 p
= find_process_by_pid(pid
);
5460 retval
= security_task_getscheduler(p
);
5465 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5466 * tasks that are on an otherwise idle runqueue:
5469 if (p
->policy
== SCHED_RR
) {
5470 time_slice
= DEF_TIMESLICE
;
5471 } else if (p
->policy
!= SCHED_FIFO
) {
5472 struct sched_entity
*se
= &p
->se
;
5473 unsigned long flags
;
5476 rq
= task_rq_lock(p
, &flags
);
5477 if (rq
->cfs
.load
.weight
)
5478 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5479 task_rq_unlock(rq
, &flags
);
5481 read_unlock(&tasklist_lock
);
5482 jiffies_to_timespec(time_slice
, &t
);
5483 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5487 read_unlock(&tasklist_lock
);
5491 static const char stat_nam
[] = "RSDTtZX";
5493 void sched_show_task(struct task_struct
*p
)
5495 unsigned long free
= 0;
5498 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5499 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5500 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5501 #if BITS_PER_LONG == 32
5502 if (state
== TASK_RUNNING
)
5503 printk(KERN_CONT
" running ");
5505 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5507 if (state
== TASK_RUNNING
)
5508 printk(KERN_CONT
" running task ");
5510 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5512 #ifdef CONFIG_DEBUG_STACK_USAGE
5514 unsigned long *n
= end_of_stack(p
);
5517 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5520 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5521 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5523 show_stack(p
, NULL
);
5526 void show_state_filter(unsigned long state_filter
)
5528 struct task_struct
*g
, *p
;
5530 #if BITS_PER_LONG == 32
5532 " task PC stack pid father\n");
5535 " task PC stack pid father\n");
5537 read_lock(&tasklist_lock
);
5538 do_each_thread(g
, p
) {
5540 * reset the NMI-timeout, listing all files on a slow
5541 * console might take alot of time:
5543 touch_nmi_watchdog();
5544 if (!state_filter
|| (p
->state
& state_filter
))
5546 } while_each_thread(g
, p
);
5548 touch_all_softlockup_watchdogs();
5550 #ifdef CONFIG_SCHED_DEBUG
5551 sysrq_sched_debug_show();
5553 read_unlock(&tasklist_lock
);
5555 * Only show locks if all tasks are dumped:
5557 if (state_filter
== -1)
5558 debug_show_all_locks();
5561 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5563 idle
->sched_class
= &idle_sched_class
;
5567 * init_idle - set up an idle thread for a given CPU
5568 * @idle: task in question
5569 * @cpu: cpu the idle task belongs to
5571 * NOTE: this function does not set the idle thread's NEED_RESCHED
5572 * flag, to make booting more robust.
5574 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5576 struct rq
*rq
= cpu_rq(cpu
);
5577 unsigned long flags
;
5580 idle
->se
.exec_start
= sched_clock();
5582 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5583 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5584 __set_task_cpu(idle
, cpu
);
5586 spin_lock_irqsave(&rq
->lock
, flags
);
5587 rq
->curr
= rq
->idle
= idle
;
5588 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5591 spin_unlock_irqrestore(&rq
->lock
, flags
);
5593 /* Set the preempt count _outside_ the spinlocks! */
5594 #if defined(CONFIG_PREEMPT)
5595 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5597 task_thread_info(idle
)->preempt_count
= 0;
5600 * The idle tasks have their own, simple scheduling class:
5602 idle
->sched_class
= &idle_sched_class
;
5606 * In a system that switches off the HZ timer nohz_cpu_mask
5607 * indicates which cpus entered this state. This is used
5608 * in the rcu update to wait only for active cpus. For system
5609 * which do not switch off the HZ timer nohz_cpu_mask should
5610 * always be CPU_MASK_NONE.
5612 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5615 * Increase the granularity value when there are more CPUs,
5616 * because with more CPUs the 'effective latency' as visible
5617 * to users decreases. But the relationship is not linear,
5618 * so pick a second-best guess by going with the log2 of the
5621 * This idea comes from the SD scheduler of Con Kolivas:
5623 static inline void sched_init_granularity(void)
5625 unsigned int factor
= 1 + ilog2(num_online_cpus());
5626 const unsigned long limit
= 200000000;
5628 sysctl_sched_min_granularity
*= factor
;
5629 if (sysctl_sched_min_granularity
> limit
)
5630 sysctl_sched_min_granularity
= limit
;
5632 sysctl_sched_latency
*= factor
;
5633 if (sysctl_sched_latency
> limit
)
5634 sysctl_sched_latency
= limit
;
5636 sysctl_sched_wakeup_granularity
*= factor
;
5641 * This is how migration works:
5643 * 1) we queue a struct migration_req structure in the source CPU's
5644 * runqueue and wake up that CPU's migration thread.
5645 * 2) we down() the locked semaphore => thread blocks.
5646 * 3) migration thread wakes up (implicitly it forces the migrated
5647 * thread off the CPU)
5648 * 4) it gets the migration request and checks whether the migrated
5649 * task is still in the wrong runqueue.
5650 * 5) if it's in the wrong runqueue then the migration thread removes
5651 * it and puts it into the right queue.
5652 * 6) migration thread up()s the semaphore.
5653 * 7) we wake up and the migration is done.
5657 * Change a given task's CPU affinity. Migrate the thread to a
5658 * proper CPU and schedule it away if the CPU it's executing on
5659 * is removed from the allowed bitmask.
5661 * NOTE: the caller must have a valid reference to the task, the
5662 * task must not exit() & deallocate itself prematurely. The
5663 * call is not atomic; no spinlocks may be held.
5665 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5667 struct migration_req req
;
5668 unsigned long flags
;
5672 rq
= task_rq_lock(p
, &flags
);
5673 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5678 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5679 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5684 if (p
->sched_class
->set_cpus_allowed
)
5685 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5687 p
->cpus_allowed
= *new_mask
;
5688 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5691 /* Can the task run on the task's current CPU? If so, we're done */
5692 if (cpu_isset(task_cpu(p
), *new_mask
))
5695 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5696 /* Need help from migration thread: drop lock and wait. */
5697 task_rq_unlock(rq
, &flags
);
5698 wake_up_process(rq
->migration_thread
);
5699 wait_for_completion(&req
.done
);
5700 tlb_migrate_finish(p
->mm
);
5704 task_rq_unlock(rq
, &flags
);
5708 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5711 * Move (not current) task off this cpu, onto dest cpu. We're doing
5712 * this because either it can't run here any more (set_cpus_allowed()
5713 * away from this CPU, or CPU going down), or because we're
5714 * attempting to rebalance this task on exec (sched_exec).
5716 * So we race with normal scheduler movements, but that's OK, as long
5717 * as the task is no longer on this CPU.
5719 * Returns non-zero if task was successfully migrated.
5721 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5723 struct rq
*rq_dest
, *rq_src
;
5726 if (unlikely(cpu_is_offline(dest_cpu
)))
5729 rq_src
= cpu_rq(src_cpu
);
5730 rq_dest
= cpu_rq(dest_cpu
);
5732 double_rq_lock(rq_src
, rq_dest
);
5733 /* Already moved. */
5734 if (task_cpu(p
) != src_cpu
)
5736 /* Affinity changed (again). */
5737 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5740 on_rq
= p
->se
.on_rq
;
5742 deactivate_task(rq_src
, p
, 0);
5744 set_task_cpu(p
, dest_cpu
);
5746 activate_task(rq_dest
, p
, 0);
5747 check_preempt_curr(rq_dest
, p
);
5751 double_rq_unlock(rq_src
, rq_dest
);
5756 * migration_thread - this is a highprio system thread that performs
5757 * thread migration by bumping thread off CPU then 'pushing' onto
5760 static int migration_thread(void *data
)
5762 int cpu
= (long)data
;
5766 BUG_ON(rq
->migration_thread
!= current
);
5768 set_current_state(TASK_INTERRUPTIBLE
);
5769 while (!kthread_should_stop()) {
5770 struct migration_req
*req
;
5771 struct list_head
*head
;
5773 spin_lock_irq(&rq
->lock
);
5775 if (cpu_is_offline(cpu
)) {
5776 spin_unlock_irq(&rq
->lock
);
5780 if (rq
->active_balance
) {
5781 active_load_balance(rq
, cpu
);
5782 rq
->active_balance
= 0;
5785 head
= &rq
->migration_queue
;
5787 if (list_empty(head
)) {
5788 spin_unlock_irq(&rq
->lock
);
5790 set_current_state(TASK_INTERRUPTIBLE
);
5793 req
= list_entry(head
->next
, struct migration_req
, list
);
5794 list_del_init(head
->next
);
5796 spin_unlock(&rq
->lock
);
5797 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5800 complete(&req
->done
);
5802 __set_current_state(TASK_RUNNING
);
5806 /* Wait for kthread_stop */
5807 set_current_state(TASK_INTERRUPTIBLE
);
5808 while (!kthread_should_stop()) {
5810 set_current_state(TASK_INTERRUPTIBLE
);
5812 __set_current_state(TASK_RUNNING
);
5816 #ifdef CONFIG_HOTPLUG_CPU
5818 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5822 local_irq_disable();
5823 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5829 * Figure out where task on dead CPU should go, use force if necessary.
5830 * NOTE: interrupts should be disabled by the caller
5832 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5834 unsigned long flags
;
5841 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5842 cpus_and(mask
, mask
, p
->cpus_allowed
);
5843 dest_cpu
= any_online_cpu(mask
);
5845 /* On any allowed CPU? */
5846 if (dest_cpu
>= nr_cpu_ids
)
5847 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5849 /* No more Mr. Nice Guy. */
5850 if (dest_cpu
>= nr_cpu_ids
) {
5851 cpumask_t cpus_allowed
;
5853 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
5855 * Try to stay on the same cpuset, where the
5856 * current cpuset may be a subset of all cpus.
5857 * The cpuset_cpus_allowed_locked() variant of
5858 * cpuset_cpus_allowed() will not block. It must be
5859 * called within calls to cpuset_lock/cpuset_unlock.
5861 rq
= task_rq_lock(p
, &flags
);
5862 p
->cpus_allowed
= cpus_allowed
;
5863 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5864 task_rq_unlock(rq
, &flags
);
5867 * Don't tell them about moving exiting tasks or
5868 * kernel threads (both mm NULL), since they never
5871 if (p
->mm
&& printk_ratelimit()) {
5872 printk(KERN_INFO
"process %d (%s) no "
5873 "longer affine to cpu%d\n",
5874 task_pid_nr(p
), p
->comm
, dead_cpu
);
5877 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5881 * While a dead CPU has no uninterruptible tasks queued at this point,
5882 * it might still have a nonzero ->nr_uninterruptible counter, because
5883 * for performance reasons the counter is not stricly tracking tasks to
5884 * their home CPUs. So we just add the counter to another CPU's counter,
5885 * to keep the global sum constant after CPU-down:
5887 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5889 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
5890 unsigned long flags
;
5892 local_irq_save(flags
);
5893 double_rq_lock(rq_src
, rq_dest
);
5894 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5895 rq_src
->nr_uninterruptible
= 0;
5896 double_rq_unlock(rq_src
, rq_dest
);
5897 local_irq_restore(flags
);
5900 /* Run through task list and migrate tasks from the dead cpu. */
5901 static void migrate_live_tasks(int src_cpu
)
5903 struct task_struct
*p
, *t
;
5905 read_lock(&tasklist_lock
);
5907 do_each_thread(t
, p
) {
5911 if (task_cpu(p
) == src_cpu
)
5912 move_task_off_dead_cpu(src_cpu
, p
);
5913 } while_each_thread(t
, p
);
5915 read_unlock(&tasklist_lock
);
5919 * Schedules idle task to be the next runnable task on current CPU.
5920 * It does so by boosting its priority to highest possible.
5921 * Used by CPU offline code.
5923 void sched_idle_next(void)
5925 int this_cpu
= smp_processor_id();
5926 struct rq
*rq
= cpu_rq(this_cpu
);
5927 struct task_struct
*p
= rq
->idle
;
5928 unsigned long flags
;
5930 /* cpu has to be offline */
5931 BUG_ON(cpu_online(this_cpu
));
5934 * Strictly not necessary since rest of the CPUs are stopped by now
5935 * and interrupts disabled on the current cpu.
5937 spin_lock_irqsave(&rq
->lock
, flags
);
5939 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5941 update_rq_clock(rq
);
5942 activate_task(rq
, p
, 0);
5944 spin_unlock_irqrestore(&rq
->lock
, flags
);
5948 * Ensures that the idle task is using init_mm right before its cpu goes
5951 void idle_task_exit(void)
5953 struct mm_struct
*mm
= current
->active_mm
;
5955 BUG_ON(cpu_online(smp_processor_id()));
5958 switch_mm(mm
, &init_mm
, current
);
5962 /* called under rq->lock with disabled interrupts */
5963 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5965 struct rq
*rq
= cpu_rq(dead_cpu
);
5967 /* Must be exiting, otherwise would be on tasklist. */
5968 BUG_ON(!p
->exit_state
);
5970 /* Cannot have done final schedule yet: would have vanished. */
5971 BUG_ON(p
->state
== TASK_DEAD
);
5976 * Drop lock around migration; if someone else moves it,
5977 * that's OK. No task can be added to this CPU, so iteration is
5980 spin_unlock_irq(&rq
->lock
);
5981 move_task_off_dead_cpu(dead_cpu
, p
);
5982 spin_lock_irq(&rq
->lock
);
5987 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5988 static void migrate_dead_tasks(unsigned int dead_cpu
)
5990 struct rq
*rq
= cpu_rq(dead_cpu
);
5991 struct task_struct
*next
;
5994 if (!rq
->nr_running
)
5996 update_rq_clock(rq
);
5997 next
= pick_next_task(rq
, rq
->curr
);
6000 migrate_dead(dead_cpu
, next
);
6004 #endif /* CONFIG_HOTPLUG_CPU */
6006 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6008 static struct ctl_table sd_ctl_dir
[] = {
6010 .procname
= "sched_domain",
6016 static struct ctl_table sd_ctl_root
[] = {
6018 .ctl_name
= CTL_KERN
,
6019 .procname
= "kernel",
6021 .child
= sd_ctl_dir
,
6026 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6028 struct ctl_table
*entry
=
6029 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6034 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6036 struct ctl_table
*entry
;
6039 * In the intermediate directories, both the child directory and
6040 * procname are dynamically allocated and could fail but the mode
6041 * will always be set. In the lowest directory the names are
6042 * static strings and all have proc handlers.
6044 for (entry
= *tablep
; entry
->mode
; entry
++) {
6046 sd_free_ctl_entry(&entry
->child
);
6047 if (entry
->proc_handler
== NULL
)
6048 kfree(entry
->procname
);
6056 set_table_entry(struct ctl_table
*entry
,
6057 const char *procname
, void *data
, int maxlen
,
6058 mode_t mode
, proc_handler
*proc_handler
)
6060 entry
->procname
= procname
;
6062 entry
->maxlen
= maxlen
;
6064 entry
->proc_handler
= proc_handler
;
6067 static struct ctl_table
*
6068 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6070 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
6075 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6076 sizeof(long), 0644, proc_doulongvec_minmax
);
6077 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6078 sizeof(long), 0644, proc_doulongvec_minmax
);
6079 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6080 sizeof(int), 0644, proc_dointvec_minmax
);
6081 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6082 sizeof(int), 0644, proc_dointvec_minmax
);
6083 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6084 sizeof(int), 0644, proc_dointvec_minmax
);
6085 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6086 sizeof(int), 0644, proc_dointvec_minmax
);
6087 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6088 sizeof(int), 0644, proc_dointvec_minmax
);
6089 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6090 sizeof(int), 0644, proc_dointvec_minmax
);
6091 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6092 sizeof(int), 0644, proc_dointvec_minmax
);
6093 set_table_entry(&table
[9], "cache_nice_tries",
6094 &sd
->cache_nice_tries
,
6095 sizeof(int), 0644, proc_dointvec_minmax
);
6096 set_table_entry(&table
[10], "flags", &sd
->flags
,
6097 sizeof(int), 0644, proc_dointvec_minmax
);
6098 /* &table[11] is terminator */
6103 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6105 struct ctl_table
*entry
, *table
;
6106 struct sched_domain
*sd
;
6107 int domain_num
= 0, i
;
6110 for_each_domain(cpu
, sd
)
6112 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6117 for_each_domain(cpu
, sd
) {
6118 snprintf(buf
, 32, "domain%d", i
);
6119 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6121 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6128 static struct ctl_table_header
*sd_sysctl_header
;
6129 static void register_sched_domain_sysctl(void)
6131 int i
, cpu_num
= num_online_cpus();
6132 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6135 WARN_ON(sd_ctl_dir
[0].child
);
6136 sd_ctl_dir
[0].child
= entry
;
6141 for_each_online_cpu(i
) {
6142 snprintf(buf
, 32, "cpu%d", i
);
6143 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6145 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6149 WARN_ON(sd_sysctl_header
);
6150 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6153 /* may be called multiple times per register */
6154 static void unregister_sched_domain_sysctl(void)
6156 if (sd_sysctl_header
)
6157 unregister_sysctl_table(sd_sysctl_header
);
6158 sd_sysctl_header
= NULL
;
6159 if (sd_ctl_dir
[0].child
)
6160 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6163 static void register_sched_domain_sysctl(void)
6166 static void unregister_sched_domain_sysctl(void)
6171 static void set_rq_online(struct rq
*rq
)
6174 const struct sched_class
*class;
6176 cpu_set(rq
->cpu
, rq
->rd
->online
);
6179 for_each_class(class) {
6180 if (class->rq_online
)
6181 class->rq_online(rq
);
6186 static void set_rq_offline(struct rq
*rq
)
6189 const struct sched_class
*class;
6191 for_each_class(class) {
6192 if (class->rq_offline
)
6193 class->rq_offline(rq
);
6196 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6202 * migration_call - callback that gets triggered when a CPU is added.
6203 * Here we can start up the necessary migration thread for the new CPU.
6205 static int __cpuinit
6206 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6208 struct task_struct
*p
;
6209 int cpu
= (long)hcpu
;
6210 unsigned long flags
;
6215 case CPU_UP_PREPARE
:
6216 case CPU_UP_PREPARE_FROZEN
:
6217 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6220 kthread_bind(p
, cpu
);
6221 /* Must be high prio: stop_machine expects to yield to it. */
6222 rq
= task_rq_lock(p
, &flags
);
6223 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6224 task_rq_unlock(rq
, &flags
);
6225 cpu_rq(cpu
)->migration_thread
= p
;
6229 case CPU_ONLINE_FROZEN
:
6230 /* Strictly unnecessary, as first user will wake it. */
6231 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6233 /* Update our root-domain */
6235 spin_lock_irqsave(&rq
->lock
, flags
);
6237 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6241 spin_unlock_irqrestore(&rq
->lock
, flags
);
6244 #ifdef CONFIG_HOTPLUG_CPU
6245 case CPU_UP_CANCELED
:
6246 case CPU_UP_CANCELED_FROZEN
:
6247 if (!cpu_rq(cpu
)->migration_thread
)
6249 /* Unbind it from offline cpu so it can run. Fall thru. */
6250 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6251 any_online_cpu(cpu_online_map
));
6252 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6253 cpu_rq(cpu
)->migration_thread
= NULL
;
6257 case CPU_DEAD_FROZEN
:
6258 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6259 migrate_live_tasks(cpu
);
6261 kthread_stop(rq
->migration_thread
);
6262 rq
->migration_thread
= NULL
;
6263 /* Idle task back to normal (off runqueue, low prio) */
6264 spin_lock_irq(&rq
->lock
);
6265 update_rq_clock(rq
);
6266 deactivate_task(rq
, rq
->idle
, 0);
6267 rq
->idle
->static_prio
= MAX_PRIO
;
6268 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6269 rq
->idle
->sched_class
= &idle_sched_class
;
6270 migrate_dead_tasks(cpu
);
6271 spin_unlock_irq(&rq
->lock
);
6273 migrate_nr_uninterruptible(rq
);
6274 BUG_ON(rq
->nr_running
!= 0);
6277 * No need to migrate the tasks: it was best-effort if
6278 * they didn't take sched_hotcpu_mutex. Just wake up
6281 spin_lock_irq(&rq
->lock
);
6282 while (!list_empty(&rq
->migration_queue
)) {
6283 struct migration_req
*req
;
6285 req
= list_entry(rq
->migration_queue
.next
,
6286 struct migration_req
, list
);
6287 list_del_init(&req
->list
);
6288 complete(&req
->done
);
6290 spin_unlock_irq(&rq
->lock
);
6294 case CPU_DYING_FROZEN
:
6295 /* Update our root-domain */
6297 spin_lock_irqsave(&rq
->lock
, flags
);
6299 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6302 spin_unlock_irqrestore(&rq
->lock
, flags
);
6309 /* Register at highest priority so that task migration (migrate_all_tasks)
6310 * happens before everything else.
6312 static struct notifier_block __cpuinitdata migration_notifier
= {
6313 .notifier_call
= migration_call
,
6317 void __init
migration_init(void)
6319 void *cpu
= (void *)(long)smp_processor_id();
6322 /* Start one for the boot CPU: */
6323 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6324 BUG_ON(err
== NOTIFY_BAD
);
6325 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6326 register_cpu_notifier(&migration_notifier
);
6332 #ifdef CONFIG_SCHED_DEBUG
6334 static inline const char *sd_level_to_string(enum sched_domain_level lvl
)
6347 case SD_LV_ALLNODES
:
6356 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6357 cpumask_t
*groupmask
)
6359 struct sched_group
*group
= sd
->groups
;
6362 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6363 cpus_clear(*groupmask
);
6365 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6367 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6368 printk("does not load-balance\n");
6370 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6375 printk(KERN_CONT
"span %s level %s\n",
6376 str
, sd_level_to_string(sd
->level
));
6378 if (!cpu_isset(cpu
, sd
->span
)) {
6379 printk(KERN_ERR
"ERROR: domain->span does not contain "
6382 if (!cpu_isset(cpu
, group
->cpumask
)) {
6383 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6387 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6391 printk(KERN_ERR
"ERROR: group is NULL\n");
6395 if (!group
->__cpu_power
) {
6396 printk(KERN_CONT
"\n");
6397 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6402 if (!cpus_weight(group
->cpumask
)) {
6403 printk(KERN_CONT
"\n");
6404 printk(KERN_ERR
"ERROR: empty group\n");
6408 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6409 printk(KERN_CONT
"\n");
6410 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6414 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6416 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6417 printk(KERN_CONT
" %s", str
);
6419 group
= group
->next
;
6420 } while (group
!= sd
->groups
);
6421 printk(KERN_CONT
"\n");
6423 if (!cpus_equal(sd
->span
, *groupmask
))
6424 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6426 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6427 printk(KERN_ERR
"ERROR: parent span is not a superset "
6428 "of domain->span\n");
6432 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6434 cpumask_t
*groupmask
;
6438 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6442 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6444 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6446 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6451 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6460 #else /* !CONFIG_SCHED_DEBUG */
6461 # define sched_domain_debug(sd, cpu) do { } while (0)
6462 #endif /* CONFIG_SCHED_DEBUG */
6464 static int sd_degenerate(struct sched_domain
*sd
)
6466 if (cpus_weight(sd
->span
) == 1)
6469 /* Following flags need at least 2 groups */
6470 if (sd
->flags
& (SD_LOAD_BALANCE
|
6471 SD_BALANCE_NEWIDLE
|
6475 SD_SHARE_PKG_RESOURCES
)) {
6476 if (sd
->groups
!= sd
->groups
->next
)
6480 /* Following flags don't use groups */
6481 if (sd
->flags
& (SD_WAKE_IDLE
|
6490 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6492 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6494 if (sd_degenerate(parent
))
6497 if (!cpus_equal(sd
->span
, parent
->span
))
6500 /* Does parent contain flags not in child? */
6501 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6502 if (cflags
& SD_WAKE_AFFINE
)
6503 pflags
&= ~SD_WAKE_BALANCE
;
6504 /* Flags needing groups don't count if only 1 group in parent */
6505 if (parent
->groups
== parent
->groups
->next
) {
6506 pflags
&= ~(SD_LOAD_BALANCE
|
6507 SD_BALANCE_NEWIDLE
|
6511 SD_SHARE_PKG_RESOURCES
);
6513 if (~cflags
& pflags
)
6519 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6521 unsigned long flags
;
6523 spin_lock_irqsave(&rq
->lock
, flags
);
6526 struct root_domain
*old_rd
= rq
->rd
;
6528 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6531 cpu_clear(rq
->cpu
, old_rd
->span
);
6533 if (atomic_dec_and_test(&old_rd
->refcount
))
6537 atomic_inc(&rd
->refcount
);
6540 cpu_set(rq
->cpu
, rd
->span
);
6541 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6544 spin_unlock_irqrestore(&rq
->lock
, flags
);
6547 static void init_rootdomain(struct root_domain
*rd
)
6549 memset(rd
, 0, sizeof(*rd
));
6551 cpus_clear(rd
->span
);
6552 cpus_clear(rd
->online
);
6554 cpupri_init(&rd
->cpupri
);
6557 static void init_defrootdomain(void)
6559 init_rootdomain(&def_root_domain
);
6560 atomic_set(&def_root_domain
.refcount
, 1);
6563 static struct root_domain
*alloc_rootdomain(void)
6565 struct root_domain
*rd
;
6567 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6571 init_rootdomain(rd
);
6577 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6578 * hold the hotplug lock.
6581 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6583 struct rq
*rq
= cpu_rq(cpu
);
6584 struct sched_domain
*tmp
;
6586 /* Remove the sched domains which do not contribute to scheduling. */
6587 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6588 struct sched_domain
*parent
= tmp
->parent
;
6591 if (sd_parent_degenerate(tmp
, parent
)) {
6592 tmp
->parent
= parent
->parent
;
6594 parent
->parent
->child
= tmp
;
6598 if (sd
&& sd_degenerate(sd
)) {
6604 sched_domain_debug(sd
, cpu
);
6606 rq_attach_root(rq
, rd
);
6607 rcu_assign_pointer(rq
->sd
, sd
);
6610 /* cpus with isolated domains */
6611 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6613 /* Setup the mask of cpus configured for isolated domains */
6614 static int __init
isolated_cpu_setup(char *str
)
6616 int ints
[NR_CPUS
], i
;
6618 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6619 cpus_clear(cpu_isolated_map
);
6620 for (i
= 1; i
<= ints
[0]; i
++)
6621 if (ints
[i
] < NR_CPUS
)
6622 cpu_set(ints
[i
], cpu_isolated_map
);
6626 __setup("isolcpus=", isolated_cpu_setup
);
6629 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6630 * to a function which identifies what group(along with sched group) a CPU
6631 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6632 * (due to the fact that we keep track of groups covered with a cpumask_t).
6634 * init_sched_build_groups will build a circular linked list of the groups
6635 * covered by the given span, and will set each group's ->cpumask correctly,
6636 * and ->cpu_power to 0.
6639 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6640 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6641 struct sched_group
**sg
,
6642 cpumask_t
*tmpmask
),
6643 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6645 struct sched_group
*first
= NULL
, *last
= NULL
;
6648 cpus_clear(*covered
);
6650 for_each_cpu_mask(i
, *span
) {
6651 struct sched_group
*sg
;
6652 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6655 if (cpu_isset(i
, *covered
))
6658 cpus_clear(sg
->cpumask
);
6659 sg
->__cpu_power
= 0;
6661 for_each_cpu_mask(j
, *span
) {
6662 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6665 cpu_set(j
, *covered
);
6666 cpu_set(j
, sg
->cpumask
);
6677 #define SD_NODES_PER_DOMAIN 16
6682 * find_next_best_node - find the next node to include in a sched_domain
6683 * @node: node whose sched_domain we're building
6684 * @used_nodes: nodes already in the sched_domain
6686 * Find the next node to include in a given scheduling domain. Simply
6687 * finds the closest node not already in the @used_nodes map.
6689 * Should use nodemask_t.
6691 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6693 int i
, n
, val
, min_val
, best_node
= 0;
6697 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6698 /* Start at @node */
6699 n
= (node
+ i
) % MAX_NUMNODES
;
6701 if (!nr_cpus_node(n
))
6704 /* Skip already used nodes */
6705 if (node_isset(n
, *used_nodes
))
6708 /* Simple min distance search */
6709 val
= node_distance(node
, n
);
6711 if (val
< min_val
) {
6717 node_set(best_node
, *used_nodes
);
6722 * sched_domain_node_span - get a cpumask for a node's sched_domain
6723 * @node: node whose cpumask we're constructing
6724 * @span: resulting cpumask
6726 * Given a node, construct a good cpumask for its sched_domain to span. It
6727 * should be one that prevents unnecessary balancing, but also spreads tasks
6730 static void sched_domain_node_span(int node
, cpumask_t
*span
)
6732 nodemask_t used_nodes
;
6733 node_to_cpumask_ptr(nodemask
, node
);
6737 nodes_clear(used_nodes
);
6739 cpus_or(*span
, *span
, *nodemask
);
6740 node_set(node
, used_nodes
);
6742 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6743 int next_node
= find_next_best_node(node
, &used_nodes
);
6745 node_to_cpumask_ptr_next(nodemask
, next_node
);
6746 cpus_or(*span
, *span
, *nodemask
);
6749 #endif /* CONFIG_NUMA */
6751 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6754 * SMT sched-domains:
6756 #ifdef CONFIG_SCHED_SMT
6757 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6758 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6761 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6765 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6768 #endif /* CONFIG_SCHED_SMT */
6771 * multi-core sched-domains:
6773 #ifdef CONFIG_SCHED_MC
6774 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6775 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6776 #endif /* CONFIG_SCHED_MC */
6778 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6780 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6785 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6786 cpus_and(*mask
, *mask
, *cpu_map
);
6787 group
= first_cpu(*mask
);
6789 *sg
= &per_cpu(sched_group_core
, group
);
6792 #elif defined(CONFIG_SCHED_MC)
6794 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6798 *sg
= &per_cpu(sched_group_core
, cpu
);
6803 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6804 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6807 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6811 #ifdef CONFIG_SCHED_MC
6812 *mask
= cpu_coregroup_map(cpu
);
6813 cpus_and(*mask
, *mask
, *cpu_map
);
6814 group
= first_cpu(*mask
);
6815 #elif defined(CONFIG_SCHED_SMT)
6816 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6817 cpus_and(*mask
, *mask
, *cpu_map
);
6818 group
= first_cpu(*mask
);
6823 *sg
= &per_cpu(sched_group_phys
, group
);
6829 * The init_sched_build_groups can't handle what we want to do with node
6830 * groups, so roll our own. Now each node has its own list of groups which
6831 * gets dynamically allocated.
6833 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6834 static struct sched_group
***sched_group_nodes_bycpu
;
6836 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6837 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6839 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6840 struct sched_group
**sg
, cpumask_t
*nodemask
)
6844 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6845 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6846 group
= first_cpu(*nodemask
);
6849 *sg
= &per_cpu(sched_group_allnodes
, group
);
6853 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6855 struct sched_group
*sg
= group_head
;
6861 for_each_cpu_mask(j
, sg
->cpumask
) {
6862 struct sched_domain
*sd
;
6864 sd
= &per_cpu(phys_domains
, j
);
6865 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6867 * Only add "power" once for each
6873 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6876 } while (sg
!= group_head
);
6878 #endif /* CONFIG_NUMA */
6881 /* Free memory allocated for various sched_group structures */
6882 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6886 for_each_cpu_mask(cpu
, *cpu_map
) {
6887 struct sched_group
**sched_group_nodes
6888 = sched_group_nodes_bycpu
[cpu
];
6890 if (!sched_group_nodes
)
6893 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6894 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6896 *nodemask
= node_to_cpumask(i
);
6897 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6898 if (cpus_empty(*nodemask
))
6908 if (oldsg
!= sched_group_nodes
[i
])
6911 kfree(sched_group_nodes
);
6912 sched_group_nodes_bycpu
[cpu
] = NULL
;
6915 #else /* !CONFIG_NUMA */
6916 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6919 #endif /* CONFIG_NUMA */
6922 * Initialize sched groups cpu_power.
6924 * cpu_power indicates the capacity of sched group, which is used while
6925 * distributing the load between different sched groups in a sched domain.
6926 * Typically cpu_power for all the groups in a sched domain will be same unless
6927 * there are asymmetries in the topology. If there are asymmetries, group
6928 * having more cpu_power will pickup more load compared to the group having
6931 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6932 * the maximum number of tasks a group can handle in the presence of other idle
6933 * or lightly loaded groups in the same sched domain.
6935 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6937 struct sched_domain
*child
;
6938 struct sched_group
*group
;
6940 WARN_ON(!sd
|| !sd
->groups
);
6942 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6947 sd
->groups
->__cpu_power
= 0;
6950 * For perf policy, if the groups in child domain share resources
6951 * (for example cores sharing some portions of the cache hierarchy
6952 * or SMT), then set this domain groups cpu_power such that each group
6953 * can handle only one task, when there are other idle groups in the
6954 * same sched domain.
6956 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6958 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6959 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6964 * add cpu_power of each child group to this groups cpu_power
6966 group
= child
->groups
;
6968 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6969 group
= group
->next
;
6970 } while (group
!= child
->groups
);
6974 * Initializers for schedule domains
6975 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6978 #define SD_INIT(sd, type) sd_init_##type(sd)
6979 #define SD_INIT_FUNC(type) \
6980 static noinline void sd_init_##type(struct sched_domain *sd) \
6982 memset(sd, 0, sizeof(*sd)); \
6983 *sd = SD_##type##_INIT; \
6984 sd->level = SD_LV_##type; \
6989 SD_INIT_FUNC(ALLNODES
)
6992 #ifdef CONFIG_SCHED_SMT
6993 SD_INIT_FUNC(SIBLING
)
6995 #ifdef CONFIG_SCHED_MC
7000 * To minimize stack usage kmalloc room for cpumasks and share the
7001 * space as the usage in build_sched_domains() dictates. Used only
7002 * if the amount of space is significant.
7005 cpumask_t tmpmask
; /* make this one first */
7008 cpumask_t this_sibling_map
;
7009 cpumask_t this_core_map
;
7011 cpumask_t send_covered
;
7014 cpumask_t domainspan
;
7016 cpumask_t notcovered
;
7021 #define SCHED_CPUMASK_ALLOC 1
7022 #define SCHED_CPUMASK_FREE(v) kfree(v)
7023 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7025 #define SCHED_CPUMASK_ALLOC 0
7026 #define SCHED_CPUMASK_FREE(v)
7027 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7030 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7031 ((unsigned long)(a) + offsetof(struct allmasks, v))
7033 static int default_relax_domain_level
= -1;
7035 static int __init
setup_relax_domain_level(char *str
)
7039 val
= simple_strtoul(str
, NULL
, 0);
7040 if (val
< SD_LV_MAX
)
7041 default_relax_domain_level
= val
;
7045 __setup("relax_domain_level=", setup_relax_domain_level
);
7047 static void set_domain_attribute(struct sched_domain
*sd
,
7048 struct sched_domain_attr
*attr
)
7052 if (!attr
|| attr
->relax_domain_level
< 0) {
7053 if (default_relax_domain_level
< 0)
7056 request
= default_relax_domain_level
;
7058 request
= attr
->relax_domain_level
;
7059 if (request
< sd
->level
) {
7060 /* turn off idle balance on this domain */
7061 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7063 /* turn on idle balance on this domain */
7064 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7069 * Build sched domains for a given set of cpus and attach the sched domains
7070 * to the individual cpus
7072 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7073 struct sched_domain_attr
*attr
)
7076 struct root_domain
*rd
;
7077 SCHED_CPUMASK_DECLARE(allmasks
);
7080 struct sched_group
**sched_group_nodes
= NULL
;
7081 int sd_allnodes
= 0;
7084 * Allocate the per-node list of sched groups
7086 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
7088 if (!sched_group_nodes
) {
7089 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7094 rd
= alloc_rootdomain();
7096 printk(KERN_WARNING
"Cannot alloc root domain\n");
7098 kfree(sched_group_nodes
);
7103 #if SCHED_CPUMASK_ALLOC
7104 /* get space for all scratch cpumask variables */
7105 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7107 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7110 kfree(sched_group_nodes
);
7115 tmpmask
= (cpumask_t
*)allmasks
;
7119 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7123 * Set up domains for cpus specified by the cpu_map.
7125 for_each_cpu_mask(i
, *cpu_map
) {
7126 struct sched_domain
*sd
= NULL
, *p
;
7127 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7129 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7130 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7133 if (cpus_weight(*cpu_map
) >
7134 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7135 sd
= &per_cpu(allnodes_domains
, i
);
7136 SD_INIT(sd
, ALLNODES
);
7137 set_domain_attribute(sd
, attr
);
7138 sd
->span
= *cpu_map
;
7139 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7145 sd
= &per_cpu(node_domains
, i
);
7147 set_domain_attribute(sd
, attr
);
7148 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7152 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7156 sd
= &per_cpu(phys_domains
, i
);
7158 set_domain_attribute(sd
, attr
);
7159 sd
->span
= *nodemask
;
7163 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7165 #ifdef CONFIG_SCHED_MC
7167 sd
= &per_cpu(core_domains
, i
);
7169 set_domain_attribute(sd
, attr
);
7170 sd
->span
= cpu_coregroup_map(i
);
7171 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7174 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7177 #ifdef CONFIG_SCHED_SMT
7179 sd
= &per_cpu(cpu_domains
, i
);
7180 SD_INIT(sd
, SIBLING
);
7181 set_domain_attribute(sd
, attr
);
7182 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7183 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7186 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7190 #ifdef CONFIG_SCHED_SMT
7191 /* Set up CPU (sibling) groups */
7192 for_each_cpu_mask(i
, *cpu_map
) {
7193 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7194 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7196 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7197 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7198 if (i
!= first_cpu(*this_sibling_map
))
7201 init_sched_build_groups(this_sibling_map
, cpu_map
,
7203 send_covered
, tmpmask
);
7207 #ifdef CONFIG_SCHED_MC
7208 /* Set up multi-core groups */
7209 for_each_cpu_mask(i
, *cpu_map
) {
7210 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7211 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7213 *this_core_map
= cpu_coregroup_map(i
);
7214 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7215 if (i
!= first_cpu(*this_core_map
))
7218 init_sched_build_groups(this_core_map
, cpu_map
,
7220 send_covered
, tmpmask
);
7224 /* Set up physical groups */
7225 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7226 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7227 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7229 *nodemask
= node_to_cpumask(i
);
7230 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7231 if (cpus_empty(*nodemask
))
7234 init_sched_build_groups(nodemask
, cpu_map
,
7236 send_covered
, tmpmask
);
7240 /* Set up node groups */
7242 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7244 init_sched_build_groups(cpu_map
, cpu_map
,
7245 &cpu_to_allnodes_group
,
7246 send_covered
, tmpmask
);
7249 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7250 /* Set up node groups */
7251 struct sched_group
*sg
, *prev
;
7252 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7253 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7254 SCHED_CPUMASK_VAR(covered
, allmasks
);
7257 *nodemask
= node_to_cpumask(i
);
7258 cpus_clear(*covered
);
7260 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7261 if (cpus_empty(*nodemask
)) {
7262 sched_group_nodes
[i
] = NULL
;
7266 sched_domain_node_span(i
, domainspan
);
7267 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7269 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7271 printk(KERN_WARNING
"Can not alloc domain group for "
7275 sched_group_nodes
[i
] = sg
;
7276 for_each_cpu_mask(j
, *nodemask
) {
7277 struct sched_domain
*sd
;
7279 sd
= &per_cpu(node_domains
, j
);
7282 sg
->__cpu_power
= 0;
7283 sg
->cpumask
= *nodemask
;
7285 cpus_or(*covered
, *covered
, *nodemask
);
7288 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
7289 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7290 int n
= (i
+ j
) % MAX_NUMNODES
;
7291 node_to_cpumask_ptr(pnodemask
, n
);
7293 cpus_complement(*notcovered
, *covered
);
7294 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7295 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7296 if (cpus_empty(*tmpmask
))
7299 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7300 if (cpus_empty(*tmpmask
))
7303 sg
= kmalloc_node(sizeof(struct sched_group
),
7307 "Can not alloc domain group for node %d\n", j
);
7310 sg
->__cpu_power
= 0;
7311 sg
->cpumask
= *tmpmask
;
7312 sg
->next
= prev
->next
;
7313 cpus_or(*covered
, *covered
, *tmpmask
);
7320 /* Calculate CPU power for physical packages and nodes */
7321 #ifdef CONFIG_SCHED_SMT
7322 for_each_cpu_mask(i
, *cpu_map
) {
7323 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7325 init_sched_groups_power(i
, sd
);
7328 #ifdef CONFIG_SCHED_MC
7329 for_each_cpu_mask(i
, *cpu_map
) {
7330 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7332 init_sched_groups_power(i
, sd
);
7336 for_each_cpu_mask(i
, *cpu_map
) {
7337 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7339 init_sched_groups_power(i
, sd
);
7343 for (i
= 0; i
< MAX_NUMNODES
; i
++)
7344 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7347 struct sched_group
*sg
;
7349 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7351 init_numa_sched_groups_power(sg
);
7355 /* Attach the domains */
7356 for_each_cpu_mask(i
, *cpu_map
) {
7357 struct sched_domain
*sd
;
7358 #ifdef CONFIG_SCHED_SMT
7359 sd
= &per_cpu(cpu_domains
, i
);
7360 #elif defined(CONFIG_SCHED_MC)
7361 sd
= &per_cpu(core_domains
, i
);
7363 sd
= &per_cpu(phys_domains
, i
);
7365 cpu_attach_domain(sd
, rd
, i
);
7368 SCHED_CPUMASK_FREE((void *)allmasks
);
7373 free_sched_groups(cpu_map
, tmpmask
);
7374 SCHED_CPUMASK_FREE((void *)allmasks
);
7379 static int build_sched_domains(const cpumask_t
*cpu_map
)
7381 return __build_sched_domains(cpu_map
, NULL
);
7384 static cpumask_t
*doms_cur
; /* current sched domains */
7385 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7386 static struct sched_domain_attr
*dattr_cur
;
7387 /* attribues of custom domains in 'doms_cur' */
7390 * Special case: If a kmalloc of a doms_cur partition (array of
7391 * cpumask_t) fails, then fallback to a single sched domain,
7392 * as determined by the single cpumask_t fallback_doms.
7394 static cpumask_t fallback_doms
;
7396 void __attribute__((weak
)) arch_update_cpu_topology(void)
7401 * Free current domain masks.
7402 * Called after all cpus are attached to NULL domain.
7404 static void free_sched_domains(void)
7407 if (doms_cur
!= &fallback_doms
)
7409 doms_cur
= &fallback_doms
;
7413 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7414 * For now this just excludes isolated cpus, but could be used to
7415 * exclude other special cases in the future.
7417 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7421 arch_update_cpu_topology();
7423 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7425 doms_cur
= &fallback_doms
;
7426 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7428 err
= build_sched_domains(doms_cur
);
7429 register_sched_domain_sysctl();
7434 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7437 free_sched_groups(cpu_map
, tmpmask
);
7441 * Detach sched domains from a group of cpus specified in cpu_map
7442 * These cpus will now be attached to the NULL domain
7444 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7449 unregister_sched_domain_sysctl();
7451 for_each_cpu_mask(i
, *cpu_map
)
7452 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7453 synchronize_sched();
7454 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7457 /* handle null as "default" */
7458 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7459 struct sched_domain_attr
*new, int idx_new
)
7461 struct sched_domain_attr tmp
;
7468 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7469 new ? (new + idx_new
) : &tmp
,
7470 sizeof(struct sched_domain_attr
));
7474 * Partition sched domains as specified by the 'ndoms_new'
7475 * cpumasks in the array doms_new[] of cpumasks. This compares
7476 * doms_new[] to the current sched domain partitioning, doms_cur[].
7477 * It destroys each deleted domain and builds each new domain.
7479 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7480 * The masks don't intersect (don't overlap.) We should setup one
7481 * sched domain for each mask. CPUs not in any of the cpumasks will
7482 * not be load balanced. If the same cpumask appears both in the
7483 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7486 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7487 * ownership of it and will kfree it when done with it. If the caller
7488 * failed the kmalloc call, then it can pass in doms_new == NULL,
7489 * and partition_sched_domains() will fallback to the single partition
7492 * Call with hotplug lock held
7494 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7495 struct sched_domain_attr
*dattr_new
)
7499 mutex_lock(&sched_domains_mutex
);
7501 /* always unregister in case we don't destroy any domains */
7502 unregister_sched_domain_sysctl();
7504 if (doms_new
== NULL
) {
7506 doms_new
= &fallback_doms
;
7507 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7511 /* Destroy deleted domains */
7512 for (i
= 0; i
< ndoms_cur
; i
++) {
7513 for (j
= 0; j
< ndoms_new
; j
++) {
7514 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7515 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7518 /* no match - a current sched domain not in new doms_new[] */
7519 detach_destroy_domains(doms_cur
+ i
);
7524 /* Build new domains */
7525 for (i
= 0; i
< ndoms_new
; i
++) {
7526 for (j
= 0; j
< ndoms_cur
; j
++) {
7527 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7528 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7531 /* no match - add a new doms_new */
7532 __build_sched_domains(doms_new
+ i
,
7533 dattr_new
? dattr_new
+ i
: NULL
);
7538 /* Remember the new sched domains */
7539 if (doms_cur
!= &fallback_doms
)
7541 kfree(dattr_cur
); /* kfree(NULL) is safe */
7542 doms_cur
= doms_new
;
7543 dattr_cur
= dattr_new
;
7544 ndoms_cur
= ndoms_new
;
7546 register_sched_domain_sysctl();
7548 mutex_unlock(&sched_domains_mutex
);
7551 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7552 int arch_reinit_sched_domains(void)
7557 mutex_lock(&sched_domains_mutex
);
7558 detach_destroy_domains(&cpu_online_map
);
7559 free_sched_domains();
7560 err
= arch_init_sched_domains(&cpu_online_map
);
7561 mutex_unlock(&sched_domains_mutex
);
7567 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7571 if (buf
[0] != '0' && buf
[0] != '1')
7575 sched_smt_power_savings
= (buf
[0] == '1');
7577 sched_mc_power_savings
= (buf
[0] == '1');
7579 ret
= arch_reinit_sched_domains();
7581 return ret
? ret
: count
;
7584 #ifdef CONFIG_SCHED_MC
7585 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
7587 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7589 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
7590 const char *buf
, size_t count
)
7592 return sched_power_savings_store(buf
, count
, 0);
7594 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
7595 sched_mc_power_savings_store
);
7598 #ifdef CONFIG_SCHED_SMT
7599 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
7601 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7603 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
7604 const char *buf
, size_t count
)
7606 return sched_power_savings_store(buf
, count
, 1);
7608 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
7609 sched_smt_power_savings_store
);
7612 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7616 #ifdef CONFIG_SCHED_SMT
7618 err
= sysfs_create_file(&cls
->kset
.kobj
,
7619 &attr_sched_smt_power_savings
.attr
);
7621 #ifdef CONFIG_SCHED_MC
7622 if (!err
&& mc_capable())
7623 err
= sysfs_create_file(&cls
->kset
.kobj
,
7624 &attr_sched_mc_power_savings
.attr
);
7628 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7631 * Force a reinitialization of the sched domains hierarchy. The domains
7632 * and groups cannot be updated in place without racing with the balancing
7633 * code, so we temporarily attach all running cpus to the NULL domain
7634 * which will prevent rebalancing while the sched domains are recalculated.
7636 static int update_sched_domains(struct notifier_block
*nfb
,
7637 unsigned long action
, void *hcpu
)
7639 int cpu
= (int)(long)hcpu
;
7642 case CPU_DOWN_PREPARE
:
7643 case CPU_DOWN_PREPARE_FROZEN
:
7644 disable_runtime(cpu_rq(cpu
));
7646 case CPU_UP_PREPARE
:
7647 case CPU_UP_PREPARE_FROZEN
:
7648 detach_destroy_domains(&cpu_online_map
);
7649 free_sched_domains();
7653 case CPU_DOWN_FAILED
:
7654 case CPU_DOWN_FAILED_FROZEN
:
7656 case CPU_ONLINE_FROZEN
:
7657 enable_runtime(cpu_rq(cpu
));
7659 case CPU_UP_CANCELED
:
7660 case CPU_UP_CANCELED_FROZEN
:
7662 case CPU_DEAD_FROZEN
:
7664 * Fall through and re-initialise the domains.
7671 #ifndef CONFIG_CPUSETS
7673 * Create default domain partitioning if cpusets are disabled.
7674 * Otherwise we let cpusets rebuild the domains based on the
7678 /* The hotplug lock is already held by cpu_up/cpu_down */
7679 arch_init_sched_domains(&cpu_online_map
);
7685 void __init
sched_init_smp(void)
7687 cpumask_t non_isolated_cpus
;
7689 #if defined(CONFIG_NUMA)
7690 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7692 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7695 mutex_lock(&sched_domains_mutex
);
7696 arch_init_sched_domains(&cpu_online_map
);
7697 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7698 if (cpus_empty(non_isolated_cpus
))
7699 cpu_set(smp_processor_id(), non_isolated_cpus
);
7700 mutex_unlock(&sched_domains_mutex
);
7702 /* XXX: Theoretical race here - CPU may be hotplugged now */
7703 hotcpu_notifier(update_sched_domains
, 0);
7706 /* Move init over to a non-isolated CPU */
7707 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7709 sched_init_granularity();
7712 void __init
sched_init_smp(void)
7714 sched_init_granularity();
7716 #endif /* CONFIG_SMP */
7718 int in_sched_functions(unsigned long addr
)
7720 return in_lock_functions(addr
) ||
7721 (addr
>= (unsigned long)__sched_text_start
7722 && addr
< (unsigned long)__sched_text_end
);
7725 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7727 cfs_rq
->tasks_timeline
= RB_ROOT
;
7728 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7729 #ifdef CONFIG_FAIR_GROUP_SCHED
7732 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7735 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7737 struct rt_prio_array
*array
;
7740 array
= &rt_rq
->active
;
7741 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7742 INIT_LIST_HEAD(array
->queue
+ i
);
7743 __clear_bit(i
, array
->bitmap
);
7745 /* delimiter for bitsearch: */
7746 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7748 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7749 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7752 rt_rq
->rt_nr_migratory
= 0;
7753 rt_rq
->overloaded
= 0;
7757 rt_rq
->rt_throttled
= 0;
7758 rt_rq
->rt_runtime
= 0;
7759 spin_lock_init(&rt_rq
->rt_runtime_lock
);
7761 #ifdef CONFIG_RT_GROUP_SCHED
7762 rt_rq
->rt_nr_boosted
= 0;
7767 #ifdef CONFIG_FAIR_GROUP_SCHED
7768 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7769 struct sched_entity
*se
, int cpu
, int add
,
7770 struct sched_entity
*parent
)
7772 struct rq
*rq
= cpu_rq(cpu
);
7773 tg
->cfs_rq
[cpu
] = cfs_rq
;
7774 init_cfs_rq(cfs_rq
, rq
);
7777 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7780 /* se could be NULL for init_task_group */
7785 se
->cfs_rq
= &rq
->cfs
;
7787 se
->cfs_rq
= parent
->my_q
;
7790 se
->load
.weight
= tg
->shares
;
7791 se
->load
.inv_weight
= 0;
7792 se
->parent
= parent
;
7796 #ifdef CONFIG_RT_GROUP_SCHED
7797 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7798 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7799 struct sched_rt_entity
*parent
)
7801 struct rq
*rq
= cpu_rq(cpu
);
7803 tg
->rt_rq
[cpu
] = rt_rq
;
7804 init_rt_rq(rt_rq
, rq
);
7806 rt_rq
->rt_se
= rt_se
;
7807 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7809 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7811 tg
->rt_se
[cpu
] = rt_se
;
7816 rt_se
->rt_rq
= &rq
->rt
;
7818 rt_se
->rt_rq
= parent
->my_q
;
7820 rt_se
->my_q
= rt_rq
;
7821 rt_se
->parent
= parent
;
7822 INIT_LIST_HEAD(&rt_se
->run_list
);
7826 void __init
sched_init(void)
7829 unsigned long alloc_size
= 0, ptr
;
7831 #ifdef CONFIG_FAIR_GROUP_SCHED
7832 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7834 #ifdef CONFIG_RT_GROUP_SCHED
7835 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7837 #ifdef CONFIG_USER_SCHED
7841 * As sched_init() is called before page_alloc is setup,
7842 * we use alloc_bootmem().
7845 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
7847 #ifdef CONFIG_FAIR_GROUP_SCHED
7848 init_task_group
.se
= (struct sched_entity
**)ptr
;
7849 ptr
+= nr_cpu_ids
* sizeof(void **);
7851 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7852 ptr
+= nr_cpu_ids
* sizeof(void **);
7854 #ifdef CONFIG_USER_SCHED
7855 root_task_group
.se
= (struct sched_entity
**)ptr
;
7856 ptr
+= nr_cpu_ids
* sizeof(void **);
7858 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7859 ptr
+= nr_cpu_ids
* sizeof(void **);
7860 #endif /* CONFIG_USER_SCHED */
7861 #endif /* CONFIG_FAIR_GROUP_SCHED */
7862 #ifdef CONFIG_RT_GROUP_SCHED
7863 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7864 ptr
+= nr_cpu_ids
* sizeof(void **);
7866 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7867 ptr
+= nr_cpu_ids
* sizeof(void **);
7869 #ifdef CONFIG_USER_SCHED
7870 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7871 ptr
+= nr_cpu_ids
* sizeof(void **);
7873 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7874 ptr
+= nr_cpu_ids
* sizeof(void **);
7875 #endif /* CONFIG_USER_SCHED */
7876 #endif /* CONFIG_RT_GROUP_SCHED */
7880 init_defrootdomain();
7883 init_rt_bandwidth(&def_rt_bandwidth
,
7884 global_rt_period(), global_rt_runtime());
7886 #ifdef CONFIG_RT_GROUP_SCHED
7887 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7888 global_rt_period(), global_rt_runtime());
7889 #ifdef CONFIG_USER_SCHED
7890 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7891 global_rt_period(), RUNTIME_INF
);
7892 #endif /* CONFIG_USER_SCHED */
7893 #endif /* CONFIG_RT_GROUP_SCHED */
7895 #ifdef CONFIG_GROUP_SCHED
7896 list_add(&init_task_group
.list
, &task_groups
);
7897 INIT_LIST_HEAD(&init_task_group
.children
);
7899 #ifdef CONFIG_USER_SCHED
7900 INIT_LIST_HEAD(&root_task_group
.children
);
7901 init_task_group
.parent
= &root_task_group
;
7902 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
7903 #endif /* CONFIG_USER_SCHED */
7904 #endif /* CONFIG_GROUP_SCHED */
7906 for_each_possible_cpu(i
) {
7910 spin_lock_init(&rq
->lock
);
7911 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7913 init_cfs_rq(&rq
->cfs
, rq
);
7914 init_rt_rq(&rq
->rt
, rq
);
7915 #ifdef CONFIG_FAIR_GROUP_SCHED
7916 init_task_group
.shares
= init_task_group_load
;
7917 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7918 #ifdef CONFIG_CGROUP_SCHED
7920 * How much cpu bandwidth does init_task_group get?
7922 * In case of task-groups formed thr' the cgroup filesystem, it
7923 * gets 100% of the cpu resources in the system. This overall
7924 * system cpu resource is divided among the tasks of
7925 * init_task_group and its child task-groups in a fair manner,
7926 * based on each entity's (task or task-group's) weight
7927 * (se->load.weight).
7929 * In other words, if init_task_group has 10 tasks of weight
7930 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7931 * then A0's share of the cpu resource is:
7933 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7935 * We achieve this by letting init_task_group's tasks sit
7936 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7938 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7939 #elif defined CONFIG_USER_SCHED
7940 root_task_group
.shares
= NICE_0_LOAD
;
7941 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
7943 * In case of task-groups formed thr' the user id of tasks,
7944 * init_task_group represents tasks belonging to root user.
7945 * Hence it forms a sibling of all subsequent groups formed.
7946 * In this case, init_task_group gets only a fraction of overall
7947 * system cpu resource, based on the weight assigned to root
7948 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7949 * by letting tasks of init_task_group sit in a separate cfs_rq
7950 * (init_cfs_rq) and having one entity represent this group of
7951 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7953 init_tg_cfs_entry(&init_task_group
,
7954 &per_cpu(init_cfs_rq
, i
),
7955 &per_cpu(init_sched_entity
, i
), i
, 1,
7956 root_task_group
.se
[i
]);
7959 #endif /* CONFIG_FAIR_GROUP_SCHED */
7961 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7962 #ifdef CONFIG_RT_GROUP_SCHED
7963 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7964 #ifdef CONFIG_CGROUP_SCHED
7965 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7966 #elif defined CONFIG_USER_SCHED
7967 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
7968 init_tg_rt_entry(&init_task_group
,
7969 &per_cpu(init_rt_rq
, i
),
7970 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
7971 root_task_group
.rt_se
[i
]);
7975 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7976 rq
->cpu_load
[j
] = 0;
7980 rq
->active_balance
= 0;
7981 rq
->next_balance
= jiffies
;
7985 rq
->migration_thread
= NULL
;
7986 INIT_LIST_HEAD(&rq
->migration_queue
);
7987 rq_attach_root(rq
, &def_root_domain
);
7990 atomic_set(&rq
->nr_iowait
, 0);
7993 set_load_weight(&init_task
);
7995 #ifdef CONFIG_PREEMPT_NOTIFIERS
7996 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8000 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
8003 #ifdef CONFIG_RT_MUTEXES
8004 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8008 * The boot idle thread does lazy MMU switching as well:
8010 atomic_inc(&init_mm
.mm_count
);
8011 enter_lazy_tlb(&init_mm
, current
);
8014 * Make us the idle thread. Technically, schedule() should not be
8015 * called from this thread, however somewhere below it might be,
8016 * but because we are the idle thread, we just pick up running again
8017 * when this runqueue becomes "idle".
8019 init_idle(current
, smp_processor_id());
8021 * During early bootup we pretend to be a normal task:
8023 current
->sched_class
= &fair_sched_class
;
8025 scheduler_running
= 1;
8028 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8029 void __might_sleep(char *file
, int line
)
8032 static unsigned long prev_jiffy
; /* ratelimiting */
8034 if ((in_atomic() || irqs_disabled()) &&
8035 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
8036 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8038 prev_jiffy
= jiffies
;
8039 printk(KERN_ERR
"BUG: sleeping function called from invalid"
8040 " context at %s:%d\n", file
, line
);
8041 printk("in_atomic():%d, irqs_disabled():%d\n",
8042 in_atomic(), irqs_disabled());
8043 debug_show_held_locks(current
);
8044 if (irqs_disabled())
8045 print_irqtrace_events(current
);
8050 EXPORT_SYMBOL(__might_sleep
);
8053 #ifdef CONFIG_MAGIC_SYSRQ
8054 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8058 update_rq_clock(rq
);
8059 on_rq
= p
->se
.on_rq
;
8061 deactivate_task(rq
, p
, 0);
8062 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8064 activate_task(rq
, p
, 0);
8065 resched_task(rq
->curr
);
8069 void normalize_rt_tasks(void)
8071 struct task_struct
*g
, *p
;
8072 unsigned long flags
;
8075 read_lock_irqsave(&tasklist_lock
, flags
);
8076 do_each_thread(g
, p
) {
8078 * Only normalize user tasks:
8083 p
->se
.exec_start
= 0;
8084 #ifdef CONFIG_SCHEDSTATS
8085 p
->se
.wait_start
= 0;
8086 p
->se
.sleep_start
= 0;
8087 p
->se
.block_start
= 0;
8092 * Renice negative nice level userspace
8095 if (TASK_NICE(p
) < 0 && p
->mm
)
8096 set_user_nice(p
, 0);
8100 spin_lock(&p
->pi_lock
);
8101 rq
= __task_rq_lock(p
);
8103 normalize_task(rq
, p
);
8105 __task_rq_unlock(rq
);
8106 spin_unlock(&p
->pi_lock
);
8107 } while_each_thread(g
, p
);
8109 read_unlock_irqrestore(&tasklist_lock
, flags
);
8112 #endif /* CONFIG_MAGIC_SYSRQ */
8116 * These functions are only useful for the IA64 MCA handling.
8118 * They can only be called when the whole system has been
8119 * stopped - every CPU needs to be quiescent, and no scheduling
8120 * activity can take place. Using them for anything else would
8121 * be a serious bug, and as a result, they aren't even visible
8122 * under any other configuration.
8126 * curr_task - return the current task for a given cpu.
8127 * @cpu: the processor in question.
8129 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8131 struct task_struct
*curr_task(int cpu
)
8133 return cpu_curr(cpu
);
8137 * set_curr_task - set the current task for a given cpu.
8138 * @cpu: the processor in question.
8139 * @p: the task pointer to set.
8141 * Description: This function must only be used when non-maskable interrupts
8142 * are serviced on a separate stack. It allows the architecture to switch the
8143 * notion of the current task on a cpu in a non-blocking manner. This function
8144 * must be called with all CPU's synchronized, and interrupts disabled, the
8145 * and caller must save the original value of the current task (see
8146 * curr_task() above) and restore that value before reenabling interrupts and
8147 * re-starting the system.
8149 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8151 void set_curr_task(int cpu
, struct task_struct
*p
)
8158 #ifdef CONFIG_FAIR_GROUP_SCHED
8159 static void free_fair_sched_group(struct task_group
*tg
)
8163 for_each_possible_cpu(i
) {
8165 kfree(tg
->cfs_rq
[i
]);
8175 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8177 struct cfs_rq
*cfs_rq
;
8178 struct sched_entity
*se
, *parent_se
;
8182 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8185 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8189 tg
->shares
= NICE_0_LOAD
;
8191 for_each_possible_cpu(i
) {
8194 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8195 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8199 se
= kmalloc_node(sizeof(struct sched_entity
),
8200 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8204 parent_se
= parent
? parent
->se
[i
] : NULL
;
8205 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8214 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8216 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8217 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8220 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8222 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8224 #else /* !CONFG_FAIR_GROUP_SCHED */
8225 static inline void free_fair_sched_group(struct task_group
*tg
)
8230 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8235 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8239 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8242 #endif /* CONFIG_FAIR_GROUP_SCHED */
8244 #ifdef CONFIG_RT_GROUP_SCHED
8245 static void free_rt_sched_group(struct task_group
*tg
)
8249 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8251 for_each_possible_cpu(i
) {
8253 kfree(tg
->rt_rq
[i
]);
8255 kfree(tg
->rt_se
[i
]);
8263 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8265 struct rt_rq
*rt_rq
;
8266 struct sched_rt_entity
*rt_se
, *parent_se
;
8270 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8273 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8277 init_rt_bandwidth(&tg
->rt_bandwidth
,
8278 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8280 for_each_possible_cpu(i
) {
8283 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8284 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8288 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8289 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8293 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8294 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8303 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8305 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8306 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8309 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8311 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8313 #else /* !CONFIG_RT_GROUP_SCHED */
8314 static inline void free_rt_sched_group(struct task_group
*tg
)
8319 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8324 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8328 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8331 #endif /* CONFIG_RT_GROUP_SCHED */
8333 #ifdef CONFIG_GROUP_SCHED
8334 static void free_sched_group(struct task_group
*tg
)
8336 free_fair_sched_group(tg
);
8337 free_rt_sched_group(tg
);
8341 /* allocate runqueue etc for a new task group */
8342 struct task_group
*sched_create_group(struct task_group
*parent
)
8344 struct task_group
*tg
;
8345 unsigned long flags
;
8348 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8350 return ERR_PTR(-ENOMEM
);
8352 if (!alloc_fair_sched_group(tg
, parent
))
8355 if (!alloc_rt_sched_group(tg
, parent
))
8358 spin_lock_irqsave(&task_group_lock
, flags
);
8359 for_each_possible_cpu(i
) {
8360 register_fair_sched_group(tg
, i
);
8361 register_rt_sched_group(tg
, i
);
8363 list_add_rcu(&tg
->list
, &task_groups
);
8365 WARN_ON(!parent
); /* root should already exist */
8367 tg
->parent
= parent
;
8368 list_add_rcu(&tg
->siblings
, &parent
->children
);
8369 INIT_LIST_HEAD(&tg
->children
);
8370 spin_unlock_irqrestore(&task_group_lock
, flags
);
8375 free_sched_group(tg
);
8376 return ERR_PTR(-ENOMEM
);
8379 /* rcu callback to free various structures associated with a task group */
8380 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8382 /* now it should be safe to free those cfs_rqs */
8383 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8386 /* Destroy runqueue etc associated with a task group */
8387 void sched_destroy_group(struct task_group
*tg
)
8389 unsigned long flags
;
8392 spin_lock_irqsave(&task_group_lock
, flags
);
8393 for_each_possible_cpu(i
) {
8394 unregister_fair_sched_group(tg
, i
);
8395 unregister_rt_sched_group(tg
, i
);
8397 list_del_rcu(&tg
->list
);
8398 list_del_rcu(&tg
->siblings
);
8399 spin_unlock_irqrestore(&task_group_lock
, flags
);
8401 /* wait for possible concurrent references to cfs_rqs complete */
8402 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8405 /* change task's runqueue when it moves between groups.
8406 * The caller of this function should have put the task in its new group
8407 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8408 * reflect its new group.
8410 void sched_move_task(struct task_struct
*tsk
)
8413 unsigned long flags
;
8416 rq
= task_rq_lock(tsk
, &flags
);
8418 update_rq_clock(rq
);
8420 running
= task_current(rq
, tsk
);
8421 on_rq
= tsk
->se
.on_rq
;
8424 dequeue_task(rq
, tsk
, 0);
8425 if (unlikely(running
))
8426 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8428 set_task_rq(tsk
, task_cpu(tsk
));
8430 #ifdef CONFIG_FAIR_GROUP_SCHED
8431 if (tsk
->sched_class
->moved_group
)
8432 tsk
->sched_class
->moved_group(tsk
);
8435 if (unlikely(running
))
8436 tsk
->sched_class
->set_curr_task(rq
);
8438 enqueue_task(rq
, tsk
, 0);
8440 task_rq_unlock(rq
, &flags
);
8442 #endif /* CONFIG_GROUP_SCHED */
8444 #ifdef CONFIG_FAIR_GROUP_SCHED
8445 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8447 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8452 dequeue_entity(cfs_rq
, se
, 0);
8454 se
->load
.weight
= shares
;
8455 se
->load
.inv_weight
= 0;
8458 enqueue_entity(cfs_rq
, se
, 0);
8461 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8463 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8464 struct rq
*rq
= cfs_rq
->rq
;
8465 unsigned long flags
;
8467 spin_lock_irqsave(&rq
->lock
, flags
);
8468 __set_se_shares(se
, shares
);
8469 spin_unlock_irqrestore(&rq
->lock
, flags
);
8472 static DEFINE_MUTEX(shares_mutex
);
8474 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8477 unsigned long flags
;
8480 * We can't change the weight of the root cgroup.
8485 if (shares
< MIN_SHARES
)
8486 shares
= MIN_SHARES
;
8487 else if (shares
> MAX_SHARES
)
8488 shares
= MAX_SHARES
;
8490 mutex_lock(&shares_mutex
);
8491 if (tg
->shares
== shares
)
8494 spin_lock_irqsave(&task_group_lock
, flags
);
8495 for_each_possible_cpu(i
)
8496 unregister_fair_sched_group(tg
, i
);
8497 list_del_rcu(&tg
->siblings
);
8498 spin_unlock_irqrestore(&task_group_lock
, flags
);
8500 /* wait for any ongoing reference to this group to finish */
8501 synchronize_sched();
8504 * Now we are free to modify the group's share on each cpu
8505 * w/o tripping rebalance_share or load_balance_fair.
8507 tg
->shares
= shares
;
8508 for_each_possible_cpu(i
) {
8512 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8513 set_se_shares(tg
->se
[i
], shares
);
8517 * Enable load balance activity on this group, by inserting it back on
8518 * each cpu's rq->leaf_cfs_rq_list.
8520 spin_lock_irqsave(&task_group_lock
, flags
);
8521 for_each_possible_cpu(i
)
8522 register_fair_sched_group(tg
, i
);
8523 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8524 spin_unlock_irqrestore(&task_group_lock
, flags
);
8526 mutex_unlock(&shares_mutex
);
8530 unsigned long sched_group_shares(struct task_group
*tg
)
8536 #ifdef CONFIG_RT_GROUP_SCHED
8538 * Ensure that the real time constraints are schedulable.
8540 static DEFINE_MUTEX(rt_constraints_mutex
);
8542 static unsigned long to_ratio(u64 period
, u64 runtime
)
8544 if (runtime
== RUNTIME_INF
)
8547 return div64_u64(runtime
<< 16, period
);
8550 #ifdef CONFIG_CGROUP_SCHED
8551 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8553 struct task_group
*tgi
, *parent
= tg
->parent
;
8554 unsigned long total
= 0;
8557 if (global_rt_period() < period
)
8560 return to_ratio(period
, runtime
) <
8561 to_ratio(global_rt_period(), global_rt_runtime());
8564 if (ktime_to_ns(parent
->rt_bandwidth
.rt_period
) < period
)
8568 list_for_each_entry_rcu(tgi
, &parent
->children
, siblings
) {
8572 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8573 tgi
->rt_bandwidth
.rt_runtime
);
8577 return total
+ to_ratio(period
, runtime
) <=
8578 to_ratio(ktime_to_ns(parent
->rt_bandwidth
.rt_period
),
8579 parent
->rt_bandwidth
.rt_runtime
);
8581 #elif defined CONFIG_USER_SCHED
8582 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8584 struct task_group
*tgi
;
8585 unsigned long total
= 0;
8586 unsigned long global_ratio
=
8587 to_ratio(global_rt_period(), global_rt_runtime());
8590 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
8594 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8595 tgi
->rt_bandwidth
.rt_runtime
);
8599 return total
+ to_ratio(period
, runtime
) < global_ratio
;
8603 /* Must be called with tasklist_lock held */
8604 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8606 struct task_struct
*g
, *p
;
8607 do_each_thread(g
, p
) {
8608 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8610 } while_each_thread(g
, p
);
8614 static int tg_set_bandwidth(struct task_group
*tg
,
8615 u64 rt_period
, u64 rt_runtime
)
8619 mutex_lock(&rt_constraints_mutex
);
8620 read_lock(&tasklist_lock
);
8621 if (rt_runtime
== 0 && tg_has_rt_tasks(tg
)) {
8625 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
8630 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8631 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8632 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8634 for_each_possible_cpu(i
) {
8635 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8637 spin_lock(&rt_rq
->rt_runtime_lock
);
8638 rt_rq
->rt_runtime
= rt_runtime
;
8639 spin_unlock(&rt_rq
->rt_runtime_lock
);
8641 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8643 read_unlock(&tasklist_lock
);
8644 mutex_unlock(&rt_constraints_mutex
);
8649 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8651 u64 rt_runtime
, rt_period
;
8653 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8654 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8655 if (rt_runtime_us
< 0)
8656 rt_runtime
= RUNTIME_INF
;
8658 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8661 long sched_group_rt_runtime(struct task_group
*tg
)
8665 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8668 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8669 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8670 return rt_runtime_us
;
8673 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8675 u64 rt_runtime
, rt_period
;
8677 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8678 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8680 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8683 long sched_group_rt_period(struct task_group
*tg
)
8687 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8688 do_div(rt_period_us
, NSEC_PER_USEC
);
8689 return rt_period_us
;
8692 static int sched_rt_global_constraints(void)
8694 struct task_group
*tg
= &root_task_group
;
8695 u64 rt_runtime
, rt_period
;
8698 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8699 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8701 mutex_lock(&rt_constraints_mutex
);
8702 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
))
8704 mutex_unlock(&rt_constraints_mutex
);
8708 #else /* !CONFIG_RT_GROUP_SCHED */
8709 static int sched_rt_global_constraints(void)
8711 unsigned long flags
;
8714 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8715 for_each_possible_cpu(i
) {
8716 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8718 spin_lock(&rt_rq
->rt_runtime_lock
);
8719 rt_rq
->rt_runtime
= global_rt_runtime();
8720 spin_unlock(&rt_rq
->rt_runtime_lock
);
8722 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8726 #endif /* CONFIG_RT_GROUP_SCHED */
8728 int sched_rt_handler(struct ctl_table
*table
, int write
,
8729 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
8733 int old_period
, old_runtime
;
8734 static DEFINE_MUTEX(mutex
);
8737 old_period
= sysctl_sched_rt_period
;
8738 old_runtime
= sysctl_sched_rt_runtime
;
8740 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
8742 if (!ret
&& write
) {
8743 ret
= sched_rt_global_constraints();
8745 sysctl_sched_rt_period
= old_period
;
8746 sysctl_sched_rt_runtime
= old_runtime
;
8748 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8749 def_rt_bandwidth
.rt_period
=
8750 ns_to_ktime(global_rt_period());
8753 mutex_unlock(&mutex
);
8758 #ifdef CONFIG_CGROUP_SCHED
8760 /* return corresponding task_group object of a cgroup */
8761 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8763 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8764 struct task_group
, css
);
8767 static struct cgroup_subsys_state
*
8768 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8770 struct task_group
*tg
, *parent
;
8772 if (!cgrp
->parent
) {
8773 /* This is early initialization for the top cgroup */
8774 init_task_group
.css
.cgroup
= cgrp
;
8775 return &init_task_group
.css
;
8778 parent
= cgroup_tg(cgrp
->parent
);
8779 tg
= sched_create_group(parent
);
8781 return ERR_PTR(-ENOMEM
);
8783 /* Bind the cgroup to task_group object we just created */
8784 tg
->css
.cgroup
= cgrp
;
8790 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8792 struct task_group
*tg
= cgroup_tg(cgrp
);
8794 sched_destroy_group(tg
);
8798 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8799 struct task_struct
*tsk
)
8801 #ifdef CONFIG_RT_GROUP_SCHED
8802 /* Don't accept realtime tasks when there is no way for them to run */
8803 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
8806 /* We don't support RT-tasks being in separate groups */
8807 if (tsk
->sched_class
!= &fair_sched_class
)
8815 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8816 struct cgroup
*old_cont
, struct task_struct
*tsk
)
8818 sched_move_task(tsk
);
8821 #ifdef CONFIG_FAIR_GROUP_SCHED
8822 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8825 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8828 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8830 struct task_group
*tg
= cgroup_tg(cgrp
);
8832 return (u64
) tg
->shares
;
8834 #endif /* CONFIG_FAIR_GROUP_SCHED */
8836 #ifdef CONFIG_RT_GROUP_SCHED
8837 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8840 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8843 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8845 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8848 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8851 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8854 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8856 return sched_group_rt_period(cgroup_tg(cgrp
));
8858 #endif /* CONFIG_RT_GROUP_SCHED */
8860 static struct cftype cpu_files
[] = {
8861 #ifdef CONFIG_FAIR_GROUP_SCHED
8864 .read_u64
= cpu_shares_read_u64
,
8865 .write_u64
= cpu_shares_write_u64
,
8868 #ifdef CONFIG_RT_GROUP_SCHED
8870 .name
= "rt_runtime_us",
8871 .read_s64
= cpu_rt_runtime_read
,
8872 .write_s64
= cpu_rt_runtime_write
,
8875 .name
= "rt_period_us",
8876 .read_u64
= cpu_rt_period_read_uint
,
8877 .write_u64
= cpu_rt_period_write_uint
,
8882 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8884 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8887 struct cgroup_subsys cpu_cgroup_subsys
= {
8889 .create
= cpu_cgroup_create
,
8890 .destroy
= cpu_cgroup_destroy
,
8891 .can_attach
= cpu_cgroup_can_attach
,
8892 .attach
= cpu_cgroup_attach
,
8893 .populate
= cpu_cgroup_populate
,
8894 .subsys_id
= cpu_cgroup_subsys_id
,
8898 #endif /* CONFIG_CGROUP_SCHED */
8900 #ifdef CONFIG_CGROUP_CPUACCT
8903 * CPU accounting code for task groups.
8905 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8906 * (balbir@in.ibm.com).
8909 /* track cpu usage of a group of tasks */
8911 struct cgroup_subsys_state css
;
8912 /* cpuusage holds pointer to a u64-type object on every cpu */
8916 struct cgroup_subsys cpuacct_subsys
;
8918 /* return cpu accounting group corresponding to this container */
8919 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8921 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8922 struct cpuacct
, css
);
8925 /* return cpu accounting group to which this task belongs */
8926 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8928 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8929 struct cpuacct
, css
);
8932 /* create a new cpu accounting group */
8933 static struct cgroup_subsys_state
*cpuacct_create(
8934 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8936 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8939 return ERR_PTR(-ENOMEM
);
8941 ca
->cpuusage
= alloc_percpu(u64
);
8942 if (!ca
->cpuusage
) {
8944 return ERR_PTR(-ENOMEM
);
8950 /* destroy an existing cpu accounting group */
8952 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8954 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8956 free_percpu(ca
->cpuusage
);
8960 /* return total cpu usage (in nanoseconds) of a group */
8961 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8963 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8964 u64 totalcpuusage
= 0;
8967 for_each_possible_cpu(i
) {
8968 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8971 * Take rq->lock to make 64-bit addition safe on 32-bit
8974 spin_lock_irq(&cpu_rq(i
)->lock
);
8975 totalcpuusage
+= *cpuusage
;
8976 spin_unlock_irq(&cpu_rq(i
)->lock
);
8979 return totalcpuusage
;
8982 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8985 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8994 for_each_possible_cpu(i
) {
8995 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8997 spin_lock_irq(&cpu_rq(i
)->lock
);
8999 spin_unlock_irq(&cpu_rq(i
)->lock
);
9005 static struct cftype files
[] = {
9008 .read_u64
= cpuusage_read
,
9009 .write_u64
= cpuusage_write
,
9013 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9015 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9019 * charge this task's execution time to its accounting group.
9021 * called with rq->lock held.
9023 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9027 if (!cpuacct_subsys
.active
)
9032 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9034 *cpuusage
+= cputime
;
9038 struct cgroup_subsys cpuacct_subsys
= {
9040 .create
= cpuacct_create
,
9041 .destroy
= cpuacct_destroy
,
9042 .populate
= cpuacct_populate
,
9043 .subsys_id
= cpuacct_subsys_id
,
9045 #endif /* CONFIG_CGROUP_CPUACCT */